Switched to luigi, parsing out FLAC metadata.

29 files changed, 18883 insertions, 8334 deletions

diff --git a/Makefile b/Makefile
index 14527e7..761b428 100644
--- a/Makefile
+++ b/Makefile
@@ -5,24 +5,19 @@ cflags = -I./ -g $(includes) $(defines) \
 lflags = -lX11 -lm
 target = twinkplayer
 includes =
-defines = -Dplat_x11 \
+defines = \
 	-Dplat_posix \
 	-Dplat_pulse \
 	-Dis64bit=1 \
 	-Dplat_x86 \
 	-DDEBUG
+luigi = luigi.o
 
 sources = \
-	cfgparse.c \
-	font.c \
 	library.c \
 	main.c \
 	memory.c \
-	plat.c \
-	rcache.c \
-	rect.c \
-	render.c \
-	ui.c
+	plat.c
 
 objects = $(sources:%.c=%.o)
 
@@ -31,12 +26,15 @@ all: $(target)
 $(objects): %.o : %.c
 	$(compiler) -MMD -MF $(basename $@).d $(cflags) -o $@ -c $<
 
-$(target): $(objects)
-	$(linker) $(objects) -o $@ $(lflags)
+$(luigi): luigi.c
+	gcc -MMD -MF $(basename $@).d -DUI_LINUX -w -std=c11 -g -o $@ -c $<
+
+$(target): $(objects) $(luigi)
+	$(linker) $(objects) $(luigi) -o $@ $(lflags)
 
 clean:
 	rm *.d
 	rm *.o
-	rm $(target)
+	rm -f $(target)
 
 -include $(sources:%.c=%.d)
diff --git a/cfgparse.c b/cfgparse.c
deleted file mode 100644
index 6d1668f..0000000
--- a/cfgparse.c
+++ /dev/null
@@ -1,233 +0,0 @@
-#include <stdio.h>
-#include <stdlib.h>
-#include <string.h>
-
-#include "cfgparse.h"
-
-static const char* next_line(const char* line) {
-	const char* c;
-	for (c = line; *c != '\n'; c++) {
-		if (!*c) return 0;
-	}
-	return c + 1;
-}
-
-static int is_object(
-	const char* line,
-	const char** start,
-	int* len
-) {
-	const char* c;
-	if (line[0] != '[') return 0;
-	*start = line + 1;
-	for (
-		c = line, *len = 0;
-		*c && *c != '\n';
-		c++, *len = *len + 1
-	) if (*c == ']') {
-		*len = *len - 1;
-		return 1;
-	}
-	return 0;
-}
-
-static int is_iden(char c) {
-	return
-		(c >= 'a' && c <= 'z') ||
-		(c >= 'A' && c <= 'Z') ||
-		(c == '_');
-}
-
-static int is_prop(
-	const char* line,
-	const char** start,
-	int* len
-) {
-	const char* c;
-	*start = line;
-	for (
-		c = line, *len = 0;
-		*c && *c != '\n';
-		c++, *len = *len + 1
-	) if (*c == ':') {
-		*len = *len;
-		return 1;
-	} else if (!is_iden(*c)) return 0;
-	return 0;
-}
-
-static void fill_prop_string(
-	cfg_Prop* p,
-	Arena* arena,
-	const char* start,
-	int len
-) {
-	p->type = cfg_type_string;
-	p->as.str = arena_alloc(arena, len + 1);
-	memcpy(p->as.str, start, len);
-	p->as.str[len] = 0;
-}
-
-static void fill_prop_number(
-	cfg_Prop* p,
-	const char* start,
-	int len
-) {
-	(void)len;
-	p->type = cfg_type_int;
-	p->as.num = (int)strtol(start, 0, 10);
-}
-
-static void fill_prop_bool(
-	cfg_Prop* p,
-	const char* start,
-	int len
-) {
-	(void)len;
-	p->type = cfg_type_int;
-	if (!memcmp("true", start, 4))
-		p->as.num = 1;
-	else
-		p->as.num = 0;
-}
-
-static void fill_prop_data(
-	cfg_Prop* p,
-	Arena* arena,
-	const char* start,
-	int len
-) {
-	int i, size;
-	char s[3];
-	s[2] = 0;
-	size = len / 2;
-	p->type = cfg_type_data;
-	p->as.data.data = arena_alloc(arena, size);
-	p->as.data.size = size;
-	for (i = 0; i < size; i++) {
-		s[0] = start[i * 2];
-		s[1] = start[i * 2 + 1];
-		sscanf(s, "%x", (unsigned*)&p->as.data.data[i]);
-	}
-}
-
-static void fill_prop_val(
-	cfg_Prop* p,
-	Arena* arena,
-	const char* start
-) {
-	const char* c;
-	int len;
-	for (; *start && (*start == ' ' || *start == '\t'); start++);
-	for (len = 0, c = start; *c && *c != '\n'; c++, len++);
-	switch (start[0]) {
-		case '"':
-			fill_prop_string(p, arena, start + 1, len - 1);
-			break;
-		case '0': case '1': case '2':
-		case '3': case '4': case '5':
-		case '6': case '7': case '8':
-		case '9': case '-':
-			fill_prop_number(p, start, len);
-			break;
-		case 'f': case 't':
-			fill_prop_bool(p, start, len);
-			break;
-		case '$':
-			fill_prop_data(p, arena, start + 1, len - 1);
-			break;
-		default:
-			p->type = cfg_type_none;
-			break;
-	}
-}
-
-cfg_Object* cfg_parse(const char* src, Arena* arena) {
-	const char* line, * start;
-	int len;
-	cfg_Object* root = 0, * obj = 0, * pobj;
-	cfg_Prop* prop = 0, * pprop;
-	for (line = src; line; line = next_line(line)) {
-		len = 0;
-		if (is_object(line, &start, &len) || line == src) {
-			pobj = obj;
-			obj = arena_alloc(arena, sizeof *obj);
-			obj->prop_count = 0;
-			obj->props = 0;
-			obj->next = 0;
-			memcpy(obj->name, start, len);
-			obj->name[len] = 0;
-			if (!root)
-				root = obj;
-			else
-				pobj->next = obj;
-		} else if (!obj) continue;
-		if (is_prop(line, &start, &len)) {
-			pprop = prop;
-			prop = arena_alloc(arena, sizeof *prop);
-			prop->next = 0;
-			memcpy(prop->name, start, len);
-			prop->name[len] = 0;
-			fill_prop_val(prop, arena, line + len + 1);
-			obj->prop_count++;
-			if (!obj->props)
-				obj->props = prop;
-			else
-				pprop->next = prop;
-		}
-	}
-	return root;
-}
-
-const cfg_Prop* find_prop(
-	const cfg_Object* obj,
-	const char* name
-) {
-	const cfg_Prop* p;
-	for (p = obj->props; p; p = p->next) {
-		if (!strcmp(p->name, name)) {
-			return p;
-		}
-	}
-	return 0;
-}
-
-const cfg_Prop* find_prop_of(
-	const cfg_Object* obj,
-	const char* name,
-	cfg_Type type
-) {
-	const cfg_Prop* p;
-	for (p = obj->props; p; p = p->next) {
-		if (p->type == type && !strcmp(p->name, name)) {
-			return p;
-		}
-	}
-	return 0;
-}
-
-int find_int_default(
-	const cfg_Object* obj,
-	const char* name,
-	int def
-) {
-	const cfg_Prop* prop;
-	prop = find_prop_of(obj, name, cfg_type_int);
-	if (prop) {
-		return prop->as.num;
-	}
-	return def;
-}
-
-const char* find_string_default(
-	const cfg_Object* obj,
-	const char* name,
-	const char* def
-) {
-	const cfg_Prop* prop;
-	prop = find_prop_of(obj, name, cfg_type_string);
-	if (prop) {
-		return prop->as.str;
-	}
-	return def;
-}
diff --git a/cfgparse.h b/cfgparse.h
deleted file mode 100644
index f8eeedb..0000000
--- a/cfgparse.h
+++ /dev/null
@@ -1,58 +0,0 @@
-#ifndef cfgparse_h
-#define cfgparse_h
-
-#include "memory.h"
-
-typedef enum {
-	cfg_type_int,
-	cfg_type_string,
-	cfg_type_data,
-	cfg_type_none
-} cfg_Type;
-
-typedef struct {
-	int size;
-	char* data;
-} cfg_Data;
-
-typedef struct cfg_Prop {
-	char name[28];
-	cfg_Type type;
-	union {
-		int num;
-		cfg_Data data;
-		char* str;
-	} as;
-	struct cfg_Prop* next;
-} cfg_Prop;
-
-typedef struct cfg_Object {
-	char name[28];
-	int prop_count;
-	cfg_Prop* props;
-	struct cfg_Object* next;
-} cfg_Object;
-
-cfg_Object* cfg_parse(const char* src, Arena* arena);
-const cfg_Prop* find_prop(
-	const cfg_Object* obj,
-	const char* name
-);
-const cfg_Prop* find_prop_of(
-	const cfg_Object* obj,
-	const char* name,
-	cfg_Type type
-);
-
-int find_int_default(
-	const cfg_Object* obj,
-	const char* name,
-	int def
-);
-const char* find_string_default(
-	const cfg_Object* obj,
-	const char* name,
-	const char* def
-);
-
-#endif
diff --git a/config.h b/config.h
index 612984f..e19aa30 100644
--- a/config.h
+++ b/config.h
@@ -4,30 +4,12 @@
 #define app_name "Twink Player"
 #define default_window_w 1280
 #define default_window_h 960
-#define ms_per_frame 40 /* 25 FPS */
 
-#define max_songs 1024
-#define song_filename_max 256
-#define song_name_max 256
-#define library_folder "/home/quou/music/"
-
-/* font */
-#define default_font_location "/usr/share/fonts/TTF/DejaVuSans.ttf"
-#define default_font_size 20
-
-/* theme */
-#define theme_text_colour       make_colour(0x000000, 0xff)
-#define theme_background_colour make_colour(0xffffff, 0xff)
-#define theme_outline_colour    make_colour(0x000000, 0xff)
-#define theme_hovered_colour    make_colour(0xcacaca, 0xff)
-#define theme_active_colour     make_colour(0xf0f8ff, 0xff)
-#define theme_outline_width     2
-#define theme_padding           2
-
-/* from here on are parametres that are not designed to
- * be changed. */
-#define memory_heap_size  (1024 * 1024 * 8)
+#define memory_heap_size  (1024 * 1024 * 512)
 #define memory_size       (memory_heap_size + (1024 * 1024 * 8))
+#define library_memory_size (1024 * 1024 * 256)
+
+#define library_path "/home/quou/music/"
 
 #if defined(is64bit) && is64bit
 #define allocation_default_alignment 8
@@ -35,14 +17,4 @@
 #define allocation_default_alignment 4
 #endif
 
-#define fbits 9
-
-#define rcache_enable 1
-#define rcache_default_cell_w 128
-#define rcache_default_cell_h 128
-#define rcache_gw (default_window_w / rcache_default_cell_w)
-#define rcache_gh (default_window_w / rcache_default_cell_h)
-#define rcache_cmd_buf_size (1024 * 1024)
-#define rcache_hash_seed 2166136261
-
 #endif
diff --git a/dr_flac.h b/dr_flac.h
new file mode 100644
index 0000000..14324cf
--- /dev/null
+++ b/dr_flac.h
@@ -0,0 +1,12536 @@
+/*
+FLAC audio decoder. Choice of public domain or MIT-0. See license statements at the end of this file.
+dr_flac - v0.12.42 - 2023-11-02
+
+David Reid - mackron@gmail.com
+
+GitHub: https://github.com/mackron/dr_libs
+*/
+
+/*
+RELEASE NOTES - v0.12.0
+=======================
+Version 0.12.0 has breaking API changes including changes to the existing API and the removal of deprecated APIs.
+
+
+Improved Client-Defined Memory Allocation
+-----------------------------------------
+The main change with this release is the addition of a more flexible way of implementing custom memory allocation routines. The
+existing system of DRFLAC_MALLOC, DRFLAC_REALLOC and DRFLAC_FREE are still in place and will be used by default when no custom
+allocation callbacks are specified.
+
+To use the new system, you pass in a pointer to a drflac_allocation_callbacks object to drflac_open() and family, like this:
+
+    void* my_malloc(size_t sz, void* pUserData)
+    {
+        return malloc(sz);
+    }
+    void* my_realloc(void* p, size_t sz, void* pUserData)
+    {
+        return realloc(p, sz);
+    }
+    void my_free(void* p, void* pUserData)
+    {
+        free(p);
+    }
+
+    ...
+
+    drflac_allocation_callbacks allocationCallbacks;
+    allocationCallbacks.pUserData = &myData;
+    allocationCallbacks.onMalloc  = my_malloc;
+    allocationCallbacks.onRealloc = my_realloc;
+    allocationCallbacks.onFree    = my_free;
+    drflac* pFlac = drflac_open_file("my_file.flac", &allocationCallbacks);
+
+The advantage of this new system is that it allows you to specify user data which will be passed in to the allocation routines.
+
+Passing in null for the allocation callbacks object will cause dr_flac to use defaults which is the same as DRFLAC_MALLOC,
+DRFLAC_REALLOC and DRFLAC_FREE and the equivalent of how it worked in previous versions.
+
+Every API that opens a drflac object now takes this extra parameter. These include the following:
+
+    drflac_open()
+    drflac_open_relaxed()
+    drflac_open_with_metadata()
+    drflac_open_with_metadata_relaxed()
+    drflac_open_file()
+    drflac_open_file_with_metadata()
+    drflac_open_memory()
+    drflac_open_memory_with_metadata()
+    drflac_open_and_read_pcm_frames_s32()
+    drflac_open_and_read_pcm_frames_s16()
+    drflac_open_and_read_pcm_frames_f32()
+    drflac_open_file_and_read_pcm_frames_s32()
+    drflac_open_file_and_read_pcm_frames_s16()
+    drflac_open_file_and_read_pcm_frames_f32()
+    drflac_open_memory_and_read_pcm_frames_s32()
+    drflac_open_memory_and_read_pcm_frames_s16()
+    drflac_open_memory_and_read_pcm_frames_f32()
+
+
+
+Optimizations
+-------------
+Seeking performance has been greatly improved. A new binary search based seeking algorithm has been introduced which significantly
+improves performance over the brute force method which was used when no seek table was present. Seek table based seeking also takes
+advantage of the new binary search seeking system to further improve performance there as well. Note that this depends on CRC which
+means it will be disabled when DR_FLAC_NO_CRC is used.
+
+The SSE4.1 pipeline has been cleaned up and optimized. You should see some improvements with decoding speed of 24-bit files in
+particular. 16-bit streams should also see some improvement.
+
+drflac_read_pcm_frames_s16() has been optimized. Previously this sat on top of drflac_read_pcm_frames_s32() and performed it's s32
+to s16 conversion in a second pass. This is now all done in a single pass. This includes SSE2 and ARM NEON optimized paths.
+
+A minor optimization has been implemented for drflac_read_pcm_frames_s32(). This will now use an SSE2 optimized pipeline for stereo
+channel reconstruction which is the last part of the decoding process.
+
+The ARM build has seen a few improvements. The CLZ (count leading zeroes) and REV (byte swap) instructions are now used when
+compiling with GCC and Clang which is achieved using inline assembly. The CLZ instruction requires ARM architecture version 5 at
+compile time and the REV instruction requires ARM architecture version 6.
+
+An ARM NEON optimized pipeline has been implemented. To enable this you'll need to add -mfpu=neon to the command line when compiling.
+
+
+Removed APIs
+------------
+The following APIs were deprecated in version 0.11.0 and have been completely removed in version 0.12.0:
+
+    drflac_read_s32()                   -> drflac_read_pcm_frames_s32()
+    drflac_read_s16()                   -> drflac_read_pcm_frames_s16()
+    drflac_read_f32()                   -> drflac_read_pcm_frames_f32()
+    drflac_seek_to_sample()             -> drflac_seek_to_pcm_frame()
+    drflac_open_and_decode_s32()        -> drflac_open_and_read_pcm_frames_s32()
+    drflac_open_and_decode_s16()        -> drflac_open_and_read_pcm_frames_s16()
+    drflac_open_and_decode_f32()        -> drflac_open_and_read_pcm_frames_f32()
+    drflac_open_and_decode_file_s32()   -> drflac_open_file_and_read_pcm_frames_s32()
+    drflac_open_and_decode_file_s16()   -> drflac_open_file_and_read_pcm_frames_s16()
+    drflac_open_and_decode_file_f32()   -> drflac_open_file_and_read_pcm_frames_f32()
+    drflac_open_and_decode_memory_s32() -> drflac_open_memory_and_read_pcm_frames_s32()
+    drflac_open_and_decode_memory_s16() -> drflac_open_memory_and_read_pcm_frames_s16()
+    drflac_open_and_decode_memory_f32() -> drflac_open_memroy_and_read_pcm_frames_f32()
+
+Prior versions of dr_flac operated on a per-sample basis whereas now it operates on PCM frames. The removed APIs all relate
+to the old per-sample APIs. You now need to use the "pcm_frame" versions.
+*/
+
+
+/*
+Introduction
+============
+dr_flac is a single file library. To use it, do something like the following in one .c file.
+
+    ```c
+    #define DR_FLAC_IMPLEMENTATION
+    #include "dr_flac.h"
+    ```
+
+You can then #include this file in other parts of the program as you would with any other header file. To decode audio data, do something like the following:
+
+    ```c
+    drflac* pFlac = drflac_open_file("MySong.flac", NULL);
+    if (pFlac == NULL) {
+        // Failed to open FLAC file
+    }
+
+    drflac_int32* pSamples = malloc(pFlac->totalPCMFrameCount * pFlac->channels * sizeof(drflac_int32));
+    drflac_uint64 numberOfInterleavedSamplesActuallyRead = drflac_read_pcm_frames_s32(pFlac, pFlac->totalPCMFrameCount, pSamples);
+    ```
+
+The drflac object represents the decoder. It is a transparent type so all the information you need, such as the number of channels and the bits per sample,
+should be directly accessible - just make sure you don't change their values. Samples are always output as interleaved signed 32-bit PCM. In the example above
+a native FLAC stream was opened, however dr_flac has seamless support for Ogg encapsulated FLAC streams as well.
+
+You do not need to decode the entire stream in one go - you just specify how many samples you'd like at any given time and the decoder will give you as many
+samples as it can, up to the amount requested. Later on when you need the next batch of samples, just call it again. Example:
+
+    ```c
+    while (drflac_read_pcm_frames_s32(pFlac, chunkSizeInPCMFrames, pChunkSamples) > 0) {
+        do_something();
+    }
+    ```
+
+You can seek to a specific PCM frame with `drflac_seek_to_pcm_frame()`.
+
+If you just want to quickly decode an entire FLAC file in one go you can do something like this:
+
+    ```c
+    unsigned int channels;
+    unsigned int sampleRate;
+    drflac_uint64 totalPCMFrameCount;
+    drflac_int32* pSampleData = drflac_open_file_and_read_pcm_frames_s32("MySong.flac", &channels, &sampleRate, &totalPCMFrameCount, NULL);
+    if (pSampleData == NULL) {
+        // Failed to open and decode FLAC file.
+    }
+
+    ...
+
+    drflac_free(pSampleData, NULL);
+    ```
+
+You can read samples as signed 16-bit integer and 32-bit floating-point PCM with the *_s16() and *_f32() family of APIs respectively, but note that these
+should be considered lossy.
+
+
+If you need access to metadata (album art, etc.), use `drflac_open_with_metadata()`, `drflac_open_file_with_metdata()` or `drflac_open_memory_with_metadata()`.
+The rationale for keeping these APIs separate is that they're slightly slower than the normal versions and also just a little bit harder to use. dr_flac
+reports metadata to the application through the use of a callback, and every metadata block is reported before `drflac_open_with_metdata()` returns.
+
+The main opening APIs (`drflac_open()`, etc.) will fail if the header is not present. The presents a problem in certain scenarios such as broadcast style
+streams or internet radio where the header may not be present because the user has started playback mid-stream. To handle this, use the relaxed APIs:
+    
+    `drflac_open_relaxed()`
+    `drflac_open_with_metadata_relaxed()`
+
+It is not recommended to use these APIs for file based streams because a missing header would usually indicate a corrupt or perverse file. In addition, these
+APIs can take a long time to initialize because they may need to spend a lot of time finding the first frame.
+
+
+
+Build Options
+=============
+#define these options before including this file.
+
+#define DR_FLAC_NO_STDIO
+  Disable `drflac_open_file()` and family.
+
+#define DR_FLAC_NO_OGG
+  Disables support for Ogg/FLAC streams.
+
+#define DR_FLAC_BUFFER_SIZE <number>
+  Defines the size of the internal buffer to store data from onRead(). This buffer is used to reduce the number of calls back to the client for more data.
+  Larger values means more memory, but better performance. My tests show diminishing returns after about 4KB (which is the default). Consider reducing this if
+  you have a very efficient implementation of onRead(), or increase it if it's very inefficient. Must be a multiple of 8.
+
+#define DR_FLAC_NO_CRC
+  Disables CRC checks. This will offer a performance boost when CRC is unnecessary. This will disable binary search seeking. When seeking, the seek table will
+  be used if available. Otherwise the seek will be performed using brute force.
+
+#define DR_FLAC_NO_SIMD
+  Disables SIMD optimizations (SSE on x86/x64 architectures, NEON on ARM architectures). Use this if you are having compatibility issues with your compiler.
+
+#define DR_FLAC_NO_WCHAR
+  Disables all functions ending with `_w`. Use this if your compiler does not provide wchar.h. Not required if DR_FLAC_NO_STDIO is also defined.
+
+
+
+Notes
+=====
+- dr_flac does not support changing the sample rate nor channel count mid stream.
+- dr_flac is not thread-safe, but its APIs can be called from any thread so long as you do your own synchronization.
+- When using Ogg encapsulation, a corrupted metadata block will result in `drflac_open_with_metadata()` and `drflac_open()` returning inconsistent samples due
+  to differences in corrupted stream recorvery logic between the two APIs.
+*/
+
+#ifndef dr_flac_h
+#define dr_flac_h
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#define DRFLAC_STRINGIFY(x)      #x
+#define DRFLAC_XSTRINGIFY(x)     DRFLAC_STRINGIFY(x)
+
+#define DRFLAC_VERSION_MAJOR     0
+#define DRFLAC_VERSION_MINOR     12
+#define DRFLAC_VERSION_REVISION  42
+#define DRFLAC_VERSION_STRING    DRFLAC_XSTRINGIFY(DRFLAC_VERSION_MAJOR) "." DRFLAC_XSTRINGIFY(DRFLAC_VERSION_MINOR) "." DRFLAC_XSTRINGIFY(DRFLAC_VERSION_REVISION)
+
+#include <stddef.h> /* For size_t. */
+
+/* Sized Types */
+typedef   signed char           drflac_int8;
+typedef unsigned char           drflac_uint8;
+typedef   signed short          drflac_int16;
+typedef unsigned short          drflac_uint16;
+typedef   signed int            drflac_int32;
+typedef unsigned int            drflac_uint32;
+#if defined(_MSC_VER) && !defined(__clang__)
+    typedef   signed __int64    drflac_int64;
+    typedef unsigned __int64    drflac_uint64;
+#else
+    #if defined(__clang__) || (defined(__GNUC__) && (__GNUC__ > 4 || (__GNUC__ == 4 && __GNUC_MINOR__ >= 6)))
+        #pragma GCC diagnostic push
+        #pragma GCC diagnostic ignored "-Wlong-long"
+        #if defined(__clang__)
+            #pragma GCC diagnostic ignored "-Wc++11-long-long"
+        #endif
+    #endif
+    typedef   signed long long  drflac_int64;
+    typedef unsigned long long  drflac_uint64;
+    #if defined(__clang__) || (defined(__GNUC__) && (__GNUC__ > 4 || (__GNUC__ == 4 && __GNUC_MINOR__ >= 6)))
+        #pragma GCC diagnostic pop
+    #endif
+#endif
+#if defined(__LP64__) || defined(_WIN64) || (defined(__x86_64__) && !defined(__ILP32__)) || defined(_M_X64) || defined(__ia64) || defined(_M_IA64) || defined(__aarch64__) || defined(_M_ARM64) || defined(__powerpc64__)
+    typedef drflac_uint64       drflac_uintptr;
+#else
+    typedef drflac_uint32       drflac_uintptr;
+#endif
+typedef drflac_uint8            drflac_bool8;
+typedef drflac_uint32           drflac_bool32;
+#define DRFLAC_TRUE             1
+#define DRFLAC_FALSE            0
+/* End Sized Types */
+
+/* Decorations */
+#if !defined(DRFLAC_API)
+    #if defined(DRFLAC_DLL)
+        #if defined(_WIN32)
+            #define DRFLAC_DLL_IMPORT  __declspec(dllimport)
+            #define DRFLAC_DLL_EXPORT  __declspec(dllexport)
+            #define DRFLAC_DLL_PRIVATE static
+        #else
+            #if defined(__GNUC__) && __GNUC__ >= 4
+                #define DRFLAC_DLL_IMPORT  __attribute__((visibility("default")))
+                #define DRFLAC_DLL_EXPORT  __attribute__((visibility("default")))
+                #define DRFLAC_DLL_PRIVATE __attribute__((visibility("hidden")))
+            #else
+                #define DRFLAC_DLL_IMPORT
+                #define DRFLAC_DLL_EXPORT
+                #define DRFLAC_DLL_PRIVATE static
+            #endif
+        #endif
+
+        #if defined(DR_FLAC_IMPLEMENTATION) || defined(DRFLAC_IMPLEMENTATION)
+            #define DRFLAC_API  DRFLAC_DLL_EXPORT
+        #else
+            #define DRFLAC_API  DRFLAC_DLL_IMPORT
+        #endif
+        #define DRFLAC_PRIVATE DRFLAC_DLL_PRIVATE
+    #else
+        #define DRFLAC_API extern
+        #define DRFLAC_PRIVATE static
+    #endif
+#endif
+/* End Decorations */
+
+#if defined(_MSC_VER) && _MSC_VER >= 1700   /* Visual Studio 2012 */
+    #define DRFLAC_DEPRECATED       __declspec(deprecated)
+#elif (defined(__GNUC__) && __GNUC__ >= 4)  /* GCC 4 */
+    #define DRFLAC_DEPRECATED       __attribute__((deprecated))
+#elif defined(__has_feature)                /* Clang */
+    #if __has_feature(attribute_deprecated)
+        #define DRFLAC_DEPRECATED   __attribute__((deprecated))
+    #else
+        #define DRFLAC_DEPRECATED
+    #endif
+#else
+    #define DRFLAC_DEPRECATED
+#endif
+
+DRFLAC_API void drflac_version(drflac_uint32* pMajor, drflac_uint32* pMinor, drflac_uint32* pRevision);
+DRFLAC_API const char* drflac_version_string(void);
+
+/* Allocation Callbacks */
+typedef struct
+{
+    void* pUserData;
+    void* (* onMalloc)(size_t sz, void* pUserData);
+    void* (* onRealloc)(void* p, size_t sz, void* pUserData);
+    void  (* onFree)(void* p, void* pUserData);
+} drflac_allocation_callbacks;
+/* End Allocation Callbacks */
+
+/*
+As data is read from the client it is placed into an internal buffer for fast access. This controls the size of that buffer. Larger values means more speed,
+but also more memory. In my testing there is diminishing returns after about 4KB, but you can fiddle with this to suit your own needs. Must be a multiple of 8.
+*/
+#ifndef DR_FLAC_BUFFER_SIZE
+#define DR_FLAC_BUFFER_SIZE   4096
+#endif
+
+
+/* Architecture Detection */
+#if defined(_WIN64) || defined(_LP64) || defined(__LP64__)
+#define DRFLAC_64BIT
+#endif
+
+#if defined(__x86_64__) || defined(_M_X64)
+    #define DRFLAC_X64
+#elif defined(__i386) || defined(_M_IX86)
+    #define DRFLAC_X86
+#elif defined(__arm__) || defined(_M_ARM) || defined(__arm64) || defined(__arm64__) || defined(__aarch64__) || defined(_M_ARM64)
+    #define DRFLAC_ARM
+#endif
+/* End Architecture Detection */
+
+
+#ifdef DRFLAC_64BIT
+typedef drflac_uint64 drflac_cache_t;
+#else
+typedef drflac_uint32 drflac_cache_t;
+#endif
+
+/* The various metadata block types. */
+#define DRFLAC_METADATA_BLOCK_TYPE_STREAMINFO       0
+#define DRFLAC_METADATA_BLOCK_TYPE_PADDING          1
+#define DRFLAC_METADATA_BLOCK_TYPE_APPLICATION      2
+#define DRFLAC_METADATA_BLOCK_TYPE_SEEKTABLE        3
+#define DRFLAC_METADATA_BLOCK_TYPE_VORBIS_COMMENT   4
+#define DRFLAC_METADATA_BLOCK_TYPE_CUESHEET         5
+#define DRFLAC_METADATA_BLOCK_TYPE_PICTURE          6
+#define DRFLAC_METADATA_BLOCK_TYPE_INVALID          127
+
+/* The various picture types specified in the PICTURE block. */
+#define DRFLAC_PICTURE_TYPE_OTHER                   0
+#define DRFLAC_PICTURE_TYPE_FILE_ICON               1
+#define DRFLAC_PICTURE_TYPE_OTHER_FILE_ICON         2
+#define DRFLAC_PICTURE_TYPE_COVER_FRONT             3
+#define DRFLAC_PICTURE_TYPE_COVER_BACK              4
+#define DRFLAC_PICTURE_TYPE_LEAFLET_PAGE            5
+#define DRFLAC_PICTURE_TYPE_MEDIA                   6
+#define DRFLAC_PICTURE_TYPE_LEAD_ARTIST             7
+#define DRFLAC_PICTURE_TYPE_ARTIST                  8
+#define DRFLAC_PICTURE_TYPE_CONDUCTOR               9
+#define DRFLAC_PICTURE_TYPE_BAND                    10
+#define DRFLAC_PICTURE_TYPE_COMPOSER                11
+#define DRFLAC_PICTURE_TYPE_LYRICIST                12
+#define DRFLAC_PICTURE_TYPE_RECORDING_LOCATION      13
+#define DRFLAC_PICTURE_TYPE_DURING_RECORDING        14
+#define DRFLAC_PICTURE_TYPE_DURING_PERFORMANCE      15
+#define DRFLAC_PICTURE_TYPE_SCREEN_CAPTURE          16
+#define DRFLAC_PICTURE_TYPE_BRIGHT_COLORED_FISH     17
+#define DRFLAC_PICTURE_TYPE_ILLUSTRATION            18
+#define DRFLAC_PICTURE_TYPE_BAND_LOGOTYPE           19
+#define DRFLAC_PICTURE_TYPE_PUBLISHER_LOGOTYPE      20
+
+typedef enum
+{
+    drflac_container_native,
+    drflac_container_ogg,
+    drflac_container_unknown
+} drflac_container;
+
+typedef enum
+{
+    drflac_seek_origin_start,
+    drflac_seek_origin_current
+} drflac_seek_origin;
+
+/* The order of members in this structure is important because we map this directly to the raw data within the SEEKTABLE metadata block. */
+typedef struct
+{
+    drflac_uint64 firstPCMFrame;
+    drflac_uint64 flacFrameOffset;   /* The offset from the first byte of the header of the first frame. */
+    drflac_uint16 pcmFrameCount;
+} drflac_seekpoint;
+
+typedef struct
+{
+    drflac_uint16 minBlockSizeInPCMFrames;
+    drflac_uint16 maxBlockSizeInPCMFrames;
+    drflac_uint32 minFrameSizeInPCMFrames;
+    drflac_uint32 maxFrameSizeInPCMFrames;
+    drflac_uint32 sampleRate;
+    drflac_uint8  channels;
+    drflac_uint8  bitsPerSample;
+    drflac_uint64 totalPCMFrameCount;
+    drflac_uint8  md5[16];
+} drflac_streaminfo;
+
+typedef struct
+{
+    /*
+    The metadata type. Use this to know how to interpret the data below. Will be set to one of the
+    DRFLAC_METADATA_BLOCK_TYPE_* tokens.
+    */
+    drflac_uint32 type;
+
+    /*
+    A pointer to the raw data. This points to a temporary buffer so don't hold on to it. It's best to
+    not modify the contents of this buffer. Use the structures below for more meaningful and structured
+    information about the metadata. It's possible for this to be null.
+    */
+    const void* pRawData;
+
+    /* The size in bytes of the block and the buffer pointed to by pRawData if it's non-NULL. */
+    drflac_uint32 rawDataSize;
+
+    union
+    {
+        drflac_streaminfo streaminfo;
+
+        struct
+        {
+            int unused;
+        } padding;
+
+        struct
+        {
+            drflac_uint32 id;
+            const void* pData;
+            drflac_uint32 dataSize;
+        } application;
+
+        struct
+        {
+            drflac_uint32 seekpointCount;
+            const drflac_seekpoint* pSeekpoints;
+        } seektable;
+
+        struct
+        {
+            drflac_uint32 vendorLength;
+            const char* vendor;
+            drflac_uint32 commentCount;
+            const void* pComments;
+        } vorbis_comment;
+
+        struct
+        {
+            char catalog[128];
+            drflac_uint64 leadInSampleCount;
+            drflac_bool32 isCD;
+            drflac_uint8 trackCount;
+            const void* pTrackData;
+        } cuesheet;
+
+        struct
+        {
+            drflac_uint32 type;
+            drflac_uint32 mimeLength;
+            const char* mime;
+            drflac_uint32 descriptionLength;
+            const char* description;
+            drflac_uint32 width;
+            drflac_uint32 height;
+            drflac_uint32 colorDepth;
+            drflac_uint32 indexColorCount;
+            drflac_uint32 pictureDataSize;
+            const drflac_uint8* pPictureData;
+        } picture;
+    } data;
+} drflac_metadata;
+
+
+/*
+Callback for when data needs to be read from the client.
+
+
+Parameters
+----------
+pUserData (in)
+    The user data that was passed to drflac_open() and family.
+
+pBufferOut (out)
+    The output buffer.
+
+bytesToRead (in)
+    The number of bytes to read.
+
+
+Return Value
+------------
+The number of bytes actually read.
+
+
+Remarks
+-------
+A return value of less than bytesToRead indicates the end of the stream. Do _not_ return from this callback until either the entire bytesToRead is filled or
+you have reached the end of the stream.
+*/
+typedef size_t (* drflac_read_proc)(void* pUserData, void* pBufferOut, size_t bytesToRead);
+
+/*
+Callback for when data needs to be seeked.
+
+
+Parameters
+----------
+pUserData (in)
+    The user data that was passed to drflac_open() and family.
+
+offset (in)
+    The number of bytes to move, relative to the origin. Will never be negative.
+
+origin (in)
+    The origin of the seek - the current position or the start of the stream.
+
+
+Return Value
+------------
+Whether or not the seek was successful.
+
+
+Remarks
+-------
+The offset will never be negative. Whether or not it is relative to the beginning or current position is determined by the "origin" parameter which will be
+either drflac_seek_origin_start or drflac_seek_origin_current.
+
+When seeking to a PCM frame using drflac_seek_to_pcm_frame(), dr_flac may call this with an offset beyond the end of the FLAC stream. This needs to be detected
+and handled by returning DRFLAC_FALSE.
+*/
+typedef drflac_bool32 (* drflac_seek_proc)(void* pUserData, int offset, drflac_seek_origin origin);
+
+/*
+Callback for when a metadata block is read.
+
+
+Parameters
+----------
+pUserData (in)
+    The user data that was passed to drflac_open() and family.
+
+pMetadata (in)
+    A pointer to a structure containing the data of the metadata block.
+
+
+Remarks
+-------
+Use pMetadata->type to determine which metadata block is being handled and how to read the data. This
+will be set to one of the DRFLAC_METADATA_BLOCK_TYPE_* tokens.
+*/
+typedef void (* drflac_meta_proc)(void* pUserData, drflac_metadata* pMetadata);
+
+
+/* Structure for internal use. Only used for decoders opened with drflac_open_memory. */
+typedef struct
+{
+    const drflac_uint8* data;
+    size_t dataSize;
+    size_t currentReadPos;
+} drflac__memory_stream;
+
+/* Structure for internal use. Used for bit streaming. */
+typedef struct
+{
+    /* The function to call when more data needs to be read. */
+    drflac_read_proc onRead;
+
+    /* The function to call when the current read position needs to be moved. */
+    drflac_seek_proc onSeek;
+
+    /* The user data to pass around to onRead and onSeek. */
+    void* pUserData;
+
+
+    /*
+    The number of unaligned bytes in the L2 cache. This will always be 0 until the end of the stream is hit. At the end of the
+    stream there will be a number of bytes that don't cleanly fit in an L1 cache line, so we use this variable to know whether
+    or not the bistreamer needs to run on a slower path to read those last bytes. This will never be more than sizeof(drflac_cache_t).
+    */
+    size_t unalignedByteCount;
+
+    /* The content of the unaligned bytes. */
+    drflac_cache_t unalignedCache;
+
+    /* The index of the next valid cache line in the "L2" cache. */
+    drflac_uint32 nextL2Line;
+
+    /* The number of bits that have been consumed by the cache. This is used to determine how many valid bits are remaining. */
+    drflac_uint32 consumedBits;
+
+    /*
+    The cached data which was most recently read from the client. There are two levels of cache. Data flows as such:
+    Client -> L2 -> L1. The L2 -> L1 movement is aligned and runs on a fast path in just a few instructions.
+    */
+    drflac_cache_t cacheL2[DR_FLAC_BUFFER_SIZE/sizeof(drflac_cache_t)];
+    drflac_cache_t cache;
+
+    /*
+    CRC-16. This is updated whenever bits are read from the bit stream. Manually set this to 0 to reset the CRC. For FLAC, this
+    is reset to 0 at the beginning of each frame.
+    */
+    drflac_uint16 crc16;
+    drflac_cache_t crc16Cache;              /* A cache for optimizing CRC calculations. This is filled when when the L1 cache is reloaded. */
+    drflac_uint32 crc16CacheIgnoredBytes;   /* The number of bytes to ignore when updating the CRC-16 from the CRC-16 cache. */
+} drflac_bs;
+
+typedef struct
+{
+    /* The type of the subframe: SUBFRAME_CONSTANT, SUBFRAME_VERBATIM, SUBFRAME_FIXED or SUBFRAME_LPC. */
+    drflac_uint8 subframeType;
+
+    /* The number of wasted bits per sample as specified by the sub-frame header. */
+    drflac_uint8 wastedBitsPerSample;
+
+    /* The order to use for the prediction stage for SUBFRAME_FIXED and SUBFRAME_LPC. */
+    drflac_uint8 lpcOrder;
+
+    /* A pointer to the buffer containing the decoded samples in the subframe. This pointer is an offset from drflac::pExtraData. */
+    drflac_int32* pSamplesS32;
+} drflac_subframe;
+
+typedef struct
+{
+    /*
+    If the stream uses variable block sizes, this will be set to the index of the first PCM frame. If fixed block sizes are used, this will
+    always be set to 0. This is 64-bit because the decoded PCM frame number will be 36 bits.
+    */
+    drflac_uint64 pcmFrameNumber;
+
+    /*
+    If the stream uses fixed block sizes, this will be set to the frame number. If variable block sizes are used, this will always be 0. This
+    is 32-bit because in fixed block sizes, the maximum frame number will be 31 bits.
+    */
+    drflac_uint32 flacFrameNumber;
+
+    /* The sample rate of this frame. */
+    drflac_uint32 sampleRate;
+
+    /* The number of PCM frames in each sub-frame within this frame. */
+    drflac_uint16 blockSizeInPCMFrames;
+
+    /*
+    The channel assignment of this frame. This is not always set to the channel count. If interchannel decorrelation is being used this
+    will be set to DRFLAC_CHANNEL_ASSIGNMENT_LEFT_SIDE, DRFLAC_CHANNEL_ASSIGNMENT_RIGHT_SIDE or DRFLAC_CHANNEL_ASSIGNMENT_MID_SIDE.
+    */
+    drflac_uint8 channelAssignment;
+
+    /* The number of bits per sample within this frame. */
+    drflac_uint8 bitsPerSample;
+
+    /* The frame's CRC. */
+    drflac_uint8 crc8;
+} drflac_frame_header;
+
+typedef struct
+{
+    /* The header. */
+    drflac_frame_header header;
+
+    /*
+    The number of PCM frames left to be read in this FLAC frame. This is initially set to the block size. As PCM frames are read,
+    this will be decremented. When it reaches 0, the decoder will see this frame as fully consumed and load the next frame.
+    */
+    drflac_uint32 pcmFramesRemaining;
+
+    /* The list of sub-frames within the frame. There is one sub-frame for each channel, and there's a maximum of 8 channels. */
+    drflac_subframe subframes[8];
+} drflac_frame;
+
+typedef struct
+{
+    /* The function to call when a metadata block is read. */
+    drflac_meta_proc onMeta;
+
+    /* The user data posted to the metadata callback function. */
+    void* pUserDataMD;
+
+    /* Memory allocation callbacks. */
+    drflac_allocation_callbacks allocationCallbacks;
+
+
+    /* The sample rate. Will be set to something like 44100. */
+    drflac_uint32 sampleRate;
+
+    /*
+    The number of channels. This will be set to 1 for monaural streams, 2 for stereo, etc. Maximum 8. This is set based on the
+    value specified in the STREAMINFO block.
+    */
+    drflac_uint8 channels;
+
+    /* The bits per sample. Will be set to something like 16, 24, etc. */
+    drflac_uint8 bitsPerSample;
+
+    /* The maximum block size, in samples. This number represents the number of samples in each channel (not combined). */
+    drflac_uint16 maxBlockSizeInPCMFrames;
+
+    /*
+    The total number of PCM Frames making up the stream. Can be 0 in which case it's still a valid stream, but just means
+    the total PCM frame count is unknown. Likely the case with streams like internet radio.
+    */
+    drflac_uint64 totalPCMFrameCount;
+
+
+    /* The container type. This is set based on whether or not the decoder was opened from a native or Ogg stream. */
+    drflac_container container;
+
+    /* The number of seekpoints in the seektable. */
+    drflac_uint32 seekpointCount;
+
+
+    /* Information about the frame the decoder is currently sitting on. */
+    drflac_frame currentFLACFrame;
+
+
+    /* The index of the PCM frame the decoder is currently sitting on. This is only used for seeking. */
+    drflac_uint64 currentPCMFrame;
+
+    /* The position of the first FLAC frame in the stream. This is only ever used for seeking. */
+    drflac_uint64 firstFLACFramePosInBytes;
+
+
+    /* A hack to avoid a malloc() when opening a decoder with drflac_open_memory(). */
+    drflac__memory_stream memoryStream;
+
+
+    /* A pointer to the decoded sample data. This is an offset of pExtraData. */
+    drflac_int32* pDecodedSamples;
+
+    /* A pointer to the seek table. This is an offset of pExtraData, or NULL if there is no seek table. */
+    drflac_seekpoint* pSeekpoints;
+
+    /* Internal use only. Only used with Ogg containers. Points to a drflac_oggbs object. This is an offset of pExtraData. */
+    void* _oggbs;
+
+    /* Internal use only. Used for profiling and testing different seeking modes. */
+    drflac_bool32 _noSeekTableSeek    : 1;
+    drflac_bool32 _noBinarySearchSeek : 1;
+    drflac_bool32 _noBruteForceSeek   : 1;
+
+    /* The bit streamer. The raw FLAC data is fed through this object. */
+    drflac_bs bs;
+
+    /* Variable length extra data. We attach this to the end of the object so we can avoid unnecessary mallocs. */
+    drflac_uint8 pExtraData[1];
+} drflac;
+
+
+/*
+Opens a FLAC decoder.
+
+
+Parameters
+----------
+onRead (in)
+    The function to call when data needs to be read from the client.
+
+onSeek (in)
+    The function to call when the read position of the client data needs to move.
+
+pUserData (in, optional)
+    A pointer to application defined data that will be passed to onRead and onSeek.
+
+pAllocationCallbacks (in, optional)
+    A pointer to application defined callbacks for managing memory allocations.
+
+
+Return Value
+------------
+Returns a pointer to an object representing the decoder.
+
+
+Remarks
+-------
+Close the decoder with `drflac_close()`.
+
+`pAllocationCallbacks` can be NULL in which case it will use `DRFLAC_MALLOC`, `DRFLAC_REALLOC` and `DRFLAC_FREE`.
+
+This function will automatically detect whether or not you are attempting to open a native or Ogg encapsulated FLAC, both of which should work seamlessly
+without any manual intervention. Ogg encapsulation also works with multiplexed streams which basically means it can play FLAC encoded audio tracks in videos.
+
+This is the lowest level function for opening a FLAC stream. You can also use `drflac_open_file()` and `drflac_open_memory()` to open the stream from a file or
+from a block of memory respectively.
+
+The STREAMINFO block must be present for this to succeed. Use `drflac_open_relaxed()` to open a FLAC stream where the header may not be present.
+
+Use `drflac_open_with_metadata()` if you need access to metadata.
+
+
+Seek Also
+---------
+drflac_open_file()
+drflac_open_memory()
+drflac_open_with_metadata()
+drflac_close()
+*/
+DRFLAC_API drflac* drflac_open(drflac_read_proc onRead, drflac_seek_proc onSeek, void* pUserData, const drflac_allocation_callbacks* pAllocationCallbacks);
+
+/*
+Opens a FLAC stream with relaxed validation of the header block.
+
+
+Parameters
+----------
+onRead (in)
+    The function to call when data needs to be read from the client.
+
+onSeek (in)
+    The function to call when the read position of the client data needs to move.
+
+container (in)
+    Whether or not the FLAC stream is encapsulated using standard FLAC encapsulation or Ogg encapsulation.
+
+pUserData (in, optional)
+    A pointer to application defined data that will be passed to onRead and onSeek.
+
+pAllocationCallbacks (in, optional)
+    A pointer to application defined callbacks for managing memory allocations.
+
+
+Return Value
+------------
+A pointer to an object representing the decoder.
+
+
+Remarks
+-------
+The same as drflac_open(), except attempts to open the stream even when a header block is not present.
+
+Because the header is not necessarily available, the caller must explicitly define the container (Native or Ogg). Do not set this to `drflac_container_unknown`
+as that is for internal use only.
+
+Opening in relaxed mode will continue reading data from onRead until it finds a valid frame. If a frame is never found it will continue forever. To abort,
+force your `onRead` callback to return 0, which dr_flac will use as an indicator that the end of the stream was found.
+
+Use `drflac_open_with_metadata_relaxed()` if you need access to metadata.
+*/
+DRFLAC_API drflac* drflac_open_relaxed(drflac_read_proc onRead, drflac_seek_proc onSeek, drflac_container container, void* pUserData, const drflac_allocation_callbacks* pAllocationCallbacks);
+
+/*
+Opens a FLAC decoder and notifies the caller of the metadata chunks (album art, etc.).
+
+
+Parameters
+----------
+onRead (in)
+    The function to call when data needs to be read from the client.
+
+onSeek (in)
+    The function to call when the read position of the client data needs to move.
+
+onMeta (in)
+    The function to call for every metadata block.
+
+pUserData (in, optional)
+    A pointer to application defined data that will be passed to onRead, onSeek and onMeta.
+
+pAllocationCallbacks (in, optional)
+    A pointer to application defined callbacks for managing memory allocations.
+
+
+Return Value
+------------
+A pointer to an object representing the decoder.
+
+
+Remarks
+-------
+Close the decoder with `drflac_close()`.
+
+`pAllocationCallbacks` can be NULL in which case it will use `DRFLAC_MALLOC`, `DRFLAC_REALLOC` and `DRFLAC_FREE`.
+
+This is slower than `drflac_open()`, so avoid this one if you don't need metadata. Internally, this will allocate and free memory on the heap for every
+metadata block except for STREAMINFO and PADDING blocks.
+
+The caller is notified of the metadata via the `onMeta` callback. All metadata blocks will be handled before the function returns. This callback takes a
+pointer to a `drflac_metadata` object which is a union containing the data of all relevant metadata blocks. Use the `type` member to discriminate against
+the different metadata types.
+
+The STREAMINFO block must be present for this to succeed. Use `drflac_open_with_metadata_relaxed()` to open a FLAC stream where the header may not be present.
+
+Note that this will behave inconsistently with `drflac_open()` if the stream is an Ogg encapsulated stream and a metadata block is corrupted. This is due to
+the way the Ogg stream recovers from corrupted pages. When `drflac_open_with_metadata()` is being used, the open routine will try to read the contents of the
+metadata block, whereas `drflac_open()` will simply seek past it (for the sake of efficiency). This inconsistency can result in different samples being
+returned depending on whether or not the stream is being opened with metadata.
+
+
+Seek Also
+---------
+drflac_open_file_with_metadata()
+drflac_open_memory_with_metadata()
+drflac_open()
+drflac_close()
+*/
+DRFLAC_API drflac* drflac_open_with_metadata(drflac_read_proc onRead, drflac_seek_proc onSeek, drflac_meta_proc onMeta, void* pUserData, const drflac_allocation_callbacks* pAllocationCallbacks);
+
+/*
+The same as drflac_open_with_metadata(), except attempts to open the stream even when a header block is not present.
+
+See Also
+--------
+drflac_open_with_metadata()
+drflac_open_relaxed()
+*/
+DRFLAC_API drflac* drflac_open_with_metadata_relaxed(drflac_read_proc onRead, drflac_seek_proc onSeek, drflac_meta_proc onMeta, drflac_container container, void* pUserData, const drflac_allocation_callbacks* pAllocationCallbacks);
+
+/*
+Closes the given FLAC decoder.
+
+
+Parameters
+----------
+pFlac (in)
+    The decoder to close.
+
+
+Remarks
+-------
+This will destroy the decoder object.
+
+
+See Also
+--------
+drflac_open()
+drflac_open_with_metadata()
+drflac_open_file()
+drflac_open_file_w()
+drflac_open_file_with_metadata()
+drflac_open_file_with_metadata_w()
+drflac_open_memory()
+drflac_open_memory_with_metadata()
+*/
+DRFLAC_API void drflac_close(drflac* pFlac);
+
+
+/*
+Reads sample data from the given FLAC decoder, output as interleaved signed 32-bit PCM.
+
+
+Parameters
+----------
+pFlac (in)
+    The decoder.
+
+framesToRead (in)
+    The number of PCM frames to read.
+
+pBufferOut (out, optional)
+    A pointer to the buffer that will receive the decoded samples.
+
+
+Return Value
+------------
+Returns the number of PCM frames actually read. If the return value is less than `framesToRead` it has reached the end.
+
+
+Remarks
+-------
+pBufferOut can be null, in which case the call will act as a seek, and the return value will be the number of frames seeked.
+*/
+DRFLAC_API drflac_uint64 drflac_read_pcm_frames_s32(drflac* pFlac, drflac_uint64 framesToRead, drflac_int32* pBufferOut);
+
+
+/*
+Reads sample data from the given FLAC decoder, output as interleaved signed 16-bit PCM.
+
+
+Parameters
+----------
+pFlac (in)
+    The decoder.
+
+framesToRead (in)
+    The number of PCM frames to read.
+
+pBufferOut (out, optional)
+    A pointer to the buffer that will receive the decoded samples.
+
+
+Return Value
+------------
+Returns the number of PCM frames actually read. If the return value is less than `framesToRead` it has reached the end.
+
+
+Remarks
+-------
+pBufferOut can be null, in which case the call will act as a seek, and the return value will be the number of frames seeked.
+
+Note that this is lossy for streams where the bits per sample is larger than 16.
+*/
+DRFLAC_API drflac_uint64 drflac_read_pcm_frames_s16(drflac* pFlac, drflac_uint64 framesToRead, drflac_int16* pBufferOut);
+
+/*
+Reads sample data from the given FLAC decoder, output as interleaved 32-bit floating point PCM.
+
+
+Parameters
+----------
+pFlac (in)
+    The decoder.
+
+framesToRead (in)
+    The number of PCM frames to read.
+
+pBufferOut (out, optional)
+    A pointer to the buffer that will receive the decoded samples.
+
+
+Return Value
+------------
+Returns the number of PCM frames actually read. If the return value is less than `framesToRead` it has reached the end.
+
+
+Remarks
+-------
+pBufferOut can be null, in which case the call will act as a seek, and the return value will be the number of frames seeked.
+
+Note that this should be considered lossy due to the nature of floating point numbers not being able to exactly represent every possible number.
+*/
+DRFLAC_API drflac_uint64 drflac_read_pcm_frames_f32(drflac* pFlac, drflac_uint64 framesToRead, float* pBufferOut);
+
+/*
+Seeks to the PCM frame at the given index.
+
+
+Parameters
+----------
+pFlac (in)
+    The decoder.
+
+pcmFrameIndex (in)
+    The index of the PCM frame to seek to. See notes below.
+
+
+Return Value
+-------------
+`DRFLAC_TRUE` if successful; `DRFLAC_FALSE` otherwise.
+*/
+DRFLAC_API drflac_bool32 drflac_seek_to_pcm_frame(drflac* pFlac, drflac_uint64 pcmFrameIndex);
+
+
+
+#ifndef DR_FLAC_NO_STDIO
+/*
+Opens a FLAC decoder from the file at the given path.
+
+
+Parameters
+----------
+pFileName (in)
+    The path of the file to open, either absolute or relative to the current directory.
+
+pAllocationCallbacks (in, optional)
+    A pointer to application defined callbacks for managing memory allocations.
+
+
+Return Value
+------------
+A pointer to an object representing the decoder.
+
+
+Remarks
+-------
+Close the decoder with drflac_close().
+
+
+Remarks
+-------
+This will hold a handle to the file until the decoder is closed with drflac_close(). Some platforms will restrict the number of files a process can have open
+at any given time, so keep this mind if you have many decoders open at the same time.
+
+
+See Also
+--------
+drflac_open_file_with_metadata()
+drflac_open()
+drflac_close()
+*/
+DRFLAC_API drflac* drflac_open_file(const char* pFileName, const drflac_allocation_callbacks* pAllocationCallbacks);
+DRFLAC_API drflac* drflac_open_file_w(const wchar_t* pFileName, const drflac_allocation_callbacks* pAllocationCallbacks);
+
+/*
+Opens a FLAC decoder from the file at the given path and notifies the caller of the metadata chunks (album art, etc.)
+
+
+Parameters
+----------
+pFileName (in)
+    The path of the file to open, either absolute or relative to the current directory.
+
+pAllocationCallbacks (in, optional)
+    A pointer to application defined callbacks for managing memory allocations.
+
+onMeta (in)
+    The callback to fire for each metadata block.
+
+pUserData (in)
+    A pointer to the user data to pass to the metadata callback.
+
+pAllocationCallbacks (in)
+    A pointer to application defined callbacks for managing memory allocations.
+
+
+Remarks
+-------
+Look at the documentation for drflac_open_with_metadata() for more information on how metadata is handled.
+
+
+See Also
+--------
+drflac_open_with_metadata()
+drflac_open()
+drflac_close()
+*/
+DRFLAC_API drflac* drflac_open_file_with_metadata(const char* pFileName, drflac_meta_proc onMeta, void* pUserData, const drflac_allocation_callbacks* pAllocationCallbacks);
+DRFLAC_API drflac* drflac_open_file_with_metadata_w(const wchar_t* pFileName, drflac_meta_proc onMeta, void* pUserData, const drflac_allocation_callbacks* pAllocationCallbacks);
+#endif
+
+/*
+Opens a FLAC decoder from a pre-allocated block of memory
+
+
+Parameters
+----------
+pData (in)
+    A pointer to the raw encoded FLAC data.
+
+dataSize (in)
+    The size in bytes of `data`.
+
+pAllocationCallbacks (in)
+    A pointer to application defined callbacks for managing memory allocations.
+
+
+Return Value
+------------
+A pointer to an object representing the decoder.
+
+
+Remarks
+-------
+This does not create a copy of the data. It is up to the application to ensure the buffer remains valid for the lifetime of the decoder.
+
+
+See Also
+--------
+drflac_open()
+drflac_close()
+*/
+DRFLAC_API drflac* drflac_open_memory(const void* pData, size_t dataSize, const drflac_allocation_callbacks* pAllocationCallbacks);
+
+/*
+Opens a FLAC decoder from a pre-allocated block of memory and notifies the caller of the metadata chunks (album art, etc.)
+
+
+Parameters
+----------
+pData (in)
+    A pointer to the raw encoded FLAC data.
+
+dataSize (in)
+    The size in bytes of `data`.
+
+onMeta (in)
+    The callback to fire for each metadata block.
+
+pUserData (in)
+    A pointer to the user data to pass to the metadata callback.
+
+pAllocationCallbacks (in)
+    A pointer to application defined callbacks for managing memory allocations.
+
+
+Remarks
+-------
+Look at the documentation for drflac_open_with_metadata() for more information on how metadata is handled.
+
+
+See Also
+-------
+drflac_open_with_metadata()
+drflac_open()
+drflac_close()
+*/
+DRFLAC_API drflac* drflac_open_memory_with_metadata(const void* pData, size_t dataSize, drflac_meta_proc onMeta, void* pUserData, const drflac_allocation_callbacks* pAllocationCallbacks);
+
+
+
+/* High Level APIs */
+
+/*
+Opens a FLAC stream from the given callbacks and fully decodes it in a single operation. The return value is a
+pointer to the sample data as interleaved signed 32-bit PCM. The returned data must be freed with drflac_free().
+
+You can pass in custom memory allocation callbacks via the pAllocationCallbacks parameter. This can be NULL in which
+case it will use DRFLAC_MALLOC, DRFLAC_REALLOC and DRFLAC_FREE.
+
+Sometimes a FLAC file won't keep track of the total sample count. In this situation the function will continuously
+read samples into a dynamically sized buffer on the heap until no samples are left.
+
+Do not call this function on a broadcast type of stream (like internet radio streams and whatnot).
+*/
+DRFLAC_API drflac_int32* drflac_open_and_read_pcm_frames_s32(drflac_read_proc onRead, drflac_seek_proc onSeek, void* pUserData, unsigned int* channels, unsigned int* sampleRate, drflac_uint64* totalPCMFrameCount, const drflac_allocation_callbacks* pAllocationCallbacks);
+
+/* Same as drflac_open_and_read_pcm_frames_s32(), except returns signed 16-bit integer samples. */
+DRFLAC_API drflac_int16* drflac_open_and_read_pcm_frames_s16(drflac_read_proc onRead, drflac_seek_proc onSeek, void* pUserData, unsigned int* channels, unsigned int* sampleRate, drflac_uint64* totalPCMFrameCount, const drflac_allocation_callbacks* pAllocationCallbacks);
+
+/* Same as drflac_open_and_read_pcm_frames_s32(), except returns 32-bit floating-point samples. */
+DRFLAC_API float* drflac_open_and_read_pcm_frames_f32(drflac_read_proc onRead, drflac_seek_proc onSeek, void* pUserData, unsigned int* channels, unsigned int* sampleRate, drflac_uint64* totalPCMFrameCount, const drflac_allocation_callbacks* pAllocationCallbacks);
+
+#ifndef DR_FLAC_NO_STDIO
+/* Same as drflac_open_and_read_pcm_frames_s32() except opens the decoder from a file. */
+DRFLAC_API drflac_int32* drflac_open_file_and_read_pcm_frames_s32(const char* filename, unsigned int* channels, unsigned int* sampleRate, drflac_uint64* totalPCMFrameCount, const drflac_allocation_callbacks* pAllocationCallbacks);
+
+/* Same as drflac_open_file_and_read_pcm_frames_s32(), except returns signed 16-bit integer samples. */
+DRFLAC_API drflac_int16* drflac_open_file_and_read_pcm_frames_s16(const char* filename, unsigned int* channels, unsigned int* sampleRate, drflac_uint64* totalPCMFrameCount, const drflac_allocation_callbacks* pAllocationCallbacks);
+
+/* Same as drflac_open_file_and_read_pcm_frames_s32(), except returns 32-bit floating-point samples. */
+DRFLAC_API float* drflac_open_file_and_read_pcm_frames_f32(const char* filename, unsigned int* channels, unsigned int* sampleRate, drflac_uint64* totalPCMFrameCount, const drflac_allocation_callbacks* pAllocationCallbacks);
+#endif
+
+/* Same as drflac_open_and_read_pcm_frames_s32() except opens the decoder from a block of memory. */
+DRFLAC_API drflac_int32* drflac_open_memory_and_read_pcm_frames_s32(const void* data, size_t dataSize, unsigned int* channels, unsigned int* sampleRate, drflac_uint64* totalPCMFrameCount, const drflac_allocation_callbacks* pAllocationCallbacks);
+
+/* Same as drflac_open_memory_and_read_pcm_frames_s32(), except returns signed 16-bit integer samples. */
+DRFLAC_API drflac_int16* drflac_open_memory_and_read_pcm_frames_s16(const void* data, size_t dataSize, unsigned int* channels, unsigned int* sampleRate, drflac_uint64* totalPCMFrameCount, const drflac_allocation_callbacks* pAllocationCallbacks);
+
+/* Same as drflac_open_memory_and_read_pcm_frames_s32(), except returns 32-bit floating-point samples. */
+DRFLAC_API float* drflac_open_memory_and_read_pcm_frames_f32(const void* data, size_t dataSize, unsigned int* channels, unsigned int* sampleRate, drflac_uint64* totalPCMFrameCount, const drflac_allocation_callbacks* pAllocationCallbacks);
+
+/*
+Frees memory that was allocated internally by dr_flac.
+
+Set pAllocationCallbacks to the same object that was passed to drflac_open_*_and_read_pcm_frames_*(). If you originally passed in NULL, pass in NULL for this.
+*/
+DRFLAC_API void drflac_free(void* p, const drflac_allocation_callbacks* pAllocationCallbacks);
+
+
+/* Structure representing an iterator for vorbis comments in a VORBIS_COMMENT metadata block. */
+typedef struct
+{
+    drflac_uint32 countRemaining;
+    const char* pRunningData;
+} drflac_vorbis_comment_iterator;
+
+/*
+Initializes a vorbis comment iterator. This can be used for iterating over the vorbis comments in a VORBIS_COMMENT
+metadata block.
+*/
+DRFLAC_API void drflac_init_vorbis_comment_iterator(drflac_vorbis_comment_iterator* pIter, drflac_uint32 commentCount, const void* pComments);
+
+/*
+Goes to the next vorbis comment in the given iterator. If null is returned it means there are no more comments. The
+returned string is NOT null terminated.
+*/
+DRFLAC_API const char* drflac_next_vorbis_comment(drflac_vorbis_comment_iterator* pIter, drflac_uint32* pCommentLengthOut);
+
+
+/* Structure representing an iterator for cuesheet tracks in a CUESHEET metadata block. */
+typedef struct
+{
+    drflac_uint32 countRemaining;
+    const char* pRunningData;
+} drflac_cuesheet_track_iterator;
+
+/* The order of members here is important because we map this directly to the raw data within the CUESHEET metadata block. */
+typedef struct
+{
+    drflac_uint64 offset;
+    drflac_uint8 index;
+    drflac_uint8 reserved[3];
+} drflac_cuesheet_track_index;
+
+typedef struct
+{
+    drflac_uint64 offset;
+    drflac_uint8 trackNumber;
+    char ISRC[12];
+    drflac_bool8 isAudio;
+    drflac_bool8 preEmphasis;
+    drflac_uint8 indexCount;
+    const drflac_cuesheet_track_index* pIndexPoints;
+} drflac_cuesheet_track;
+
+/*
+Initializes a cuesheet track iterator. This can be used for iterating over the cuesheet tracks in a CUESHEET metadata
+block.
+*/
+DRFLAC_API void drflac_init_cuesheet_track_iterator(drflac_cuesheet_track_iterator* pIter, drflac_uint32 trackCount, const void* pTrackData);
+
+/* Goes to the next cuesheet track in the given iterator. If DRFLAC_FALSE is returned it means there are no more comments. */
+DRFLAC_API drflac_bool32 drflac_next_cuesheet_track(drflac_cuesheet_track_iterator* pIter, drflac_cuesheet_track* pCuesheetTrack);
+
+
+#ifdef __cplusplus
+}
+#endif
+#endif  /* dr_flac_h */
+
+
+/************************************************************************************************************************************************************
+ ************************************************************************************************************************************************************
+
+ IMPLEMENTATION
+
+ ************************************************************************************************************************************************************
+ ************************************************************************************************************************************************************/
+#if defined(DR_FLAC_IMPLEMENTATION) || defined(DRFLAC_IMPLEMENTATION)
+#ifndef dr_flac_c
+#define dr_flac_c
+
+/* Disable some annoying warnings. */
+#if defined(__clang__) || (defined(__GNUC__) && (__GNUC__ > 4 || (__GNUC__ == 4 && __GNUC_MINOR__ >= 6)))
+    #pragma GCC diagnostic push
+    #if __GNUC__ >= 7
+    #pragma GCC diagnostic ignored "-Wimplicit-fallthrough"
+    #endif
+#endif
+
+#ifdef __linux__
+    #ifndef _BSD_SOURCE
+        #define _BSD_SOURCE
+    #endif
+    #ifndef _DEFAULT_SOURCE
+        #define _DEFAULT_SOURCE
+    #endif
+    #ifndef __USE_BSD
+        #define __USE_BSD
+    #endif
+    #include <endian.h>
+#endif
+
+#include <stdlib.h>
+#include <string.h>
+
+/* Inline */
+#ifdef _MSC_VER
+    #define DRFLAC_INLINE __forceinline
+#elif defined(__GNUC__)
+    /*
+    I've had a bug report where GCC is emitting warnings about functions possibly not being inlineable. This warning happens when
+    the __attribute__((always_inline)) attribute is defined without an "inline" statement. I think therefore there must be some
+    case where "__inline__" is not always defined, thus the compiler emitting these warnings. When using -std=c89 or -ansi on the
+    command line, we cannot use the "inline" keyword and instead need to use "__inline__". In an attempt to work around this issue
+    I am using "__inline__" only when we're compiling in strict ANSI mode.
+    */
+    #if defined(__STRICT_ANSI__)
+        #define DRFLAC_GNUC_INLINE_HINT __inline__
+    #else
+        #define DRFLAC_GNUC_INLINE_HINT inline
+    #endif
+
+    #if (__GNUC__ > 3 || (__GNUC__ == 3 && __GNUC_MINOR__ >= 2)) || defined(__clang__)
+        #define DRFLAC_INLINE DRFLAC_GNUC_INLINE_HINT __attribute__((always_inline))
+    #else
+        #define DRFLAC_INLINE DRFLAC_GNUC_INLINE_HINT
+    #endif
+#elif defined(__WATCOMC__)
+    #define DRFLAC_INLINE __inline
+#else
+    #define DRFLAC_INLINE
+#endif
+/* End Inline */
+
+/*
+Intrinsics Support
+
+There's a bug in GCC 4.2.x which results in an incorrect compilation error when using _mm_slli_epi32() where it complains with
+
+    "error: shift must be an immediate"
+
+Unfortuantely dr_flac depends on this for a few things so we're just going to disable SSE on GCC 4.2 and below.
+*/
+#if !defined(DR_FLAC_NO_SIMD)
+    #if defined(DRFLAC_X64) || defined(DRFLAC_X86)
+        #if defined(_MSC_VER) && !defined(__clang__)
+            /* MSVC. */
+            #if _MSC_VER >= 1400 && !defined(DRFLAC_NO_SSE2)    /* 2005 */
+                #define DRFLAC_SUPPORT_SSE2
+            #endif
+            #if _MSC_VER >= 1600 && !defined(DRFLAC_NO_SSE41)   /* 2010 */
+                #define DRFLAC_SUPPORT_SSE41
+            #endif
+        #elif defined(__clang__) || (defined(__GNUC__) && (__GNUC__ > 4 || (__GNUC__ == 4 && __GNUC_MINOR__ >= 3)))
+            /* Assume GNUC-style. */
+            #if defined(__SSE2__) && !defined(DRFLAC_NO_SSE2)
+                #define DRFLAC_SUPPORT_SSE2
+            #endif
+            #if defined(__SSE4_1__) && !defined(DRFLAC_NO_SSE41)
+                #define DRFLAC_SUPPORT_SSE41
+            #endif
+        #endif
+
+        /* If at this point we still haven't determined compiler support for the intrinsics just fall back to __has_include. */
+        #if !defined(__GNUC__) && !defined(__clang__) && defined(__has_include)
+            #if !defined(DRFLAC_SUPPORT_SSE2) && !defined(DRFLAC_NO_SSE2) && __has_include(<emmintrin.h>)
+                #define DRFLAC_SUPPORT_SSE2
+            #endif
+            #if !defined(DRFLAC_SUPPORT_SSE41) && !defined(DRFLAC_NO_SSE41) && __has_include(<smmintrin.h>)
+                #define DRFLAC_SUPPORT_SSE41
+            #endif
+        #endif
+
+        #if defined(DRFLAC_SUPPORT_SSE41)
+            #include <smmintrin.h>
+        #elif defined(DRFLAC_SUPPORT_SSE2)
+            #include <emmintrin.h>
+        #endif
+    #endif
+
+    #if defined(DRFLAC_ARM)
+        #if !defined(DRFLAC_NO_NEON) && (defined(__ARM_NEON) || defined(__aarch64__) || defined(_M_ARM64))
+            #define DRFLAC_SUPPORT_NEON
+            #include <arm_neon.h>
+        #endif
+    #endif
+#endif
+
+/* Compile-time CPU feature support. */
+#if !defined(DR_FLAC_NO_SIMD) && (defined(DRFLAC_X86) || defined(DRFLAC_X64))
+    #if defined(_MSC_VER) && !defined(__clang__)
+        #if _MSC_VER >= 1400
+            #include <intrin.h>
+            static void drflac__cpuid(int info[4], int fid)
+            {
+                __cpuid(info, fid);
+            }
+        #else
+            #define DRFLAC_NO_CPUID
+        #endif
+    #else
+        #if defined(__GNUC__) || defined(__clang__)
+            static void drflac__cpuid(int info[4], int fid)
+            {
+                /*
+                It looks like the -fPIC option uses the ebx register which GCC complains about. We can work around this by just using a different register, the
+                specific register of which I'm letting the compiler decide on. The "k" prefix is used to specify a 32-bit register. The {...} syntax is for
+                supporting different assembly dialects.
+
+                What's basically happening is that we're saving and restoring the ebx register manually.
+                */
+                #if defined(DRFLAC_X86) && defined(__PIC__)
+                    __asm__ __volatile__ (
+                        "xchg{l} {%%}ebx, %k1;"
+                        "cpuid;"
+                        "xchg{l} {%%}ebx, %k1;"
+                        : "=a"(info[0]), "=&r"(info[1]), "=c"(info[2]), "=d"(info[3]) : "a"(fid), "c"(0)
+                    );
+                #else
+                    __asm__ __volatile__ (
+                        "cpuid" : "=a"(info[0]), "=b"(info[1]), "=c"(info[2]), "=d"(info[3]) : "a"(fid), "c"(0)
+                    );
+                #endif
+            }
+        #else
+            #define DRFLAC_NO_CPUID
+        #endif
+    #endif
+#else
+    #define DRFLAC_NO_CPUID
+#endif
+
+static DRFLAC_INLINE drflac_bool32 drflac_has_sse2(void)
+{
+#if defined(DRFLAC_SUPPORT_SSE2)
+    #if (defined(DRFLAC_X64) || defined(DRFLAC_X86)) && !defined(DRFLAC_NO_SSE2)
+        #if defined(DRFLAC_X64)
+            return DRFLAC_TRUE;    /* 64-bit targets always support SSE2. */
+        #elif (defined(_M_IX86_FP) && _M_IX86_FP == 2) || defined(__SSE2__)
+            return DRFLAC_TRUE;    /* If the compiler is allowed to freely generate SSE2 code we can assume support. */
+        #else
+            #if defined(DRFLAC_NO_CPUID)
+                return DRFLAC_FALSE;
+            #else
+                int info[4];
+                drflac__cpuid(info, 1);
+                return (info[3] & (1 << 26)) != 0;
+            #endif
+        #endif
+    #else
+        return DRFLAC_FALSE;       /* SSE2 is only supported on x86 and x64 architectures. */
+    #endif
+#else
+    return DRFLAC_FALSE;           /* No compiler support. */
+#endif
+}
+
+static DRFLAC_INLINE drflac_bool32 drflac_has_sse41(void)
+{
+#if defined(DRFLAC_SUPPORT_SSE41)
+    #if (defined(DRFLAC_X64) || defined(DRFLAC_X86)) && !defined(DRFLAC_NO_SSE41)
+        #if defined(__SSE4_1__) || defined(__AVX__)
+            return DRFLAC_TRUE;    /* If the compiler is allowed to freely generate SSE41 code we can assume support. */
+        #else
+            #if defined(DRFLAC_NO_CPUID)
+                return DRFLAC_FALSE;
+            #else
+                int info[4];
+                drflac__cpuid(info, 1);
+                return (info[2] & (1 << 19)) != 0;
+            #endif
+        #endif
+    #else
+        return DRFLAC_FALSE;       /* SSE41 is only supported on x86 and x64 architectures. */
+    #endif
+#else
+    return DRFLAC_FALSE;           /* No compiler support. */
+#endif
+}
+
+
+#if defined(_MSC_VER) && _MSC_VER >= 1500 && (defined(DRFLAC_X86) || defined(DRFLAC_X64)) && !defined(__clang__)
+    #define DRFLAC_HAS_LZCNT_INTRINSIC
+#elif (defined(__GNUC__) && ((__GNUC__ > 4) || (__GNUC__ == 4 && __GNUC_MINOR__ >= 7)))
+    #define DRFLAC_HAS_LZCNT_INTRINSIC
+#elif defined(__clang__)
+    #if defined(__has_builtin)
+        #if __has_builtin(__builtin_clzll) || __has_builtin(__builtin_clzl)
+            #define DRFLAC_HAS_LZCNT_INTRINSIC
+        #endif
+    #endif
+#endif
+
+#if defined(_MSC_VER) && _MSC_VER >= 1400 && !defined(__clang__)
+    #define DRFLAC_HAS_BYTESWAP16_INTRINSIC
+    #define DRFLAC_HAS_BYTESWAP32_INTRINSIC
+    #define DRFLAC_HAS_BYTESWAP64_INTRINSIC
+#elif defined(__clang__)
+    #if defined(__has_builtin)
+        #if __has_builtin(__builtin_bswap16)
+            #define DRFLAC_HAS_BYTESWAP16_INTRINSIC
+        #endif
+        #if __has_builtin(__builtin_bswap32)
+            #define DRFLAC_HAS_BYTESWAP32_INTRINSIC
+        #endif
+        #if __has_builtin(__builtin_bswap64)
+            #define DRFLAC_HAS_BYTESWAP64_INTRINSIC
+        #endif
+    #endif
+#elif defined(__GNUC__)
+    #if ((__GNUC__ > 4) || (__GNUC__ == 4 && __GNUC_MINOR__ >= 3))
+        #define DRFLAC_HAS_BYTESWAP32_INTRINSIC
+        #define DRFLAC_HAS_BYTESWAP64_INTRINSIC
+    #endif
+    #if ((__GNUC__ > 4) || (__GNUC__ == 4 && __GNUC_MINOR__ >= 8))
+        #define DRFLAC_HAS_BYTESWAP16_INTRINSIC
+    #endif
+#elif defined(__WATCOMC__) && defined(__386__)
+    #define DRFLAC_HAS_BYTESWAP16_INTRINSIC
+    #define DRFLAC_HAS_BYTESWAP32_INTRINSIC
+    #define DRFLAC_HAS_BYTESWAP64_INTRINSIC
+    extern __inline drflac_uint16 _watcom_bswap16(drflac_uint16);
+    extern __inline drflac_uint32 _watcom_bswap32(drflac_uint32);
+    extern __inline drflac_uint64 _watcom_bswap64(drflac_uint64);
+#pragma aux _watcom_bswap16 = \
+    "xchg al, ah" \
+    parm  [ax]    \
+    value [ax]    \
+    modify nomemory;
+#pragma aux _watcom_bswap32 = \
+    "bswap eax" \
+    parm  [eax] \
+    value [eax] \
+    modify nomemory;
+#pragma aux _watcom_bswap64 = \
+    "bswap eax"     \
+    "bswap edx"     \
+    "xchg eax,edx"  \
+    parm [eax edx]  \
+    value [eax edx] \
+    modify nomemory;
+#endif
+
+
+/* Standard library stuff. */
+#ifndef DRFLAC_ASSERT
+#include <assert.h>
+#define DRFLAC_ASSERT(expression)           assert(expression)
+#endif
+#ifndef DRFLAC_MALLOC
+#define DRFLAC_MALLOC(sz)                   malloc((sz))
+#endif
+#ifndef DRFLAC_REALLOC
+#define DRFLAC_REALLOC(p, sz)               realloc((p), (sz))
+#endif
+#ifndef DRFLAC_FREE
+#define DRFLAC_FREE(p)                      free((p))
+#endif
+#ifndef DRFLAC_COPY_MEMORY
+#define DRFLAC_COPY_MEMORY(dst, src, sz)    memcpy((dst), (src), (sz))
+#endif
+#ifndef DRFLAC_ZERO_MEMORY
+#define DRFLAC_ZERO_MEMORY(p, sz)           memset((p), 0, (sz))
+#endif
+#ifndef DRFLAC_ZERO_OBJECT
+#define DRFLAC_ZERO_OBJECT(p)               DRFLAC_ZERO_MEMORY((p), sizeof(*(p)))
+#endif
+
+#define DRFLAC_MAX_SIMD_VECTOR_SIZE                     64  /* 64 for AVX-512 in the future. */
+
+/* Result Codes */
+typedef drflac_int32 drflac_result;
+#define DRFLAC_SUCCESS                                   0
+#define DRFLAC_ERROR                                    -1   /* A generic error. */
+#define DRFLAC_INVALID_ARGS                             -2
+#define DRFLAC_INVALID_OPERATION                        -3
+#define DRFLAC_OUT_OF_MEMORY                            -4
+#define DRFLAC_OUT_OF_RANGE                             -5
+#define DRFLAC_ACCESS_DENIED                            -6
+#define DRFLAC_DOES_NOT_EXIST                           -7
+#define DRFLAC_ALREADY_EXISTS                           -8
+#define DRFLAC_TOO_MANY_OPEN_FILES                      -9
+#define DRFLAC_INVALID_FILE                             -10
+#define DRFLAC_TOO_BIG                                  -11
+#define DRFLAC_PATH_TOO_LONG                            -12
+#define DRFLAC_NAME_TOO_LONG                            -13
+#define DRFLAC_NOT_DIRECTORY                            -14
+#define DRFLAC_IS_DIRECTORY                             -15
+#define DRFLAC_DIRECTORY_NOT_EMPTY                      -16
+#define DRFLAC_END_OF_FILE                              -17
+#define DRFLAC_NO_SPACE                                 -18
+#define DRFLAC_BUSY                                     -19
+#define DRFLAC_IO_ERROR                                 -20
+#define DRFLAC_INTERRUPT                                -21
+#define DRFLAC_UNAVAILABLE                              -22
+#define DRFLAC_ALREADY_IN_USE                           -23
+#define DRFLAC_BAD_ADDRESS                              -24
+#define DRFLAC_BAD_SEEK                                 -25
+#define DRFLAC_BAD_PIPE                                 -26
+#define DRFLAC_DEADLOCK                                 -27
+#define DRFLAC_TOO_MANY_LINKS                           -28
+#define DRFLAC_NOT_IMPLEMENTED                          -29
+#define DRFLAC_NO_MESSAGE                               -30
+#define DRFLAC_BAD_MESSAGE                              -31
+#define DRFLAC_NO_DATA_AVAILABLE                        -32
+#define DRFLAC_INVALID_DATA                             -33
+#define DRFLAC_TIMEOUT                                  -34
+#define DRFLAC_NO_NETWORK                               -35
+#define DRFLAC_NOT_UNIQUE                               -36
+#define DRFLAC_NOT_SOCKET                               -37
+#define DRFLAC_NO_ADDRESS                               -38
+#define DRFLAC_BAD_PROTOCOL                             -39
+#define DRFLAC_PROTOCOL_UNAVAILABLE                     -40
+#define DRFLAC_PROTOCOL_NOT_SUPPORTED                   -41
+#define DRFLAC_PROTOCOL_FAMILY_NOT_SUPPORTED            -42
+#define DRFLAC_ADDRESS_FAMILY_NOT_SUPPORTED             -43
+#define DRFLAC_SOCKET_NOT_SUPPORTED                     -44
+#define DRFLAC_CONNECTION_RESET                         -45
+#define DRFLAC_ALREADY_CONNECTED                        -46
+#define DRFLAC_NOT_CONNECTED                            -47
+#define DRFLAC_CONNECTION_REFUSED                       -48
+#define DRFLAC_NO_HOST                                  -49
+#define DRFLAC_IN_PROGRESS                              -50
+#define DRFLAC_CANCELLED                                -51
+#define DRFLAC_MEMORY_ALREADY_MAPPED                    -52
+#define DRFLAC_AT_END                                   -53
+
+#define DRFLAC_CRC_MISMATCH                             -100
+/* End Result Codes */
+
+
+#define DRFLAC_SUBFRAME_CONSTANT                        0
+#define DRFLAC_SUBFRAME_VERBATIM                        1
+#define DRFLAC_SUBFRAME_FIXED                           8
+#define DRFLAC_SUBFRAME_LPC                             32
+#define DRFLAC_SUBFRAME_RESERVED                        255
+
+#define DRFLAC_RESIDUAL_CODING_METHOD_PARTITIONED_RICE  0
+#define DRFLAC_RESIDUAL_CODING_METHOD_PARTITIONED_RICE2 1
+
+#define DRFLAC_CHANNEL_ASSIGNMENT_INDEPENDENT           0
+#define DRFLAC_CHANNEL_ASSIGNMENT_LEFT_SIDE             8
+#define DRFLAC_CHANNEL_ASSIGNMENT_RIGHT_SIDE            9
+#define DRFLAC_CHANNEL_ASSIGNMENT_MID_SIDE              10
+
+#define DRFLAC_SEEKPOINT_SIZE_IN_BYTES                  18
+#define DRFLAC_CUESHEET_TRACK_SIZE_IN_BYTES             36
+#define DRFLAC_CUESHEET_TRACK_INDEX_SIZE_IN_BYTES       12
+
+#define drflac_align(x, a)                              ((((x) + (a) - 1) / (a)) * (a))
+
+
+DRFLAC_API void drflac_version(drflac_uint32* pMajor, drflac_uint32* pMinor, drflac_uint32* pRevision)
+{
+    if (pMajor) {
+        *pMajor = DRFLAC_VERSION_MAJOR;
+    }
+
+    if (pMinor) {
+        *pMinor = DRFLAC_VERSION_MINOR;
+    }
+
+    if (pRevision) {
+        *pRevision = DRFLAC_VERSION_REVISION;
+    }
+}
+
+DRFLAC_API const char* drflac_version_string(void)
+{
+    return DRFLAC_VERSION_STRING;
+}
+
+
+/* CPU caps. */
+#if defined(__has_feature)
+    #if __has_feature(thread_sanitizer)
+        #define DRFLAC_NO_THREAD_SANITIZE __attribute__((no_sanitize("thread")))
+    #else
+        #define DRFLAC_NO_THREAD_SANITIZE
+    #endif
+#else
+    #define DRFLAC_NO_THREAD_SANITIZE
+#endif
+
+#if defined(DRFLAC_HAS_LZCNT_INTRINSIC)
+static drflac_bool32 drflac__gIsLZCNTSupported = DRFLAC_FALSE;
+#endif
+
+#ifndef DRFLAC_NO_CPUID
+static drflac_bool32 drflac__gIsSSE2Supported  = DRFLAC_FALSE;
+static drflac_bool32 drflac__gIsSSE41Supported = DRFLAC_FALSE;
+
+/*
+I've had a bug report that Clang's ThreadSanitizer presents a warning in this function. Having reviewed this, this does
+actually make sense. However, since CPU caps should never differ for a running process, I don't think the trade off of
+complicating internal API's by passing around CPU caps versus just disabling the warnings is worthwhile. I'm therefore
+just going to disable these warnings. This is disabled via the DRFLAC_NO_THREAD_SANITIZE attribute.
+*/
+DRFLAC_NO_THREAD_SANITIZE static void drflac__init_cpu_caps(void)
+{
+    static drflac_bool32 isCPUCapsInitialized = DRFLAC_FALSE;
+
+    if (!isCPUCapsInitialized) {
+        /* LZCNT */
+#if defined(DRFLAC_HAS_LZCNT_INTRINSIC)
+        int info[4] = {0};
+        drflac__cpuid(info, 0x80000001);
+        drflac__gIsLZCNTSupported = (info[2] & (1 << 5)) != 0;
+#endif
+
+        /* SSE2 */
+        drflac__gIsSSE2Supported = drflac_has_sse2();
+
+        /* SSE4.1 */
+        drflac__gIsSSE41Supported = drflac_has_sse41();
+
+        /* Initialized. */
+        isCPUCapsInitialized = DRFLAC_TRUE;
+    }
+}
+#else
+static drflac_bool32 drflac__gIsNEONSupported  = DRFLAC_FALSE;
+
+static DRFLAC_INLINE drflac_bool32 drflac__has_neon(void)
+{
+#if defined(DRFLAC_SUPPORT_NEON)
+    #if defined(DRFLAC_ARM) && !defined(DRFLAC_NO_NEON)
+        #if (defined(__ARM_NEON) || defined(__aarch64__) || defined(_M_ARM64))
+            return DRFLAC_TRUE;    /* If the compiler is allowed to freely generate NEON code we can assume support. */
+        #else
+            /* TODO: Runtime check. */
+            return DRFLAC_FALSE;
+        #endif
+    #else
+        return DRFLAC_FALSE;       /* NEON is only supported on ARM architectures. */
+    #endif
+#else
+    return DRFLAC_FALSE;           /* No compiler support. */
+#endif
+}
+
+DRFLAC_NO_THREAD_SANITIZE static void drflac__init_cpu_caps(void)
+{
+    drflac__gIsNEONSupported = drflac__has_neon();
+
+#if defined(DRFLAC_HAS_LZCNT_INTRINSIC) && defined(DRFLAC_ARM) && (defined(__ARM_ARCH) && __ARM_ARCH >= 5)
+    drflac__gIsLZCNTSupported = DRFLAC_TRUE;
+#endif
+}
+#endif
+
+
+/* Endian Management */
+static DRFLAC_INLINE drflac_bool32 drflac__is_little_endian(void)
+{
+#if defined(DRFLAC_X86) || defined(DRFLAC_X64)
+    return DRFLAC_TRUE;
+#elif defined(__BYTE_ORDER) && defined(__LITTLE_ENDIAN) && __BYTE_ORDER == __LITTLE_ENDIAN
+    return DRFLAC_TRUE;
+#else
+    int n = 1;
+    return (*(char*)&n) == 1;
+#endif
+}
+
+static DRFLAC_INLINE drflac_uint16 drflac__swap_endian_uint16(drflac_uint16 n)
+{
+#ifdef DRFLAC_HAS_BYTESWAP16_INTRINSIC
+    #if defined(_MSC_VER) && !defined(__clang__)
+        return _byteswap_ushort(n);
+    #elif defined(__GNUC__) || defined(__clang__)
+        return __builtin_bswap16(n);
+    #elif defined(__WATCOMC__) && defined(__386__)
+        return _watcom_bswap16(n);
+    #else
+        #error "This compiler does not support the byte swap intrinsic."
+    #endif
+#else
+    return ((n & 0xFF00) >> 8) |
+           ((n & 0x00FF) << 8);
+#endif
+}
+
+static DRFLAC_INLINE drflac_uint32 drflac__swap_endian_uint32(drflac_uint32 n)
+{
+#ifdef DRFLAC_HAS_BYTESWAP32_INTRINSIC
+    #if defined(_MSC_VER) && !defined(__clang__)
+        return _byteswap_ulong(n);
+    #elif defined(__GNUC__) || defined(__clang__)
+        #if defined(DRFLAC_ARM) && (defined(__ARM_ARCH) && __ARM_ARCH >= 6) && !defined(__ARM_ARCH_6M__) && !defined(DRFLAC_64BIT)   /* <-- 64-bit inline assembly has not been tested, so disabling for now. */
+            /* Inline assembly optimized implementation for ARM. In my testing, GCC does not generate optimized code with __builtin_bswap32(). */
+            drflac_uint32 r;
+            __asm__ __volatile__ (
+            #if defined(DRFLAC_64BIT)
+                "rev %w[out], %w[in]" : [out]"=r"(r) : [in]"r"(n)   /* <-- This is untested. If someone in the community could test this, that would be appreciated! */
+            #else
+                "rev %[out], %[in]" : [out]"=r"(r) : [in]"r"(n)
+            #endif
+            );
+            return r;
+        #else
+            return __builtin_bswap32(n);
+        #endif
+    #elif defined(__WATCOMC__) && defined(__386__)
+        return _watcom_bswap32(n);
+    #else
+        #error "This compiler does not support the byte swap intrinsic."
+    #endif
+#else
+    return ((n & 0xFF000000) >> 24) |
+           ((n & 0x00FF0000) >>  8) |
+           ((n & 0x0000FF00) <<  8) |
+           ((n & 0x000000FF) << 24);
+#endif
+}
+
+static DRFLAC_INLINE drflac_uint64 drflac__swap_endian_uint64(drflac_uint64 n)
+{
+#ifdef DRFLAC_HAS_BYTESWAP64_INTRINSIC
+    #if defined(_MSC_VER) && !defined(__clang__)
+        return _byteswap_uint64(n);
+    #elif defined(__GNUC__) || defined(__clang__)
+        return __builtin_bswap64(n);
+    #elif defined(__WATCOMC__) && defined(__386__)
+        return _watcom_bswap64(n);
+    #else
+        #error "This compiler does not support the byte swap intrinsic."
+    #endif
+#else
+    /* Weird "<< 32" bitshift is required for C89 because it doesn't support 64-bit constants. Should be optimized out by a good compiler. */
+    return ((n & ((drflac_uint64)0xFF000000 << 32)) >> 56) |
+           ((n & ((drflac_uint64)0x00FF0000 << 32)) >> 40) |
+           ((n & ((drflac_uint64)0x0000FF00 << 32)) >> 24) |
+           ((n & ((drflac_uint64)0x000000FF << 32)) >>  8) |
+           ((n & ((drflac_uint64)0xFF000000      )) <<  8) |
+           ((n & ((drflac_uint64)0x00FF0000      )) << 24) |
+           ((n & ((drflac_uint64)0x0000FF00      )) << 40) |
+           ((n & ((drflac_uint64)0x000000FF      )) << 56);
+#endif
+}
+
+
+static DRFLAC_INLINE drflac_uint16 drflac__be2host_16(drflac_uint16 n)
+{
+    if (drflac__is_little_endian()) {
+        return drflac__swap_endian_uint16(n);
+    }
+
+    return n;
+}
+
+static DRFLAC_INLINE drflac_uint32 drflac__be2host_32(drflac_uint32 n)
+{
+    if (drflac__is_little_endian()) {
+        return drflac__swap_endian_uint32(n);
+    }
+
+    return n;
+}
+
+static DRFLAC_INLINE drflac_uint32 drflac__be2host_32_ptr_unaligned(const void* pData)
+{
+    const drflac_uint8* pNum = (drflac_uint8*)pData;
+    return *(pNum) << 24 | *(pNum+1) << 16 | *(pNum+2) << 8 | *(pNum+3);
+}
+
+static DRFLAC_INLINE drflac_uint64 drflac__be2host_64(drflac_uint64 n)
+{
+    if (drflac__is_little_endian()) {
+        return drflac__swap_endian_uint64(n);
+    }
+
+    return n;
+}
+
+
+static DRFLAC_INLINE drflac_uint32 drflac__le2host_32(drflac_uint32 n)
+{
+    if (!drflac__is_little_endian()) {
+        return drflac__swap_endian_uint32(n);
+    }
+
+    return n;
+}
+
+static DRFLAC_INLINE drflac_uint32 drflac__le2host_32_ptr_unaligned(const void* pData)
+{
+    const drflac_uint8* pNum = (drflac_uint8*)pData;
+    return *pNum | *(pNum+1) << 8 |  *(pNum+2) << 16 | *(pNum+3) << 24;
+}
+
+
+static DRFLAC_INLINE drflac_uint32 drflac__unsynchsafe_32(drflac_uint32 n)
+{
+    drflac_uint32 result = 0;
+    result |= (n & 0x7F000000) >> 3;
+    result |= (n & 0x007F0000) >> 2;
+    result |= (n & 0x00007F00) >> 1;
+    result |= (n & 0x0000007F) >> 0;
+
+    return result;
+}
+
+
+
+/* The CRC code below is based on this document: http://zlib.net/crc_v3.txt */
+static drflac_uint8 drflac__crc8_table[] = {
+    0x00, 0x07, 0x0E, 0x09, 0x1C, 0x1B, 0x12, 0x15, 0x38, 0x3F, 0x36, 0x31, 0x24, 0x23, 0x2A, 0x2D,
+    0x70, 0x77, 0x7E, 0x79, 0x6C, 0x6B, 0x62, 0x65, 0x48, 0x4F, 0x46, 0x41, 0x54, 0x53, 0x5A, 0x5D,
+    0xE0, 0xE7, 0xEE, 0xE9, 0xFC, 0xFB, 0xF2, 0xF5, 0xD8, 0xDF, 0xD6, 0xD1, 0xC4, 0xC3, 0xCA, 0xCD,
+    0x90, 0x97, 0x9E, 0x99, 0x8C, 0x8B, 0x82, 0x85, 0xA8, 0xAF, 0xA6, 0xA1, 0xB4, 0xB3, 0xBA, 0xBD,
+    0xC7, 0xC0, 0xC9, 0xCE, 0xDB, 0xDC, 0xD5, 0xD2, 0xFF, 0xF8, 0xF1, 0xF6, 0xE3, 0xE4, 0xED, 0xEA,
+    0xB7, 0xB0, 0xB9, 0xBE, 0xAB, 0xAC, 0xA5, 0xA2, 0x8F, 0x88, 0x81, 0x86, 0x93, 0x94, 0x9D, 0x9A,
+    0x27, 0x20, 0x29, 0x2E, 0x3B, 0x3C, 0x35, 0x32, 0x1F, 0x18, 0x11, 0x16, 0x03, 0x04, 0x0D, 0x0A,
+    0x57, 0x50, 0x59, 0x5E, 0x4B, 0x4C, 0x45, 0x42, 0x6F, 0x68, 0x61, 0x66, 0x73, 0x74, 0x7D, 0x7A,
+    0x89, 0x8E, 0x87, 0x80, 0x95, 0x92, 0x9B, 0x9C, 0xB1, 0xB6, 0xBF, 0xB8, 0xAD, 0xAA, 0xA3, 0xA4,
+    0xF9, 0xFE, 0xF7, 0xF0, 0xE5, 0xE2, 0xEB, 0xEC, 0xC1, 0xC6, 0xCF, 0xC8, 0xDD, 0xDA, 0xD3, 0xD4,
+    0x69, 0x6E, 0x67, 0x60, 0x75, 0x72, 0x7B, 0x7C, 0x51, 0x56, 0x5F, 0x58, 0x4D, 0x4A, 0x43, 0x44,
+    0x19, 0x1E, 0x17, 0x10, 0x05, 0x02, 0x0B, 0x0C, 0x21, 0x26, 0x2F, 0x28, 0x3D, 0x3A, 0x33, 0x34,
+    0x4E, 0x49, 0x40, 0x47, 0x52, 0x55, 0x5C, 0x5B, 0x76, 0x71, 0x78, 0x7F, 0x6A, 0x6D, 0x64, 0x63,
+    0x3E, 0x39, 0x30, 0x37, 0x22, 0x25, 0x2C, 0x2B, 0x06, 0x01, 0x08, 0x0F, 0x1A, 0x1D, 0x14, 0x13,
+    0xAE, 0xA9, 0xA0, 0xA7, 0xB2, 0xB5, 0xBC, 0xBB, 0x96, 0x91, 0x98, 0x9F, 0x8A, 0x8D, 0x84, 0x83,
+    0xDE, 0xD9, 0xD0, 0xD7, 0xC2, 0xC5, 0xCC, 0xCB, 0xE6, 0xE1, 0xE8, 0xEF, 0xFA, 0xFD, 0xF4, 0xF3
+};
+
+static drflac_uint16 drflac__crc16_table[] = {
+    0x0000, 0x8005, 0x800F, 0x000A, 0x801B, 0x001E, 0x0014, 0x8011,
+    0x8033, 0x0036, 0x003C, 0x8039, 0x0028, 0x802D, 0x8027, 0x0022,
+    0x8063, 0x0066, 0x006C, 0x8069, 0x0078, 0x807D, 0x8077, 0x0072,
+    0x0050, 0x8055, 0x805F, 0x005A, 0x804B, 0x004E, 0x0044, 0x8041,
+    0x80C3, 0x00C6, 0x00CC, 0x80C9, 0x00D8, 0x80DD, 0x80D7, 0x00D2,
+    0x00F0, 0x80F5, 0x80FF, 0x00FA, 0x80EB, 0x00EE, 0x00E4, 0x80E1,
+    0x00A0, 0x80A5, 0x80AF, 0x00AA, 0x80BB, 0x00BE, 0x00B4, 0x80B1,
+    0x8093, 0x0096, 0x009C, 0x8099, 0x0088, 0x808D, 0x8087, 0x0082,
+    0x8183, 0x0186, 0x018C, 0x8189, 0x0198, 0x819D, 0x8197, 0x0192,
+    0x01B0, 0x81B5, 0x81BF, 0x01BA, 0x81AB, 0x01AE, 0x01A4, 0x81A1,
+    0x01E0, 0x81E5, 0x81EF, 0x01EA, 0x81FB, 0x01FE, 0x01F4, 0x81F1,
+    0x81D3, 0x01D6, 0x01DC, 0x81D9, 0x01C8, 0x81CD, 0x81C7, 0x01C2,
+    0x0140, 0x8145, 0x814F, 0x014A, 0x815B, 0x015E, 0x0154, 0x8151,
+    0x8173, 0x0176, 0x017C, 0x8179, 0x0168, 0x816D, 0x8167, 0x0162,
+    0x8123, 0x0126, 0x012C, 0x8129, 0x0138, 0x813D, 0x8137, 0x0132,
+    0x0110, 0x8115, 0x811F, 0x011A, 0x810B, 0x010E, 0x0104, 0x8101,
+    0x8303, 0x0306, 0x030C, 0x8309, 0x0318, 0x831D, 0x8317, 0x0312,
+    0x0330, 0x8335, 0x833F, 0x033A, 0x832B, 0x032E, 0x0324, 0x8321,
+    0x0360, 0x8365, 0x836F, 0x036A, 0x837B, 0x037E, 0x0374, 0x8371,
+    0x8353, 0x0356, 0x035C, 0x8359, 0x0348, 0x834D, 0x8347, 0x0342,
+    0x03C0, 0x83C5, 0x83CF, 0x03CA, 0x83DB, 0x03DE, 0x03D4, 0x83D1,
+    0x83F3, 0x03F6, 0x03FC, 0x83F9, 0x03E8, 0x83ED, 0x83E7, 0x03E2,
+    0x83A3, 0x03A6, 0x03AC, 0x83A9, 0x03B8, 0x83BD, 0x83B7, 0x03B2,
+    0x0390, 0x8395, 0x839F, 0x039A, 0x838B, 0x038E, 0x0384, 0x8381,
+    0x0280, 0x8285, 0x828F, 0x028A, 0x829B, 0x029E, 0x0294, 0x8291,
+    0x82B3, 0x02B6, 0x02BC, 0x82B9, 0x02A8, 0x82AD, 0x82A7, 0x02A2,
+    0x82E3, 0x02E6, 0x02EC, 0x82E9, 0x02F8, 0x82FD, 0x82F7, 0x02F2,
+    0x02D0, 0x82D5, 0x82DF, 0x02DA, 0x82CB, 0x02CE, 0x02C4, 0x82C1,
+    0x8243, 0x0246, 0x024C, 0x8249, 0x0258, 0x825D, 0x8257, 0x0252,
+    0x0270, 0x8275, 0x827F, 0x027A, 0x826B, 0x026E, 0x0264, 0x8261,
+    0x0220, 0x8225, 0x822F, 0x022A, 0x823B, 0x023E, 0x0234, 0x8231,
+    0x8213, 0x0216, 0x021C, 0x8219, 0x0208, 0x820D, 0x8207, 0x0202
+};
+
+static DRFLAC_INLINE drflac_uint8 drflac_crc8_byte(drflac_uint8 crc, drflac_uint8 data)
+{
+    return drflac__crc8_table[crc ^ data];
+}
+
+static DRFLAC_INLINE drflac_uint8 drflac_crc8(drflac_uint8 crc, drflac_uint32 data, drflac_uint32 count)
+{
+#ifdef DR_FLAC_NO_CRC
+    (void)crc;
+    (void)data;
+    (void)count;
+    return 0;
+#else
+#if 0
+    /* REFERENCE (use of this implementation requires an explicit flush by doing "drflac_crc8(crc, 0, 8);") */
+    drflac_uint8 p = 0x07;
+    for (int i = count-1; i >= 0; --i) {
+        drflac_uint8 bit = (data & (1 << i)) >> i;
+        if (crc & 0x80) {
+            crc = ((crc << 1) | bit) ^ p;
+        } else {
+            crc = ((crc << 1) | bit);
+        }
+    }
+    return crc;
+#else
+    drflac_uint32 wholeBytes;
+    drflac_uint32 leftoverBits;
+    drflac_uint64 leftoverDataMask;
+
+    static drflac_uint64 leftoverDataMaskTable[8] = {
+        0x00, 0x01, 0x03, 0x07, 0x0F, 0x1F, 0x3F, 0x7F
+    };
+
+    DRFLAC_ASSERT(count <= 32);
+
+    wholeBytes = count >> 3;
+    leftoverBits = count - (wholeBytes*8);
+    leftoverDataMask = leftoverDataMaskTable[leftoverBits];
+
+    switch (wholeBytes) {
+        case 4: crc = drflac_crc8_byte(crc, (drflac_uint8)((data & (0xFF000000UL << leftoverBits)) >> (24 + leftoverBits)));
+        case 3: crc = drflac_crc8_byte(crc, (drflac_uint8)((data & (0x00FF0000UL << leftoverBits)) >> (16 + leftoverBits)));
+        case 2: crc = drflac_crc8_byte(crc, (drflac_uint8)((data & (0x0000FF00UL << leftoverBits)) >> ( 8 + leftoverBits)));
+        case 1: crc = drflac_crc8_byte(crc, (drflac_uint8)((data & (0x000000FFUL << leftoverBits)) >> ( 0 + leftoverBits)));
+        case 0: if (leftoverBits > 0) crc = (drflac_uint8)((crc << leftoverBits) ^ drflac__crc8_table[(crc >> (8 - leftoverBits)) ^ (data & leftoverDataMask)]);
+    }
+    return crc;
+#endif
+#endif
+}
+
+static DRFLAC_INLINE drflac_uint16 drflac_crc16_byte(drflac_uint16 crc, drflac_uint8 data)
+{
+    return (crc << 8) ^ drflac__crc16_table[(drflac_uint8)(crc >> 8) ^ data];
+}
+
+static DRFLAC_INLINE drflac_uint16 drflac_crc16_cache(drflac_uint16 crc, drflac_cache_t data)
+{
+#ifdef DRFLAC_64BIT
+    crc = drflac_crc16_byte(crc, (drflac_uint8)((data >> 56) & 0xFF));
+    crc = drflac_crc16_byte(crc, (drflac_uint8)((data >> 48) & 0xFF));
+    crc = drflac_crc16_byte(crc, (drflac_uint8)((data >> 40) & 0xFF));
+    crc = drflac_crc16_byte(crc, (drflac_uint8)((data >> 32) & 0xFF));
+#endif
+    crc = drflac_crc16_byte(crc, (drflac_uint8)((data >> 24) & 0xFF));
+    crc = drflac_crc16_byte(crc, (drflac_uint8)((data >> 16) & 0xFF));
+    crc = drflac_crc16_byte(crc, (drflac_uint8)((data >>  8) & 0xFF));
+    crc = drflac_crc16_byte(crc, (drflac_uint8)((data >>  0) & 0xFF));
+
+    return crc;
+}
+
+static DRFLAC_INLINE drflac_uint16 drflac_crc16_bytes(drflac_uint16 crc, drflac_cache_t data, drflac_uint32 byteCount)
+{
+    switch (byteCount)
+    {
+#ifdef DRFLAC_64BIT
+    case 8: crc = drflac_crc16_byte(crc, (drflac_uint8)((data >> 56) & 0xFF));
+    case 7: crc = drflac_crc16_byte(crc, (drflac_uint8)((data >> 48) & 0xFF));
+    case 6: crc = drflac_crc16_byte(crc, (drflac_uint8)((data >> 40) & 0xFF));
+    case 5: crc = drflac_crc16_byte(crc, (drflac_uint8)((data >> 32) & 0xFF));
+#endif
+    case 4: crc = drflac_crc16_byte(crc, (drflac_uint8)((data >> 24) & 0xFF));
+    case 3: crc = drflac_crc16_byte(crc, (drflac_uint8)((data >> 16) & 0xFF));
+    case 2: crc = drflac_crc16_byte(crc, (drflac_uint8)((data >>  8) & 0xFF));
+    case 1: crc = drflac_crc16_byte(crc, (drflac_uint8)((data >>  0) & 0xFF));
+    }
+
+    return crc;
+}
+
+#if 0
+static DRFLAC_INLINE drflac_uint16 drflac_crc16__32bit(drflac_uint16 crc, drflac_uint32 data, drflac_uint32 count)
+{
+#ifdef DR_FLAC_NO_CRC
+    (void)crc;
+    (void)data;
+    (void)count;
+    return 0;
+#else
+#if 0
+    /* REFERENCE (use of this implementation requires an explicit flush by doing "drflac_crc16(crc, 0, 16);") */
+    drflac_uint16 p = 0x8005;
+    for (int i = count-1; i >= 0; --i) {
+        drflac_uint16 bit = (data & (1ULL << i)) >> i;
+        if (r & 0x8000) {
+            r = ((r << 1) | bit) ^ p;
+        } else {
+            r = ((r << 1) | bit);
+        }
+    }
+
+    return crc;
+#else
+    drflac_uint32 wholeBytes;
+    drflac_uint32 leftoverBits;
+    drflac_uint64 leftoverDataMask;
+
+    static drflac_uint64 leftoverDataMaskTable[8] = {
+        0x00, 0x01, 0x03, 0x07, 0x0F, 0x1F, 0x3F, 0x7F
+    };
+
+    DRFLAC_ASSERT(count <= 64);
+
+    wholeBytes = count >> 3;
+    leftoverBits = count & 7;
+    leftoverDataMask = leftoverDataMaskTable[leftoverBits];
+
+    switch (wholeBytes) {
+        default:
+        case 4: crc = drflac_crc16_byte(crc, (drflac_uint8)((data & (0xFF000000UL << leftoverBits)) >> (24 + leftoverBits)));
+        case 3: crc = drflac_crc16_byte(crc, (drflac_uint8)((data & (0x00FF0000UL << leftoverBits)) >> (16 + leftoverBits)));
+        case 2: crc = drflac_crc16_byte(crc, (drflac_uint8)((data & (0x0000FF00UL << leftoverBits)) >> ( 8 + leftoverBits)));
+        case 1: crc = drflac_crc16_byte(crc, (drflac_uint8)((data & (0x000000FFUL << leftoverBits)) >> ( 0 + leftoverBits)));
+        case 0: if (leftoverBits > 0) crc = (crc << leftoverBits) ^ drflac__crc16_table[(crc >> (16 - leftoverBits)) ^ (data & leftoverDataMask)];
+    }
+    return crc;
+#endif
+#endif
+}
+
+static DRFLAC_INLINE drflac_uint16 drflac_crc16__64bit(drflac_uint16 crc, drflac_uint64 data, drflac_uint32 count)
+{
+#ifdef DR_FLAC_NO_CRC
+    (void)crc;
+    (void)data;
+    (void)count;
+    return 0;
+#else
+    drflac_uint32 wholeBytes;
+    drflac_uint32 leftoverBits;
+    drflac_uint64 leftoverDataMask;
+
+    static drflac_uint64 leftoverDataMaskTable[8] = {
+        0x00, 0x01, 0x03, 0x07, 0x0F, 0x1F, 0x3F, 0x7F
+    };
+
+    DRFLAC_ASSERT(count <= 64);
+
+    wholeBytes = count >> 3;
+    leftoverBits = count & 7;
+    leftoverDataMask = leftoverDataMaskTable[leftoverBits];
+
+    switch (wholeBytes) {
+        default:
+        case 8: crc = drflac_crc16_byte(crc, (drflac_uint8)((data & (((drflac_uint64)0xFF000000 << 32) << leftoverBits)) >> (56 + leftoverBits)));    /* Weird "<< 32" bitshift is required for C89 because it doesn't support 64-bit constants. Should be optimized out by a good compiler. */
+        case 7: crc = drflac_crc16_byte(crc, (drflac_uint8)((data & (((drflac_uint64)0x00FF0000 << 32) << leftoverBits)) >> (48 + leftoverBits)));
+        case 6: crc = drflac_crc16_byte(crc, (drflac_uint8)((data & (((drflac_uint64)0x0000FF00 << 32) << leftoverBits)) >> (40 + leftoverBits)));
+        case 5: crc = drflac_crc16_byte(crc, (drflac_uint8)((data & (((drflac_uint64)0x000000FF << 32) << leftoverBits)) >> (32 + leftoverBits)));
+        case 4: crc = drflac_crc16_byte(crc, (drflac_uint8)((data & (((drflac_uint64)0xFF000000      ) << leftoverBits)) >> (24 + leftoverBits)));
+        case 3: crc = drflac_crc16_byte(crc, (drflac_uint8)((data & (((drflac_uint64)0x00FF0000      ) << leftoverBits)) >> (16 + leftoverBits)));
+        case 2: crc = drflac_crc16_byte(crc, (drflac_uint8)((data & (((drflac_uint64)0x0000FF00      ) << leftoverBits)) >> ( 8 + leftoverBits)));
+        case 1: crc = drflac_crc16_byte(crc, (drflac_uint8)((data & (((drflac_uint64)0x000000FF      ) << leftoverBits)) >> ( 0 + leftoverBits)));
+        case 0: if (leftoverBits > 0) crc = (crc << leftoverBits) ^ drflac__crc16_table[(crc >> (16 - leftoverBits)) ^ (data & leftoverDataMask)];
+    }
+    return crc;
+#endif
+}
+
+
+static DRFLAC_INLINE drflac_uint16 drflac_crc16(drflac_uint16 crc, drflac_cache_t data, drflac_uint32 count)
+{
+#ifdef DRFLAC_64BIT
+    return drflac_crc16__64bit(crc, data, count);
+#else
+    return drflac_crc16__32bit(crc, data, count);
+#endif
+}
+#endif
+
+
+#ifdef DRFLAC_64BIT
+#define drflac__be2host__cache_line drflac__be2host_64
+#else
+#define drflac__be2host__cache_line drflac__be2host_32
+#endif
+
+/*
+BIT READING ATTEMPT #2
+
+This uses a 32- or 64-bit bit-shifted cache - as bits are read, the cache is shifted such that the first valid bit is sitting
+on the most significant bit. It uses the notion of an L1 and L2 cache (borrowed from CPU architecture), where the L1 cache
+is a 32- or 64-bit unsigned integer (depending on whether or not a 32- or 64-bit build is being compiled) and the L2 is an
+array of "cache lines", with each cache line being the same size as the L1. The L2 is a buffer of about 4KB and is where data
+from onRead() is read into.
+*/
+#define DRFLAC_CACHE_L1_SIZE_BYTES(bs)                      (sizeof((bs)->cache))
+#define DRFLAC_CACHE_L1_SIZE_BITS(bs)                       (sizeof((bs)->cache)*8)
+#define DRFLAC_CACHE_L1_BITS_REMAINING(bs)                  (DRFLAC_CACHE_L1_SIZE_BITS(bs) - (bs)->consumedBits)
+#define DRFLAC_CACHE_L1_SELECTION_MASK(_bitCount)           (~((~(drflac_cache_t)0) >> (_bitCount)))
+#define DRFLAC_CACHE_L1_SELECTION_SHIFT(bs, _bitCount)      (DRFLAC_CACHE_L1_SIZE_BITS(bs) - (_bitCount))
+#define DRFLAC_CACHE_L1_SELECT(bs, _bitCount)               (((bs)->cache) & DRFLAC_CACHE_L1_SELECTION_MASK(_bitCount))
+#define DRFLAC_CACHE_L1_SELECT_AND_SHIFT(bs, _bitCount)     (DRFLAC_CACHE_L1_SELECT((bs), (_bitCount)) >>  DRFLAC_CACHE_L1_SELECTION_SHIFT((bs), (_bitCount)))
+#define DRFLAC_CACHE_L1_SELECT_AND_SHIFT_SAFE(bs, _bitCount)(DRFLAC_CACHE_L1_SELECT((bs), (_bitCount)) >> (DRFLAC_CACHE_L1_SELECTION_SHIFT((bs), (_bitCount)) & (DRFLAC_CACHE_L1_SIZE_BITS(bs)-1)))
+#define DRFLAC_CACHE_L2_SIZE_BYTES(bs)                      (sizeof((bs)->cacheL2))
+#define DRFLAC_CACHE_L2_LINE_COUNT(bs)                      (DRFLAC_CACHE_L2_SIZE_BYTES(bs) / sizeof((bs)->cacheL2[0]))
+#define DRFLAC_CACHE_L2_LINES_REMAINING(bs)                 (DRFLAC_CACHE_L2_LINE_COUNT(bs) - (bs)->nextL2Line)
+
+
+#ifndef DR_FLAC_NO_CRC
+static DRFLAC_INLINE void drflac__reset_crc16(drflac_bs* bs)
+{
+    bs->crc16 = 0;
+    bs->crc16CacheIgnoredBytes = bs->consumedBits >> 3;
+}
+
+static DRFLAC_INLINE void drflac__update_crc16(drflac_bs* bs)
+{
+    if (bs->crc16CacheIgnoredBytes == 0) {
+        bs->crc16 = drflac_crc16_cache(bs->crc16, bs->crc16Cache);
+    } else {
+        bs->crc16 = drflac_crc16_bytes(bs->crc16, bs->crc16Cache, DRFLAC_CACHE_L1_SIZE_BYTES(bs) - bs->crc16CacheIgnoredBytes);
+        bs->crc16CacheIgnoredBytes = 0;
+    }
+}
+
+static DRFLAC_INLINE drflac_uint16 drflac__flush_crc16(drflac_bs* bs)
+{
+    /* We should never be flushing in a situation where we are not aligned on a byte boundary. */
+    DRFLAC_ASSERT((DRFLAC_CACHE_L1_BITS_REMAINING(bs) & 7) == 0);
+
+    /*
+    The bits that were read from the L1 cache need to be accumulated. The number of bytes needing to be accumulated is determined
+    by the number of bits that have been consumed.
+    */
+    if (DRFLAC_CACHE_L1_BITS_REMAINING(bs) == 0) {
+        drflac__update_crc16(bs);
+    } else {
+        /* We only accumulate the consumed bits. */
+        bs->crc16 = drflac_crc16_bytes(bs->crc16, bs->crc16Cache >> DRFLAC_CACHE_L1_BITS_REMAINING(bs), (bs->consumedBits >> 3) - bs->crc16CacheIgnoredBytes);
+
+        /*
+        The bits that we just accumulated should never be accumulated again. We need to keep track of how many bytes were accumulated
+        so we can handle that later.
+        */
+        bs->crc16CacheIgnoredBytes = bs->consumedBits >> 3;
+    }
+
+    return bs->crc16;
+}
+#endif
+
+static DRFLAC_INLINE drflac_bool32 drflac__reload_l1_cache_from_l2(drflac_bs* bs)
+{
+    size_t bytesRead;
+    size_t alignedL1LineCount;
+
+    /* Fast path. Try loading straight from L2. */
+    if (bs->nextL2Line < DRFLAC_CACHE_L2_LINE_COUNT(bs)) {
+        bs->cache = bs->cacheL2[bs->nextL2Line++];
+        return DRFLAC_TRUE;
+    }
+
+    /*
+    If we get here it means we've run out of data in the L2 cache. We'll need to fetch more from the client, if there's
+    any left.
+    */
+    if (bs->unalignedByteCount > 0) {
+        return DRFLAC_FALSE;   /* If we have any unaligned bytes it means there's no more aligned bytes left in the client. */
+    }
+
+    bytesRead = bs->onRead(bs->pUserData, bs->cacheL2, DRFLAC_CACHE_L2_SIZE_BYTES(bs));
+
+    bs->nextL2Line = 0;
+    if (bytesRead == DRFLAC_CACHE_L2_SIZE_BYTES(bs)) {
+        bs->cache = bs->cacheL2[bs->nextL2Line++];
+        return DRFLAC_TRUE;
+    }
+
+
+    /*
+    If we get here it means we were unable to retrieve enough data to fill the entire L2 cache. It probably
+    means we've just reached the end of the file. We need to move the valid data down to the end of the buffer
+    and adjust the index of the next line accordingly. Also keep in mind that the L2 cache must be aligned to
+    the size of the L1 so we'll need to seek backwards by any misaligned bytes.
+    */
+    alignedL1LineCount = bytesRead / DRFLAC_CACHE_L1_SIZE_BYTES(bs);
+
+    /* We need to keep track of any unaligned bytes for later use. */
+    bs->unalignedByteCount = bytesRead - (alignedL1LineCount * DRFLAC_CACHE_L1_SIZE_BYTES(bs));
+    if (bs->unalignedByteCount > 0) {
+        bs->unalignedCache = bs->cacheL2[alignedL1LineCount];
+    }
+
+    if (alignedL1LineCount > 0) {
+        size_t offset = DRFLAC_CACHE_L2_LINE_COUNT(bs) - alignedL1LineCount;
+        size_t i;
+        for (i = alignedL1LineCount; i > 0; --i) {
+            bs->cacheL2[i-1 + offset] = bs->cacheL2[i-1];
+        }
+
+        bs->nextL2Line = (drflac_uint32)offset;
+        bs->cache = bs->cacheL2[bs->nextL2Line++];
+        return DRFLAC_TRUE;
+    } else {
+        /* If we get into this branch it means we weren't able to load any L1-aligned data. */
+        bs->nextL2Line = DRFLAC_CACHE_L2_LINE_COUNT(bs);
+        return DRFLAC_FALSE;
+    }
+}
+
+static drflac_bool32 drflac__reload_cache(drflac_bs* bs)
+{
+    size_t bytesRead;
+
+#ifndef DR_FLAC_NO_CRC
+    drflac__update_crc16(bs);
+#endif
+
+    /* Fast path. Try just moving the next value in the L2 cache to the L1 cache. */
+    if (drflac__reload_l1_cache_from_l2(bs)) {
+        bs->cache = drflac__be2host__cache_line(bs->cache);
+        bs->consumedBits = 0;
+#ifndef DR_FLAC_NO_CRC
+        bs->crc16Cache = bs->cache;
+#endif
+        return DRFLAC_TRUE;
+    }
+
+    /* Slow path. */
+
+    /*
+    If we get here it means we have failed to load the L1 cache from the L2. Likely we've just reached the end of the stream and the last
+    few bytes did not meet the alignment requirements for the L2 cache. In this case we need to fall back to a slower path and read the
+    data from the unaligned cache.
+    */
+    bytesRead = bs->unalignedByteCount;
+    if (bytesRead == 0) {
+        bs->consumedBits = DRFLAC_CACHE_L1_SIZE_BITS(bs);   /* <-- The stream has been exhausted, so marked the bits as consumed. */
+        return DRFLAC_FALSE;
+    }
+
+    DRFLAC_ASSERT(bytesRead < DRFLAC_CACHE_L1_SIZE_BYTES(bs));
+    bs->consumedBits = (drflac_uint32)(DRFLAC_CACHE_L1_SIZE_BYTES(bs) - bytesRead) * 8;
+
+    bs->cache = drflac__be2host__cache_line(bs->unalignedCache);
+    bs->cache &= DRFLAC_CACHE_L1_SELECTION_MASK(DRFLAC_CACHE_L1_BITS_REMAINING(bs));    /* <-- Make sure the consumed bits are always set to zero. Other parts of the library depend on this property. */
+    bs->unalignedByteCount = 0;     /* <-- At this point the unaligned bytes have been moved into the cache and we thus have no more unaligned bytes. */
+
+#ifndef DR_FLAC_NO_CRC
+    bs->crc16Cache = bs->cache >> bs->consumedBits;
+    bs->crc16CacheIgnoredBytes = bs->consumedBits >> 3;
+#endif
+    return DRFLAC_TRUE;
+}
+
+static void drflac__reset_cache(drflac_bs* bs)
+{
+    bs->nextL2Line   = DRFLAC_CACHE_L2_LINE_COUNT(bs);  /* <-- This clears the L2 cache. */
+    bs->consumedBits = DRFLAC_CACHE_L1_SIZE_BITS(bs);   /* <-- This clears the L1 cache. */
+    bs->cache = 0;
+    bs->unalignedByteCount = 0;                         /* <-- This clears the trailing unaligned bytes. */
+    bs->unalignedCache = 0;
+
+#ifndef DR_FLAC_NO_CRC
+    bs->crc16Cache = 0;
+    bs->crc16CacheIgnoredBytes = 0;
+#endif
+}
+
+
+static DRFLAC_INLINE drflac_bool32 drflac__read_uint32(drflac_bs* bs, unsigned int bitCount, drflac_uint32* pResultOut)
+{
+    DRFLAC_ASSERT(bs != NULL);
+    DRFLAC_ASSERT(pResultOut != NULL);
+    DRFLAC_ASSERT(bitCount > 0);
+    DRFLAC_ASSERT(bitCount <= 32);
+
+    if (bs->consumedBits == DRFLAC_CACHE_L1_SIZE_BITS(bs)) {
+        if (!drflac__reload_cache(bs)) {
+            return DRFLAC_FALSE;
+        }
+    }
+
+    if (bitCount <= DRFLAC_CACHE_L1_BITS_REMAINING(bs)) {
+        /*
+        If we want to load all 32-bits from a 32-bit cache we need to do it slightly differently because we can't do
+        a 32-bit shift on a 32-bit integer. This will never be the case on 64-bit caches, so we can have a slightly
+        more optimal solution for this.
+        */
+#ifdef DRFLAC_64BIT
+        *pResultOut = (drflac_uint32)DRFLAC_CACHE_L1_SELECT_AND_SHIFT(bs, bitCount);
+        bs->consumedBits += bitCount;
+        bs->cache <<= bitCount;
+#else
+        if (bitCount < DRFLAC_CACHE_L1_SIZE_BITS(bs)) {
+            *pResultOut = (drflac_uint32)DRFLAC_CACHE_L1_SELECT_AND_SHIFT(bs, bitCount);
+            bs->consumedBits += bitCount;
+            bs->cache <<= bitCount;
+        } else {
+            /* Cannot shift by 32-bits, so need to do it differently. */
+            *pResultOut = (drflac_uint32)bs->cache;
+            bs->consumedBits = DRFLAC_CACHE_L1_SIZE_BITS(bs);
+            bs->cache = 0;
+        }
+#endif
+
+        return DRFLAC_TRUE;
+    } else {
+        /* It straddles the cached data. It will never cover more than the next chunk. We just read the number in two parts and combine them. */
+        drflac_uint32 bitCountHi = DRFLAC_CACHE_L1_BITS_REMAINING(bs);
+        drflac_uint32 bitCountLo = bitCount - bitCountHi;
+        drflac_uint32 resultHi;
+
+        DRFLAC_ASSERT(bitCountHi > 0);
+        DRFLAC_ASSERT(bitCountHi < 32);
+        resultHi = (drflac_uint32)DRFLAC_CACHE_L1_SELECT_AND_SHIFT(bs, bitCountHi);
+
+        if (!drflac__reload_cache(bs)) {
+            return DRFLAC_FALSE;
+        }
+        if (bitCountLo > DRFLAC_CACHE_L1_BITS_REMAINING(bs)) {
+            /* This happens when we get to end of stream */
+            return DRFLAC_FALSE;
+        }
+
+        *pResultOut = (resultHi << bitCountLo) | (drflac_uint32)DRFLAC_CACHE_L1_SELECT_AND_SHIFT(bs, bitCountLo);
+        bs->consumedBits += bitCountLo;
+        bs->cache <<= bitCountLo;
+        return DRFLAC_TRUE;
+    }
+}
+
+static drflac_bool32 drflac__read_int32(drflac_bs* bs, unsigned int bitCount, drflac_int32* pResult)
+{
+    drflac_uint32 result;
+
+    DRFLAC_ASSERT(bs != NULL);
+    DRFLAC_ASSERT(pResult != NULL);
+    DRFLAC_ASSERT(bitCount > 0);
+    DRFLAC_ASSERT(bitCount <= 32);
+
+    if (!drflac__read_uint32(bs, bitCount, &result)) {
+        return DRFLAC_FALSE;
+    }
+
+    /* Do not attempt to shift by 32 as it's undefined. */
+    if (bitCount < 32) {
+        drflac_uint32 signbit;
+        signbit = ((result >> (bitCount-1)) & 0x01);
+        result |= (~signbit + 1) << bitCount;
+    }
+
+    *pResult = (drflac_int32)result;
+    return DRFLAC_TRUE;
+}
+
+#ifdef DRFLAC_64BIT
+static drflac_bool32 drflac__read_uint64(drflac_bs* bs, unsigned int bitCount, drflac_uint64* pResultOut)
+{
+    drflac_uint32 resultHi;
+    drflac_uint32 resultLo;
+
+    DRFLAC_ASSERT(bitCount <= 64);
+    DRFLAC_ASSERT(bitCount >  32);
+
+    if (!drflac__read_uint32(bs, bitCount - 32, &resultHi)) {
+        return DRFLAC_FALSE;
+    }
+
+    if (!drflac__read_uint32(bs, 32, &resultLo)) {
+        return DRFLAC_FALSE;
+    }
+
+    *pResultOut = (((drflac_uint64)resultHi) << 32) | ((drflac_uint64)resultLo);
+    return DRFLAC_TRUE;
+}
+#endif
+
+/* Function below is unused, but leaving it here in case I need to quickly add it again. */
+#if 0
+static drflac_bool32 drflac__read_int64(drflac_bs* bs, unsigned int bitCount, drflac_int64* pResultOut)
+{
+    drflac_uint64 result;
+    drflac_uint64 signbit;
+
+    DRFLAC_ASSERT(bitCount <= 64);
+
+    if (!drflac__read_uint64(bs, bitCount, &result)) {
+        return DRFLAC_FALSE;
+    }
+
+    signbit = ((result >> (bitCount-1)) & 0x01);
+    result |= (~signbit + 1) << bitCount;
+
+    *pResultOut = (drflac_int64)result;
+    return DRFLAC_TRUE;
+}
+#endif
+
+static drflac_bool32 drflac__read_uint16(drflac_bs* bs, unsigned int bitCount, drflac_uint16* pResult)
+{
+    drflac_uint32 result;
+
+    DRFLAC_ASSERT(bs != NULL);
+    DRFLAC_ASSERT(pResult != NULL);
+    DRFLAC_ASSERT(bitCount > 0);
+    DRFLAC_ASSERT(bitCount <= 16);
+
+    if (!drflac__read_uint32(bs, bitCount, &result)) {
+        return DRFLAC_FALSE;
+    }
+
+    *pResult = (drflac_uint16)result;
+    return DRFLAC_TRUE;
+}
+
+#if 0
+static drflac_bool32 drflac__read_int16(drflac_bs* bs, unsigned int bitCount, drflac_int16* pResult)
+{
+    drflac_int32 result;
+
+    DRFLAC_ASSERT(bs != NULL);
+    DRFLAC_ASSERT(pResult != NULL);
+    DRFLAC_ASSERT(bitCount > 0);
+    DRFLAC_ASSERT(bitCount <= 16);
+
+    if (!drflac__read_int32(bs, bitCount, &result)) {
+        return DRFLAC_FALSE;
+    }
+
+    *pResult = (drflac_int16)result;
+    return DRFLAC_TRUE;
+}
+#endif
+
+static drflac_bool32 drflac__read_uint8(drflac_bs* bs, unsigned int bitCount, drflac_uint8* pResult)
+{
+    drflac_uint32 result;
+
+    DRFLAC_ASSERT(bs != NULL);
+    DRFLAC_ASSERT(pResult != NULL);
+    DRFLAC_ASSERT(bitCount > 0);
+    DRFLAC_ASSERT(bitCount <= 8);
+
+    if (!drflac__read_uint32(bs, bitCount, &result)) {
+        return DRFLAC_FALSE;
+    }
+
+    *pResult = (drflac_uint8)result;
+    return DRFLAC_TRUE;
+}
+
+static drflac_bool32 drflac__read_int8(drflac_bs* bs, unsigned int bitCount, drflac_int8* pResult)
+{
+    drflac_int32 result;
+
+    DRFLAC_ASSERT(bs != NULL);
+    DRFLAC_ASSERT(pResult != NULL);
+    DRFLAC_ASSERT(bitCount > 0);
+    DRFLAC_ASSERT(bitCount <= 8);
+
+    if (!drflac__read_int32(bs, bitCount, &result)) {
+        return DRFLAC_FALSE;
+    }
+
+    *pResult = (drflac_int8)result;
+    return DRFLAC_TRUE;
+}
+
+
+static drflac_bool32 drflac__seek_bits(drflac_bs* bs, size_t bitsToSeek)
+{
+    if (bitsToSeek <= DRFLAC_CACHE_L1_BITS_REMAINING(bs)) {
+        bs->consumedBits += (drflac_uint32)bitsToSeek;
+        bs->cache <<= bitsToSeek;
+        return DRFLAC_TRUE;
+    } else {
+        /* It straddles the cached data. This function isn't called too frequently so I'm favouring simplicity here. */
+        bitsToSeek       -= DRFLAC_CACHE_L1_BITS_REMAINING(bs);
+        bs->consumedBits += DRFLAC_CACHE_L1_BITS_REMAINING(bs);
+        bs->cache         = 0;
+
+        /* Simple case. Seek in groups of the same number as bits that fit within a cache line. */
+#ifdef DRFLAC_64BIT
+        while (bitsToSeek >= DRFLAC_CACHE_L1_SIZE_BITS(bs)) {
+            drflac_uint64 bin;
+            if (!drflac__read_uint64(bs, DRFLAC_CACHE_L1_SIZE_BITS(bs), &bin)) {
+                return DRFLAC_FALSE;
+            }
+            bitsToSeek -= DRFLAC_CACHE_L1_SIZE_BITS(bs);
+        }
+#else
+        while (bitsToSeek >= DRFLAC_CACHE_L1_SIZE_BITS(bs)) {
+            drflac_uint32 bin;
+            if (!drflac__read_uint32(bs, DRFLAC_CACHE_L1_SIZE_BITS(bs), &bin)) {
+                return DRFLAC_FALSE;
+            }
+            bitsToSeek -= DRFLAC_CACHE_L1_SIZE_BITS(bs);
+        }
+#endif
+
+        /* Whole leftover bytes. */
+        while (bitsToSeek >= 8) {
+            drflac_uint8 bin;
+            if (!drflac__read_uint8(bs, 8, &bin)) {
+                return DRFLAC_FALSE;
+            }
+            bitsToSeek -= 8;
+        }
+
+        /* Leftover bits. */
+        if (bitsToSeek > 0) {
+            drflac_uint8 bin;
+            if (!drflac__read_uint8(bs, (drflac_uint32)bitsToSeek, &bin)) {
+                return DRFLAC_FALSE;
+            }
+            bitsToSeek = 0; /* <-- Necessary for the assert below. */
+        }
+
+        DRFLAC_ASSERT(bitsToSeek == 0);
+        return DRFLAC_TRUE;
+    }
+}
+
+
+/* This function moves the bit streamer to the first bit after the sync code (bit 15 of the of the frame header). It will also update the CRC-16. */
+static drflac_bool32 drflac__find_and_seek_to_next_sync_code(drflac_bs* bs)
+{
+    DRFLAC_ASSERT(bs != NULL);
+
+    /*
+    The sync code is always aligned to 8 bits. This is convenient for us because it means we can do byte-aligned movements. The first
+    thing to do is align to the next byte.
+    */
+    if (!drflac__seek_bits(bs, DRFLAC_CACHE_L1_BITS_REMAINING(bs) & 7)) {
+        return DRFLAC_FALSE;
+    }
+
+    for (;;) {
+        drflac_uint8 hi;
+
+#ifndef DR_FLAC_NO_CRC
+        drflac__reset_crc16(bs);
+#endif
+
+        if (!drflac__read_uint8(bs, 8, &hi)) {
+            return DRFLAC_FALSE;
+        }
+
+        if (hi == 0xFF) {
+            drflac_uint8 lo;
+            if (!drflac__read_uint8(bs, 6, &lo)) {
+                return DRFLAC_FALSE;
+            }
+
+            if (lo == 0x3E) {
+                return DRFLAC_TRUE;
+            } else {
+                if (!drflac__seek_bits(bs, DRFLAC_CACHE_L1_BITS_REMAINING(bs) & 7)) {
+                    return DRFLAC_FALSE;
+                }
+            }
+        }
+    }
+
+    /* Should never get here. */
+    /*return DRFLAC_FALSE;*/
+}
+
+
+#if defined(DRFLAC_HAS_LZCNT_INTRINSIC)
+#define DRFLAC_IMPLEMENT_CLZ_LZCNT
+#endif
+#if  defined(_MSC_VER) && _MSC_VER >= 1400 && (defined(DRFLAC_X64) || defined(DRFLAC_X86)) && !defined(__clang__)
+#define DRFLAC_IMPLEMENT_CLZ_MSVC
+#endif
+#if  defined(__WATCOMC__) && defined(__386__)
+#define DRFLAC_IMPLEMENT_CLZ_WATCOM
+#endif
+#ifdef __MRC__
+#include <intrinsics.h>
+#define DRFLAC_IMPLEMENT_CLZ_MRC
+#endif
+
+static DRFLAC_INLINE drflac_uint32 drflac__clz_software(drflac_cache_t x)
+{
+    drflac_uint32 n;
+    static drflac_uint32 clz_table_4[] = {
+        0,
+        4,
+        3, 3,
+        2, 2, 2, 2,
+        1, 1, 1, 1, 1, 1, 1, 1
+    };
+
+    if (x == 0) {
+        return sizeof(x)*8;
+    }
+
+    n = clz_table_4[x >> (sizeof(x)*8 - 4)];
+    if (n == 0) {
+#ifdef DRFLAC_64BIT
+        if ((x & ((drflac_uint64)0xFFFFFFFF << 32)) == 0) { n  = 32; x <<= 32; }
+        if ((x & ((drflac_uint64)0xFFFF0000 << 32)) == 0) { n += 16; x <<= 16; }
+        if ((x & ((drflac_uint64)0xFF000000 << 32)) == 0) { n += 8;  x <<= 8;  }
+        if ((x & ((drflac_uint64)0xF0000000 << 32)) == 0) { n += 4;  x <<= 4;  }
+#else
+        if ((x & 0xFFFF0000) == 0) { n  = 16; x <<= 16; }
+        if ((x & 0xFF000000) == 0) { n += 8;  x <<= 8;  }
+        if ((x & 0xF0000000) == 0) { n += 4;  x <<= 4;  }
+#endif
+        n += clz_table_4[x >> (sizeof(x)*8 - 4)];
+    }
+
+    return n - 1;
+}
+
+#ifdef DRFLAC_IMPLEMENT_CLZ_LZCNT
+static DRFLAC_INLINE drflac_bool32 drflac__is_lzcnt_supported(void)
+{
+    /* Fast compile time check for ARM. */
+#if defined(DRFLAC_HAS_LZCNT_INTRINSIC) && defined(DRFLAC_ARM) && (defined(__ARM_ARCH) && __ARM_ARCH >= 5)
+    return DRFLAC_TRUE;
+#elif defined(__MRC__)
+    return DRFLAC_TRUE;
+#else
+    /* If the compiler itself does not support the intrinsic then we'll need to return false. */
+    #ifdef DRFLAC_HAS_LZCNT_INTRINSIC
+        return drflac__gIsLZCNTSupported;
+    #else
+        return DRFLAC_FALSE;
+    #endif
+#endif
+}
+
+static DRFLAC_INLINE drflac_uint32 drflac__clz_lzcnt(drflac_cache_t x)
+{
+    /*
+    It's critical for competitive decoding performance that this function be highly optimal. With MSVC we can use the __lzcnt64() and __lzcnt() intrinsics
+    to achieve good performance, however on GCC and Clang it's a little bit more annoying. The __builtin_clzl() and __builtin_clzll() intrinsics leave
+    it undefined as to the return value when `x` is 0. We need this to be well defined as returning 32 or 64, depending on whether or not it's a 32- or
+    64-bit build. To work around this we would need to add a conditional to check for the x = 0 case, but this creates unnecessary inefficiency. To work
+    around this problem I have written some inline assembly to emit the LZCNT (x86) or CLZ (ARM) instruction directly which removes the need to include
+    the conditional. This has worked well in the past, but for some reason Clang's MSVC compatible driver, clang-cl, does not seem to be handling this
+    in the same way as the normal Clang driver. It seems that `clang-cl` is just outputting the wrong results sometimes, maybe due to some register
+    getting clobbered?
+
+    I'm not sure if this is a bug with dr_flac's inlined assembly (most likely), a bug in `clang-cl` or just a misunderstanding on my part with inline
+    assembly rules for `clang-cl`. If somebody can identify an error in dr_flac's inlined assembly I'm happy to get that fixed.
+
+    Fortunately there is an easy workaround for this. Clang implements MSVC-specific intrinsics for compatibility. It also defines _MSC_VER for extra
+    compatibility. We can therefore just check for _MSC_VER and use the MSVC intrinsic which, fortunately for us, Clang supports. It would still be nice
+    to know how to fix the inlined assembly for correctness sake, however.
+    */
+
+#if defined(_MSC_VER) /*&& !defined(__clang__)*/    /* <-- Intentionally wanting Clang to use the MSVC __lzcnt64/__lzcnt intrinsics due to above ^. */
+    #ifdef DRFLAC_64BIT
+        return (drflac_uint32)__lzcnt64(x);
+    #else
+        return (drflac_uint32)__lzcnt(x);
+    #endif
+#else
+    #if defined(__GNUC__) || defined(__clang__)
+        #if defined(DRFLAC_X64)
+            {
+                drflac_uint64 r;
+                __asm__ __volatile__ (
+                    "lzcnt{ %1, %0| %0, %1}" : "=r"(r) : "r"(x) : "cc"
+                );
+
+                return (drflac_uint32)r;
+            }
+        #elif defined(DRFLAC_X86)
+            {
+                drflac_uint32 r;
+                __asm__ __volatile__ (
+                    "lzcnt{l %1, %0| %0, %1}" : "=r"(r) : "r"(x) : "cc"
+                );
+
+                return r;
+            }
+        #elif defined(DRFLAC_ARM) && (defined(__ARM_ARCH) && __ARM_ARCH >= 5) && !defined(__ARM_ARCH_6M__) && !defined(DRFLAC_64BIT)   /* <-- I haven't tested 64-bit inline assembly, so only enabling this for the 32-bit build for now. */
+            {
+                unsigned int r;
+                __asm__ __volatile__ (
+                #if defined(DRFLAC_64BIT)
+                    "clz %w[out], %w[in]" : [out]"=r"(r) : [in]"r"(x)   /* <-- This is untested. If someone in the community could test this, that would be appreciated! */
+                #else
+                    "clz %[out], %[in]" : [out]"=r"(r) : [in]"r"(x)
+                #endif
+                );
+
+                return r;
+            }
+        #else
+            if (x == 0) {
+                return sizeof(x)*8;
+            }
+            #ifdef DRFLAC_64BIT
+                return (drflac_uint32)__builtin_clzll((drflac_uint64)x);
+            #else
+                return (drflac_uint32)__builtin_clzl((drflac_uint32)x);
+            #endif
+        #endif
+    #else
+        /* Unsupported compiler. */
+        #error "This compiler does not support the lzcnt intrinsic."
+    #endif
+#endif
+}
+#endif
+
+#ifdef DRFLAC_IMPLEMENT_CLZ_MSVC
+#include <intrin.h> /* For BitScanReverse(). */
+
+static DRFLAC_INLINE drflac_uint32 drflac__clz_msvc(drflac_cache_t x)
+{
+    drflac_uint32 n;
+
+    if (x == 0) {
+        return sizeof(x)*8;
+    }
+
+#ifdef DRFLAC_64BIT
+    _BitScanReverse64((unsigned long*)&n, x);
+#else
+    _BitScanReverse((unsigned long*)&n, x);
+#endif
+    return sizeof(x)*8 - n - 1;
+}
+#endif
+
+#ifdef DRFLAC_IMPLEMENT_CLZ_WATCOM
+static __inline drflac_uint32 drflac__clz_watcom (drflac_uint32);
+#ifdef DRFLAC_IMPLEMENT_CLZ_WATCOM_LZCNT
+/* Use the LZCNT instruction (only available on some processors since the 2010s). */
+#pragma aux drflac__clz_watcom_lzcnt = \
+    "db 0F3h, 0Fh, 0BDh, 0C0h" /* lzcnt eax, eax */ \
+    parm [eax] \
+    value [eax] \
+    modify nomemory;
+#else
+/* Use the 386+-compatible implementation. */
+#pragma aux drflac__clz_watcom = \
+    "bsr eax, eax" \
+    "xor eax, 31" \
+    parm [eax] nomemory \
+    value [eax] \
+    modify exact [eax] nomemory;
+#endif
+#endif
+
+static DRFLAC_INLINE drflac_uint32 drflac__clz(drflac_cache_t x)
+{
+#ifdef DRFLAC_IMPLEMENT_CLZ_LZCNT
+    if (drflac__is_lzcnt_supported()) {
+        return drflac__clz_lzcnt(x);
+    } else
+#endif
+    {
+#ifdef DRFLAC_IMPLEMENT_CLZ_MSVC
+        return drflac__clz_msvc(x);
+#elif defined(DRFLAC_IMPLEMENT_CLZ_WATCOM_LZCNT)
+        return drflac__clz_watcom_lzcnt(x);
+#elif defined(DRFLAC_IMPLEMENT_CLZ_WATCOM)
+        return (x == 0) ? sizeof(x)*8 : drflac__clz_watcom(x);
+#elif defined(__MRC__)
+        return __cntlzw(x);
+#else
+        return drflac__clz_software(x);
+#endif
+    }
+}
+
+
+static DRFLAC_INLINE drflac_bool32 drflac__seek_past_next_set_bit(drflac_bs* bs, unsigned int* pOffsetOut)
+{
+    drflac_uint32 zeroCounter = 0;
+    drflac_uint32 setBitOffsetPlus1;
+
+    while (bs->cache == 0) {
+        zeroCounter += (drflac_uint32)DRFLAC_CACHE_L1_BITS_REMAINING(bs);
+        if (!drflac__reload_cache(bs)) {
+            return DRFLAC_FALSE;
+        }
+    }
+
+    if (bs->cache == 1) {
+        /* Not catching this would lead to undefined behaviour: a shift of a 32-bit number by 32 or more is undefined */
+        *pOffsetOut = zeroCounter + (drflac_uint32)DRFLAC_CACHE_L1_BITS_REMAINING(bs) - 1;
+        if (!drflac__reload_cache(bs)) {
+            return DRFLAC_FALSE;
+        }
+
+        return DRFLAC_TRUE;
+    }
+
+    setBitOffsetPlus1 = drflac__clz(bs->cache);
+    setBitOffsetPlus1 += 1;
+
+    if (setBitOffsetPlus1 > DRFLAC_CACHE_L1_BITS_REMAINING(bs)) {
+        /* This happens when we get to end of stream */
+        return DRFLAC_FALSE;
+    }
+
+    bs->consumedBits += setBitOffsetPlus1;
+    bs->cache <<= setBitOffsetPlus1;
+
+    *pOffsetOut = zeroCounter + setBitOffsetPlus1 - 1;
+    return DRFLAC_TRUE;
+}
+
+
+
+static drflac_bool32 drflac__seek_to_byte(drflac_bs* bs, drflac_uint64 offsetFromStart)
+{
+    DRFLAC_ASSERT(bs != NULL);
+    DRFLAC_ASSERT(offsetFromStart > 0);
+
+    /*
+    Seeking from the start is not quite as trivial as it sounds because the onSeek callback takes a signed 32-bit integer (which
+    is intentional because it simplifies the implementation of the onSeek callbacks), however offsetFromStart is unsigned 64-bit.
+    To resolve we just need to do an initial seek from the start, and then a series of offset seeks to make up the remainder.
+    */
+    if (offsetFromStart > 0x7FFFFFFF) {
+        drflac_uint64 bytesRemaining = offsetFromStart;
+        if (!bs->onSeek(bs->pUserData, 0x7FFFFFFF, drflac_seek_origin_start)) {
+            return DRFLAC_FALSE;
+        }
+        bytesRemaining -= 0x7FFFFFFF;
+
+        while (bytesRemaining > 0x7FFFFFFF) {
+            if (!bs->onSeek(bs->pUserData, 0x7FFFFFFF, drflac_seek_origin_current)) {
+                return DRFLAC_FALSE;
+            }
+            bytesRemaining -= 0x7FFFFFFF;
+        }
+
+        if (bytesRemaining > 0) {
+            if (!bs->onSeek(bs->pUserData, (int)bytesRemaining, drflac_seek_origin_current)) {
+                return DRFLAC_FALSE;
+            }
+        }
+    } else {
+        if (!bs->onSeek(bs->pUserData, (int)offsetFromStart, drflac_seek_origin_start)) {
+            return DRFLAC_FALSE;
+        }
+    }
+
+    /* The cache should be reset to force a reload of fresh data from the client. */
+    drflac__reset_cache(bs);
+    return DRFLAC_TRUE;
+}
+
+
+static drflac_result drflac__read_utf8_coded_number(drflac_bs* bs, drflac_uint64* pNumberOut, drflac_uint8* pCRCOut)
+{
+    drflac_uint8 crc;
+    drflac_uint64 result;
+    drflac_uint8 utf8[7] = {0};
+    int byteCount;
+    int i;
+
+    DRFLAC_ASSERT(bs != NULL);
+    DRFLAC_ASSERT(pNumberOut != NULL);
+    DRFLAC_ASSERT(pCRCOut != NULL);
+
+    crc = *pCRCOut;
+
+    if (!drflac__read_uint8(bs, 8, utf8)) {
+        *pNumberOut = 0;
+        return DRFLAC_AT_END;
+    }
+    crc = drflac_crc8(crc, utf8[0], 8);
+
+    if ((utf8[0] & 0x80) == 0) {
+        *pNumberOut = utf8[0];
+        *pCRCOut = crc;
+        return DRFLAC_SUCCESS;
+    }
+
+    /*byteCount = 1;*/
+    if ((utf8[0] & 0xE0) == 0xC0) {
+        byteCount = 2;
+    } else if ((utf8[0] & 0xF0) == 0xE0) {
+        byteCount = 3;
+    } else if ((utf8[0] & 0xF8) == 0xF0) {
+        byteCount = 4;
+    } else if ((utf8[0] & 0xFC) == 0xF8) {
+        byteCount = 5;
+    } else if ((utf8[0] & 0xFE) == 0xFC) {
+        byteCount = 6;
+    } else if ((utf8[0] & 0xFF) == 0xFE) {
+        byteCount = 7;
+    } else {
+        *pNumberOut = 0;
+        return DRFLAC_CRC_MISMATCH;     /* Bad UTF-8 encoding. */
+    }
+
+    /* Read extra bytes. */
+    DRFLAC_ASSERT(byteCount > 1);
+
+    result = (drflac_uint64)(utf8[0] & (0xFF >> (byteCount + 1)));
+    for (i = 1; i < byteCount; ++i) {
+        if (!drflac__read_uint8(bs, 8, utf8 + i)) {
+            *pNumberOut = 0;
+            return DRFLAC_AT_END;
+        }
+        crc = drflac_crc8(crc, utf8[i], 8);
+
+        result = (result << 6) | (utf8[i] & 0x3F);
+    }
+
+    *pNumberOut = result;
+    *pCRCOut = crc;
+    return DRFLAC_SUCCESS;
+}
+
+
+static DRFLAC_INLINE drflac_uint32 drflac__ilog2_u32(drflac_uint32 x)
+{
+#if 1   /* Needs optimizing. */
+    drflac_uint32 result = 0;
+    while (x > 0) {
+        result += 1;
+        x >>= 1;
+    }
+
+    return result;
+#endif
+}
+
+static DRFLAC_INLINE drflac_bool32 drflac__use_64_bit_prediction(drflac_uint32 bitsPerSample, drflac_uint32 order, drflac_uint32 precision)
+{
+    /* https://web.archive.org/web/20220205005724/https://github.com/ietf-wg-cellar/flac-specification/blob/37a49aa48ba4ba12e8757badfc59c0df35435fec/rfc_backmatter.md */
+    return bitsPerSample + precision + drflac__ilog2_u32(order) > 32;
+}
+
+
+/*
+The next two functions are responsible for calculating the prediction.
+
+When the bits per sample is >16 we need to use 64-bit integer arithmetic because otherwise we'll run out of precision. It's
+safe to assume this will be slower on 32-bit platforms so we use a more optimal solution when the bits per sample is <=16.
+*/
+#if defined(__clang__)
+__attribute__((no_sanitize("signed-integer-overflow")))
+#endif
+static DRFLAC_INLINE drflac_int32 drflac__calculate_prediction_32(drflac_uint32 order, drflac_int32 shift, const drflac_int32* coefficients, drflac_int32* pDecodedSamples)
+{
+    drflac_int32 prediction = 0;
+
+    DRFLAC_ASSERT(order <= 32);
+
+    /* 32-bit version. */
+
+    /* VC++ optimizes this to a single jmp. I've not yet verified this for other compilers. */
+    switch (order)
+    {
+    case 32: prediction += coefficients[31] * pDecodedSamples[-32];
+    case 31: prediction += coefficients[30] * pDecodedSamples[-31];
+    case 30: prediction += coefficients[29] * pDecodedSamples[-30];
+    case 29: prediction += coefficients[28] * pDecodedSamples[-29];
+    case 28: prediction += coefficients[27] * pDecodedSamples[-28];
+    case 27: prediction += coefficients[26] * pDecodedSamples[-27];
+    case 26: prediction += coefficients[25] * pDecodedSamples[-26];
+    case 25: prediction += coefficients[24] * pDecodedSamples[-25];
+    case 24: prediction += coefficients[23] * pDecodedSamples[-24];
+    case 23: prediction += coefficients[22] * pDecodedSamples[-23];
+    case 22: prediction += coefficients[21] * pDecodedSamples[-22];
+    case 21: prediction += coefficients[20] * pDecodedSamples[-21];
+    case 20: prediction += coefficients[19] * pDecodedSamples[-20];
+    case 19: prediction += coefficients[18] * pDecodedSamples[-19];
+    case 18: prediction += coefficients[17] * pDecodedSamples[-18];
+    case 17: prediction += coefficients[16] * pDecodedSamples[-17];
+    case 16: prediction += coefficients[15] * pDecodedSamples[-16];
+    case 15: prediction += coefficients[14] * pDecodedSamples[-15];
+    case 14: prediction += coefficients[13] * pDecodedSamples[-14];
+    case 13: prediction += coefficients[12] * pDecodedSamples[-13];
+    case 12: prediction += coefficients[11] * pDecodedSamples[-12];
+    case 11: prediction += coefficients[10] * pDecodedSamples[-11];
+    case 10: prediction += coefficients[ 9] * pDecodedSamples[-10];
+    case  9: prediction += coefficients[ 8] * pDecodedSamples[- 9];
+    case  8: prediction += coefficients[ 7] * pDecodedSamples[- 8];
+    case  7: prediction += coefficients[ 6] * pDecodedSamples[- 7];
+    case  6: prediction += coefficients[ 5] * pDecodedSamples[- 6];
+    case  5: prediction += coefficients[ 4] * pDecodedSamples[- 5];
+    case  4: prediction += coefficients[ 3] * pDecodedSamples[- 4];
+    case  3: prediction += coefficients[ 2] * pDecodedSamples[- 3];
+    case  2: prediction += coefficients[ 1] * pDecodedSamples[- 2];
+    case  1: prediction += coefficients[ 0] * pDecodedSamples[- 1];
+    }
+
+    return (drflac_int32)(prediction >> shift);
+}
+
+static DRFLAC_INLINE drflac_int32 drflac__calculate_prediction_64(drflac_uint32 order, drflac_int32 shift, const drflac_int32* coefficients, drflac_int32* pDecodedSamples)
+{
+    drflac_int64 prediction;
+
+    DRFLAC_ASSERT(order <= 32);
+
+    /* 64-bit version. */
+
+    /* This method is faster on the 32-bit build when compiling with VC++. See note below. */
+#ifndef DRFLAC_64BIT
+    if (order == 8)
+    {
+        prediction  = coefficients[0] * (drflac_int64)pDecodedSamples[-1];
+        prediction += coefficients[1] * (drflac_int64)pDecodedSamples[-2];
+        prediction += coefficients[2] * (drflac_int64)pDecodedSamples[-3];
+        prediction += coefficients[3] * (drflac_int64)pDecodedSamples[-4];
+        prediction += coefficients[4] * (drflac_int64)pDecodedSamples[-5];
+        prediction += coefficients[5] * (drflac_int64)pDecodedSamples[-6];
+        prediction += coefficients[6] * (drflac_int64)pDecodedSamples[-7];
+        prediction += coefficients[7] * (drflac_int64)pDecodedSamples[-8];
+    }
+    else if (order == 7)
+    {
+        prediction  = coefficients[0] * (drflac_int64)pDecodedSamples[-1];
+        prediction += coefficients[1] * (drflac_int64)pDecodedSamples[-2];
+        prediction += coefficients[2] * (drflac_int64)pDecodedSamples[-3];
+        prediction += coefficients[3] * (drflac_int64)pDecodedSamples[-4];
+        prediction += coefficients[4] * (drflac_int64)pDecodedSamples[-5];
+        prediction += coefficients[5] * (drflac_int64)pDecodedSamples[-6];
+        prediction += coefficients[6] * (drflac_int64)pDecodedSamples[-7];
+    }
+    else if (order == 3)
+    {
+        prediction  = coefficients[0] * (drflac_int64)pDecodedSamples[-1];
+        prediction += coefficients[1] * (drflac_int64)pDecodedSamples[-2];
+        prediction += coefficients[2] * (drflac_int64)pDecodedSamples[-3];
+    }
+    else if (order == 6)
+    {
+        prediction  = coefficients[0] * (drflac_int64)pDecodedSamples[-1];
+        prediction += coefficients[1] * (drflac_int64)pDecodedSamples[-2];
+        prediction += coefficients[2] * (drflac_int64)pDecodedSamples[-3];
+        prediction += coefficients[3] * (drflac_int64)pDecodedSamples[-4];
+        prediction += coefficients[4] * (drflac_int64)pDecodedSamples[-5];
+        prediction += coefficients[5] * (drflac_int64)pDecodedSamples[-6];
+    }
+    else if (order == 5)
+    {
+        prediction  = coefficients[0] * (drflac_int64)pDecodedSamples[-1];
+        prediction += coefficients[1] * (drflac_int64)pDecodedSamples[-2];
+        prediction += coefficients[2] * (drflac_int64)pDecodedSamples[-3];
+        prediction += coefficients[3] * (drflac_int64)pDecodedSamples[-4];
+        prediction += coefficients[4] * (drflac_int64)pDecodedSamples[-5];
+    }
+    else if (order == 4)
+    {
+        prediction  = coefficients[0] * (drflac_int64)pDecodedSamples[-1];
+        prediction += coefficients[1] * (drflac_int64)pDecodedSamples[-2];
+        prediction += coefficients[2] * (drflac_int64)pDecodedSamples[-3];
+        prediction += coefficients[3] * (drflac_int64)pDecodedSamples[-4];
+    }
+    else if (order == 12)
+    {
+        prediction  = coefficients[0]  * (drflac_int64)pDecodedSamples[-1];
+        prediction += coefficients[1]  * (drflac_int64)pDecodedSamples[-2];
+        prediction += coefficients[2]  * (drflac_int64)pDecodedSamples[-3];
+        prediction += coefficients[3]  * (drflac_int64)pDecodedSamples[-4];
+        prediction += coefficients[4]  * (drflac_int64)pDecodedSamples[-5];
+        prediction += coefficients[5]  * (drflac_int64)pDecodedSamples[-6];
+        prediction += coefficients[6]  * (drflac_int64)pDecodedSamples[-7];
+        prediction += coefficients[7]  * (drflac_int64)pDecodedSamples[-8];
+        prediction += coefficients[8]  * (drflac_int64)pDecodedSamples[-9];
+        prediction += coefficients[9]  * (drflac_int64)pDecodedSamples[-10];
+        prediction += coefficients[10] * (drflac_int64)pDecodedSamples[-11];
+        prediction += coefficients[11] * (drflac_int64)pDecodedSamples[-12];
+    }
+    else if (order == 2)
+    {
+        prediction  = coefficients[0] * (drflac_int64)pDecodedSamples[-1];
+        prediction += coefficients[1] * (drflac_int64)pDecodedSamples[-2];
+    }
+    else if (order == 1)
+    {
+        prediction = coefficients[0] * (drflac_int64)pDecodedSamples[-1];
+    }
+    else if (order == 10)
+    {
+        prediction  = coefficients[0]  * (drflac_int64)pDecodedSamples[-1];
+        prediction += coefficients[1]  * (drflac_int64)pDecodedSamples[-2];
+        prediction += coefficients[2]  * (drflac_int64)pDecodedSamples[-3];
+        prediction += coefficients[3]  * (drflac_int64)pDecodedSamples[-4];
+        prediction += coefficients[4]  * (drflac_int64)pDecodedSamples[-5];
+        prediction += coefficients[5]  * (drflac_int64)pDecodedSamples[-6];
+        prediction += coefficients[6]  * (drflac_int64)pDecodedSamples[-7];
+        prediction += coefficients[7]  * (drflac_int64)pDecodedSamples[-8];
+        prediction += coefficients[8]  * (drflac_int64)pDecodedSamples[-9];
+        prediction += coefficients[9]  * (drflac_int64)pDecodedSamples[-10];
+    }
+    else if (order == 9)
+    {
+        prediction  = coefficients[0]  * (drflac_int64)pDecodedSamples[-1];
+        prediction += coefficients[1]  * (drflac_int64)pDecodedSamples[-2];
+        prediction += coefficients[2]  * (drflac_int64)pDecodedSamples[-3];
+        prediction += coefficients[3]  * (drflac_int64)pDecodedSamples[-4];
+        prediction += coefficients[4]  * (drflac_int64)pDecodedSamples[-5];
+        prediction += coefficients[5]  * (drflac_int64)pDecodedSamples[-6];
+        prediction += coefficients[6]  * (drflac_int64)pDecodedSamples[-7];
+        prediction += coefficients[7]  * (drflac_int64)pDecodedSamples[-8];
+        prediction += coefficients[8]  * (drflac_int64)pDecodedSamples[-9];
+    }
+    else if (order == 11)
+    {
+        prediction  = coefficients[0]  * (drflac_int64)pDecodedSamples[-1];
+        prediction += coefficients[1]  * (drflac_int64)pDecodedSamples[-2];
+        prediction += coefficients[2]  * (drflac_int64)pDecodedSamples[-3];
+        prediction += coefficients[3]  * (drflac_int64)pDecodedSamples[-4];
+        prediction += coefficients[4]  * (drflac_int64)pDecodedSamples[-5];
+        prediction += coefficients[5]  * (drflac_int64)pDecodedSamples[-6];
+        prediction += coefficients[6]  * (drflac_int64)pDecodedSamples[-7];
+        prediction += coefficients[7]  * (drflac_int64)pDecodedSamples[-8];
+        prediction += coefficients[8]  * (drflac_int64)pDecodedSamples[-9];
+        prediction += coefficients[9]  * (drflac_int64)pDecodedSamples[-10];
+        prediction += coefficients[10] * (drflac_int64)pDecodedSamples[-11];
+    }
+    else
+    {
+        int j;
+
+        prediction = 0;
+        for (j = 0; j < (int)order; ++j) {
+            prediction += coefficients[j] * (drflac_int64)pDecodedSamples[-j-1];
+        }
+    }
+#endif
+
+    /*
+    VC++ optimizes this to a single jmp instruction, but only the 64-bit build. The 32-bit build generates less efficient code for some
+    reason. The ugly version above is faster so we'll just switch between the two depending on the target platform.
+    */
+#ifdef DRFLAC_64BIT
+    prediction = 0;
+    switch (order)
+    {
+    case 32: prediction += coefficients[31] * (drflac_int64)pDecodedSamples[-32];
+    case 31: prediction += coefficients[30] * (drflac_int64)pDecodedSamples[-31];
+    case 30: prediction += coefficients[29] * (drflac_int64)pDecodedSamples[-30];
+    case 29: prediction += coefficients[28] * (drflac_int64)pDecodedSamples[-29];
+    case 28: prediction += coefficients[27] * (drflac_int64)pDecodedSamples[-28];
+    case 27: prediction += coefficients[26] * (drflac_int64)pDecodedSamples[-27];
+    case 26: prediction += coefficients[25] * (drflac_int64)pDecodedSamples[-26];
+    case 25: prediction += coefficients[24] * (drflac_int64)pDecodedSamples[-25];
+    case 24: prediction += coefficients[23] * (drflac_int64)pDecodedSamples[-24];
+    case 23: prediction += coefficients[22] * (drflac_int64)pDecodedSamples[-23];
+    case 22: prediction += coefficients[21] * (drflac_int64)pDecodedSamples[-22];
+    case 21: prediction += coefficients[20] * (drflac_int64)pDecodedSamples[-21];
+    case 20: prediction += coefficients[19] * (drflac_int64)pDecodedSamples[-20];
+    case 19: prediction += coefficients[18] * (drflac_int64)pDecodedSamples[-19];
+    case 18: prediction += coefficients[17] * (drflac_int64)pDecodedSamples[-18];
+    case 17: prediction += coefficients[16] * (drflac_int64)pDecodedSamples[-17];
+    case 16: prediction += coefficients[15] * (drflac_int64)pDecodedSamples[-16];
+    case 15: prediction += coefficients[14] * (drflac_int64)pDecodedSamples[-15];
+    case 14: prediction += coefficients[13] * (drflac_int64)pDecodedSamples[-14];
+    case 13: prediction += coefficients[12] * (drflac_int64)pDecodedSamples[-13];
+    case 12: prediction += coefficients[11] * (drflac_int64)pDecodedSamples[-12];
+    case 11: prediction += coefficients[10] * (drflac_int64)pDecodedSamples[-11];
+    case 10: prediction += coefficients[ 9] * (drflac_int64)pDecodedSamples[-10];
+    case  9: prediction += coefficients[ 8] * (drflac_int64)pDecodedSamples[- 9];
+    case  8: prediction += coefficients[ 7] * (drflac_int64)pDecodedSamples[- 8];
+    case  7: prediction += coefficients[ 6] * (drflac_int64)pDecodedSamples[- 7];
+    case  6: prediction += coefficients[ 5] * (drflac_int64)pDecodedSamples[- 6];
+    case  5: prediction += coefficients[ 4] * (drflac_int64)pDecodedSamples[- 5];
+    case  4: prediction += coefficients[ 3] * (drflac_int64)pDecodedSamples[- 4];
+    case  3: prediction += coefficients[ 2] * (drflac_int64)pDecodedSamples[- 3];
+    case  2: prediction += coefficients[ 1] * (drflac_int64)pDecodedSamples[- 2];
+    case  1: prediction += coefficients[ 0] * (drflac_int64)pDecodedSamples[- 1];
+    }
+#endif
+
+    return (drflac_int32)(prediction >> shift);
+}
+
+
+#if 0
+/*
+Reference implementation for reading and decoding samples with residual. This is intentionally left unoptimized for the
+sake of readability and should only be used as a reference.
+*/
+static drflac_bool32 drflac__decode_samples_with_residual__rice__reference(drflac_bs* bs, drflac_uint32 bitsPerSample, drflac_uint32 count, drflac_uint8 riceParam, drflac_uint32 lpcOrder, drflac_int32 lpcShift, drflac_uint32 lpcPrecision, const drflac_int32* coefficients, drflac_int32* pSamplesOut)
+{
+    drflac_uint32 i;
+
+    DRFLAC_ASSERT(bs != NULL);
+    DRFLAC_ASSERT(pSamplesOut != NULL);
+
+    for (i = 0; i < count; ++i) {
+        drflac_uint32 zeroCounter = 0;
+        for (;;) {
+            drflac_uint8 bit;
+            if (!drflac__read_uint8(bs, 1, &bit)) {
+                return DRFLAC_FALSE;
+            }
+
+            if (bit == 0) {
+                zeroCounter += 1;
+            } else {
+                break;
+            }
+        }
+
+        drflac_uint32 decodedRice;
+        if (riceParam > 0) {
+            if (!drflac__read_uint32(bs, riceParam, &decodedRice)) {
+                return DRFLAC_FALSE;
+            }
+        } else {
+            decodedRice = 0;
+        }
+
+        decodedRice |= (zeroCounter << riceParam);
+        if ((decodedRice & 0x01)) {
+            decodedRice = ~(decodedRice >> 1);
+        } else {
+            decodedRice =  (decodedRice >> 1);
+        }
+
+
+        if (drflac__use_64_bit_prediction(bitsPerSample, lpcOrder, lpcPrecision)) {
+            pSamplesOut[i] = decodedRice + drflac__calculate_prediction_64(lpcOrder, lpcShift, coefficients, pSamplesOut + i);
+        } else {
+            pSamplesOut[i] = decodedRice + drflac__calculate_prediction_32(lpcOrder, lpcShift, coefficients, pSamplesOut + i);
+        }
+    }
+
+    return DRFLAC_TRUE;
+}
+#endif
+
+#if 0
+static drflac_bool32 drflac__read_rice_parts__reference(drflac_bs* bs, drflac_uint8 riceParam, drflac_uint32* pZeroCounterOut, drflac_uint32* pRiceParamPartOut)
+{
+    drflac_uint32 zeroCounter = 0;
+    drflac_uint32 decodedRice;
+
+    for (;;) {
+        drflac_uint8 bit;
+        if (!drflac__read_uint8(bs, 1, &bit)) {
+            return DRFLAC_FALSE;
+        }
+
+        if (bit == 0) {
+            zeroCounter += 1;
+        } else {
+            break;
+        }
+    }
+
+    if (riceParam > 0) {
+        if (!drflac__read_uint32(bs, riceParam, &decodedRice)) {
+            return DRFLAC_FALSE;
+        }
+    } else {
+        decodedRice = 0;
+    }
+
+    *pZeroCounterOut = zeroCounter;
+    *pRiceParamPartOut = decodedRice;
+    return DRFLAC_TRUE;
+}
+#endif
+
+#if 0
+static DRFLAC_INLINE drflac_bool32 drflac__read_rice_parts(drflac_bs* bs, drflac_uint8 riceParam, drflac_uint32* pZeroCounterOut, drflac_uint32* pRiceParamPartOut)
+{
+    drflac_cache_t riceParamMask;
+    drflac_uint32 zeroCounter;
+    drflac_uint32 setBitOffsetPlus1;
+    drflac_uint32 riceParamPart;
+    drflac_uint32 riceLength;
+
+    DRFLAC_ASSERT(riceParam > 0);   /* <-- riceParam should never be 0. drflac__read_rice_parts__param_equals_zero() should be used instead for this case. */
+
+    riceParamMask = DRFLAC_CACHE_L1_SELECTION_MASK(riceParam);
+
+    zeroCounter = 0;
+    while (bs->cache == 0) {
+        zeroCounter += (drflac_uint32)DRFLAC_CACHE_L1_BITS_REMAINING(bs);
+        if (!drflac__reload_cache(bs)) {
+            return DRFLAC_FALSE;
+        }
+    }
+
+    setBitOffsetPlus1 = drflac__clz(bs->cache);
+    zeroCounter += setBitOffsetPlus1;
+    setBitOffsetPlus1 += 1;
+
+    riceLength = setBitOffsetPlus1 + riceParam;
+    if (riceLength < DRFLAC_CACHE_L1_BITS_REMAINING(bs)) {
+        riceParamPart = (drflac_uint32)((bs->cache & (riceParamMask >> setBitOffsetPlus1)) >> DRFLAC_CACHE_L1_SELECTION_SHIFT(bs, riceLength));
+
+        bs->consumedBits += riceLength;
+        bs->cache <<= riceLength;
+    } else {
+        drflac_uint32 bitCountLo;
+        drflac_cache_t resultHi;
+
+        bs->consumedBits += riceLength;
+        bs->cache <<= setBitOffsetPlus1 & (DRFLAC_CACHE_L1_SIZE_BITS(bs)-1);    /* <-- Equivalent to "if (setBitOffsetPlus1 < DRFLAC_CACHE_L1_SIZE_BITS(bs)) { bs->cache <<= setBitOffsetPlus1; }" */
+
+        /* It straddles the cached data. It will never cover more than the next chunk. We just read the number in two parts and combine them. */
+        bitCountLo = bs->consumedBits - DRFLAC_CACHE_L1_SIZE_BITS(bs);
+        resultHi = DRFLAC_CACHE_L1_SELECT_AND_SHIFT(bs, riceParam);  /* <-- Use DRFLAC_CACHE_L1_SELECT_AND_SHIFT_SAFE() if ever this function allows riceParam=0. */
+
+        if (bs->nextL2Line < DRFLAC_CACHE_L2_LINE_COUNT(bs)) {
+#ifndef DR_FLAC_NO_CRC
+            drflac__update_crc16(bs);
+#endif
+            bs->cache = drflac__be2host__cache_line(bs->cacheL2[bs->nextL2Line++]);
+            bs->consumedBits = 0;
+#ifndef DR_FLAC_NO_CRC
+            bs->crc16Cache = bs->cache;
+#endif
+        } else {
+            /* Slow path. We need to fetch more data from the client. */
+            if (!drflac__reload_cache(bs)) {
+                return DRFLAC_FALSE;
+            }
+            if (bitCountLo > DRFLAC_CACHE_L1_BITS_REMAINING(bs)) {
+                /* This happens when we get to end of stream */
+                return DRFLAC_FALSE;
+            }
+        }
+
+        riceParamPart = (drflac_uint32)(resultHi | DRFLAC_CACHE_L1_SELECT_AND_SHIFT_SAFE(bs, bitCountLo));
+
+        bs->consumedBits += bitCountLo;
+        bs->cache <<= bitCountLo;
+    }
+
+    pZeroCounterOut[0] = zeroCounter;
+    pRiceParamPartOut[0] = riceParamPart;
+
+    return DRFLAC_TRUE;
+}
+#endif
+
+static DRFLAC_INLINE drflac_bool32 drflac__read_rice_parts_x1(drflac_bs* bs, drflac_uint8 riceParam, drflac_uint32* pZeroCounterOut, drflac_uint32* pRiceParamPartOut)
+{
+    drflac_uint32  riceParamPlus1 = riceParam + 1;
+    /*drflac_cache_t riceParamPlus1Mask  = DRFLAC_CACHE_L1_SELECTION_MASK(riceParamPlus1);*/
+    drflac_uint32  riceParamPlus1Shift = DRFLAC_CACHE_L1_SELECTION_SHIFT(bs, riceParamPlus1);
+    drflac_uint32  riceParamPlus1MaxConsumedBits = DRFLAC_CACHE_L1_SIZE_BITS(bs) - riceParamPlus1;
+
+    /*
+    The idea here is to use local variables for the cache in an attempt to encourage the compiler to store them in registers. I have
+    no idea how this will work in practice...
+    */
+    drflac_cache_t bs_cache = bs->cache;
+    drflac_uint32  bs_consumedBits = bs->consumedBits;
+
+    /* The first thing to do is find the first unset bit. Most likely a bit will be set in the current cache line. */
+    drflac_uint32  lzcount = drflac__clz(bs_cache);
+    if (lzcount < sizeof(bs_cache)*8) {
+        pZeroCounterOut[0] = lzcount;
+
+        /*
+        It is most likely that the riceParam part (which comes after the zero counter) is also on this cache line. When extracting
+        this, we include the set bit from the unary coded part because it simplifies cache management. This bit will be handled
+        outside of this function at a higher level.
+        */
+    extract_rice_param_part:
+        bs_cache       <<= lzcount;
+        bs_consumedBits += lzcount;
+
+        if (bs_consumedBits <= riceParamPlus1MaxConsumedBits) {
+            /* Getting here means the rice parameter part is wholly contained within the current cache line. */
+            pRiceParamPartOut[0] = (drflac_uint32)(bs_cache >> riceParamPlus1Shift);
+            bs_cache       <<= riceParamPlus1;
+            bs_consumedBits += riceParamPlus1;
+        } else {
+            drflac_uint32 riceParamPartHi;
+            drflac_uint32 riceParamPartLo;
+            drflac_uint32 riceParamPartLoBitCount;
+
+            /*
+            Getting here means the rice parameter part straddles the cache line. We need to read from the tail of the current cache
+            line, reload the cache, and then combine it with the head of the next cache line.
+            */
+
+            /* Grab the high part of the rice parameter part. */
+            riceParamPartHi = (drflac_uint32)(bs_cache >> riceParamPlus1Shift);
+
+            /* Before reloading the cache we need to grab the size in bits of the low part. */
+            riceParamPartLoBitCount = bs_consumedBits - riceParamPlus1MaxConsumedBits;
+            DRFLAC_ASSERT(riceParamPartLoBitCount > 0 && riceParamPartLoBitCount < 32);
+
+            /* Now reload the cache. */
+            if (bs->nextL2Line < DRFLAC_CACHE_L2_LINE_COUNT(bs)) {
+            #ifndef DR_FLAC_NO_CRC
+                drflac__update_crc16(bs);
+            #endif
+                bs_cache = drflac__be2host__cache_line(bs->cacheL2[bs->nextL2Line++]);
+                bs_consumedBits = riceParamPartLoBitCount;
+            #ifndef DR_FLAC_NO_CRC
+                bs->crc16Cache = bs_cache;
+            #endif
+            } else {
+                /* Slow path. We need to fetch more data from the client. */
+                if (!drflac__reload_cache(bs)) {
+                    return DRFLAC_FALSE;
+                }
+                if (riceParamPartLoBitCount > DRFLAC_CACHE_L1_BITS_REMAINING(bs)) {
+                    /* This happens when we get to end of stream */
+                    return DRFLAC_FALSE;
+                }
+
+                bs_cache = bs->cache;
+                bs_consumedBits = bs->consumedBits + riceParamPartLoBitCount;
+            }
+
+            /* We should now have enough information to construct the rice parameter part. */
+            riceParamPartLo = (drflac_uint32)(bs_cache >> (DRFLAC_CACHE_L1_SELECTION_SHIFT(bs, riceParamPartLoBitCount)));
+            pRiceParamPartOut[0] = riceParamPartHi | riceParamPartLo;
+
+            bs_cache <<= riceParamPartLoBitCount;
+        }
+    } else {
+        /*
+        Getting here means there are no bits set on the cache line. This is a less optimal case because we just wasted a call
+        to drflac__clz() and we need to reload the cache.
+        */
+        drflac_uint32 zeroCounter = (drflac_uint32)(DRFLAC_CACHE_L1_SIZE_BITS(bs) - bs_consumedBits);
+        for (;;) {
+            if (bs->nextL2Line < DRFLAC_CACHE_L2_LINE_COUNT(bs)) {
+            #ifndef DR_FLAC_NO_CRC
+                drflac__update_crc16(bs);
+            #endif
+                bs_cache = drflac__be2host__cache_line(bs->cacheL2[bs->nextL2Line++]);
+                bs_consumedBits = 0;
+            #ifndef DR_FLAC_NO_CRC
+                bs->crc16Cache = bs_cache;
+            #endif
+            } else {
+                /* Slow path. We need to fetch more data from the client. */
+                if (!drflac__reload_cache(bs)) {
+                    return DRFLAC_FALSE;
+                }
+
+                bs_cache = bs->cache;
+                bs_consumedBits = bs->consumedBits;
+            }
+
+            lzcount = drflac__clz(bs_cache);
+            zeroCounter += lzcount;
+
+            if (lzcount < sizeof(bs_cache)*8) {
+                break;
+            }
+        }
+
+        pZeroCounterOut[0] = zeroCounter;
+        goto extract_rice_param_part;
+    }
+
+    /* Make sure the cache is restored at the end of it all. */
+    bs->cache = bs_cache;
+    bs->consumedBits = bs_consumedBits;
+
+    return DRFLAC_TRUE;
+}
+
+static DRFLAC_INLINE drflac_bool32 drflac__seek_rice_parts(drflac_bs* bs, drflac_uint8 riceParam)
+{
+    drflac_uint32  riceParamPlus1 = riceParam + 1;
+    drflac_uint32  riceParamPlus1MaxConsumedBits = DRFLAC_CACHE_L1_SIZE_BITS(bs) - riceParamPlus1;
+
+    /*
+    The idea here is to use local variables for the cache in an attempt to encourage the compiler to store them in registers. I have
+    no idea how this will work in practice...
+    */
+    drflac_cache_t bs_cache = bs->cache;
+    drflac_uint32  bs_consumedBits = bs->consumedBits;
+
+    /* The first thing to do is find the first unset bit. Most likely a bit will be set in the current cache line. */
+    drflac_uint32  lzcount = drflac__clz(bs_cache);
+    if (lzcount < sizeof(bs_cache)*8) {
+        /*
+        It is most likely that the riceParam part (which comes after the zero counter) is also on this cache line. When extracting
+        this, we include the set bit from the unary coded part because it simplifies cache management. This bit will be handled
+        outside of this function at a higher level.
+        */
+    extract_rice_param_part:
+        bs_cache       <<= lzcount;
+        bs_consumedBits += lzcount;
+
+        if (bs_consumedBits <= riceParamPlus1MaxConsumedBits) {
+            /* Getting here means the rice parameter part is wholly contained within the current cache line. */
+            bs_cache       <<= riceParamPlus1;
+            bs_consumedBits += riceParamPlus1;
+        } else {
+            /*
+            Getting here means the rice parameter part straddles the cache line. We need to read from the tail of the current cache
+            line, reload the cache, and then combine it with the head of the next cache line.
+            */
+
+            /* Before reloading the cache we need to grab the size in bits of the low part. */
+            drflac_uint32 riceParamPartLoBitCount = bs_consumedBits - riceParamPlus1MaxConsumedBits;
+            DRFLAC_ASSERT(riceParamPartLoBitCount > 0 && riceParamPartLoBitCount < 32);
+
+            /* Now reload the cache. */
+            if (bs->nextL2Line < DRFLAC_CACHE_L2_LINE_COUNT(bs)) {
+            #ifndef DR_FLAC_NO_CRC
+                drflac__update_crc16(bs);
+            #endif
+                bs_cache = drflac__be2host__cache_line(bs->cacheL2[bs->nextL2Line++]);
+                bs_consumedBits = riceParamPartLoBitCount;
+            #ifndef DR_FLAC_NO_CRC
+                bs->crc16Cache = bs_cache;
+            #endif
+            } else {
+                /* Slow path. We need to fetch more data from the client. */
+                if (!drflac__reload_cache(bs)) {
+                    return DRFLAC_FALSE;
+                }
+
+                if (riceParamPartLoBitCount > DRFLAC_CACHE_L1_BITS_REMAINING(bs)) {
+                    /* This happens when we get to end of stream */
+                    return DRFLAC_FALSE;
+                }
+
+                bs_cache = bs->cache;
+                bs_consumedBits = bs->consumedBits + riceParamPartLoBitCount;
+            }
+
+            bs_cache <<= riceParamPartLoBitCount;
+        }
+    } else {
+        /*
+        Getting here means there are no bits set on the cache line. This is a less optimal case because we just wasted a call
+        to drflac__clz() and we need to reload the cache.
+        */
+        for (;;) {
+            if (bs->nextL2Line < DRFLAC_CACHE_L2_LINE_COUNT(bs)) {
+            #ifndef DR_FLAC_NO_CRC
+                drflac__update_crc16(bs);
+            #endif
+                bs_cache = drflac__be2host__cache_line(bs->cacheL2[bs->nextL2Line++]);
+                bs_consumedBits = 0;
+            #ifndef DR_FLAC_NO_CRC
+                bs->crc16Cache = bs_cache;
+            #endif
+            } else {
+                /* Slow path. We need to fetch more data from the client. */
+                if (!drflac__reload_cache(bs)) {
+                    return DRFLAC_FALSE;
+                }
+
+                bs_cache = bs->cache;
+                bs_consumedBits = bs->consumedBits;
+            }
+
+            lzcount = drflac__clz(bs_cache);
+            if (lzcount < sizeof(bs_cache)*8) {
+                break;
+            }
+        }
+
+        goto extract_rice_param_part;
+    }
+
+    /* Make sure the cache is restored at the end of it all. */
+    bs->cache = bs_cache;
+    bs->consumedBits = bs_consumedBits;
+
+    return DRFLAC_TRUE;
+}
+
+
+static drflac_bool32 drflac__decode_samples_with_residual__rice__scalar_zeroorder(drflac_bs* bs, drflac_uint32 bitsPerSample, drflac_uint32 count, drflac_uint8 riceParam, drflac_uint32 order, drflac_int32 shift, const drflac_int32* coefficients, drflac_int32* pSamplesOut)
+{
+    drflac_uint32 t[2] = {0x00000000, 0xFFFFFFFF};
+    drflac_uint32 zeroCountPart0;
+    drflac_uint32 riceParamPart0;
+    drflac_uint32 riceParamMask;
+    drflac_uint32 i;
+
+    DRFLAC_ASSERT(bs != NULL);
+    DRFLAC_ASSERT(pSamplesOut != NULL);
+
+    (void)bitsPerSample;
+    (void)order;
+    (void)shift;
+    (void)coefficients;
+
+    riceParamMask  = (drflac_uint32)~((~0UL) << riceParam);
+
+    i = 0;
+    while (i < count) {
+        /* Rice extraction. */
+        if (!drflac__read_rice_parts_x1(bs, riceParam, &zeroCountPart0, &riceParamPart0)) {
+            return DRFLAC_FALSE;
+        }
+
+        /* Rice reconstruction. */
+        riceParamPart0 &= riceParamMask;
+        riceParamPart0 |= (zeroCountPart0 << riceParam);
+        riceParamPart0  = (riceParamPart0 >> 1) ^ t[riceParamPart0 & 0x01];
+
+        pSamplesOut[i] = riceParamPart0;
+
+        i += 1;
+    }
+
+    return DRFLAC_TRUE;
+}
+
+static drflac_bool32 drflac__decode_samples_with_residual__rice__scalar(drflac_bs* bs, drflac_uint32 bitsPerSample, drflac_uint32 count, drflac_uint8 riceParam, drflac_uint32 lpcOrder, drflac_int32 lpcShift, drflac_uint32 lpcPrecision, const drflac_int32* coefficients, drflac_int32* pSamplesOut)
+{
+    drflac_uint32 t[2] = {0x00000000, 0xFFFFFFFF};
+    drflac_uint32 zeroCountPart0 = 0;
+    drflac_uint32 zeroCountPart1 = 0;
+    drflac_uint32 zeroCountPart2 = 0;
+    drflac_uint32 zeroCountPart3 = 0;
+    drflac_uint32 riceParamPart0 = 0;
+    drflac_uint32 riceParamPart1 = 0;
+    drflac_uint32 riceParamPart2 = 0;
+    drflac_uint32 riceParamPart3 = 0;
+    drflac_uint32 riceParamMask;
+    const drflac_int32* pSamplesOutEnd;
+    drflac_uint32 i;
+
+    DRFLAC_ASSERT(bs != NULL);
+    DRFLAC_ASSERT(pSamplesOut != NULL);
+
+    if (lpcOrder == 0) {
+        return drflac__decode_samples_with_residual__rice__scalar_zeroorder(bs, bitsPerSample, count, riceParam, lpcOrder, lpcShift, coefficients, pSamplesOut);
+    }
+
+    riceParamMask  = (drflac_uint32)~((~0UL) << riceParam);
+    pSamplesOutEnd = pSamplesOut + (count & ~3);
+
+    if (drflac__use_64_bit_prediction(bitsPerSample, lpcOrder, lpcPrecision)) {
+        while (pSamplesOut < pSamplesOutEnd) {
+            /*
+            Rice extraction. It's faster to do this one at a time against local variables than it is to use the x4 version
+            against an array. Not sure why, but perhaps it's making more efficient use of registers?
+            */
+            if (!drflac__read_rice_parts_x1(bs, riceParam, &zeroCountPart0, &riceParamPart0) ||
+                !drflac__read_rice_parts_x1(bs, riceParam, &zeroCountPart1, &riceParamPart1) ||
+                !drflac__read_rice_parts_x1(bs, riceParam, &zeroCountPart2, &riceParamPart2) ||
+                !drflac__read_rice_parts_x1(bs, riceParam, &zeroCountPart3, &riceParamPart3)) {
+                return DRFLAC_FALSE;
+            }
+
+            riceParamPart0 &= riceParamMask;
+            riceParamPart1 &= riceParamMask;
+            riceParamPart2 &= riceParamMask;
+            riceParamPart3 &= riceParamMask;
+
+            riceParamPart0 |= (zeroCountPart0 << riceParam);
+            riceParamPart1 |= (zeroCountPart1 << riceParam);
+            riceParamPart2 |= (zeroCountPart2 << riceParam);
+            riceParamPart3 |= (zeroCountPart3 << riceParam);
+
+            riceParamPart0  = (riceParamPart0 >> 1) ^ t[riceParamPart0 & 0x01];
+            riceParamPart1  = (riceParamPart1 >> 1) ^ t[riceParamPart1 & 0x01];
+            riceParamPart2  = (riceParamPart2 >> 1) ^ t[riceParamPart2 & 0x01];
+            riceParamPart3  = (riceParamPart3 >> 1) ^ t[riceParamPart3 & 0x01];
+
+            pSamplesOut[0] = riceParamPart0 + drflac__calculate_prediction_64(lpcOrder, lpcShift, coefficients, pSamplesOut + 0);
+            pSamplesOut[1] = riceParamPart1 + drflac__calculate_prediction_64(lpcOrder, lpcShift, coefficients, pSamplesOut + 1);
+            pSamplesOut[2] = riceParamPart2 + drflac__calculate_prediction_64(lpcOrder, lpcShift, coefficients, pSamplesOut + 2);
+            pSamplesOut[3] = riceParamPart3 + drflac__calculate_prediction_64(lpcOrder, lpcShift, coefficients, pSamplesOut + 3);
+
+            pSamplesOut += 4;
+        }
+    } else {
+        while (pSamplesOut < pSamplesOutEnd) {
+            if (!drflac__read_rice_parts_x1(bs, riceParam, &zeroCountPart0, &riceParamPart0) ||
+                !drflac__read_rice_parts_x1(bs, riceParam, &zeroCountPart1, &riceParamPart1) ||
+                !drflac__read_rice_parts_x1(bs, riceParam, &zeroCountPart2, &riceParamPart2) ||
+                !drflac__read_rice_parts_x1(bs, riceParam, &zeroCountPart3, &riceParamPart3)) {
+                return DRFLAC_FALSE;
+            }
+
+            riceParamPart0 &= riceParamMask;
+            riceParamPart1 &= riceParamMask;
+            riceParamPart2 &= riceParamMask;
+            riceParamPart3 &= riceParamMask;
+
+            riceParamPart0 |= (zeroCountPart0 << riceParam);
+            riceParamPart1 |= (zeroCountPart1 << riceParam);
+            riceParamPart2 |= (zeroCountPart2 << riceParam);
+            riceParamPart3 |= (zeroCountPart3 << riceParam);
+
+            riceParamPart0  = (riceParamPart0 >> 1) ^ t[riceParamPart0 & 0x01];
+            riceParamPart1  = (riceParamPart1 >> 1) ^ t[riceParamPart1 & 0x01];
+            riceParamPart2  = (riceParamPart2 >> 1) ^ t[riceParamPart2 & 0x01];
+            riceParamPart3  = (riceParamPart3 >> 1) ^ t[riceParamPart3 & 0x01];
+
+            pSamplesOut[0] = riceParamPart0 + drflac__calculate_prediction_32(lpcOrder, lpcShift, coefficients, pSamplesOut + 0);
+            pSamplesOut[1] = riceParamPart1 + drflac__calculate_prediction_32(lpcOrder, lpcShift, coefficients, pSamplesOut + 1);
+            pSamplesOut[2] = riceParamPart2 + drflac__calculate_prediction_32(lpcOrder, lpcShift, coefficients, pSamplesOut + 2);
+            pSamplesOut[3] = riceParamPart3 + drflac__calculate_prediction_32(lpcOrder, lpcShift, coefficients, pSamplesOut + 3);
+
+            pSamplesOut += 4;
+        }
+    }
+
+    i = (count & ~3);
+    while (i < count) {
+        /* Rice extraction. */
+        if (!drflac__read_rice_parts_x1(bs, riceParam, &zeroCountPart0, &riceParamPart0)) {
+            return DRFLAC_FALSE;
+        }
+
+        /* Rice reconstruction. */
+        riceParamPart0 &= riceParamMask;
+        riceParamPart0 |= (zeroCountPart0 << riceParam);
+        riceParamPart0  = (riceParamPart0 >> 1) ^ t[riceParamPart0 & 0x01];
+        /*riceParamPart0  = (riceParamPart0 >> 1) ^ (~(riceParamPart0 & 0x01) + 1);*/
+
+        /* Sample reconstruction. */
+        if (drflac__use_64_bit_prediction(bitsPerSample, lpcOrder, lpcPrecision)) {
+            pSamplesOut[0] = riceParamPart0 + drflac__calculate_prediction_64(lpcOrder, lpcShift, coefficients, pSamplesOut + 0);
+        } else {
+            pSamplesOut[0] = riceParamPart0 + drflac__calculate_prediction_32(lpcOrder, lpcShift, coefficients, pSamplesOut + 0);
+        }
+
+        i += 1;
+        pSamplesOut += 1;
+    }
+
+    return DRFLAC_TRUE;
+}
+
+#if defined(DRFLAC_SUPPORT_SSE2)
+static DRFLAC_INLINE __m128i drflac__mm_packs_interleaved_epi32(__m128i a, __m128i b)
+{
+    __m128i r;
+
+    /* Pack. */
+    r = _mm_packs_epi32(a, b);
+
+    /* a3a2 a1a0 b3b2 b1b0 -> a3a2 b3b2 a1a0 b1b0 */
+    r = _mm_shuffle_epi32(r, _MM_SHUFFLE(3, 1, 2, 0));
+
+    /* a3a2 b3b2 a1a0 b1b0 -> a3b3 a2b2 a1b1 a0b0 */
+    r = _mm_shufflehi_epi16(r, _MM_SHUFFLE(3, 1, 2, 0));
+    r = _mm_shufflelo_epi16(r, _MM_SHUFFLE(3, 1, 2, 0));
+
+    return r;
+}
+#endif
+
+#if defined(DRFLAC_SUPPORT_SSE41)
+static DRFLAC_INLINE __m128i drflac__mm_not_si128(__m128i a)
+{
+    return _mm_xor_si128(a, _mm_cmpeq_epi32(_mm_setzero_si128(), _mm_setzero_si128()));
+}
+
+static DRFLAC_INLINE __m128i drflac__mm_hadd_epi32(__m128i x)
+{
+    __m128i x64 = _mm_add_epi32(x, _mm_shuffle_epi32(x, _MM_SHUFFLE(1, 0, 3, 2)));
+    __m128i x32 = _mm_shufflelo_epi16(x64, _MM_SHUFFLE(1, 0, 3, 2));
+    return _mm_add_epi32(x64, x32);
+}
+
+static DRFLAC_INLINE __m128i drflac__mm_hadd_epi64(__m128i x)
+{
+    return _mm_add_epi64(x, _mm_shuffle_epi32(x, _MM_SHUFFLE(1, 0, 3, 2)));
+}
+
+static DRFLAC_INLINE __m128i drflac__mm_srai_epi64(__m128i x, int count)
+{
+    /*
+    To simplify this we are assuming count < 32. This restriction allows us to work on a low side and a high side. The low side
+    is shifted with zero bits, whereas the right side is shifted with sign bits.
+    */
+    __m128i lo = _mm_srli_epi64(x, count);
+    __m128i hi = _mm_srai_epi32(x, count);
+
+    hi = _mm_and_si128(hi, _mm_set_epi32(0xFFFFFFFF, 0, 0xFFFFFFFF, 0));    /* The high part needs to have the low part cleared. */
+
+    return _mm_or_si128(lo, hi);
+}
+
+static drflac_bool32 drflac__decode_samples_with_residual__rice__sse41_32(drflac_bs* bs, drflac_uint32 count, drflac_uint8 riceParam, drflac_uint32 order, drflac_int32 shift, const drflac_int32* coefficients, drflac_int32* pSamplesOut)
+{
+    int i;
+    drflac_uint32 riceParamMask;
+    drflac_int32* pDecodedSamples    = pSamplesOut;
+    drflac_int32* pDecodedSamplesEnd = pSamplesOut + (count & ~3);
+    drflac_uint32 zeroCountParts0 = 0;
+    drflac_uint32 zeroCountParts1 = 0;
+    drflac_uint32 zeroCountParts2 = 0;
+    drflac_uint32 zeroCountParts3 = 0;
+    drflac_uint32 riceParamParts0 = 0;
+    drflac_uint32 riceParamParts1 = 0;
+    drflac_uint32 riceParamParts2 = 0;
+    drflac_uint32 riceParamParts3 = 0;
+    __m128i coefficients128_0;
+    __m128i coefficients128_4;
+    __m128i coefficients128_8;
+    __m128i samples128_0;
+    __m128i samples128_4;
+    __m128i samples128_8;
+    __m128i riceParamMask128;
+
+    const drflac_uint32 t[2] = {0x00000000, 0xFFFFFFFF};
+
+    riceParamMask    = (drflac_uint32)~((~0UL) << riceParam);
+    riceParamMask128 = _mm_set1_epi32(riceParamMask);
+
+    /* Pre-load. */
+    coefficients128_0 = _mm_setzero_si128();
+    coefficients128_4 = _mm_setzero_si128();
+    coefficients128_8 = _mm_setzero_si128();
+
+    samples128_0 = _mm_setzero_si128();
+    samples128_4 = _mm_setzero_si128();
+    samples128_8 = _mm_setzero_si128();
+
+    /*
+    Pre-loading the coefficients and prior samples is annoying because we need to ensure we don't try reading more than
+    what's available in the input buffers. It would be convenient to use a fall-through switch to do this, but this results
+    in strict aliasing warnings with GCC. To work around this I'm just doing something hacky. This feels a bit convoluted
+    so I think there's opportunity for this to be simplified.
+    */
+#if 1
+    {
+        int runningOrder = order;
+
+        /* 0 - 3. */
+        if (runningOrder >= 4) {
+            coefficients128_0 = _mm_loadu_si128((const __m128i*)(coefficients + 0));
+            samples128_0      = _mm_loadu_si128((const __m128i*)(pSamplesOut  - 4));
+            runningOrder -= 4;
+        } else {
+            switch (runningOrder) {
+                case 3: coefficients128_0 = _mm_set_epi32(0, coefficients[2], coefficients[1], coefficients[0]); samples128_0 = _mm_set_epi32(pSamplesOut[-1], pSamplesOut[-2], pSamplesOut[-3], 0); break;
+                case 2: coefficients128_0 = _mm_set_epi32(0, 0,               coefficients[1], coefficients[0]); samples128_0 = _mm_set_epi32(pSamplesOut[-1], pSamplesOut[-2], 0,               0); break;
+                case 1: coefficients128_0 = _mm_set_epi32(0, 0,               0,               coefficients[0]); samples128_0 = _mm_set_epi32(pSamplesOut[-1], 0,               0,               0); break;
+            }
+            runningOrder = 0;
+        }
+
+        /* 4 - 7 */
+        if (runningOrder >= 4) {
+            coefficients128_4 = _mm_loadu_si128((const __m128i*)(coefficients + 4));
+            samples128_4      = _mm_loadu_si128((const __m128i*)(pSamplesOut  - 8));
+            runningOrder -= 4;
+        } else {
+            switch (runningOrder) {
+                case 3: coefficients128_4 = _mm_set_epi32(0, coefficients[6], coefficients[5], coefficients[4]); samples128_4 = _mm_set_epi32(pSamplesOut[-5], pSamplesOut[-6], pSamplesOut[-7], 0); break;
+                case 2: coefficients128_4 = _mm_set_epi32(0, 0,               coefficients[5], coefficients[4]); samples128_4 = _mm_set_epi32(pSamplesOut[-5], pSamplesOut[-6], 0,               0); break;
+                case 1: coefficients128_4 = _mm_set_epi32(0, 0,               0,               coefficients[4]); samples128_4 = _mm_set_epi32(pSamplesOut[-5], 0,               0,               0); break;
+            }
+            runningOrder = 0;
+        }
+
+        /* 8 - 11 */
+        if (runningOrder == 4) {
+            coefficients128_8 = _mm_loadu_si128((const __m128i*)(coefficients + 8));
+            samples128_8      = _mm_loadu_si128((const __m128i*)(pSamplesOut  - 12));
+            runningOrder -= 4;
+        } else {
+            switch (runningOrder) {
+                case 3: coefficients128_8 = _mm_set_epi32(0, coefficients[10], coefficients[9], coefficients[8]); samples128_8 = _mm_set_epi32(pSamplesOut[-9], pSamplesOut[-10], pSamplesOut[-11], 0); break;
+                case 2: coefficients128_8 = _mm_set_epi32(0, 0,                coefficients[9], coefficients[8]); samples128_8 = _mm_set_epi32(pSamplesOut[-9], pSamplesOut[-10], 0,                0); break;
+                case 1: coefficients128_8 = _mm_set_epi32(0, 0,                0,               coefficients[8]); samples128_8 = _mm_set_epi32(pSamplesOut[-9], 0,                0,                0); break;
+            }
+            runningOrder = 0;
+        }
+
+        /* Coefficients need to be shuffled for our streaming algorithm below to work. Samples are already in the correct order from the loading routine above. */
+        coefficients128_0 = _mm_shuffle_epi32(coefficients128_0, _MM_SHUFFLE(0, 1, 2, 3));
+        coefficients128_4 = _mm_shuffle_epi32(coefficients128_4, _MM_SHUFFLE(0, 1, 2, 3));
+        coefficients128_8 = _mm_shuffle_epi32(coefficients128_8, _MM_SHUFFLE(0, 1, 2, 3));
+    }
+#else
+    /* This causes strict-aliasing warnings with GCC. */
+    switch (order)
+    {
+    case 12: ((drflac_int32*)&coefficients128_8)[0] = coefficients[11]; ((drflac_int32*)&samples128_8)[0] = pDecodedSamples[-12];
+    case 11: ((drflac_int32*)&coefficients128_8)[1] = coefficients[10]; ((drflac_int32*)&samples128_8)[1] = pDecodedSamples[-11];
+    case 10: ((drflac_int32*)&coefficients128_8)[2] = coefficients[ 9]; ((drflac_int32*)&samples128_8)[2] = pDecodedSamples[-10];
+    case 9:  ((drflac_int32*)&coefficients128_8)[3] = coefficients[ 8]; ((drflac_int32*)&samples128_8)[3] = pDecodedSamples[- 9];
+    case 8:  ((drflac_int32*)&coefficients128_4)[0] = coefficients[ 7]; ((drflac_int32*)&samples128_4)[0] = pDecodedSamples[- 8];
+    case 7:  ((drflac_int32*)&coefficients128_4)[1] = coefficients[ 6]; ((drflac_int32*)&samples128_4)[1] = pDecodedSamples[- 7];
+    case 6:  ((drflac_int32*)&coefficients128_4)[2] = coefficients[ 5]; ((drflac_int32*)&samples128_4)[2] = pDecodedSamples[- 6];
+    case 5:  ((drflac_int32*)&coefficients128_4)[3] = coefficients[ 4]; ((drflac_int32*)&samples128_4)[3] = pDecodedSamples[- 5];
+    case 4:  ((drflac_int32*)&coefficients128_0)[0] = coefficients[ 3]; ((drflac_int32*)&samples128_0)[0] = pDecodedSamples[- 4];
+    case 3:  ((drflac_int32*)&coefficients128_0)[1] = coefficients[ 2]; ((drflac_int32*)&samples128_0)[1] = pDecodedSamples[- 3];
+    case 2:  ((drflac_int32*)&coefficients128_0)[2] = coefficients[ 1]; ((drflac_int32*)&samples128_0)[2] = pDecodedSamples[- 2];
+    case 1:  ((drflac_int32*)&coefficients128_0)[3] = coefficients[ 0]; ((drflac_int32*)&samples128_0)[3] = pDecodedSamples[- 1];
+    }
+#endif
+
+    /* For this version we are doing one sample at a time. */
+    while (pDecodedSamples < pDecodedSamplesEnd) {
+        __m128i prediction128;
+        __m128i zeroCountPart128;
+        __m128i riceParamPart128;
+
+        if (!drflac__read_rice_parts_x1(bs, riceParam, &zeroCountParts0, &riceParamParts0) ||
+            !drflac__read_rice_parts_x1(bs, riceParam, &zeroCountParts1, &riceParamParts1) ||
+            !drflac__read_rice_parts_x1(bs, riceParam, &zeroCountParts2, &riceParamParts2) ||
+            !drflac__read_rice_parts_x1(bs, riceParam, &zeroCountParts3, &riceParamParts3)) {
+            return DRFLAC_FALSE;
+        }
+
+        zeroCountPart128 = _mm_set_epi32(zeroCountParts3, zeroCountParts2, zeroCountParts1, zeroCountParts0);
+        riceParamPart128 = _mm_set_epi32(riceParamParts3, riceParamParts2, riceParamParts1, riceParamParts0);
+
+        riceParamPart128 = _mm_and_si128(riceParamPart128, riceParamMask128);
+        riceParamPart128 = _mm_or_si128(riceParamPart128, _mm_slli_epi32(zeroCountPart128, riceParam));
+        riceParamPart128 = _mm_xor_si128(_mm_srli_epi32(riceParamPart128, 1), _mm_add_epi32(drflac__mm_not_si128(_mm_and_si128(riceParamPart128, _mm_set1_epi32(0x01))), _mm_set1_epi32(0x01)));  /* <-- SSE2 compatible */
+        /*riceParamPart128 = _mm_xor_si128(_mm_srli_epi32(riceParamPart128, 1), _mm_mullo_epi32(_mm_and_si128(riceParamPart128, _mm_set1_epi32(0x01)), _mm_set1_epi32(0xFFFFFFFF)));*/   /* <-- Only supported from SSE4.1 and is slower in my testing... */
+
+        if (order <= 4) {
+            for (i = 0; i < 4; i += 1) {
+                prediction128 = _mm_mullo_epi32(coefficients128_0, samples128_0);
+
+                /* Horizontal add and shift. */
+                prediction128 = drflac__mm_hadd_epi32(prediction128);
+                prediction128 = _mm_srai_epi32(prediction128, shift);
+                prediction128 = _mm_add_epi32(riceParamPart128, prediction128);
+
+                samples128_0 = _mm_alignr_epi8(prediction128, samples128_0, 4);
+                riceParamPart128 = _mm_alignr_epi8(_mm_setzero_si128(), riceParamPart128, 4);
+            }
+        } else if (order <= 8) {
+            for (i = 0; i < 4; i += 1) {
+                prediction128 =                              _mm_mullo_epi32(coefficients128_4, samples128_4);
+                prediction128 = _mm_add_epi32(prediction128, _mm_mullo_epi32(coefficients128_0, samples128_0));
+
+                /* Horizontal add and shift. */
+                prediction128 = drflac__mm_hadd_epi32(prediction128);
+                prediction128 = _mm_srai_epi32(prediction128, shift);
+                prediction128 = _mm_add_epi32(riceParamPart128, prediction128);
+
+                samples128_4 = _mm_alignr_epi8(samples128_0,  samples128_4, 4);
+                samples128_0 = _mm_alignr_epi8(prediction128, samples128_0, 4);
+                riceParamPart128 = _mm_alignr_epi8(_mm_setzero_si128(), riceParamPart128, 4);
+            }
+        } else {
+            for (i = 0; i < 4; i += 1) {
+                prediction128 =                              _mm_mullo_epi32(coefficients128_8, samples128_8);
+                prediction128 = _mm_add_epi32(prediction128, _mm_mullo_epi32(coefficients128_4, samples128_4));
+                prediction128 = _mm_add_epi32(prediction128, _mm_mullo_epi32(coefficients128_0, samples128_0));
+
+                /* Horizontal add and shift. */
+                prediction128 = drflac__mm_hadd_epi32(prediction128);
+                prediction128 = _mm_srai_epi32(prediction128, shift);
+                prediction128 = _mm_add_epi32(riceParamPart128, prediction128);
+
+                samples128_8 = _mm_alignr_epi8(samples128_4,  samples128_8, 4);
+                samples128_4 = _mm_alignr_epi8(samples128_0,  samples128_4, 4);
+                samples128_0 = _mm_alignr_epi8(prediction128, samples128_0, 4);
+                riceParamPart128 = _mm_alignr_epi8(_mm_setzero_si128(), riceParamPart128, 4);
+            }
+        }
+
+        /* We store samples in groups of 4. */
+        _mm_storeu_si128((__m128i*)pDecodedSamples, samples128_0);
+        pDecodedSamples += 4;
+    }
+
+    /* Make sure we process the last few samples. */
+    i = (count & ~3);
+    while (i < (int)count) {
+        /* Rice extraction. */
+        if (!drflac__read_rice_parts_x1(bs, riceParam, &zeroCountParts0, &riceParamParts0)) {
+            return DRFLAC_FALSE;
+        }
+
+        /* Rice reconstruction. */
+        riceParamParts0 &= riceParamMask;
+        riceParamParts0 |= (zeroCountParts0 << riceParam);
+        riceParamParts0  = (riceParamParts0 >> 1) ^ t[riceParamParts0 & 0x01];
+
+        /* Sample reconstruction. */
+        pDecodedSamples[0] = riceParamParts0 + drflac__calculate_prediction_32(order, shift, coefficients, pDecodedSamples);
+
+        i += 1;
+        pDecodedSamples += 1;
+    }
+
+    return DRFLAC_TRUE;
+}
+
+static drflac_bool32 drflac__decode_samples_with_residual__rice__sse41_64(drflac_bs* bs, drflac_uint32 count, drflac_uint8 riceParam, drflac_uint32 order, drflac_int32 shift, const drflac_int32* coefficients, drflac_int32* pSamplesOut)
+{
+    int i;
+    drflac_uint32 riceParamMask;
+    drflac_int32* pDecodedSamples    = pSamplesOut;
+    drflac_int32* pDecodedSamplesEnd = pSamplesOut + (count & ~3);
+    drflac_uint32 zeroCountParts0 = 0;
+    drflac_uint32 zeroCountParts1 = 0;
+    drflac_uint32 zeroCountParts2 = 0;
+    drflac_uint32 zeroCountParts3 = 0;
+    drflac_uint32 riceParamParts0 = 0;
+    drflac_uint32 riceParamParts1 = 0;
+    drflac_uint32 riceParamParts2 = 0;
+    drflac_uint32 riceParamParts3 = 0;
+    __m128i coefficients128_0;
+    __m128i coefficients128_4;
+    __m128i coefficients128_8;
+    __m128i samples128_0;
+    __m128i samples128_4;
+    __m128i samples128_8;
+    __m128i prediction128;
+    __m128i riceParamMask128;
+
+    const drflac_uint32 t[2] = {0x00000000, 0xFFFFFFFF};
+
+    DRFLAC_ASSERT(order <= 12);
+
+    riceParamMask    = (drflac_uint32)~((~0UL) << riceParam);
+    riceParamMask128 = _mm_set1_epi32(riceParamMask);
+
+    prediction128 = _mm_setzero_si128();
+
+    /* Pre-load. */
+    coefficients128_0  = _mm_setzero_si128();
+    coefficients128_4  = _mm_setzero_si128();
+    coefficients128_8  = _mm_setzero_si128();
+
+    samples128_0  = _mm_setzero_si128();
+    samples128_4  = _mm_setzero_si128();
+    samples128_8  = _mm_setzero_si128();
+
+#if 1
+    {
+        int runningOrder = order;
+
+        /* 0 - 3. */
+        if (runningOrder >= 4) {
+            coefficients128_0 = _mm_loadu_si128((const __m128i*)(coefficients + 0));
+            samples128_0      = _mm_loadu_si128((const __m128i*)(pSamplesOut  - 4));
+            runningOrder -= 4;
+        } else {
+            switch (runningOrder) {
+                case 3: coefficients128_0 = _mm_set_epi32(0, coefficients[2], coefficients[1], coefficients[0]); samples128_0 = _mm_set_epi32(pSamplesOut[-1], pSamplesOut[-2], pSamplesOut[-3], 0); break;
+                case 2: coefficients128_0 = _mm_set_epi32(0, 0,               coefficients[1], coefficients[0]); samples128_0 = _mm_set_epi32(pSamplesOut[-1], pSamplesOut[-2], 0,               0); break;
+                case 1: coefficients128_0 = _mm_set_epi32(0, 0,               0,               coefficients[0]); samples128_0 = _mm_set_epi32(pSamplesOut[-1], 0,               0,               0); break;
+            }
+            runningOrder = 0;
+        }
+
+        /* 4 - 7 */
+        if (runningOrder >= 4) {
+            coefficients128_4 = _mm_loadu_si128((const __m128i*)(coefficients + 4));
+            samples128_4      = _mm_loadu_si128((const __m128i*)(pSamplesOut  - 8));
+            runningOrder -= 4;
+        } else {
+            switch (runningOrder) {
+                case 3: coefficients128_4 = _mm_set_epi32(0, coefficients[6], coefficients[5], coefficients[4]); samples128_4 = _mm_set_epi32(pSamplesOut[-5], pSamplesOut[-6], pSamplesOut[-7], 0); break;
+                case 2: coefficients128_4 = _mm_set_epi32(0, 0,               coefficients[5], coefficients[4]); samples128_4 = _mm_set_epi32(pSamplesOut[-5], pSamplesOut[-6], 0,               0); break;
+                case 1: coefficients128_4 = _mm_set_epi32(0, 0,               0,               coefficients[4]); samples128_4 = _mm_set_epi32(pSamplesOut[-5], 0,               0,               0); break;
+            }
+            runningOrder = 0;
+        }
+
+        /* 8 - 11 */
+        if (runningOrder == 4) {
+            coefficients128_8 = _mm_loadu_si128((const __m128i*)(coefficients + 8));
+            samples128_8      = _mm_loadu_si128((const __m128i*)(pSamplesOut  - 12));
+            runningOrder -= 4;
+        } else {
+            switch (runningOrder) {
+                case 3: coefficients128_8 = _mm_set_epi32(0, coefficients[10], coefficients[9], coefficients[8]); samples128_8 = _mm_set_epi32(pSamplesOut[-9], pSamplesOut[-10], pSamplesOut[-11], 0); break;
+                case 2: coefficients128_8 = _mm_set_epi32(0, 0,                coefficients[9], coefficients[8]); samples128_8 = _mm_set_epi32(pSamplesOut[-9], pSamplesOut[-10], 0,                0); break;
+                case 1: coefficients128_8 = _mm_set_epi32(0, 0,                0,               coefficients[8]); samples128_8 = _mm_set_epi32(pSamplesOut[-9], 0,                0,                0); break;
+            }
+            runningOrder = 0;
+        }
+
+        /* Coefficients need to be shuffled for our streaming algorithm below to work. Samples are already in the correct order from the loading routine above. */
+        coefficients128_0 = _mm_shuffle_epi32(coefficients128_0, _MM_SHUFFLE(0, 1, 2, 3));
+        coefficients128_4 = _mm_shuffle_epi32(coefficients128_4, _MM_SHUFFLE(0, 1, 2, 3));
+        coefficients128_8 = _mm_shuffle_epi32(coefficients128_8, _MM_SHUFFLE(0, 1, 2, 3));
+    }
+#else
+    switch (order)
+    {
+    case 12: ((drflac_int32*)&coefficients128_8)[0] = coefficients[11]; ((drflac_int32*)&samples128_8)[0] = pDecodedSamples[-12];
+    case 11: ((drflac_int32*)&coefficients128_8)[1] = coefficients[10]; ((drflac_int32*)&samples128_8)[1] = pDecodedSamples[-11];
+    case 10: ((drflac_int32*)&coefficients128_8)[2] = coefficients[ 9]; ((drflac_int32*)&samples128_8)[2] = pDecodedSamples[-10];
+    case 9:  ((drflac_int32*)&coefficients128_8)[3] = coefficients[ 8]; ((drflac_int32*)&samples128_8)[3] = pDecodedSamples[- 9];
+    case 8:  ((drflac_int32*)&coefficients128_4)[0] = coefficients[ 7]; ((drflac_int32*)&samples128_4)[0] = pDecodedSamples[- 8];
+    case 7:  ((drflac_int32*)&coefficients128_4)[1] = coefficients[ 6]; ((drflac_int32*)&samples128_4)[1] = pDecodedSamples[- 7];
+    case 6:  ((drflac_int32*)&coefficients128_4)[2] = coefficients[ 5]; ((drflac_int32*)&samples128_4)[2] = pDecodedSamples[- 6];
+    case 5:  ((drflac_int32*)&coefficients128_4)[3] = coefficients[ 4]; ((drflac_int32*)&samples128_4)[3] = pDecodedSamples[- 5];
+    case 4:  ((drflac_int32*)&coefficients128_0)[0] = coefficients[ 3]; ((drflac_int32*)&samples128_0)[0] = pDecodedSamples[- 4];
+    case 3:  ((drflac_int32*)&coefficients128_0)[1] = coefficients[ 2]; ((drflac_int32*)&samples128_0)[1] = pDecodedSamples[- 3];
+    case 2:  ((drflac_int32*)&coefficients128_0)[2] = coefficients[ 1]; ((drflac_int32*)&samples128_0)[2] = pDecodedSamples[- 2];
+    case 1:  ((drflac_int32*)&coefficients128_0)[3] = coefficients[ 0]; ((drflac_int32*)&samples128_0)[3] = pDecodedSamples[- 1];
+    }
+#endif
+
+    /* For this version we are doing one sample at a time. */
+    while (pDecodedSamples < pDecodedSamplesEnd) {
+        __m128i zeroCountPart128;
+        __m128i riceParamPart128;
+
+        if (!drflac__read_rice_parts_x1(bs, riceParam, &zeroCountParts0, &riceParamParts0) ||
+            !drflac__read_rice_parts_x1(bs, riceParam, &zeroCountParts1, &riceParamParts1) ||
+            !drflac__read_rice_parts_x1(bs, riceParam, &zeroCountParts2, &riceParamParts2) ||
+            !drflac__read_rice_parts_x1(bs, riceParam, &zeroCountParts3, &riceParamParts3)) {
+            return DRFLAC_FALSE;
+        }
+
+        zeroCountPart128 = _mm_set_epi32(zeroCountParts3, zeroCountParts2, zeroCountParts1, zeroCountParts0);
+        riceParamPart128 = _mm_set_epi32(riceParamParts3, riceParamParts2, riceParamParts1, riceParamParts0);
+
+        riceParamPart128 = _mm_and_si128(riceParamPart128, riceParamMask128);
+        riceParamPart128 = _mm_or_si128(riceParamPart128, _mm_slli_epi32(zeroCountPart128, riceParam));
+        riceParamPart128 = _mm_xor_si128(_mm_srli_epi32(riceParamPart128, 1), _mm_add_epi32(drflac__mm_not_si128(_mm_and_si128(riceParamPart128, _mm_set1_epi32(1))), _mm_set1_epi32(1)));
+
+        for (i = 0; i < 4; i += 1) {
+            prediction128 = _mm_xor_si128(prediction128, prediction128);    /* Reset to 0. */
+
+            switch (order)
+            {
+            case 12:
+            case 11: prediction128 = _mm_add_epi64(prediction128, _mm_mul_epi32(_mm_shuffle_epi32(coefficients128_8, _MM_SHUFFLE(1, 1, 0, 0)), _mm_shuffle_epi32(samples128_8, _MM_SHUFFLE(1, 1, 0, 0))));
+            case 10:
+            case  9: prediction128 = _mm_add_epi64(prediction128, _mm_mul_epi32(_mm_shuffle_epi32(coefficients128_8, _MM_SHUFFLE(3, 3, 2, 2)), _mm_shuffle_epi32(samples128_8, _MM_SHUFFLE(3, 3, 2, 2))));
+            case  8:
+            case  7: prediction128 = _mm_add_epi64(prediction128, _mm_mul_epi32(_mm_shuffle_epi32(coefficients128_4, _MM_SHUFFLE(1, 1, 0, 0)), _mm_shuffle_epi32(samples128_4, _MM_SHUFFLE(1, 1, 0, 0))));
+            case  6:
+            case  5: prediction128 = _mm_add_epi64(prediction128, _mm_mul_epi32(_mm_shuffle_epi32(coefficients128_4, _MM_SHUFFLE(3, 3, 2, 2)), _mm_shuffle_epi32(samples128_4, _MM_SHUFFLE(3, 3, 2, 2))));
+            case  4:
+            case  3: prediction128 = _mm_add_epi64(prediction128, _mm_mul_epi32(_mm_shuffle_epi32(coefficients128_0, _MM_SHUFFLE(1, 1, 0, 0)), _mm_shuffle_epi32(samples128_0, _MM_SHUFFLE(1, 1, 0, 0))));
+            case  2:
+            case  1: prediction128 = _mm_add_epi64(prediction128, _mm_mul_epi32(_mm_shuffle_epi32(coefficients128_0, _MM_SHUFFLE(3, 3, 2, 2)), _mm_shuffle_epi32(samples128_0, _MM_SHUFFLE(3, 3, 2, 2))));
+            }
+
+            /* Horizontal add and shift. */
+            prediction128 = drflac__mm_hadd_epi64(prediction128);
+            prediction128 = drflac__mm_srai_epi64(prediction128, shift);
+            prediction128 = _mm_add_epi32(riceParamPart128, prediction128);
+
+            /* Our value should be sitting in prediction128[0]. We need to combine this with our SSE samples. */
+            samples128_8 = _mm_alignr_epi8(samples128_4,  samples128_8, 4);
+            samples128_4 = _mm_alignr_epi8(samples128_0,  samples128_4, 4);
+            samples128_0 = _mm_alignr_epi8(prediction128, samples128_0, 4);
+
+            /* Slide our rice parameter down so that the value in position 0 contains the next one to process. */
+            riceParamPart128 = _mm_alignr_epi8(_mm_setzero_si128(), riceParamPart128, 4);
+        }
+
+        /* We store samples in groups of 4. */
+        _mm_storeu_si128((__m128i*)pDecodedSamples, samples128_0);
+        pDecodedSamples += 4;
+    }
+
+    /* Make sure we process the last few samples. */
+    i = (count & ~3);
+    while (i < (int)count) {
+        /* Rice extraction. */
+        if (!drflac__read_rice_parts_x1(bs, riceParam, &zeroCountParts0, &riceParamParts0)) {
+            return DRFLAC_FALSE;
+        }
+
+        /* Rice reconstruction. */
+        riceParamParts0 &= riceParamMask;
+        riceParamParts0 |= (zeroCountParts0 << riceParam);
+        riceParamParts0  = (riceParamParts0 >> 1) ^ t[riceParamParts0 & 0x01];
+
+        /* Sample reconstruction. */
+        pDecodedSamples[0] = riceParamParts0 + drflac__calculate_prediction_64(order, shift, coefficients, pDecodedSamples);
+
+        i += 1;
+        pDecodedSamples += 1;
+    }
+
+    return DRFLAC_TRUE;
+}
+
+static drflac_bool32 drflac__decode_samples_with_residual__rice__sse41(drflac_bs* bs, drflac_uint32 bitsPerSample, drflac_uint32 count, drflac_uint8 riceParam, drflac_uint32 lpcOrder, drflac_int32 lpcShift, drflac_uint32 lpcPrecision, const drflac_int32* coefficients, drflac_int32* pSamplesOut)
+{
+    DRFLAC_ASSERT(bs != NULL);
+    DRFLAC_ASSERT(pSamplesOut != NULL);
+
+    /* In my testing the order is rarely > 12, so in this case I'm going to simplify the SSE implementation by only handling order <= 12. */
+    if (lpcOrder > 0 && lpcOrder <= 12) {
+        if (drflac__use_64_bit_prediction(bitsPerSample, lpcOrder, lpcPrecision)) {
+            return drflac__decode_samples_with_residual__rice__sse41_64(bs, count, riceParam, lpcOrder, lpcShift, coefficients, pSamplesOut);
+        } else {
+            return drflac__decode_samples_with_residual__rice__sse41_32(bs, count, riceParam, lpcOrder, lpcShift, coefficients, pSamplesOut);
+        }
+    } else {
+        return drflac__decode_samples_with_residual__rice__scalar(bs, bitsPerSample, count, riceParam, lpcOrder, lpcShift, lpcPrecision, coefficients, pSamplesOut);
+    }
+}
+#endif
+
+#if defined(DRFLAC_SUPPORT_NEON)
+static DRFLAC_INLINE void drflac__vst2q_s32(drflac_int32* p, int32x4x2_t x)
+{
+    vst1q_s32(p+0, x.val[0]);
+    vst1q_s32(p+4, x.val[1]);
+}
+
+static DRFLAC_INLINE void drflac__vst2q_u32(drflac_uint32* p, uint32x4x2_t x)
+{
+    vst1q_u32(p+0, x.val[0]);
+    vst1q_u32(p+4, x.val[1]);
+}
+
+static DRFLAC_INLINE void drflac__vst2q_f32(float* p, float32x4x2_t x)
+{
+    vst1q_f32(p+0, x.val[0]);
+    vst1q_f32(p+4, x.val[1]);
+}
+
+static DRFLAC_INLINE void drflac__vst2q_s16(drflac_int16* p, int16x4x2_t x)
+{
+    vst1q_s16(p, vcombine_s16(x.val[0], x.val[1]));
+}
+
+static DRFLAC_INLINE void drflac__vst2q_u16(drflac_uint16* p, uint16x4x2_t x)
+{
+    vst1q_u16(p, vcombine_u16(x.val[0], x.val[1]));
+}
+
+static DRFLAC_INLINE int32x4_t drflac__vdupq_n_s32x4(drflac_int32 x3, drflac_int32 x2, drflac_int32 x1, drflac_int32 x0)
+{
+    drflac_int32 x[4];
+    x[3] = x3;
+    x[2] = x2;
+    x[1] = x1;
+    x[0] = x0;
+    return vld1q_s32(x);
+}
+
+static DRFLAC_INLINE int32x4_t drflac__valignrq_s32_1(int32x4_t a, int32x4_t b)
+{
+    /* Equivalent to SSE's _mm_alignr_epi8(a, b, 4) */
+
+    /* Reference */
+    /*return drflac__vdupq_n_s32x4(
+        vgetq_lane_s32(a, 0),
+        vgetq_lane_s32(b, 3),
+        vgetq_lane_s32(b, 2),
+        vgetq_lane_s32(b, 1)
+    );*/
+
+    return vextq_s32(b, a, 1);
+}
+
+static DRFLAC_INLINE uint32x4_t drflac__valignrq_u32_1(uint32x4_t a, uint32x4_t b)
+{
+    /* Equivalent to SSE's _mm_alignr_epi8(a, b, 4) */
+
+    /* Reference */
+    /*return drflac__vdupq_n_s32x4(
+        vgetq_lane_s32(a, 0),
+        vgetq_lane_s32(b, 3),
+        vgetq_lane_s32(b, 2),
+        vgetq_lane_s32(b, 1)
+    );*/
+
+    return vextq_u32(b, a, 1);
+}
+
+static DRFLAC_INLINE int32x2_t drflac__vhaddq_s32(int32x4_t x)
+{
+    /* The sum must end up in position 0. */
+
+    /* Reference */
+    /*return vdupq_n_s32(
+        vgetq_lane_s32(x, 3) +
+        vgetq_lane_s32(x, 2) +
+        vgetq_lane_s32(x, 1) +
+        vgetq_lane_s32(x, 0)
+    );*/
+
+    int32x2_t r = vadd_s32(vget_high_s32(x), vget_low_s32(x));
+    return vpadd_s32(r, r);
+}
+
+static DRFLAC_INLINE int64x1_t drflac__vhaddq_s64(int64x2_t x)
+{
+    return vadd_s64(vget_high_s64(x), vget_low_s64(x));
+}
+
+static DRFLAC_INLINE int32x4_t drflac__vrevq_s32(int32x4_t x)
+{
+    /* Reference */
+    /*return drflac__vdupq_n_s32x4(
+        vgetq_lane_s32(x, 0),
+        vgetq_lane_s32(x, 1),
+        vgetq_lane_s32(x, 2),
+        vgetq_lane_s32(x, 3)
+    );*/
+
+    return vrev64q_s32(vcombine_s32(vget_high_s32(x), vget_low_s32(x)));
+}
+
+static DRFLAC_INLINE int32x4_t drflac__vnotq_s32(int32x4_t x)
+{
+    return veorq_s32(x, vdupq_n_s32(0xFFFFFFFF));
+}
+
+static DRFLAC_INLINE uint32x4_t drflac__vnotq_u32(uint32x4_t x)
+{
+    return veorq_u32(x, vdupq_n_u32(0xFFFFFFFF));
+}
+
+static drflac_bool32 drflac__decode_samples_with_residual__rice__neon_32(drflac_bs* bs, drflac_uint32 count, drflac_uint8 riceParam, drflac_uint32 order, drflac_int32 shift, const drflac_int32* coefficients, drflac_int32* pSamplesOut)
+{
+    int i;
+    drflac_uint32 riceParamMask;
+    drflac_int32* pDecodedSamples    = pSamplesOut;
+    drflac_int32* pDecodedSamplesEnd = pSamplesOut + (count & ~3);
+    drflac_uint32 zeroCountParts[4];
+    drflac_uint32 riceParamParts[4];
+    int32x4_t coefficients128_0;
+    int32x4_t coefficients128_4;
+    int32x4_t coefficients128_8;
+    int32x4_t samples128_0;
+    int32x4_t samples128_4;
+    int32x4_t samples128_8;
+    uint32x4_t riceParamMask128;
+    int32x4_t riceParam128;
+    int32x2_t shift64;
+    uint32x4_t one128;
+
+    const drflac_uint32 t[2] = {0x00000000, 0xFFFFFFFF};
+
+    riceParamMask    = (drflac_uint32)~((~0UL) << riceParam);
+    riceParamMask128 = vdupq_n_u32(riceParamMask);
+
+    riceParam128 = vdupq_n_s32(riceParam);
+    shift64 = vdup_n_s32(-shift); /* Negate the shift because we'll be doing a variable shift using vshlq_s32(). */
+    one128 = vdupq_n_u32(1);
+
+    /*
+    Pre-loading the coefficients and prior samples is annoying because we need to ensure we don't try reading more than
+    what's available in the input buffers. It would be conenient to use a fall-through switch to do this, but this results
+    in strict aliasing warnings with GCC. To work around this I'm just doing something hacky. This feels a bit convoluted
+    so I think there's opportunity for this to be simplified.
+    */
+    {
+        int runningOrder = order;
+        drflac_int32 tempC[4] = {0, 0, 0, 0};
+        drflac_int32 tempS[4] = {0, 0, 0, 0};
+
+        /* 0 - 3. */
+        if (runningOrder >= 4) {
+            coefficients128_0 = vld1q_s32(coefficients + 0);
+            samples128_0      = vld1q_s32(pSamplesOut  - 4);
+            runningOrder -= 4;
+        } else {
+            switch (runningOrder) {
+                case 3: tempC[2] = coefficients[2]; tempS[1] = pSamplesOut[-3]; /* fallthrough */
+                case 2: tempC[1] = coefficients[1]; tempS[2] = pSamplesOut[-2]; /* fallthrough */
+                case 1: tempC[0] = coefficients[0]; tempS[3] = pSamplesOut[-1]; /* fallthrough */
+            }
+
+            coefficients128_0 = vld1q_s32(tempC);
+            samples128_0      = vld1q_s32(tempS);
+            runningOrder = 0;
+        }
+
+        /* 4 - 7 */
+        if (runningOrder >= 4) {
+            coefficients128_4 = vld1q_s32(coefficients + 4);
+            samples128_4      = vld1q_s32(pSamplesOut  - 8);
+            runningOrder -= 4;
+        } else {
+            switch (runningOrder) {
+                case 3: tempC[2] = coefficients[6]; tempS[1] = pSamplesOut[-7]; /* fallthrough */
+                case 2: tempC[1] = coefficients[5]; tempS[2] = pSamplesOut[-6]; /* fallthrough */
+                case 1: tempC[0] = coefficients[4]; tempS[3] = pSamplesOut[-5]; /* fallthrough */
+            }
+
+            coefficients128_4 = vld1q_s32(tempC);
+            samples128_4      = vld1q_s32(tempS);
+            runningOrder = 0;
+        }
+
+        /* 8 - 11 */
+        if (runningOrder == 4) {
+            coefficients128_8 = vld1q_s32(coefficients + 8);
+            samples128_8      = vld1q_s32(pSamplesOut  - 12);
+            runningOrder -= 4;
+        } else {
+            switch (runningOrder) {
+                case 3: tempC[2] = coefficients[10]; tempS[1] = pSamplesOut[-11]; /* fallthrough */
+                case 2: tempC[1] = coefficients[ 9]; tempS[2] = pSamplesOut[-10]; /* fallthrough */
+                case 1: tempC[0] = coefficients[ 8]; tempS[3] = pSamplesOut[- 9]; /* fallthrough */
+            }
+
+            coefficients128_8 = vld1q_s32(tempC);
+            samples128_8      = vld1q_s32(tempS);
+            runningOrder = 0;
+        }
+
+        /* Coefficients need to be shuffled for our streaming algorithm below to work. Samples are already in the correct order from the loading routine above. */
+        coefficients128_0 = drflac__vrevq_s32(coefficients128_0);
+        coefficients128_4 = drflac__vrevq_s32(coefficients128_4);
+        coefficients128_8 = drflac__vrevq_s32(coefficients128_8);
+    }
+
+    /* For this version we are doing one sample at a time. */
+    while (pDecodedSamples < pDecodedSamplesEnd) {
+        int32x4_t prediction128;
+        int32x2_t prediction64;
+        uint32x4_t zeroCountPart128;
+        uint32x4_t riceParamPart128;
+
+        if (!drflac__read_rice_parts_x1(bs, riceParam, &zeroCountParts[0], &riceParamParts[0]) ||
+            !drflac__read_rice_parts_x1(bs, riceParam, &zeroCountParts[1], &riceParamParts[1]) ||
+            !drflac__read_rice_parts_x1(bs, riceParam, &zeroCountParts[2], &riceParamParts[2]) ||
+            !drflac__read_rice_parts_x1(bs, riceParam, &zeroCountParts[3], &riceParamParts[3])) {
+            return DRFLAC_FALSE;
+        }
+
+        zeroCountPart128 = vld1q_u32(zeroCountParts);
+        riceParamPart128 = vld1q_u32(riceParamParts);
+
+        riceParamPart128 = vandq_u32(riceParamPart128, riceParamMask128);
+        riceParamPart128 = vorrq_u32(riceParamPart128, vshlq_u32(zeroCountPart128, riceParam128));
+        riceParamPart128 = veorq_u32(vshrq_n_u32(riceParamPart128, 1), vaddq_u32(drflac__vnotq_u32(vandq_u32(riceParamPart128, one128)), one128));
+
+        if (order <= 4) {
+            for (i = 0; i < 4; i += 1) {
+                prediction128 = vmulq_s32(coefficients128_0, samples128_0);
+
+                /* Horizontal add and shift. */
+                prediction64 = drflac__vhaddq_s32(prediction128);
+                prediction64 = vshl_s32(prediction64, shift64);
+                prediction64 = vadd_s32(prediction64, vget_low_s32(vreinterpretq_s32_u32(riceParamPart128)));
+
+                samples128_0 = drflac__valignrq_s32_1(vcombine_s32(prediction64, vdup_n_s32(0)), samples128_0);
+                riceParamPart128 = drflac__valignrq_u32_1(vdupq_n_u32(0), riceParamPart128);
+            }
+        } else if (order <= 8) {
+            for (i = 0; i < 4; i += 1) {
+                prediction128 =                vmulq_s32(coefficients128_4, samples128_4);
+                prediction128 = vmlaq_s32(prediction128, coefficients128_0, samples128_0);
+
+                /* Horizontal add and shift. */
+                prediction64 = drflac__vhaddq_s32(prediction128);
+                prediction64 = vshl_s32(prediction64, shift64);
+                prediction64 = vadd_s32(prediction64, vget_low_s32(vreinterpretq_s32_u32(riceParamPart128)));
+
+                samples128_4 = drflac__valignrq_s32_1(samples128_0, samples128_4);
+                samples128_0 = drflac__valignrq_s32_1(vcombine_s32(prediction64, vdup_n_s32(0)), samples128_0);
+                riceParamPart128 = drflac__valignrq_u32_1(vdupq_n_u32(0), riceParamPart128);
+            }
+        } else {
+            for (i = 0; i < 4; i += 1) {
+                prediction128 =                vmulq_s32(coefficients128_8, samples128_8);
+                prediction128 = vmlaq_s32(prediction128, coefficients128_4, samples128_4);
+                prediction128 = vmlaq_s32(prediction128, coefficients128_0, samples128_0);
+
+                /* Horizontal add and shift. */
+                prediction64 = drflac__vhaddq_s32(prediction128);
+                prediction64 = vshl_s32(prediction64, shift64);
+                prediction64 = vadd_s32(prediction64, vget_low_s32(vreinterpretq_s32_u32(riceParamPart128)));
+
+                samples128_8 = drflac__valignrq_s32_1(samples128_4, samples128_8);
+                samples128_4 = drflac__valignrq_s32_1(samples128_0, samples128_4);
+                samples128_0 = drflac__valignrq_s32_1(vcombine_s32(prediction64, vdup_n_s32(0)), samples128_0);
+                riceParamPart128 = drflac__valignrq_u32_1(vdupq_n_u32(0), riceParamPart128);
+            }
+        }
+
+        /* We store samples in groups of 4. */
+        vst1q_s32(pDecodedSamples, samples128_0);
+        pDecodedSamples += 4;
+    }
+
+    /* Make sure we process the last few samples. */
+    i = (count & ~3);
+    while (i < (int)count) {
+        /* Rice extraction. */
+        if (!drflac__read_rice_parts_x1(bs, riceParam, &zeroCountParts[0], &riceParamParts[0])) {
+            return DRFLAC_FALSE;
+        }
+
+        /* Rice reconstruction. */
+        riceParamParts[0] &= riceParamMask;
+        riceParamParts[0] |= (zeroCountParts[0] << riceParam);
+        riceParamParts[0]  = (riceParamParts[0] >> 1) ^ t[riceParamParts[0] & 0x01];
+
+        /* Sample reconstruction. */
+        pDecodedSamples[0] = riceParamParts[0] + drflac__calculate_prediction_32(order, shift, coefficients, pDecodedSamples);
+
+        i += 1;
+        pDecodedSamples += 1;
+    }
+
+    return DRFLAC_TRUE;
+}
+
+static drflac_bool32 drflac__decode_samples_with_residual__rice__neon_64(drflac_bs* bs, drflac_uint32 count, drflac_uint8 riceParam, drflac_uint32 order, drflac_int32 shift, const drflac_int32* coefficients, drflac_int32* pSamplesOut)
+{
+    int i;
+    drflac_uint32 riceParamMask;
+    drflac_int32* pDecodedSamples    = pSamplesOut;
+    drflac_int32* pDecodedSamplesEnd = pSamplesOut + (count & ~3);
+    drflac_uint32 zeroCountParts[4];
+    drflac_uint32 riceParamParts[4];
+    int32x4_t coefficients128_0;
+    int32x4_t coefficients128_4;
+    int32x4_t coefficients128_8;
+    int32x4_t samples128_0;
+    int32x4_t samples128_4;
+    int32x4_t samples128_8;
+    uint32x4_t riceParamMask128;
+    int32x4_t riceParam128;
+    int64x1_t shift64;
+    uint32x4_t one128;
+    int64x2_t prediction128 = { 0 };
+    uint32x4_t zeroCountPart128;
+    uint32x4_t riceParamPart128;
+
+    const drflac_uint32 t[2] = {0x00000000, 0xFFFFFFFF};
+
+    riceParamMask    = (drflac_uint32)~((~0UL) << riceParam);
+    riceParamMask128 = vdupq_n_u32(riceParamMask);
+
+    riceParam128 = vdupq_n_s32(riceParam);
+    shift64 = vdup_n_s64(-shift); /* Negate the shift because we'll be doing a variable shift using vshlq_s32(). */
+    one128 = vdupq_n_u32(1);
+
+    /*
+    Pre-loading the coefficients and prior samples is annoying because we need to ensure we don't try reading more than
+    what's available in the input buffers. It would be convenient to use a fall-through switch to do this, but this results
+    in strict aliasing warnings with GCC. To work around this I'm just doing something hacky. This feels a bit convoluted
+    so I think there's opportunity for this to be simplified.
+    */
+    {
+        int runningOrder = order;
+        drflac_int32 tempC[4] = {0, 0, 0, 0};
+        drflac_int32 tempS[4] = {0, 0, 0, 0};
+
+        /* 0 - 3. */
+        if (runningOrder >= 4) {
+            coefficients128_0 = vld1q_s32(coefficients + 0);
+            samples128_0      = vld1q_s32(pSamplesOut  - 4);
+            runningOrder -= 4;
+        } else {
+            switch (runningOrder) {
+                case 3: tempC[2] = coefficients[2]; tempS[1] = pSamplesOut[-3]; /* fallthrough */
+                case 2: tempC[1] = coefficients[1]; tempS[2] = pSamplesOut[-2]; /* fallthrough */
+                case 1: tempC[0] = coefficients[0]; tempS[3] = pSamplesOut[-1]; /* fallthrough */
+            }
+
+            coefficients128_0 = vld1q_s32(tempC);
+            samples128_0      = vld1q_s32(tempS);
+            runningOrder = 0;
+        }
+
+        /* 4 - 7 */
+        if (runningOrder >= 4) {
+            coefficients128_4 = vld1q_s32(coefficients + 4);
+            samples128_4      = vld1q_s32(pSamplesOut  - 8);
+            runningOrder -= 4;
+        } else {
+            switch (runningOrder) {
+                case 3: tempC[2] = coefficients[6]; tempS[1] = pSamplesOut[-7]; /* fallthrough */
+                case 2: tempC[1] = coefficients[5]; tempS[2] = pSamplesOut[-6]; /* fallthrough */
+                case 1: tempC[0] = coefficients[4]; tempS[3] = pSamplesOut[-5]; /* fallthrough */
+            }
+
+            coefficients128_4 = vld1q_s32(tempC);
+            samples128_4      = vld1q_s32(tempS);
+            runningOrder = 0;
+        }
+
+        /* 8 - 11 */
+        if (runningOrder == 4) {
+            coefficients128_8 = vld1q_s32(coefficients + 8);
+            samples128_8      = vld1q_s32(pSamplesOut  - 12);
+            runningOrder -= 4;
+        } else {
+            switch (runningOrder) {
+                case 3: tempC[2] = coefficients[10]; tempS[1] = pSamplesOut[-11]; /* fallthrough */
+                case 2: tempC[1] = coefficients[ 9]; tempS[2] = pSamplesOut[-10]; /* fallthrough */
+                case 1: tempC[0] = coefficients[ 8]; tempS[3] = pSamplesOut[- 9]; /* fallthrough */
+            }
+
+            coefficients128_8 = vld1q_s32(tempC);
+            samples128_8      = vld1q_s32(tempS);
+            runningOrder = 0;
+        }
+
+        /* Coefficients need to be shuffled for our streaming algorithm below to work. Samples are already in the correct order from the loading routine above. */
+        coefficients128_0 = drflac__vrevq_s32(coefficients128_0);
+        coefficients128_4 = drflac__vrevq_s32(coefficients128_4);
+        coefficients128_8 = drflac__vrevq_s32(coefficients128_8);
+    }
+
+    /* For this version we are doing one sample at a time. */
+    while (pDecodedSamples < pDecodedSamplesEnd) {
+        if (!drflac__read_rice_parts_x1(bs, riceParam, &zeroCountParts[0], &riceParamParts[0]) ||
+            !drflac__read_rice_parts_x1(bs, riceParam, &zeroCountParts[1], &riceParamParts[1]) ||
+            !drflac__read_rice_parts_x1(bs, riceParam, &zeroCountParts[2], &riceParamParts[2]) ||
+            !drflac__read_rice_parts_x1(bs, riceParam, &zeroCountParts[3], &riceParamParts[3])) {
+            return DRFLAC_FALSE;
+        }
+
+        zeroCountPart128 = vld1q_u32(zeroCountParts);
+        riceParamPart128 = vld1q_u32(riceParamParts);
+
+        riceParamPart128 = vandq_u32(riceParamPart128, riceParamMask128);
+        riceParamPart128 = vorrq_u32(riceParamPart128, vshlq_u32(zeroCountPart128, riceParam128));
+        riceParamPart128 = veorq_u32(vshrq_n_u32(riceParamPart128, 1), vaddq_u32(drflac__vnotq_u32(vandq_u32(riceParamPart128, one128)), one128));
+
+        for (i = 0; i < 4; i += 1) {
+            int64x1_t prediction64;
+
+            prediction128 = veorq_s64(prediction128, prediction128);    /* Reset to 0. */
+            switch (order)
+            {
+            case 12:
+            case 11: prediction128 = vaddq_s64(prediction128, vmull_s32(vget_low_s32(coefficients128_8), vget_low_s32(samples128_8)));
+            case 10:
+            case  9: prediction128 = vaddq_s64(prediction128, vmull_s32(vget_high_s32(coefficients128_8), vget_high_s32(samples128_8)));
+            case  8:
+            case  7: prediction128 = vaddq_s64(prediction128, vmull_s32(vget_low_s32(coefficients128_4), vget_low_s32(samples128_4)));
+            case  6:
+            case  5: prediction128 = vaddq_s64(prediction128, vmull_s32(vget_high_s32(coefficients128_4), vget_high_s32(samples128_4)));
+            case  4:
+            case  3: prediction128 = vaddq_s64(prediction128, vmull_s32(vget_low_s32(coefficients128_0), vget_low_s32(samples128_0)));
+            case  2:
+            case  1: prediction128 = vaddq_s64(prediction128, vmull_s32(vget_high_s32(coefficients128_0), vget_high_s32(samples128_0)));
+            }
+
+            /* Horizontal add and shift. */
+            prediction64 = drflac__vhaddq_s64(prediction128);
+            prediction64 = vshl_s64(prediction64, shift64);
+            prediction64 = vadd_s64(prediction64, vdup_n_s64(vgetq_lane_u32(riceParamPart128, 0)));
+
+            /* Our value should be sitting in prediction64[0]. We need to combine this with our SSE samples. */
+            samples128_8 = drflac__valignrq_s32_1(samples128_4, samples128_8);
+            samples128_4 = drflac__valignrq_s32_1(samples128_0, samples128_4);
+            samples128_0 = drflac__valignrq_s32_1(vcombine_s32(vreinterpret_s32_s64(prediction64), vdup_n_s32(0)), samples128_0);
+
+            /* Slide our rice parameter down so that the value in position 0 contains the next one to process. */
+            riceParamPart128 = drflac__valignrq_u32_1(vdupq_n_u32(0), riceParamPart128);
+        }
+
+        /* We store samples in groups of 4. */
+        vst1q_s32(pDecodedSamples, samples128_0);
+        pDecodedSamples += 4;
+    }
+
+    /* Make sure we process the last few samples. */
+    i = (count & ~3);
+    while (i < (int)count) {
+        /* Rice extraction. */
+        if (!drflac__read_rice_parts_x1(bs, riceParam, &zeroCountParts[0], &riceParamParts[0])) {
+            return DRFLAC_FALSE;
+        }
+
+        /* Rice reconstruction. */
+        riceParamParts[0] &= riceParamMask;
+        riceParamParts[0] |= (zeroCountParts[0] << riceParam);
+        riceParamParts[0]  = (riceParamParts[0] >> 1) ^ t[riceParamParts[0] & 0x01];
+
+        /* Sample reconstruction. */
+        pDecodedSamples[0] = riceParamParts[0] + drflac__calculate_prediction_64(order, shift, coefficients, pDecodedSamples);
+
+        i += 1;
+        pDecodedSamples += 1;
+    }
+
+    return DRFLAC_TRUE;
+}
+
+static drflac_bool32 drflac__decode_samples_with_residual__rice__neon(drflac_bs* bs, drflac_uint32 bitsPerSample, drflac_uint32 count, drflac_uint8 riceParam, drflac_uint32 lpcOrder, drflac_int32 lpcShift, drflac_uint32 lpcPrecision, const drflac_int32* coefficients, drflac_int32* pSamplesOut)
+{
+    DRFLAC_ASSERT(bs != NULL);
+    DRFLAC_ASSERT(pSamplesOut != NULL);
+
+    /* In my testing the order is rarely > 12, so in this case I'm going to simplify the NEON implementation by only handling order <= 12. */
+    if (lpcOrder > 0 && lpcOrder <= 12) {
+        if (drflac__use_64_bit_prediction(bitsPerSample, lpcOrder, lpcPrecision)) {
+            return drflac__decode_samples_with_residual__rice__neon_64(bs, count, riceParam, lpcOrder, lpcShift, coefficients, pSamplesOut);
+        } else {
+            return drflac__decode_samples_with_residual__rice__neon_32(bs, count, riceParam, lpcOrder, lpcShift, coefficients, pSamplesOut);
+        }
+    } else {
+        return drflac__decode_samples_with_residual__rice__scalar(bs, bitsPerSample, count, riceParam, lpcOrder, lpcShift, lpcPrecision, coefficients, pSamplesOut);
+    }
+}
+#endif
+
+static drflac_bool32 drflac__decode_samples_with_residual__rice(drflac_bs* bs, drflac_uint32 bitsPerSample, drflac_uint32 count, drflac_uint8 riceParam, drflac_uint32 lpcOrder, drflac_int32 lpcShift, drflac_uint32 lpcPrecision, const drflac_int32* coefficients, drflac_int32* pSamplesOut)
+{
+#if defined(DRFLAC_SUPPORT_SSE41)
+    if (drflac__gIsSSE41Supported) {
+        return drflac__decode_samples_with_residual__rice__sse41(bs, bitsPerSample, count, riceParam, lpcOrder, lpcShift, lpcPrecision, coefficients, pSamplesOut);
+    } else
+#elif defined(DRFLAC_SUPPORT_NEON)
+    if (drflac__gIsNEONSupported) {
+        return drflac__decode_samples_with_residual__rice__neon(bs, bitsPerSample, count, riceParam, lpcOrder, lpcShift, lpcPrecision, coefficients, pSamplesOut);
+    } else
+#endif
+    {
+        /* Scalar fallback. */
+    #if 0
+        return drflac__decode_samples_with_residual__rice__reference(bs, bitsPerSample, count, riceParam, lpcOrder, lpcShift, lpcPrecision, coefficients, pSamplesOut);
+    #else
+        return drflac__decode_samples_with_residual__rice__scalar(bs, bitsPerSample, count, riceParam, lpcOrder, lpcShift, lpcPrecision, coefficients, pSamplesOut);
+    #endif
+    }
+}
+
+/* Reads and seeks past a string of residual values as Rice codes. The decoder should be sitting on the first bit of the Rice codes. */
+static drflac_bool32 drflac__read_and_seek_residual__rice(drflac_bs* bs, drflac_uint32 count, drflac_uint8 riceParam)
+{
+    drflac_uint32 i;
+
+    DRFLAC_ASSERT(bs != NULL);
+
+    for (i = 0; i < count; ++i) {
+        if (!drflac__seek_rice_parts(bs, riceParam)) {
+            return DRFLAC_FALSE;
+        }
+    }
+
+    return DRFLAC_TRUE;
+}
+
+#if defined(__clang__)
+__attribute__((no_sanitize("signed-integer-overflow")))
+#endif
+static drflac_bool32 drflac__decode_samples_with_residual__unencoded(drflac_bs* bs, drflac_uint32 bitsPerSample, drflac_uint32 count, drflac_uint8 unencodedBitsPerSample, drflac_uint32 lpcOrder, drflac_int32 lpcShift, drflac_uint32 lpcPrecision, const drflac_int32* coefficients, drflac_int32* pSamplesOut)
+{
+    drflac_uint32 i;
+
+    DRFLAC_ASSERT(bs != NULL);
+    DRFLAC_ASSERT(unencodedBitsPerSample <= 31);    /* <-- unencodedBitsPerSample is a 5 bit number, so cannot exceed 31. */
+    DRFLAC_ASSERT(pSamplesOut != NULL);
+
+    for (i = 0; i < count; ++i) {
+        if (unencodedBitsPerSample > 0) {
+            if (!drflac__read_int32(bs, unencodedBitsPerSample, pSamplesOut + i)) {
+                return DRFLAC_FALSE;
+            }
+        } else {
+            pSamplesOut[i] = 0;
+        }
+
+        if (drflac__use_64_bit_prediction(bitsPerSample, lpcOrder, lpcPrecision)) {
+            pSamplesOut[i] += drflac__calculate_prediction_64(lpcOrder, lpcShift, coefficients, pSamplesOut + i);
+        } else {
+            pSamplesOut[i] += drflac__calculate_prediction_32(lpcOrder, lpcShift, coefficients, pSamplesOut + i);
+        }
+    }
+
+    return DRFLAC_TRUE;
+}
+
+
+/*
+Reads and decodes the residual for the sub-frame the decoder is currently sitting on. This function should be called
+when the decoder is sitting at the very start of the RESIDUAL block. The first <order> residuals will be ignored. The
+<blockSize> and <order> parameters are used to determine how many residual values need to be decoded.
+*/
+static drflac_bool32 drflac__decode_samples_with_residual(drflac_bs* bs, drflac_uint32 bitsPerSample, drflac_uint32 blockSize, drflac_uint32 lpcOrder, drflac_int32 lpcShift, drflac_uint32 lpcPrecision, const drflac_int32* coefficients, drflac_int32* pDecodedSamples)
+{
+    drflac_uint8 residualMethod;
+    drflac_uint8 partitionOrder;
+    drflac_uint32 samplesInPartition;
+    drflac_uint32 partitionsRemaining;
+
+    DRFLAC_ASSERT(bs != NULL);
+    DRFLAC_ASSERT(blockSize != 0);
+    DRFLAC_ASSERT(pDecodedSamples != NULL);       /* <-- Should we allow NULL, in which case we just seek past the residual rather than do a full decode? */
+
+    if (!drflac__read_uint8(bs, 2, &residualMethod)) {
+        return DRFLAC_FALSE;
+    }
+
+    if (residualMethod != DRFLAC_RESIDUAL_CODING_METHOD_PARTITIONED_RICE && residualMethod != DRFLAC_RESIDUAL_CODING_METHOD_PARTITIONED_RICE2) {
+        return DRFLAC_FALSE;    /* Unknown or unsupported residual coding method. */
+    }
+
+    /* Ignore the first <order> values. */
+    pDecodedSamples += lpcOrder;
+
+    if (!drflac__read_uint8(bs, 4, &partitionOrder)) {
+        return DRFLAC_FALSE;
+    }
+
+    /*
+    From the FLAC spec:
+      The Rice partition order in a Rice-coded residual section must be less than or equal to 8.
+    */
+    if (partitionOrder > 8) {
+        return DRFLAC_FALSE;
+    }
+
+    /* Validation check. */
+    if ((blockSize / (1 << partitionOrder)) < lpcOrder) {
+        return DRFLAC_FALSE;
+    }
+
+    samplesInPartition = (blockSize / (1 << partitionOrder)) - lpcOrder;
+    partitionsRemaining = (1 << partitionOrder);
+    for (;;) {
+        drflac_uint8 riceParam = 0;
+        if (residualMethod == DRFLAC_RESIDUAL_CODING_METHOD_PARTITIONED_RICE) {
+            if (!drflac__read_uint8(bs, 4, &riceParam)) {
+                return DRFLAC_FALSE;
+            }
+            if (riceParam == 15) {
+                riceParam = 0xFF;
+            }
+        } else if (residualMethod == DRFLAC_RESIDUAL_CODING_METHOD_PARTITIONED_RICE2) {
+            if (!drflac__read_uint8(bs, 5, &riceParam)) {
+                return DRFLAC_FALSE;
+            }
+            if (riceParam == 31) {
+                riceParam = 0xFF;
+            }
+        }
+
+        if (riceParam != 0xFF) {
+            if (!drflac__decode_samples_with_residual__rice(bs, bitsPerSample, samplesInPartition, riceParam, lpcOrder, lpcShift, lpcPrecision, coefficients, pDecodedSamples)) {
+                return DRFLAC_FALSE;
+            }
+        } else {
+            drflac_uint8 unencodedBitsPerSample = 0;
+            if (!drflac__read_uint8(bs, 5, &unencodedBitsPerSample)) {
+                return DRFLAC_FALSE;
+            }
+
+            if (!drflac__decode_samples_with_residual__unencoded(bs, bitsPerSample, samplesInPartition, unencodedBitsPerSample, lpcOrder, lpcShift, lpcPrecision, coefficients, pDecodedSamples)) {
+                return DRFLAC_FALSE;
+            }
+        }
+
+        pDecodedSamples += samplesInPartition;
+
+        if (partitionsRemaining == 1) {
+            break;
+        }
+
+        partitionsRemaining -= 1;
+
+        if (partitionOrder != 0) {
+            samplesInPartition = blockSize / (1 << partitionOrder);
+        }
+    }
+
+    return DRFLAC_TRUE;
+}
+
+/*
+Reads and seeks past the residual for the sub-frame the decoder is currently sitting on. This function should be called
+when the decoder is sitting at the very start of the RESIDUAL block. The first <order> residuals will be set to 0. The
+<blockSize> and <order> parameters are used to determine how many residual values need to be decoded.
+*/
+static drflac_bool32 drflac__read_and_seek_residual(drflac_bs* bs, drflac_uint32 blockSize, drflac_uint32 order)
+{
+    drflac_uint8 residualMethod;
+    drflac_uint8 partitionOrder;
+    drflac_uint32 samplesInPartition;
+    drflac_uint32 partitionsRemaining;
+
+    DRFLAC_ASSERT(bs != NULL);
+    DRFLAC_ASSERT(blockSize != 0);
+
+    if (!drflac__read_uint8(bs, 2, &residualMethod)) {
+        return DRFLAC_FALSE;
+    }
+
+    if (residualMethod != DRFLAC_RESIDUAL_CODING_METHOD_PARTITIONED_RICE && residualMethod != DRFLAC_RESIDUAL_CODING_METHOD_PARTITIONED_RICE2) {
+        return DRFLAC_FALSE;    /* Unknown or unsupported residual coding method. */
+    }
+
+    if (!drflac__read_uint8(bs, 4, &partitionOrder)) {
+        return DRFLAC_FALSE;
+    }
+
+    /*
+    From the FLAC spec:
+      The Rice partition order in a Rice-coded residual section must be less than or equal to 8.
+    */
+    if (partitionOrder > 8) {
+        return DRFLAC_FALSE;
+    }
+
+    /* Validation check. */
+    if ((blockSize / (1 << partitionOrder)) <= order) {
+        return DRFLAC_FALSE;
+    }
+
+    samplesInPartition = (blockSize / (1 << partitionOrder)) - order;
+    partitionsRemaining = (1 << partitionOrder);
+    for (;;)
+    {
+        drflac_uint8 riceParam = 0;
+        if (residualMethod == DRFLAC_RESIDUAL_CODING_METHOD_PARTITIONED_RICE) {
+            if (!drflac__read_uint8(bs, 4, &riceParam)) {
+                return DRFLAC_FALSE;
+            }
+            if (riceParam == 15) {
+                riceParam = 0xFF;
+            }
+        } else if (residualMethod == DRFLAC_RESIDUAL_CODING_METHOD_PARTITIONED_RICE2) {
+            if (!drflac__read_uint8(bs, 5, &riceParam)) {
+                return DRFLAC_FALSE;
+            }
+            if (riceParam == 31) {
+                riceParam = 0xFF;
+            }
+        }
+
+        if (riceParam != 0xFF) {
+            if (!drflac__read_and_seek_residual__rice(bs, samplesInPartition, riceParam)) {
+                return DRFLAC_FALSE;
+            }
+        } else {
+            drflac_uint8 unencodedBitsPerSample = 0;
+            if (!drflac__read_uint8(bs, 5, &unencodedBitsPerSample)) {
+                return DRFLAC_FALSE;
+            }
+
+            if (!drflac__seek_bits(bs, unencodedBitsPerSample * samplesInPartition)) {
+                return DRFLAC_FALSE;
+            }
+        }
+
+
+        if (partitionsRemaining == 1) {
+            break;
+        }
+
+        partitionsRemaining -= 1;
+        samplesInPartition = blockSize / (1 << partitionOrder);
+    }
+
+    return DRFLAC_TRUE;
+}
+
+
+static drflac_bool32 drflac__decode_samples__constant(drflac_bs* bs, drflac_uint32 blockSize, drflac_uint32 subframeBitsPerSample, drflac_int32* pDecodedSamples)
+{
+    drflac_uint32 i;
+
+    /* Only a single sample needs to be decoded here. */
+    drflac_int32 sample;
+    if (!drflac__read_int32(bs, subframeBitsPerSample, &sample)) {
+        return DRFLAC_FALSE;
+    }
+
+    /*
+    We don't really need to expand this, but it does simplify the process of reading samples. If this becomes a performance issue (unlikely)
+    we'll want to look at a more efficient way.
+    */
+    for (i = 0; i < blockSize; ++i) {
+        pDecodedSamples[i] = sample;
+    }
+
+    return DRFLAC_TRUE;
+}
+
+static drflac_bool32 drflac__decode_samples__verbatim(drflac_bs* bs, drflac_uint32 blockSize, drflac_uint32 subframeBitsPerSample, drflac_int32* pDecodedSamples)
+{
+    drflac_uint32 i;
+
+    for (i = 0; i < blockSize; ++i) {
+        drflac_int32 sample;
+        if (!drflac__read_int32(bs, subframeBitsPerSample, &sample)) {
+            return DRFLAC_FALSE;
+        }
+
+        pDecodedSamples[i] = sample;
+    }
+
+    return DRFLAC_TRUE;
+}
+
+static drflac_bool32 drflac__decode_samples__fixed(drflac_bs* bs, drflac_uint32 blockSize, drflac_uint32 subframeBitsPerSample, drflac_uint8 lpcOrder, drflac_int32* pDecodedSamples)
+{
+    drflac_uint32 i;
+
+    static drflac_int32 lpcCoefficientsTable[5][4] = {
+        {0,  0, 0,  0},
+        {1,  0, 0,  0},
+        {2, -1, 0,  0},
+        {3, -3, 1,  0},
+        {4, -6, 4, -1}
+    };
+
+    /* Warm up samples and coefficients. */
+    for (i = 0; i < lpcOrder; ++i) {
+        drflac_int32 sample;
+        if (!drflac__read_int32(bs, subframeBitsPerSample, &sample)) {
+            return DRFLAC_FALSE;
+        }
+
+        pDecodedSamples[i] = sample;
+    }
+
+    if (!drflac__decode_samples_with_residual(bs, subframeBitsPerSample, blockSize, lpcOrder, 0, 4, lpcCoefficientsTable[lpcOrder], pDecodedSamples)) {
+        return DRFLAC_FALSE;
+    }
+
+    return DRFLAC_TRUE;
+}
+
+static drflac_bool32 drflac__decode_samples__lpc(drflac_bs* bs, drflac_uint32 blockSize, drflac_uint32 bitsPerSample, drflac_uint8 lpcOrder, drflac_int32* pDecodedSamples)
+{
+    drflac_uint8 i;
+    drflac_uint8 lpcPrecision;
+    drflac_int8 lpcShift;
+    drflac_int32 coefficients[32];
+
+    /* Warm up samples. */
+    for (i = 0; i < lpcOrder; ++i) {
+        drflac_int32 sample;
+        if (!drflac__read_int32(bs, bitsPerSample, &sample)) {
+            return DRFLAC_FALSE;
+        }
+
+        pDecodedSamples[i] = sample;
+    }
+
+    if (!drflac__read_uint8(bs, 4, &lpcPrecision)) {
+        return DRFLAC_FALSE;
+    }
+    if (lpcPrecision == 15) {
+        return DRFLAC_FALSE;    /* Invalid. */
+    }
+    lpcPrecision += 1;
+
+    if (!drflac__read_int8(bs, 5, &lpcShift)) {
+        return DRFLAC_FALSE;
+    }
+
+    /*
+    From the FLAC specification:
+
+        Quantized linear predictor coefficient shift needed in bits (NOTE: this number is signed two's-complement)
+
+    Emphasis on the "signed two's-complement". In practice there does not seem to be any encoders nor decoders supporting negative shifts. For now dr_flac is
+    not going to support negative shifts as I don't have any reference files. However, when a reference file comes through I will consider adding support.
+    */
+    if (lpcShift < 0) {
+        return DRFLAC_FALSE;
+    }
+
+    DRFLAC_ZERO_MEMORY(coefficients, sizeof(coefficients));
+    for (i = 0; i < lpcOrder; ++i) {
+        if (!drflac__read_int32(bs, lpcPrecision, coefficients + i)) {
+            return DRFLAC_FALSE;
+        }
+    }
+
+    if (!drflac__decode_samples_with_residual(bs, bitsPerSample, blockSize, lpcOrder, lpcShift, lpcPrecision, coefficients, pDecodedSamples)) {
+        return DRFLAC_FALSE;
+    }
+
+    return DRFLAC_TRUE;
+}
+
+
+static drflac_bool32 drflac__read_next_flac_frame_header(drflac_bs* bs, drflac_uint8 streaminfoBitsPerSample, drflac_frame_header* header)
+{
+    const drflac_uint32 sampleRateTable[12]  = {0, 88200, 176400, 192000, 8000, 16000, 22050, 24000, 32000, 44100, 48000, 96000};
+    const drflac_uint8 bitsPerSampleTable[8] = {0, 8, 12, (drflac_uint8)-1, 16, 20, 24, (drflac_uint8)-1};   /* -1 = reserved. */
+
+    DRFLAC_ASSERT(bs != NULL);
+    DRFLAC_ASSERT(header != NULL);
+
+    /* Keep looping until we find a valid sync code. */
+    for (;;) {
+        drflac_uint8 crc8 = 0xCE; /* 0xCE = drflac_crc8(0, 0x3FFE, 14); */
+        drflac_uint8 reserved = 0;
+        drflac_uint8 blockingStrategy = 0;
+        drflac_uint8 blockSize = 0;
+        drflac_uint8 sampleRate = 0;
+        drflac_uint8 channelAssignment = 0;
+        drflac_uint8 bitsPerSample = 0;
+        drflac_bool32 isVariableBlockSize;
+
+        if (!drflac__find_and_seek_to_next_sync_code(bs)) {
+            return DRFLAC_FALSE;
+        }
+
+        if (!drflac__read_uint8(bs, 1, &reserved)) {
+            return DRFLAC_FALSE;
+        }
+        if (reserved == 1) {
+            continue;
+        }
+        crc8 = drflac_crc8(crc8, reserved, 1);
+
+        if (!drflac__read_uint8(bs, 1, &blockingStrategy)) {
+            return DRFLAC_FALSE;
+        }
+        crc8 = drflac_crc8(crc8, blockingStrategy, 1);
+
+        if (!drflac__read_uint8(bs, 4, &blockSize)) {
+            return DRFLAC_FALSE;
+        }
+        if (blockSize == 0) {
+            continue;
+        }
+        crc8 = drflac_crc8(crc8, blockSize, 4);
+
+        if (!drflac__read_uint8(bs, 4, &sampleRate)) {
+            return DRFLAC_FALSE;
+        }
+        crc8 = drflac_crc8(crc8, sampleRate, 4);
+
+        if (!drflac__read_uint8(bs, 4, &channelAssignment)) {
+            return DRFLAC_FALSE;
+        }
+        if (channelAssignment > 10) {
+            continue;
+        }
+        crc8 = drflac_crc8(crc8, channelAssignment, 4);
+
+        if (!drflac__read_uint8(bs, 3, &bitsPerSample)) {
+            return DRFLAC_FALSE;
+        }
+        if (bitsPerSample == 3 || bitsPerSample == 7) {
+            continue;
+        }
+        crc8 = drflac_crc8(crc8, bitsPerSample, 3);
+
+
+        if (!drflac__read_uint8(bs, 1, &reserved)) {
+            return DRFLAC_FALSE;
+        }
+        if (reserved == 1) {
+            continue;
+        }
+        crc8 = drflac_crc8(crc8, reserved, 1);
+
+
+        isVariableBlockSize = blockingStrategy == 1;
+        if (isVariableBlockSize) {
+            drflac_uint64 pcmFrameNumber;
+            drflac_result result = drflac__read_utf8_coded_number(bs, &pcmFrameNumber, &crc8);
+            if (result != DRFLAC_SUCCESS) {
+                if (result == DRFLAC_AT_END) {
+                    return DRFLAC_FALSE;
+                } else {
+                    continue;
+                }
+            }
+            header->flacFrameNumber  = 0;
+            header->pcmFrameNumber = pcmFrameNumber;
+        } else {
+            drflac_uint64 flacFrameNumber = 0;
+            drflac_result result = drflac__read_utf8_coded_number(bs, &flacFrameNumber, &crc8);
+            if (result != DRFLAC_SUCCESS) {
+                if (result == DRFLAC_AT_END) {
+                    return DRFLAC_FALSE;
+                } else {
+                    continue;
+                }
+            }
+            header->flacFrameNumber  = (drflac_uint32)flacFrameNumber;   /* <-- Safe cast. */
+            header->pcmFrameNumber = 0;
+        }
+
+
+        DRFLAC_ASSERT(blockSize > 0);
+        if (blockSize == 1) {
+            header->blockSizeInPCMFrames = 192;
+        } else if (blockSize <= 5) {
+            DRFLAC_ASSERT(blockSize >= 2);
+            header->blockSizeInPCMFrames = 576 * (1 << (blockSize - 2));
+        } else if (blockSize == 6) {
+            if (!drflac__read_uint16(bs, 8, &header->blockSizeInPCMFrames)) {
+                return DRFLAC_FALSE;
+            }
+            crc8 = drflac_crc8(crc8, header->blockSizeInPCMFrames, 8);
+            header->blockSizeInPCMFrames += 1;
+        } else if (blockSize == 7) {
+            if (!drflac__read_uint16(bs, 16, &header->blockSizeInPCMFrames)) {
+                return DRFLAC_FALSE;
+            }
+            crc8 = drflac_crc8(crc8, header->blockSizeInPCMFrames, 16);
+            if (header->blockSizeInPCMFrames == 0xFFFF) {
+                return DRFLAC_FALSE;    /* Frame is too big. This is the size of the frame minus 1. The STREAMINFO block defines the max block size which is 16-bits. Adding one will make it 17 bits and therefore too big. */
+            }
+            header->blockSizeInPCMFrames += 1;
+        } else {
+            DRFLAC_ASSERT(blockSize >= 8);
+            header->blockSizeInPCMFrames = 256 * (1 << (blockSize - 8));
+        }
+
+
+        if (sampleRate <= 11) {
+            header->sampleRate = sampleRateTable[sampleRate];
+        } else if (sampleRate == 12) {
+            if (!drflac__read_uint32(bs, 8, &header->sampleRate)) {
+                return DRFLAC_FALSE;
+            }
+            crc8 = drflac_crc8(crc8, header->sampleRate, 8);
+            header->sampleRate *= 1000;
+        } else if (sampleRate == 13) {
+            if (!drflac__read_uint32(bs, 16, &header->sampleRate)) {
+                return DRFLAC_FALSE;
+            }
+            crc8 = drflac_crc8(crc8, header->sampleRate, 16);
+        } else if (sampleRate == 14) {
+            if (!drflac__read_uint32(bs, 16, &header->sampleRate)) {
+                return DRFLAC_FALSE;
+            }
+            crc8 = drflac_crc8(crc8, header->sampleRate, 16);
+            header->sampleRate *= 10;
+        } else {
+            continue;  /* Invalid. Assume an invalid block. */
+        }
+
+
+        header->channelAssignment = channelAssignment;
+
+        header->bitsPerSample = bitsPerSampleTable[bitsPerSample];
+        if (header->bitsPerSample == 0) {
+            header->bitsPerSample = streaminfoBitsPerSample;
+        }
+
+        if (header->bitsPerSample != streaminfoBitsPerSample) {
+            /* If this subframe has a different bitsPerSample then streaminfo or the first frame, reject it */
+            return DRFLAC_FALSE;
+        }
+
+        if (!drflac__read_uint8(bs, 8, &header->crc8)) {
+            return DRFLAC_FALSE;
+        }
+
+#ifndef DR_FLAC_NO_CRC
+        if (header->crc8 != crc8) {
+            continue;    /* CRC mismatch. Loop back to the top and find the next sync code. */
+        }
+#endif
+        return DRFLAC_TRUE;
+    }
+}
+
+static drflac_bool32 drflac__read_subframe_header(drflac_bs* bs, drflac_subframe* pSubframe)
+{
+    drflac_uint8 header;
+    int type;
+
+    if (!drflac__read_uint8(bs, 8, &header)) {
+        return DRFLAC_FALSE;
+    }
+
+    /* First bit should always be 0. */
+    if ((header & 0x80) != 0) {
+        return DRFLAC_FALSE;
+    }
+
+    type = (header & 0x7E) >> 1;
+    if (type == 0) {
+        pSubframe->subframeType = DRFLAC_SUBFRAME_CONSTANT;
+    } else if (type == 1) {
+        pSubframe->subframeType = DRFLAC_SUBFRAME_VERBATIM;
+    } else {
+        if ((type & 0x20) != 0) {
+            pSubframe->subframeType = DRFLAC_SUBFRAME_LPC;
+            pSubframe->lpcOrder = (drflac_uint8)(type & 0x1F) + 1;
+        } else if ((type & 0x08) != 0) {
+            pSubframe->subframeType = DRFLAC_SUBFRAME_FIXED;
+            pSubframe->lpcOrder = (drflac_uint8)(type & 0x07);
+            if (pSubframe->lpcOrder > 4) {
+                pSubframe->subframeType = DRFLAC_SUBFRAME_RESERVED;
+                pSubframe->lpcOrder = 0;
+            }
+        } else {
+            pSubframe->subframeType = DRFLAC_SUBFRAME_RESERVED;
+        }
+    }
+
+    if (pSubframe->subframeType == DRFLAC_SUBFRAME_RESERVED) {
+        return DRFLAC_FALSE;
+    }
+
+    /* Wasted bits per sample. */
+    pSubframe->wastedBitsPerSample = 0;
+    if ((header & 0x01) == 1) {
+        unsigned int wastedBitsPerSample;
+        if (!drflac__seek_past_next_set_bit(bs, &wastedBitsPerSample)) {
+            return DRFLAC_FALSE;
+        }
+        pSubframe->wastedBitsPerSample = (drflac_uint8)wastedBitsPerSample + 1;
+    }
+
+    return DRFLAC_TRUE;
+}
+
+static drflac_bool32 drflac__decode_subframe(drflac_bs* bs, drflac_frame* frame, int subframeIndex, drflac_int32* pDecodedSamplesOut)
+{
+    drflac_subframe* pSubframe;
+    drflac_uint32 subframeBitsPerSample;
+
+    DRFLAC_ASSERT(bs != NULL);
+    DRFLAC_ASSERT(frame != NULL);
+
+    pSubframe = frame->subframes + subframeIndex;
+    if (!drflac__read_subframe_header(bs, pSubframe)) {
+        return DRFLAC_FALSE;
+    }
+
+    /* Side channels require an extra bit per sample. Took a while to figure that one out... */
+    subframeBitsPerSample = frame->header.bitsPerSample;
+    if ((frame->header.channelAssignment == DRFLAC_CHANNEL_ASSIGNMENT_LEFT_SIDE || frame->header.channelAssignment == DRFLAC_CHANNEL_ASSIGNMENT_MID_SIDE) && subframeIndex == 1) {
+        subframeBitsPerSample += 1;
+    } else if (frame->header.channelAssignment == DRFLAC_CHANNEL_ASSIGNMENT_RIGHT_SIDE && subframeIndex == 0) {
+        subframeBitsPerSample += 1;
+    }
+
+    if (subframeBitsPerSample > 32) {
+        /* libFLAC and ffmpeg reject 33-bit subframes as well */
+        return DRFLAC_FALSE;
+    }
+
+    /* Need to handle wasted bits per sample. */
+    if (pSubframe->wastedBitsPerSample >= subframeBitsPerSample) {
+        return DRFLAC_FALSE;
+    }
+    subframeBitsPerSample -= pSubframe->wastedBitsPerSample;
+
+    pSubframe->pSamplesS32 = pDecodedSamplesOut;
+
+    switch (pSubframe->subframeType)
+    {
+        case DRFLAC_SUBFRAME_CONSTANT:
+        {
+            drflac__decode_samples__constant(bs, frame->header.blockSizeInPCMFrames, subframeBitsPerSample, pSubframe->pSamplesS32);
+        } break;
+
+        case DRFLAC_SUBFRAME_VERBATIM:
+        {
+            drflac__decode_samples__verbatim(bs, frame->header.blockSizeInPCMFrames, subframeBitsPerSample, pSubframe->pSamplesS32);
+        } break;
+
+        case DRFLAC_SUBFRAME_FIXED:
+        {
+            drflac__decode_samples__fixed(bs, frame->header.blockSizeInPCMFrames, subframeBitsPerSample, pSubframe->lpcOrder, pSubframe->pSamplesS32);
+        } break;
+
+        case DRFLAC_SUBFRAME_LPC:
+        {
+            drflac__decode_samples__lpc(bs, frame->header.blockSizeInPCMFrames, subframeBitsPerSample, pSubframe->lpcOrder, pSubframe->pSamplesS32);
+        } break;
+
+        default: return DRFLAC_FALSE;
+    }
+
+    return DRFLAC_TRUE;
+}
+
+static drflac_bool32 drflac__seek_subframe(drflac_bs* bs, drflac_frame* frame, int subframeIndex)
+{
+    drflac_subframe* pSubframe;
+    drflac_uint32 subframeBitsPerSample;
+
+    DRFLAC_ASSERT(bs != NULL);
+    DRFLAC_ASSERT(frame != NULL);
+
+    pSubframe = frame->subframes + subframeIndex;
+    if (!drflac__read_subframe_header(bs, pSubframe)) {
+        return DRFLAC_FALSE;
+    }
+
+    /* Side channels require an extra bit per sample. Took a while to figure that one out... */
+    subframeBitsPerSample = frame->header.bitsPerSample;
+    if ((frame->header.channelAssignment == DRFLAC_CHANNEL_ASSIGNMENT_LEFT_SIDE || frame->header.channelAssignment == DRFLAC_CHANNEL_ASSIGNMENT_MID_SIDE) && subframeIndex == 1) {
+        subframeBitsPerSample += 1;
+    } else if (frame->header.channelAssignment == DRFLAC_CHANNEL_ASSIGNMENT_RIGHT_SIDE && subframeIndex == 0) {
+        subframeBitsPerSample += 1;
+    }
+
+    /* Need to handle wasted bits per sample. */
+    if (pSubframe->wastedBitsPerSample >= subframeBitsPerSample) {
+        return DRFLAC_FALSE;
+    }
+    subframeBitsPerSample -= pSubframe->wastedBitsPerSample;
+
+    pSubframe->pSamplesS32 = NULL;
+
+    switch (pSubframe->subframeType)
+    {
+        case DRFLAC_SUBFRAME_CONSTANT:
+        {
+            if (!drflac__seek_bits(bs, subframeBitsPerSample)) {
+                return DRFLAC_FALSE;
+            }
+        } break;
+
+        case DRFLAC_SUBFRAME_VERBATIM:
+        {
+            unsigned int bitsToSeek = frame->header.blockSizeInPCMFrames * subframeBitsPerSample;
+            if (!drflac__seek_bits(bs, bitsToSeek)) {
+                return DRFLAC_FALSE;
+            }
+        } break;
+
+        case DRFLAC_SUBFRAME_FIXED:
+        {
+            unsigned int bitsToSeek = pSubframe->lpcOrder * subframeBitsPerSample;
+            if (!drflac__seek_bits(bs, bitsToSeek)) {
+                return DRFLAC_FALSE;
+            }
+
+            if (!drflac__read_and_seek_residual(bs, frame->header.blockSizeInPCMFrames, pSubframe->lpcOrder)) {
+                return DRFLAC_FALSE;
+            }
+        } break;
+
+        case DRFLAC_SUBFRAME_LPC:
+        {
+            drflac_uint8 lpcPrecision;
+
+            unsigned int bitsToSeek = pSubframe->lpcOrder * subframeBitsPerSample;
+            if (!drflac__seek_bits(bs, bitsToSeek)) {
+                return DRFLAC_FALSE;
+            }
+
+            if (!drflac__read_uint8(bs, 4, &lpcPrecision)) {
+                return DRFLAC_FALSE;
+            }
+            if (lpcPrecision == 15) {
+                return DRFLAC_FALSE;    /* Invalid. */
+            }
+            lpcPrecision += 1;
+
+
+            bitsToSeek = (pSubframe->lpcOrder * lpcPrecision) + 5;    /* +5 for shift. */
+            if (!drflac__seek_bits(bs, bitsToSeek)) {
+                return DRFLAC_FALSE;
+            }
+
+            if (!drflac__read_and_seek_residual(bs, frame->header.blockSizeInPCMFrames, pSubframe->lpcOrder)) {
+                return DRFLAC_FALSE;
+            }
+        } break;
+
+        default: return DRFLAC_FALSE;
+    }
+
+    return DRFLAC_TRUE;
+}
+
+
+static DRFLAC_INLINE drflac_uint8 drflac__get_channel_count_from_channel_assignment(drflac_int8 channelAssignment)
+{
+    drflac_uint8 lookup[] = {1, 2, 3, 4, 5, 6, 7, 8, 2, 2, 2};
+
+    DRFLAC_ASSERT(channelAssignment <= 10);
+    return lookup[channelAssignment];
+}
+
+static drflac_result drflac__decode_flac_frame(drflac* pFlac)
+{
+    int channelCount;
+    int i;
+    drflac_uint8 paddingSizeInBits;
+    drflac_uint16 desiredCRC16;
+#ifndef DR_FLAC_NO_CRC
+    drflac_uint16 actualCRC16;
+#endif
+
+    /* This function should be called while the stream is sitting on the first byte after the frame header. */
+    DRFLAC_ZERO_MEMORY(pFlac->currentFLACFrame.subframes, sizeof(pFlac->currentFLACFrame.subframes));
+
+    /* The frame block size must never be larger than the maximum block size defined by the FLAC stream. */
+    if (pFlac->currentFLACFrame.header.blockSizeInPCMFrames > pFlac->maxBlockSizeInPCMFrames) {
+        return DRFLAC_ERROR;
+    }
+
+    /* The number of channels in the frame must match the channel count from the STREAMINFO block. */
+    channelCount = drflac__get_channel_count_from_channel_assignment(pFlac->currentFLACFrame.header.channelAssignment);
+    if (channelCount != (int)pFlac->channels) {
+        return DRFLAC_ERROR;
+    }
+
+    for (i = 0; i < channelCount; ++i) {
+        if (!drflac__decode_subframe(&pFlac->bs, &pFlac->currentFLACFrame, i, pFlac->pDecodedSamples + (pFlac->currentFLACFrame.header.blockSizeInPCMFrames * i))) {
+            return DRFLAC_ERROR;
+        }
+    }
+
+    paddingSizeInBits = (drflac_uint8)(DRFLAC_CACHE_L1_BITS_REMAINING(&pFlac->bs) & 7);
+    if (paddingSizeInBits > 0) {
+        drflac_uint8 padding = 0;
+        if (!drflac__read_uint8(&pFlac->bs, paddingSizeInBits, &padding)) {
+            return DRFLAC_AT_END;
+        }
+    }
+
+#ifndef DR_FLAC_NO_CRC
+    actualCRC16 = drflac__flush_crc16(&pFlac->bs);
+#endif
+    if (!drflac__read_uint16(&pFlac->bs, 16, &desiredCRC16)) {
+        return DRFLAC_AT_END;
+    }
+
+#ifndef DR_FLAC_NO_CRC
+    if (actualCRC16 != desiredCRC16) {
+        return DRFLAC_CRC_MISMATCH;    /* CRC mismatch. */
+    }
+#endif
+
+    pFlac->currentFLACFrame.pcmFramesRemaining = pFlac->currentFLACFrame.header.blockSizeInPCMFrames;
+
+    return DRFLAC_SUCCESS;
+}
+
+static drflac_result drflac__seek_flac_frame(drflac* pFlac)
+{
+    int channelCount;
+    int i;
+    drflac_uint16 desiredCRC16;
+#ifndef DR_FLAC_NO_CRC
+    drflac_uint16 actualCRC16;
+#endif
+
+    channelCount = drflac__get_channel_count_from_channel_assignment(pFlac->currentFLACFrame.header.channelAssignment);
+    for (i = 0; i < channelCount; ++i) {
+        if (!drflac__seek_subframe(&pFlac->bs, &pFlac->currentFLACFrame, i)) {
+            return DRFLAC_ERROR;
+        }
+    }
+
+    /* Padding. */
+    if (!drflac__seek_bits(&pFlac->bs, DRFLAC_CACHE_L1_BITS_REMAINING(&pFlac->bs) & 7)) {
+        return DRFLAC_ERROR;
+    }
+
+    /* CRC. */
+#ifndef DR_FLAC_NO_CRC
+    actualCRC16 = drflac__flush_crc16(&pFlac->bs);
+#endif
+    if (!drflac__read_uint16(&pFlac->bs, 16, &desiredCRC16)) {
+        return DRFLAC_AT_END;
+    }
+
+#ifndef DR_FLAC_NO_CRC
+    if (actualCRC16 != desiredCRC16) {
+        return DRFLAC_CRC_MISMATCH;    /* CRC mismatch. */
+    }
+#endif
+
+    return DRFLAC_SUCCESS;
+}
+
+static drflac_bool32 drflac__read_and_decode_next_flac_frame(drflac* pFlac)
+{
+    DRFLAC_ASSERT(pFlac != NULL);
+
+    for (;;) {
+        drflac_result result;
+
+        if (!drflac__read_next_flac_frame_header(&pFlac->bs, pFlac->bitsPerSample, &pFlac->currentFLACFrame.header)) {
+            return DRFLAC_FALSE;
+        }
+
+        result = drflac__decode_flac_frame(pFlac);
+        if (result != DRFLAC_SUCCESS) {
+            if (result == DRFLAC_CRC_MISMATCH) {
+                continue;   /* CRC mismatch. Skip to the next frame. */
+            } else {
+                return DRFLAC_FALSE;
+            }
+        }
+
+        return DRFLAC_TRUE;
+    }
+}
+
+static void drflac__get_pcm_frame_range_of_current_flac_frame(drflac* pFlac, drflac_uint64* pFirstPCMFrame, drflac_uint64* pLastPCMFrame)
+{
+    drflac_uint64 firstPCMFrame;
+    drflac_uint64 lastPCMFrame;
+
+    DRFLAC_ASSERT(pFlac != NULL);
+
+    firstPCMFrame = pFlac->currentFLACFrame.header.pcmFrameNumber;
+    if (firstPCMFrame == 0) {
+        firstPCMFrame = ((drflac_uint64)pFlac->currentFLACFrame.header.flacFrameNumber) * pFlac->maxBlockSizeInPCMFrames;
+    }
+
+    lastPCMFrame = firstPCMFrame + pFlac->currentFLACFrame.header.blockSizeInPCMFrames;
+    if (lastPCMFrame > 0) {
+        lastPCMFrame -= 1; /* Needs to be zero based. */
+    }
+
+    if (pFirstPCMFrame) {
+        *pFirstPCMFrame = firstPCMFrame;
+    }
+    if (pLastPCMFrame) {
+        *pLastPCMFrame = lastPCMFrame;
+    }
+}
+
+static drflac_bool32 drflac__seek_to_first_frame(drflac* pFlac)
+{
+    drflac_bool32 result;
+
+    DRFLAC_ASSERT(pFlac != NULL);
+
+    result = drflac__seek_to_byte(&pFlac->bs, pFlac->firstFLACFramePosInBytes);
+
+    DRFLAC_ZERO_MEMORY(&pFlac->currentFLACFrame, sizeof(pFlac->currentFLACFrame));
+    pFlac->currentPCMFrame = 0;
+
+    return result;
+}
+
+static DRFLAC_INLINE drflac_result drflac__seek_to_next_flac_frame(drflac* pFlac)
+{
+    /* This function should only ever be called while the decoder is sitting on the first byte past the FRAME_HEADER section. */
+    DRFLAC_ASSERT(pFlac != NULL);
+    return drflac__seek_flac_frame(pFlac);
+}
+
+
+static drflac_uint64 drflac__seek_forward_by_pcm_frames(drflac* pFlac, drflac_uint64 pcmFramesToSeek)
+{
+    drflac_uint64 pcmFramesRead = 0;
+    while (pcmFramesToSeek > 0) {
+        if (pFlac->currentFLACFrame.pcmFramesRemaining == 0) {
+            if (!drflac__read_and_decode_next_flac_frame(pFlac)) {
+                break;  /* Couldn't read the next frame, so just break from the loop and return. */
+            }
+        } else {
+            if (pFlac->currentFLACFrame.pcmFramesRemaining > pcmFramesToSeek) {
+                pcmFramesRead   += pcmFramesToSeek;
+                pFlac->currentFLACFrame.pcmFramesRemaining -= (drflac_uint32)pcmFramesToSeek;   /* <-- Safe cast. Will always be < currentFrame.pcmFramesRemaining < 65536. */
+                pcmFramesToSeek  = 0;
+            } else {
+                pcmFramesRead   += pFlac->currentFLACFrame.pcmFramesRemaining;
+                pcmFramesToSeek -= pFlac->currentFLACFrame.pcmFramesRemaining;
+                pFlac->currentFLACFrame.pcmFramesRemaining = 0;
+            }
+        }
+    }
+
+    pFlac->currentPCMFrame += pcmFramesRead;
+    return pcmFramesRead;
+}
+
+
+static drflac_bool32 drflac__seek_to_pcm_frame__brute_force(drflac* pFlac, drflac_uint64 pcmFrameIndex)
+{
+    drflac_bool32 isMidFrame = DRFLAC_FALSE;
+    drflac_uint64 runningPCMFrameCount;
+
+    DRFLAC_ASSERT(pFlac != NULL);
+
+    /* If we are seeking forward we start from the current position. Otherwise we need to start all the way from the start of the file. */
+    if (pcmFrameIndex >= pFlac->currentPCMFrame) {
+        /* Seeking forward. Need to seek from the current position. */
+        runningPCMFrameCount = pFlac->currentPCMFrame;
+
+        /* The frame header for the first frame may not yet have been read. We need to do that if necessary. */
+        if (pFlac->currentPCMFrame == 0 && pFlac->currentFLACFrame.pcmFramesRemaining == 0) {
+            if (!drflac__read_next_flac_frame_header(&pFlac->bs, pFlac->bitsPerSample, &pFlac->currentFLACFrame.header)) {
+                return DRFLAC_FALSE;
+            }
+        } else {
+            isMidFrame = DRFLAC_TRUE;
+        }
+    } else {
+        /* Seeking backwards. Need to seek from the start of the file. */
+        runningPCMFrameCount = 0;
+
+        /* Move back to the start. */
+        if (!drflac__seek_to_first_frame(pFlac)) {
+            return DRFLAC_FALSE;
+        }
+
+        /* Decode the first frame in preparation for sample-exact seeking below. */
+        if (!drflac__read_next_flac_frame_header(&pFlac->bs, pFlac->bitsPerSample, &pFlac->currentFLACFrame.header)) {
+            return DRFLAC_FALSE;
+        }
+    }
+
+    /*
+    We need to as quickly as possible find the frame that contains the target sample. To do this, we iterate over each frame and inspect its
+    header. If based on the header we can determine that the frame contains the sample, we do a full decode of that frame.
+    */
+    for (;;) {
+        drflac_uint64 pcmFrameCountInThisFLACFrame;
+        drflac_uint64 firstPCMFrameInFLACFrame = 0;
+        drflac_uint64 lastPCMFrameInFLACFrame = 0;
+
+        drflac__get_pcm_frame_range_of_current_flac_frame(pFlac, &firstPCMFrameInFLACFrame, &lastPCMFrameInFLACFrame);
+
+        pcmFrameCountInThisFLACFrame = (lastPCMFrameInFLACFrame - firstPCMFrameInFLACFrame) + 1;
+        if (pcmFrameIndex < (runningPCMFrameCount + pcmFrameCountInThisFLACFrame)) {
+            /*
+            The sample should be in this frame. We need to fully decode it, however if it's an invalid frame (a CRC mismatch), we need to pretend
+            it never existed and keep iterating.
+            */
+            drflac_uint64 pcmFramesToDecode = pcmFrameIndex - runningPCMFrameCount;
+
+            if (!isMidFrame) {
+                drflac_result result = drflac__decode_flac_frame(pFlac);
+                if (result == DRFLAC_SUCCESS) {
+                    /* The frame is valid. We just need to skip over some samples to ensure it's sample-exact. */
+                    return drflac__seek_forward_by_pcm_frames(pFlac, pcmFramesToDecode) == pcmFramesToDecode;  /* <-- If this fails, something bad has happened (it should never fail). */
+                } else {
+                    if (result == DRFLAC_CRC_MISMATCH) {
+                        goto next_iteration;   /* CRC mismatch. Pretend this frame never existed. */
+                    } else {
+                        return DRFLAC_FALSE;
+                    }
+                }
+            } else {
+                /* We started seeking mid-frame which means we need to skip the frame decoding part. */
+                return drflac__seek_forward_by_pcm_frames(pFlac, pcmFramesToDecode) == pcmFramesToDecode;
+            }
+        } else {
+            /*
+            It's not in this frame. We need to seek past the frame, but check if there was a CRC mismatch. If so, we pretend this
+            frame never existed and leave the running sample count untouched.
+            */
+            if (!isMidFrame) {
+                drflac_result result = drflac__seek_to_next_flac_frame(pFlac);
+                if (result == DRFLAC_SUCCESS) {
+                    runningPCMFrameCount += pcmFrameCountInThisFLACFrame;
+                } else {
+                    if (result == DRFLAC_CRC_MISMATCH) {
+                        goto next_iteration;   /* CRC mismatch. Pretend this frame never existed. */
+                    } else {
+                        return DRFLAC_FALSE;
+                    }
+                }
+            } else {
+                /*
+                We started seeking mid-frame which means we need to seek by reading to the end of the frame instead of with
+                drflac__seek_to_next_flac_frame() which only works if the decoder is sitting on the byte just after the frame header.
+                */
+                runningPCMFrameCount += pFlac->currentFLACFrame.pcmFramesRemaining;
+                pFlac->currentFLACFrame.pcmFramesRemaining = 0;
+                isMidFrame = DRFLAC_FALSE;
+            }
+
+            /* If we are seeking to the end of the file and we've just hit it, we're done. */
+            if (pcmFrameIndex == pFlac->totalPCMFrameCount && runningPCMFrameCount == pFlac->totalPCMFrameCount) {
+                return DRFLAC_TRUE;
+            }
+        }
+
+    next_iteration:
+        /* Grab the next frame in preparation for the next iteration. */
+        if (!drflac__read_next_flac_frame_header(&pFlac->bs, pFlac->bitsPerSample, &pFlac->currentFLACFrame.header)) {
+            return DRFLAC_FALSE;
+        }
+    }
+}
+
+
+#if !defined(DR_FLAC_NO_CRC)
+/*
+We use an average compression ratio to determine our approximate start location. FLAC files are generally about 50%-70% the size of their
+uncompressed counterparts so we'll use this as a basis. I'm going to split the middle and use a factor of 0.6 to determine the starting
+location.
+*/
+#define DRFLAC_BINARY_SEARCH_APPROX_COMPRESSION_RATIO 0.6f
+
+static drflac_bool32 drflac__seek_to_approximate_flac_frame_to_byte(drflac* pFlac, drflac_uint64 targetByte, drflac_uint64 rangeLo, drflac_uint64 rangeHi, drflac_uint64* pLastSuccessfulSeekOffset)
+{
+    DRFLAC_ASSERT(pFlac != NULL);
+    DRFLAC_ASSERT(pLastSuccessfulSeekOffset != NULL);
+    DRFLAC_ASSERT(targetByte >= rangeLo);
+    DRFLAC_ASSERT(targetByte <= rangeHi);
+
+    *pLastSuccessfulSeekOffset = pFlac->firstFLACFramePosInBytes;
+
+    for (;;) {
+        /* After rangeLo == rangeHi == targetByte fails, we need to break out. */
+        drflac_uint64 lastTargetByte = targetByte;
+
+        /* When seeking to a byte, failure probably means we've attempted to seek beyond the end of the stream. To counter this we just halve it each attempt. */
+        if (!drflac__seek_to_byte(&pFlac->bs, targetByte)) {
+            /* If we couldn't even seek to the first byte in the stream we have a problem. Just abandon the whole thing. */
+            if (targetByte == 0) {
+                drflac__seek_to_first_frame(pFlac); /* Try to recover. */
+                return DRFLAC_FALSE;
+            }
+
+            /* Halve the byte location and continue. */
+            targetByte = rangeLo + ((rangeHi - rangeLo)/2);
+            rangeHi = targetByte;
+        } else {
+            /* Getting here should mean that we have seeked to an appropriate byte. */
+
+            /* Clear the details of the FLAC frame so we don't misreport data. */
+            DRFLAC_ZERO_MEMORY(&pFlac->currentFLACFrame, sizeof(pFlac->currentFLACFrame));
+
+            /*
+            Now seek to the next FLAC frame. We need to decode the entire frame (not just the header) because it's possible for the header to incorrectly pass the
+            CRC check and return bad data. We need to decode the entire frame to be more certain. Although this seems unlikely, this has happened to me in testing
+            so it needs to stay this way for now.
+            */
+#if 1
+            if (!drflac__read_and_decode_next_flac_frame(pFlac)) {
+                /* Halve the byte location and continue. */
+                targetByte = rangeLo + ((rangeHi - rangeLo)/2);
+                rangeHi = targetByte;
+            } else {
+                break;
+            }
+#else
+            if (!drflac__read_next_flac_frame_header(&pFlac->bs, pFlac->bitsPerSample, &pFlac->currentFLACFrame.header)) {
+                /* Halve the byte location and continue. */
+                targetByte = rangeLo + ((rangeHi - rangeLo)/2);
+                rangeHi = targetByte;
+            } else {
+                break;
+            }
+#endif
+        }
+
+        /* We already tried this byte and there are no more to try, break out. */
+        if(targetByte == lastTargetByte) {
+            return DRFLAC_FALSE;
+        }
+    }
+
+    /* The current PCM frame needs to be updated based on the frame we just seeked to. */
+    drflac__get_pcm_frame_range_of_current_flac_frame(pFlac, &pFlac->currentPCMFrame, NULL);
+
+    DRFLAC_ASSERT(targetByte <= rangeHi);
+
+    *pLastSuccessfulSeekOffset = targetByte;
+    return DRFLAC_TRUE;
+}
+
+static drflac_bool32 drflac__decode_flac_frame_and_seek_forward_by_pcm_frames(drflac* pFlac, drflac_uint64 offset)
+{
+    /* This section of code would be used if we were only decoding the FLAC frame header when calling drflac__seek_to_approximate_flac_frame_to_byte(). */
+#if 0
+    if (drflac__decode_flac_frame(pFlac) != DRFLAC_SUCCESS) {
+        /* We failed to decode this frame which may be due to it being corrupt. We'll just use the next valid FLAC frame. */
+        if (drflac__read_and_decode_next_flac_frame(pFlac) == DRFLAC_FALSE) {
+            return DRFLAC_FALSE;
+        }
+    }
+#endif
+
+    return drflac__seek_forward_by_pcm_frames(pFlac, offset) == offset;
+}
+
+
+static drflac_bool32 drflac__seek_to_pcm_frame__binary_search_internal(drflac* pFlac, drflac_uint64 pcmFrameIndex, drflac_uint64 byteRangeLo, drflac_uint64 byteRangeHi)
+{
+    /* This assumes pFlac->currentPCMFrame is sitting on byteRangeLo upon entry. */
+
+    drflac_uint64 targetByte;
+    drflac_uint64 pcmRangeLo = pFlac->totalPCMFrameCount;
+    drflac_uint64 pcmRangeHi = 0;
+    drflac_uint64 lastSuccessfulSeekOffset = (drflac_uint64)-1;
+    drflac_uint64 closestSeekOffsetBeforeTargetPCMFrame = byteRangeLo;
+    drflac_uint32 seekForwardThreshold = (pFlac->maxBlockSizeInPCMFrames != 0) ? pFlac->maxBlockSizeInPCMFrames*2 : 4096;
+
+    targetByte = byteRangeLo + (drflac_uint64)(((drflac_int64)((pcmFrameIndex - pFlac->currentPCMFrame) * pFlac->channels * pFlac->bitsPerSample)/8.0f) * DRFLAC_BINARY_SEARCH_APPROX_COMPRESSION_RATIO);
+    if (targetByte > byteRangeHi) {
+        targetByte = byteRangeHi;
+    }
+
+    for (;;) {
+        if (drflac__seek_to_approximate_flac_frame_to_byte(pFlac, targetByte, byteRangeLo, byteRangeHi, &lastSuccessfulSeekOffset)) {
+            /* We found a FLAC frame. We need to check if it contains the sample we're looking for. */
+            drflac_uint64 newPCMRangeLo;
+            drflac_uint64 newPCMRangeHi;
+            drflac__get_pcm_frame_range_of_current_flac_frame(pFlac, &newPCMRangeLo, &newPCMRangeHi);
+
+            /* If we selected the same frame, it means we should be pretty close. Just decode the rest. */
+            if (pcmRangeLo == newPCMRangeLo) {
+                if (!drflac__seek_to_approximate_flac_frame_to_byte(pFlac, closestSeekOffsetBeforeTargetPCMFrame, closestSeekOffsetBeforeTargetPCMFrame, byteRangeHi, &lastSuccessfulSeekOffset)) {
+                    break;  /* Failed to seek to closest frame. */
+                }
+
+                if (drflac__decode_flac_frame_and_seek_forward_by_pcm_frames(pFlac, pcmFrameIndex - pFlac->currentPCMFrame)) {
+                    return DRFLAC_TRUE;
+                } else {
+                    break;  /* Failed to seek forward. */
+                }
+            }
+
+            pcmRangeLo = newPCMRangeLo;
+            pcmRangeHi = newPCMRangeHi;
+
+            if (pcmRangeLo <= pcmFrameIndex && pcmRangeHi >= pcmFrameIndex) {
+                /* The target PCM frame is in this FLAC frame. */
+                if (drflac__decode_flac_frame_and_seek_forward_by_pcm_frames(pFlac, pcmFrameIndex - pFlac->currentPCMFrame) ) {
+                    return DRFLAC_TRUE;
+                } else {
+                    break;  /* Failed to seek to FLAC frame. */
+                }
+            } else {
+                const float approxCompressionRatio = (drflac_int64)(lastSuccessfulSeekOffset - pFlac->firstFLACFramePosInBytes) / ((drflac_int64)(pcmRangeLo * pFlac->channels * pFlac->bitsPerSample)/8.0f);
+
+                if (pcmRangeLo > pcmFrameIndex) {
+                    /* We seeked too far forward. We need to move our target byte backward and try again. */
+                    byteRangeHi = lastSuccessfulSeekOffset;
+                    if (byteRangeLo > byteRangeHi) {
+                        byteRangeLo = byteRangeHi;
+                    }
+
+                    targetByte = byteRangeLo + ((byteRangeHi - byteRangeLo) / 2);
+                    if (targetByte < byteRangeLo) {
+                        targetByte = byteRangeLo;
+                    }
+                } else /*if (pcmRangeHi < pcmFrameIndex)*/ {
+                    /* We didn't seek far enough. We need to move our target byte forward and try again. */
+
+                    /* If we're close enough we can just seek forward. */
+                    if ((pcmFrameIndex - pcmRangeLo) < seekForwardThreshold) {
+                        if (drflac__decode_flac_frame_and_seek_forward_by_pcm_frames(pFlac, pcmFrameIndex - pFlac->currentPCMFrame)) {
+                            return DRFLAC_TRUE;
+                        } else {
+                            break;  /* Failed to seek to FLAC frame. */
+                        }
+                    } else {
+                        byteRangeLo = lastSuccessfulSeekOffset;
+                        if (byteRangeHi < byteRangeLo) {
+                            byteRangeHi = byteRangeLo;
+                        }
+
+                        targetByte = lastSuccessfulSeekOffset + (drflac_uint64)(((drflac_int64)((pcmFrameIndex-pcmRangeLo) * pFlac->channels * pFlac->bitsPerSample)/8.0f) * approxCompressionRatio);
+                        if (targetByte > byteRangeHi) {
+                            targetByte = byteRangeHi;
+                        }
+
+                        if (closestSeekOffsetBeforeTargetPCMFrame < lastSuccessfulSeekOffset) {
+                            closestSeekOffsetBeforeTargetPCMFrame = lastSuccessfulSeekOffset;
+                        }
+                    }
+                }
+            }
+        } else {
+            /* Getting here is really bad. We just recover as best we can, but moving to the first frame in the stream, and then abort. */
+            break;
+        }
+    }
+
+    drflac__seek_to_first_frame(pFlac); /* <-- Try to recover. */
+    return DRFLAC_FALSE;
+}
+
+static drflac_bool32 drflac__seek_to_pcm_frame__binary_search(drflac* pFlac, drflac_uint64 pcmFrameIndex)
+{
+    drflac_uint64 byteRangeLo;
+    drflac_uint64 byteRangeHi;
+    drflac_uint32 seekForwardThreshold = (pFlac->maxBlockSizeInPCMFrames != 0) ? pFlac->maxBlockSizeInPCMFrames*2 : 4096;
+
+    /* Our algorithm currently assumes the FLAC stream is currently sitting at the start. */
+    if (drflac__seek_to_first_frame(pFlac) == DRFLAC_FALSE) {
+        return DRFLAC_FALSE;
+    }
+
+    /* If we're close enough to the start, just move to the start and seek forward. */
+    if (pcmFrameIndex < seekForwardThreshold) {
+        return drflac__seek_forward_by_pcm_frames(pFlac, pcmFrameIndex) == pcmFrameIndex;
+    }
+
+    /*
+    Our starting byte range is the byte position of the first FLAC frame and the approximate end of the file as if it were completely uncompressed. This ensures
+    the entire file is included, even though most of the time it'll exceed the end of the actual stream. This is OK as the frame searching logic will handle it.
+    */
+    byteRangeLo = pFlac->firstFLACFramePosInBytes;
+    byteRangeHi = pFlac->firstFLACFramePosInBytes + (drflac_uint64)((drflac_int64)(pFlac->totalPCMFrameCount * pFlac->channels * pFlac->bitsPerSample)/8.0f);
+
+    return drflac__seek_to_pcm_frame__binary_search_internal(pFlac, pcmFrameIndex, byteRangeLo, byteRangeHi);
+}
+#endif  /* !DR_FLAC_NO_CRC */
+
+static drflac_bool32 drflac__seek_to_pcm_frame__seek_table(drflac* pFlac, drflac_uint64 pcmFrameIndex)
+{
+    drflac_uint32 iClosestSeekpoint = 0;
+    drflac_bool32 isMidFrame = DRFLAC_FALSE;
+    drflac_uint64 runningPCMFrameCount;
+    drflac_uint32 iSeekpoint;
+
+
+    DRFLAC_ASSERT(pFlac != NULL);
+
+    if (pFlac->pSeekpoints == NULL || pFlac->seekpointCount == 0) {
+        return DRFLAC_FALSE;
+    }
+
+    /* Do not use the seektable if pcmFramIndex is not coverd by it. */
+    if (pFlac->pSeekpoints[0].firstPCMFrame > pcmFrameIndex) {
+        return DRFLAC_FALSE;
+    }
+
+    for (iSeekpoint = 0; iSeekpoint < pFlac->seekpointCount; ++iSeekpoint) {
+        if (pFlac->pSeekpoints[iSeekpoint].firstPCMFrame >= pcmFrameIndex) {
+            break;
+        }
+
+        iClosestSeekpoint = iSeekpoint;
+    }
+
+    /* There's been cases where the seek table contains only zeros. We need to do some basic validation on the closest seekpoint. */
+    if (pFlac->pSeekpoints[iClosestSeekpoint].pcmFrameCount == 0 || pFlac->pSeekpoints[iClosestSeekpoint].pcmFrameCount > pFlac->maxBlockSizeInPCMFrames) {
+        return DRFLAC_FALSE;
+    }
+    if (pFlac->pSeekpoints[iClosestSeekpoint].firstPCMFrame > pFlac->totalPCMFrameCount && pFlac->totalPCMFrameCount > 0) {
+        return DRFLAC_FALSE;
+    }
+
+#if !defined(DR_FLAC_NO_CRC)
+    /* At this point we should know the closest seek point. We can use a binary search for this. We need to know the total sample count for this. */
+    if (pFlac->totalPCMFrameCount > 0) {
+        drflac_uint64 byteRangeLo;
+        drflac_uint64 byteRangeHi;
+
+        byteRangeHi = pFlac->firstFLACFramePosInBytes + (drflac_uint64)((drflac_int64)(pFlac->totalPCMFrameCount * pFlac->channels * pFlac->bitsPerSample)/8.0f);
+        byteRangeLo = pFlac->firstFLACFramePosInBytes + pFlac->pSeekpoints[iClosestSeekpoint].flacFrameOffset;
+
+        /*
+        If our closest seek point is not the last one, we only need to search between it and the next one. The section below calculates an appropriate starting
+        value for byteRangeHi which will clamp it appropriately.
+
+        Note that the next seekpoint must have an offset greater than the closest seekpoint because otherwise our binary search algorithm will break down. There
+        have been cases where a seektable consists of seek points where every byte offset is set to 0 which causes problems. If this happens we need to abort.
+        */
+        if (iClosestSeekpoint < pFlac->seekpointCount-1) {
+            drflac_uint32 iNextSeekpoint = iClosestSeekpoint + 1;
+
+            /* Basic validation on the seekpoints to ensure they're usable. */
+            if (pFlac->pSeekpoints[iClosestSeekpoint].flacFrameOffset >= pFlac->pSeekpoints[iNextSeekpoint].flacFrameOffset || pFlac->pSeekpoints[iNextSeekpoint].pcmFrameCount == 0) {
+                return DRFLAC_FALSE;    /* The next seekpoint doesn't look right. The seek table cannot be trusted from here. Abort. */
+            }
+
+            if (pFlac->pSeekpoints[iNextSeekpoint].firstPCMFrame != (((drflac_uint64)0xFFFFFFFF << 32) | 0xFFFFFFFF)) { /* Make sure it's not a placeholder seekpoint. */
+                byteRangeHi = pFlac->firstFLACFramePosInBytes + pFlac->pSeekpoints[iNextSeekpoint].flacFrameOffset - 1; /* byteRangeHi must be zero based. */
+            }
+        }
+
+        if (drflac__seek_to_byte(&pFlac->bs, pFlac->firstFLACFramePosInBytes + pFlac->pSeekpoints[iClosestSeekpoint].flacFrameOffset)) {
+            if (drflac__read_next_flac_frame_header(&pFlac->bs, pFlac->bitsPerSample, &pFlac->currentFLACFrame.header)) {
+                drflac__get_pcm_frame_range_of_current_flac_frame(pFlac, &pFlac->currentPCMFrame, NULL);
+
+                if (drflac__seek_to_pcm_frame__binary_search_internal(pFlac, pcmFrameIndex, byteRangeLo, byteRangeHi)) {
+                    return DRFLAC_TRUE;
+                }
+            }
+        }
+    }
+#endif  /* !DR_FLAC_NO_CRC */
+
+    /* Getting here means we need to use a slower algorithm because the binary search method failed or cannot be used. */
+
+    /*
+    If we are seeking forward and the closest seekpoint is _before_ the current sample, we just seek forward from where we are. Otherwise we start seeking
+    from the seekpoint's first sample.
+    */
+    if (pcmFrameIndex >= pFlac->currentPCMFrame && pFlac->pSeekpoints[iClosestSeekpoint].firstPCMFrame <= pFlac->currentPCMFrame) {
+        /* Optimized case. Just seek forward from where we are. */
+        runningPCMFrameCount = pFlac->currentPCMFrame;
+
+        /* The frame header for the first frame may not yet have been read. We need to do that if necessary. */
+        if (pFlac->currentPCMFrame == 0 && pFlac->currentFLACFrame.pcmFramesRemaining == 0) {
+            if (!drflac__read_next_flac_frame_header(&pFlac->bs, pFlac->bitsPerSample, &pFlac->currentFLACFrame.header)) {
+                return DRFLAC_FALSE;
+            }
+        } else {
+            isMidFrame = DRFLAC_TRUE;
+        }
+    } else {
+        /* Slower case. Seek to the start of the seekpoint and then seek forward from there. */
+        runningPCMFrameCount = pFlac->pSeekpoints[iClosestSeekpoint].firstPCMFrame;
+
+        if (!drflac__seek_to_byte(&pFlac->bs, pFlac->firstFLACFramePosInBytes + pFlac->pSeekpoints[iClosestSeekpoint].flacFrameOffset)) {
+            return DRFLAC_FALSE;
+        }
+
+        /* Grab the frame the seekpoint is sitting on in preparation for the sample-exact seeking below. */
+        if (!drflac__read_next_flac_frame_header(&pFlac->bs, pFlac->bitsPerSample, &pFlac->currentFLACFrame.header)) {
+            return DRFLAC_FALSE;
+        }
+    }
+
+    for (;;) {
+        drflac_uint64 pcmFrameCountInThisFLACFrame;
+        drflac_uint64 firstPCMFrameInFLACFrame = 0;
+        drflac_uint64 lastPCMFrameInFLACFrame = 0;
+
+        drflac__get_pcm_frame_range_of_current_flac_frame(pFlac, &firstPCMFrameInFLACFrame, &lastPCMFrameInFLACFrame);
+
+        pcmFrameCountInThisFLACFrame = (lastPCMFrameInFLACFrame - firstPCMFrameInFLACFrame) + 1;
+        if (pcmFrameIndex < (runningPCMFrameCount + pcmFrameCountInThisFLACFrame)) {
+            /*
+            The sample should be in this frame. We need to fully decode it, but if it's an invalid frame (a CRC mismatch) we need to pretend
+            it never existed and keep iterating.
+            */
+            drflac_uint64 pcmFramesToDecode = pcmFrameIndex - runningPCMFrameCount;
+
+            if (!isMidFrame) {
+                drflac_result result = drflac__decode_flac_frame(pFlac);
+                if (result == DRFLAC_SUCCESS) {
+                    /* The frame is valid. We just need to skip over some samples to ensure it's sample-exact. */
+                    return drflac__seek_forward_by_pcm_frames(pFlac, pcmFramesToDecode) == pcmFramesToDecode;  /* <-- If this fails, something bad has happened (it should never fail). */
+                } else {
+                    if (result == DRFLAC_CRC_MISMATCH) {
+                        goto next_iteration;   /* CRC mismatch. Pretend this frame never existed. */
+                    } else {
+                        return DRFLAC_FALSE;
+                    }
+                }
+            } else {
+                /* We started seeking mid-frame which means we need to skip the frame decoding part. */
+                return drflac__seek_forward_by_pcm_frames(pFlac, pcmFramesToDecode) == pcmFramesToDecode;
+            }
+        } else {
+            /*
+            It's not in this frame. We need to seek past the frame, but check if there was a CRC mismatch. If so, we pretend this
+            frame never existed and leave the running sample count untouched.
+            */
+            if (!isMidFrame) {
+                drflac_result result = drflac__seek_to_next_flac_frame(pFlac);
+                if (result == DRFLAC_SUCCESS) {
+                    runningPCMFrameCount += pcmFrameCountInThisFLACFrame;
+                } else {
+                    if (result == DRFLAC_CRC_MISMATCH) {
+                        goto next_iteration;   /* CRC mismatch. Pretend this frame never existed. */
+                    } else {
+                        return DRFLAC_FALSE;
+                    }
+                }
+            } else {
+                /*
+                We started seeking mid-frame which means we need to seek by reading to the end of the frame instead of with
+                drflac__seek_to_next_flac_frame() which only works if the decoder is sitting on the byte just after the frame header.
+                */
+                runningPCMFrameCount += pFlac->currentFLACFrame.pcmFramesRemaining;
+                pFlac->currentFLACFrame.pcmFramesRemaining = 0;
+                isMidFrame = DRFLAC_FALSE;
+            }
+
+            /* If we are seeking to the end of the file and we've just hit it, we're done. */
+            if (pcmFrameIndex == pFlac->totalPCMFrameCount && runningPCMFrameCount == pFlac->totalPCMFrameCount) {
+                return DRFLAC_TRUE;
+            }
+        }
+
+    next_iteration:
+        /* Grab the next frame in preparation for the next iteration. */
+        if (!drflac__read_next_flac_frame_header(&pFlac->bs, pFlac->bitsPerSample, &pFlac->currentFLACFrame.header)) {
+            return DRFLAC_FALSE;
+        }
+    }
+}
+
+
+#ifndef DR_FLAC_NO_OGG
+typedef struct
+{
+    drflac_uint8 capturePattern[4];  /* Should be "OggS" */
+    drflac_uint8 structureVersion;   /* Always 0. */
+    drflac_uint8 headerType;
+    drflac_uint64 granulePosition;
+    drflac_uint32 serialNumber;
+    drflac_uint32 sequenceNumber;
+    drflac_uint32 checksum;
+    drflac_uint8 segmentCount;
+    drflac_uint8 segmentTable[255];
+} drflac_ogg_page_header;
+#endif
+
+typedef struct
+{
+    drflac_read_proc onRead;
+    drflac_seek_proc onSeek;
+    drflac_meta_proc onMeta;
+    drflac_container container;
+    void* pUserData;
+    void* pUserDataMD;
+    drflac_uint32 sampleRate;
+    drflac_uint8  channels;
+    drflac_uint8  bitsPerSample;
+    drflac_uint64 totalPCMFrameCount;
+    drflac_uint16 maxBlockSizeInPCMFrames;
+    drflac_uint64 runningFilePos;
+    drflac_bool32 hasStreamInfoBlock;
+    drflac_bool32 hasMetadataBlocks;
+    drflac_bs bs;                           /* <-- A bit streamer is required for loading data during initialization. */
+    drflac_frame_header firstFrameHeader;   /* <-- The header of the first frame that was read during relaxed initalization. Only set if there is no STREAMINFO block. */
+
+#ifndef DR_FLAC_NO_OGG
+    drflac_uint32 oggSerial;
+    drflac_uint64 oggFirstBytePos;
+    drflac_ogg_page_header oggBosHeader;
+#endif
+} drflac_init_info;
+
+static DRFLAC_INLINE void drflac__decode_block_header(drflac_uint32 blockHeader, drflac_uint8* isLastBlock, drflac_uint8* blockType, drflac_uint32* blockSize)
+{
+    blockHeader = drflac__be2host_32(blockHeader);
+    *isLastBlock = (drflac_uint8)((blockHeader & 0x80000000UL) >> 31);
+    *blockType   = (drflac_uint8)((blockHeader & 0x7F000000UL) >> 24);
+    *blockSize   =                (blockHeader & 0x00FFFFFFUL);
+}
+
+static DRFLAC_INLINE drflac_bool32 drflac__read_and_decode_block_header(drflac_read_proc onRead, void* pUserData, drflac_uint8* isLastBlock, drflac_uint8* blockType, drflac_uint32* blockSize)
+{
+    drflac_uint32 blockHeader;
+
+    *blockSize = 0;
+    if (onRead(pUserData, &blockHeader, 4) != 4) {
+        return DRFLAC_FALSE;
+    }
+
+    drflac__decode_block_header(blockHeader, isLastBlock, blockType, blockSize);
+    return DRFLAC_TRUE;
+}
+
+static drflac_bool32 drflac__read_streaminfo(drflac_read_proc onRead, void* pUserData, drflac_streaminfo* pStreamInfo)
+{
+    drflac_uint32 blockSizes;
+    drflac_uint64 frameSizes = 0;
+    drflac_uint64 importantProps;
+    drflac_uint8 md5[16];
+
+    /* min/max block size. */
+    if (onRead(pUserData, &blockSizes, 4) != 4) {
+        return DRFLAC_FALSE;
+    }
+
+    /* min/max frame size. */
+    if (onRead(pUserData, &frameSizes, 6) != 6) {
+        return DRFLAC_FALSE;
+    }
+
+    /* Sample rate, channels, bits per sample and total sample count. */
+    if (onRead(pUserData, &importantProps, 8) != 8) {
+        return DRFLAC_FALSE;
+    }
+
+    /* MD5 */
+    if (onRead(pUserData, md5, sizeof(md5)) != sizeof(md5)) {
+        return DRFLAC_FALSE;
+    }
+
+    blockSizes     = drflac__be2host_32(blockSizes);
+    frameSizes     = drflac__be2host_64(frameSizes);
+    importantProps = drflac__be2host_64(importantProps);
+
+    pStreamInfo->minBlockSizeInPCMFrames = (drflac_uint16)((blockSizes & 0xFFFF0000) >> 16);
+    pStreamInfo->maxBlockSizeInPCMFrames = (drflac_uint16) (blockSizes & 0x0000FFFF);
+    pStreamInfo->minFrameSizeInPCMFrames = (drflac_uint32)((frameSizes     &  (((drflac_uint64)0x00FFFFFF << 16) << 24)) >> 40);
+    pStreamInfo->maxFrameSizeInPCMFrames = (drflac_uint32)((frameSizes     &  (((drflac_uint64)0x00FFFFFF << 16) <<  0)) >> 16);
+    pStreamInfo->sampleRate              = (drflac_uint32)((importantProps &  (((drflac_uint64)0x000FFFFF << 16) << 28)) >> 44);
+    pStreamInfo->channels                = (drflac_uint8 )((importantProps &  (((drflac_uint64)0x0000000E << 16) << 24)) >> 41) + 1;
+    pStreamInfo->bitsPerSample           = (drflac_uint8 )((importantProps &  (((drflac_uint64)0x0000001F << 16) << 20)) >> 36) + 1;
+    pStreamInfo->totalPCMFrameCount      =                ((importantProps & ((((drflac_uint64)0x0000000F << 16) << 16) | 0xFFFFFFFF)));
+    DRFLAC_COPY_MEMORY(pStreamInfo->md5, md5, sizeof(md5));
+
+    return DRFLAC_TRUE;
+}
+
+
+static void* drflac__malloc_default(size_t sz, void* pUserData)
+{
+    (void)pUserData;
+    return DRFLAC_MALLOC(sz);
+}
+
+static void* drflac__realloc_default(void* p, size_t sz, void* pUserData)
+{
+    (void)pUserData;
+    return DRFLAC_REALLOC(p, sz);
+}
+
+static void drflac__free_default(void* p, void* pUserData)
+{
+    (void)pUserData;
+    DRFLAC_FREE(p);
+}
+
+
+static void* drflac__malloc_from_callbacks(size_t sz, const drflac_allocation_callbacks* pAllocationCallbacks)
+{
+    if (pAllocationCallbacks == NULL) {
+        return NULL;
+    }
+
+    if (pAllocationCallbacks->onMalloc != NULL) {
+        return pAllocationCallbacks->onMalloc(sz, pAllocationCallbacks->pUserData);
+    }
+
+    /* Try using realloc(). */
+    if (pAllocationCallbacks->onRealloc != NULL) {
+        return pAllocationCallbacks->onRealloc(NULL, sz, pAllocationCallbacks->pUserData);
+    }
+
+    return NULL;
+}
+
+static void* drflac__realloc_from_callbacks(void* p, size_t szNew, size_t szOld, const drflac_allocation_callbacks* pAllocationCallbacks)
+{
+    if (pAllocationCallbacks == NULL) {
+        return NULL;
+    }
+
+    if (pAllocationCallbacks->onRealloc != NULL) {
+        return pAllocationCallbacks->onRealloc(p, szNew, pAllocationCallbacks->pUserData);
+    }
+
+    /* Try emulating realloc() in terms of malloc()/free(). */
+    if (pAllocationCallbacks->onMalloc != NULL && pAllocationCallbacks->onFree != NULL) {
+        void* p2;
+
+        p2 = pAllocationCallbacks->onMalloc(szNew, pAllocationCallbacks->pUserData);
+        if (p2 == NULL) {
+            return NULL;
+        }
+
+        if (p != NULL) {
+            DRFLAC_COPY_MEMORY(p2, p, szOld);
+            pAllocationCallbacks->onFree(p, pAllocationCallbacks->pUserData);
+        }
+
+        return p2;
+    }
+
+    return NULL;
+}
+
+static void drflac__free_from_callbacks(void* p, const drflac_allocation_callbacks* pAllocationCallbacks)
+{
+    if (p == NULL || pAllocationCallbacks == NULL) {
+        return;
+    }
+
+    if (pAllocationCallbacks->onFree != NULL) {
+        pAllocationCallbacks->onFree(p, pAllocationCallbacks->pUserData);
+    }
+}
+
+
+static drflac_bool32 drflac__read_and_decode_metadata(drflac_read_proc onRead, drflac_seek_proc onSeek, drflac_meta_proc onMeta, void* pUserData, void* pUserDataMD, drflac_uint64* pFirstFramePos, drflac_uint64* pSeektablePos, drflac_uint32* pSeekpointCount, drflac_allocation_callbacks* pAllocationCallbacks)
+{
+    /*
+    We want to keep track of the byte position in the stream of the seektable. At the time of calling this function we know that
+    we'll be sitting on byte 42.
+    */
+    drflac_uint64 runningFilePos = 42;
+    drflac_uint64 seektablePos   = 0;
+    drflac_uint32 seektableSize  = 0;
+
+    for (;;) {
+        drflac_metadata metadata;
+        drflac_uint8 isLastBlock = 0;
+        drflac_uint8 blockType = 0;
+        drflac_uint32 blockSize;
+        if (drflac__read_and_decode_block_header(onRead, pUserData, &isLastBlock, &blockType, &blockSize) == DRFLAC_FALSE) {
+            return DRFLAC_FALSE;
+        }
+        runningFilePos += 4;
+
+        metadata.type = blockType;
+        metadata.pRawData = NULL;
+        metadata.rawDataSize = 0;
+
+        switch (blockType)
+        {
+            case DRFLAC_METADATA_BLOCK_TYPE_APPLICATION:
+            {
+                if (blockSize < 4) {
+                    return DRFLAC_FALSE;
+                }
+
+                if (onMeta) {
+                    void* pRawData = drflac__malloc_from_callbacks(blockSize, pAllocationCallbacks);
+                    if (pRawData == NULL) {
+                        return DRFLAC_FALSE;
+                    }
+
+                    if (onRead(pUserData, pRawData, blockSize) != blockSize) {
+                        drflac__free_from_callbacks(pRawData, pAllocationCallbacks);
+                        return DRFLAC_FALSE;
+                    }
+
+                    metadata.pRawData = pRawData;
+                    metadata.rawDataSize = blockSize;
+                    metadata.data.application.id       = drflac__be2host_32(*(drflac_uint32*)pRawData);
+                    metadata.data.application.pData    = (const void*)((drflac_uint8*)pRawData + sizeof(drflac_uint32));
+                    metadata.data.application.dataSize = blockSize - sizeof(drflac_uint32);
+                    onMeta(pUserDataMD, &metadata);
+
+                    drflac__free_from_callbacks(pRawData, pAllocationCallbacks);
+                }
+            } break;
+
+            case DRFLAC_METADATA_BLOCK_TYPE_SEEKTABLE:
+            {
+                seektablePos  = runningFilePos;
+                seektableSize = blockSize;
+
+                if (onMeta) {
+                    drflac_uint32 seekpointCount;
+                    drflac_uint32 iSeekpoint;
+                    void* pRawData;
+
+                    seekpointCount = blockSize/DRFLAC_SEEKPOINT_SIZE_IN_BYTES;
+
+                    pRawData = drflac__malloc_from_callbacks(seekpointCount * sizeof(drflac_seekpoint), pAllocationCallbacks);
+                    if (pRawData == NULL) {
+                        return DRFLAC_FALSE;
+                    }
+
+                    /* We need to read seekpoint by seekpoint and do some processing. */
+                    for (iSeekpoint = 0; iSeekpoint < seekpointCount; ++iSeekpoint) {
+                        drflac_seekpoint* pSeekpoint = (drflac_seekpoint*)pRawData + iSeekpoint;
+
+                        if (onRead(pUserData, pSeekpoint, DRFLAC_SEEKPOINT_SIZE_IN_BYTES) != DRFLAC_SEEKPOINT_SIZE_IN_BYTES) {
+                            drflac__free_from_callbacks(pRawData, pAllocationCallbacks);
+                            return DRFLAC_FALSE;
+                        }
+
+                        /* Endian swap. */
+                        pSeekpoint->firstPCMFrame   = drflac__be2host_64(pSeekpoint->firstPCMFrame);
+                        pSeekpoint->flacFrameOffset = drflac__be2host_64(pSeekpoint->flacFrameOffset);
+                        pSeekpoint->pcmFrameCount   = drflac__be2host_16(pSeekpoint->pcmFrameCount);
+                    }
+
+                    metadata.pRawData = pRawData;
+                    metadata.rawDataSize = blockSize;
+                    metadata.data.seektable.seekpointCount = seekpointCount;
+                    metadata.data.seektable.pSeekpoints = (const drflac_seekpoint*)pRawData;
+
+                    onMeta(pUserDataMD, &metadata);
+
+                    drflac__free_from_callbacks(pRawData, pAllocationCallbacks);
+                }
+            } break;
+
+            case DRFLAC_METADATA_BLOCK_TYPE_VORBIS_COMMENT:
+            {
+                if (blockSize < 8) {
+                    return DRFLAC_FALSE;
+                }
+
+                if (onMeta) {
+                    void* pRawData;
+                    const char* pRunningData;
+                    const char* pRunningDataEnd;
+                    drflac_uint32 i;
+
+                    pRawData = drflac__malloc_from_callbacks(blockSize, pAllocationCallbacks);
+                    if (pRawData == NULL) {
+                        return DRFLAC_FALSE;
+                    }
+
+                    if (onRead(pUserData, pRawData, blockSize) != blockSize) {
+                        drflac__free_from_callbacks(pRawData, pAllocationCallbacks);
+                        return DRFLAC_FALSE;
+                    }
+
+                    metadata.pRawData = pRawData;
+                    metadata.rawDataSize = blockSize;
+
+                    pRunningData    = (const char*)pRawData;
+                    pRunningDataEnd = (const char*)pRawData + blockSize;
+
+                    metadata.data.vorbis_comment.vendorLength = drflac__le2host_32_ptr_unaligned(pRunningData); pRunningData += 4;
+
+                    /* Need space for the rest of the block */
+                    if ((pRunningDataEnd - pRunningData) - 4 < (drflac_int64)metadata.data.vorbis_comment.vendorLength) { /* <-- Note the order of operations to avoid overflow to a valid value */
+                        drflac__free_from_callbacks(pRawData, pAllocationCallbacks);
+                        return DRFLAC_FALSE;
+                    }
+                    metadata.data.vorbis_comment.vendor       = pRunningData;                                            pRunningData += metadata.data.vorbis_comment.vendorLength;
+                    metadata.data.vorbis_comment.commentCount = drflac__le2host_32_ptr_unaligned(pRunningData); pRunningData += 4;
+
+                    /* Need space for 'commentCount' comments after the block, which at minimum is a drflac_uint32 per comment */
+                    if ((pRunningDataEnd - pRunningData) / sizeof(drflac_uint32) < metadata.data.vorbis_comment.commentCount) { /* <-- Note the order of operations to avoid overflow to a valid value */
+                        drflac__free_from_callbacks(pRawData, pAllocationCallbacks);
+                        return DRFLAC_FALSE;
+                    }
+                    metadata.data.vorbis_comment.pComments    = pRunningData;
+
+                    /* Check that the comments section is valid before passing it to the callback */
+                    for (i = 0; i < metadata.data.vorbis_comment.commentCount; ++i) {
+                        drflac_uint32 commentLength;
+
+                        if (pRunningDataEnd - pRunningData < 4) {
+                            drflac__free_from_callbacks(pRawData, pAllocationCallbacks);
+                            return DRFLAC_FALSE;
+                        }
+
+                        commentLength = drflac__le2host_32_ptr_unaligned(pRunningData); pRunningData += 4;
+                        if (pRunningDataEnd - pRunningData < (drflac_int64)commentLength) { /* <-- Note the order of operations to avoid overflow to a valid value */
+                            drflac__free_from_callbacks(pRawData, pAllocationCallbacks);
+                            return DRFLAC_FALSE;
+                        }
+                        pRunningData += commentLength;
+                    }
+
+                    onMeta(pUserDataMD, &metadata);
+
+                    drflac__free_from_callbacks(pRawData, pAllocationCallbacks);
+                }
+            } break;
+
+            case DRFLAC_METADATA_BLOCK_TYPE_CUESHEET:
+            {
+                if (blockSize < 396) {
+                    return DRFLAC_FALSE;
+                }
+
+                if (onMeta) {
+                    void* pRawData;
+                    const char* pRunningData;
+                    const char* pRunningDataEnd;
+                    size_t bufferSize;
+                    drflac_uint8 iTrack;
+                    drflac_uint8 iIndex;
+                    void* pTrackData;
+
+                    /*
+                    This needs to be loaded in two passes. The first pass is used to calculate the size of the memory allocation
+                    we need for storing the necessary data. The second pass will fill that buffer with usable data.
+                    */
+                    pRawData = drflac__malloc_from_callbacks(blockSize, pAllocationCallbacks);
+                    if (pRawData == NULL) {
+                        return DRFLAC_FALSE;
+                    }
+
+                    if (onRead(pUserData, pRawData, blockSize) != blockSize) {
+                        drflac__free_from_callbacks(pRawData, pAllocationCallbacks);
+                        return DRFLAC_FALSE;
+                    }
+
+                    metadata.pRawData = pRawData;
+                    metadata.rawDataSize = blockSize;
+
+                    pRunningData    = (const char*)pRawData;
+                    pRunningDataEnd = (const char*)pRawData + blockSize;
+
+                    DRFLAC_COPY_MEMORY(metadata.data.cuesheet.catalog, pRunningData, 128);                              pRunningData += 128;
+                    metadata.data.cuesheet.leadInSampleCount = drflac__be2host_64(*(const drflac_uint64*)pRunningData); pRunningData += 8;
+                    metadata.data.cuesheet.isCD              = (pRunningData[0] & 0x80) != 0;                           pRunningData += 259;
+                    metadata.data.cuesheet.trackCount        = pRunningData[0];                                         pRunningData += 1;
+                    metadata.data.cuesheet.pTrackData        = NULL;    /* Will be filled later. */
+
+                    /* Pass 1: Calculate the size of the buffer for the track data. */
+                    {
+                        const char* pRunningDataSaved = pRunningData;   /* Will be restored at the end in preparation for the second pass. */
+
+                        bufferSize = metadata.data.cuesheet.trackCount * DRFLAC_CUESHEET_TRACK_SIZE_IN_BYTES;
+
+                        for (iTrack = 0; iTrack < metadata.data.cuesheet.trackCount; ++iTrack) {
+                            drflac_uint8 indexCount;
+                            drflac_uint32 indexPointSize;
+
+                            if (pRunningDataEnd - pRunningData < DRFLAC_CUESHEET_TRACK_SIZE_IN_BYTES) {
+                                drflac__free_from_callbacks(pRawData, pAllocationCallbacks);
+                                return DRFLAC_FALSE;
+                            }
+
+                            /* Skip to the index point count */
+                            pRunningData += 35;
+                            
+                            indexCount = pRunningData[0];
+                            pRunningData += 1;
+                            
+                            bufferSize += indexCount * sizeof(drflac_cuesheet_track_index);
+
+                            /* Quick validation check. */
+                            indexPointSize = indexCount * DRFLAC_CUESHEET_TRACK_INDEX_SIZE_IN_BYTES;
+                            if (pRunningDataEnd - pRunningData < (drflac_int64)indexPointSize) {
+                                drflac__free_from_callbacks(pRawData, pAllocationCallbacks);
+                                return DRFLAC_FALSE;
+                            }
+
+                            pRunningData += indexPointSize;
+                        }
+
+                        pRunningData = pRunningDataSaved;
+                    }
+
+                    /* Pass 2: Allocate a buffer and fill the data. Validation was done in the step above so can be skipped. */
+                    {
+                        char* pRunningTrackData;
+
+                        pTrackData = drflac__malloc_from_callbacks(bufferSize, pAllocationCallbacks);
+                        if (pTrackData == NULL) {
+                            drflac__free_from_callbacks(pRawData, pAllocationCallbacks);
+                            return DRFLAC_FALSE;
+                        }
+
+                        pRunningTrackData = (char*)pTrackData;
+
+                        for (iTrack = 0; iTrack < metadata.data.cuesheet.trackCount; ++iTrack) {
+                            drflac_uint8 indexCount;
+
+                            DRFLAC_COPY_MEMORY(pRunningTrackData, pRunningData, DRFLAC_CUESHEET_TRACK_SIZE_IN_BYTES);
+                            pRunningData      += DRFLAC_CUESHEET_TRACK_SIZE_IN_BYTES-1; /* Skip forward, but not beyond the last byte in the CUESHEET_TRACK block which is the index count. */
+                            pRunningTrackData += DRFLAC_CUESHEET_TRACK_SIZE_IN_BYTES-1;
+
+                            /* Grab the index count for the next part. */
+                            indexCount = pRunningData[0];
+                            pRunningData      += 1;
+                            pRunningTrackData += 1;
+
+                            /* Extract each track index. */
+                            for (iIndex = 0; iIndex < indexCount; ++iIndex) {
+                                drflac_cuesheet_track_index* pTrackIndex = (drflac_cuesheet_track_index*)pRunningTrackData;
+
+                                DRFLAC_COPY_MEMORY(pRunningTrackData, pRunningData, DRFLAC_CUESHEET_TRACK_INDEX_SIZE_IN_BYTES);
+                                pRunningData      += DRFLAC_CUESHEET_TRACK_INDEX_SIZE_IN_BYTES;
+                                pRunningTrackData += sizeof(drflac_cuesheet_track_index);
+
+                                pTrackIndex->offset = drflac__be2host_64(pTrackIndex->offset);
+                            }
+                        }
+
+                        metadata.data.cuesheet.pTrackData = pTrackData;
+                    }
+
+                    /* The original data is no longer needed. */
+                    drflac__free_from_callbacks(pRawData, pAllocationCallbacks);
+                    pRawData = NULL;
+
+                    onMeta(pUserDataMD, &metadata);
+
+                    drflac__free_from_callbacks(pTrackData, pAllocationCallbacks);
+                    pTrackData = NULL;
+                }
+            } break;
+
+            case DRFLAC_METADATA_BLOCK_TYPE_PICTURE:
+            {
+                if (blockSize < 32) {
+                    return DRFLAC_FALSE;
+                }
+
+                if (onMeta) {
+                    void* pRawData;
+                    const char* pRunningData;
+                    const char* pRunningDataEnd;
+
+                    pRawData = drflac__malloc_from_callbacks(blockSize, pAllocationCallbacks);
+                    if (pRawData == NULL) {
+                        return DRFLAC_FALSE;
+                    }
+
+                    if (onRead(pUserData, pRawData, blockSize) != blockSize) {
+                        drflac__free_from_callbacks(pRawData, pAllocationCallbacks);
+                        return DRFLAC_FALSE;
+                    }
+
+                    metadata.pRawData = pRawData;
+                    metadata.rawDataSize = blockSize;
+
+                    pRunningData    = (const char*)pRawData;
+                    pRunningDataEnd = (const char*)pRawData + blockSize;
+
+                    metadata.data.picture.type       = drflac__be2host_32_ptr_unaligned(pRunningData); pRunningData += 4;
+                    metadata.data.picture.mimeLength = drflac__be2host_32_ptr_unaligned(pRunningData); pRunningData += 4;
+
+                    /* Need space for the rest of the block */
+                    if ((pRunningDataEnd - pRunningData) - 24 < (drflac_int64)metadata.data.picture.mimeLength) { /* <-- Note the order of operations to avoid overflow to a valid value */
+                        drflac__free_from_callbacks(pRawData, pAllocationCallbacks);
+                        return DRFLAC_FALSE;
+                    }
+                    metadata.data.picture.mime              = pRunningData;                                   pRunningData += metadata.data.picture.mimeLength;
+                    metadata.data.picture.descriptionLength = drflac__be2host_32_ptr_unaligned(pRunningData); pRunningData += 4;
+
+                    /* Need space for the rest of the block */
+                    if ((pRunningDataEnd - pRunningData) - 20 < (drflac_int64)metadata.data.picture.descriptionLength) { /* <-- Note the order of operations to avoid overflow to a valid value */
+                        drflac__free_from_callbacks(pRawData, pAllocationCallbacks);
+                        return DRFLAC_FALSE;
+                    }
+                    metadata.data.picture.description     = pRunningData;                                   pRunningData += metadata.data.picture.descriptionLength;
+                    metadata.data.picture.width           = drflac__be2host_32_ptr_unaligned(pRunningData); pRunningData += 4;
+                    metadata.data.picture.height          = drflac__be2host_32_ptr_unaligned(pRunningData); pRunningData += 4;
+                    metadata.data.picture.colorDepth      = drflac__be2host_32_ptr_unaligned(pRunningData); pRunningData += 4;
+                    metadata.data.picture.indexColorCount = drflac__be2host_32_ptr_unaligned(pRunningData); pRunningData += 4;
+                    metadata.data.picture.pictureDataSize = drflac__be2host_32_ptr_unaligned(pRunningData); pRunningData += 4;
+                    metadata.data.picture.pPictureData    = (const drflac_uint8*)pRunningData;
+
+                    /* Need space for the picture after the block */
+                    if (pRunningDataEnd - pRunningData < (drflac_int64)metadata.data.picture.pictureDataSize) { /* <-- Note the order of operations to avoid overflow to a valid value */
+                        drflac__free_from_callbacks(pRawData, pAllocationCallbacks);
+                        return DRFLAC_FALSE;
+                    }
+
+                    onMeta(pUserDataMD, &metadata);
+
+                    drflac__free_from_callbacks(pRawData, pAllocationCallbacks);
+                }
+            } break;
+
+            case DRFLAC_METADATA_BLOCK_TYPE_PADDING:
+            {
+                if (onMeta) {
+                    metadata.data.padding.unused = 0;
+
+                    /* Padding doesn't have anything meaningful in it, so just skip over it, but make sure the caller is aware of it by firing the callback. */
+                    if (!onSeek(pUserData, blockSize, drflac_seek_origin_current)) {
+                        isLastBlock = DRFLAC_TRUE;  /* An error occurred while seeking. Attempt to recover by treating this as the last block which will in turn terminate the loop. */
+                    } else {
+                        onMeta(pUserDataMD, &metadata);
+                    }
+                }
+            } break;
+
+            case DRFLAC_METADATA_BLOCK_TYPE_INVALID:
+            {
+                /* Invalid chunk. Just skip over this one. */
+                if (onMeta) {
+                    if (!onSeek(pUserData, blockSize, drflac_seek_origin_current)) {
+                        isLastBlock = DRFLAC_TRUE;  /* An error occurred while seeking. Attempt to recover by treating this as the last block which will in turn terminate the loop. */
+                    }
+                }
+            } break;
+
+            default:
+            {
+                /*
+                It's an unknown chunk, but not necessarily invalid. There's a chance more metadata blocks might be defined later on, so we
+                can at the very least report the chunk to the application and let it look at the raw data.
+                */
+                if (onMeta) {
+                    void* pRawData = drflac__malloc_from_callbacks(blockSize, pAllocationCallbacks);
+                    if (pRawData == NULL) {
+                        return DRFLAC_FALSE;
+                    }
+
+                    if (onRead(pUserData, pRawData, blockSize) != blockSize) {
+                        drflac__free_from_callbacks(pRawData, pAllocationCallbacks);
+                        return DRFLAC_FALSE;
+                    }
+
+                    metadata.pRawData = pRawData;
+                    metadata.rawDataSize = blockSize;
+                    onMeta(pUserDataMD, &metadata);
+
+                    drflac__free_from_callbacks(pRawData, pAllocationCallbacks);
+                }
+            } break;
+        }
+
+        /* If we're not handling metadata, just skip over the block. If we are, it will have been handled earlier in the switch statement above. */
+        if (onMeta == NULL && blockSize > 0) {
+            if (!onSeek(pUserData, blockSize, drflac_seek_origin_current)) {
+                isLastBlock = DRFLAC_TRUE;
+            }
+        }
+
+        runningFilePos += blockSize;
+        if (isLastBlock) {
+            break;
+        }
+    }
+
+    *pSeektablePos   = seektablePos;
+    *pSeekpointCount = seektableSize / DRFLAC_SEEKPOINT_SIZE_IN_BYTES;
+    *pFirstFramePos  = runningFilePos;
+
+    return DRFLAC_TRUE;
+}
+
+static drflac_bool32 drflac__init_private__native(drflac_init_info* pInit, drflac_read_proc onRead, drflac_seek_proc onSeek, drflac_meta_proc onMeta, void* pUserData, void* pUserDataMD, drflac_bool32 relaxed)
+{
+    /* Pre Condition: The bit stream should be sitting just past the 4-byte id header. */
+
+    drflac_uint8 isLastBlock;
+    drflac_uint8 blockType;
+    drflac_uint32 blockSize;
+
+    (void)onSeek;
+
+    pInit->container = drflac_container_native;
+
+    /* The first metadata block should be the STREAMINFO block. */
+    if (!drflac__read_and_decode_block_header(onRead, pUserData, &isLastBlock, &blockType, &blockSize)) {
+        return DRFLAC_FALSE;
+    }
+
+    if (blockType != DRFLAC_METADATA_BLOCK_TYPE_STREAMINFO || blockSize != 34) {
+        if (!relaxed) {
+            /* We're opening in strict mode and the first block is not the STREAMINFO block. Error. */
+            return DRFLAC_FALSE;
+        } else {
+            /*
+            Relaxed mode. To open from here we need to just find the first frame and set the sample rate, etc. to whatever is defined
+            for that frame.
+            */
+            pInit->hasStreamInfoBlock = DRFLAC_FALSE;
+            pInit->hasMetadataBlocks  = DRFLAC_FALSE;
+
+            if (!drflac__read_next_flac_frame_header(&pInit->bs, 0, &pInit->firstFrameHeader)) {
+                return DRFLAC_FALSE;    /* Couldn't find a frame. */
+            }
+
+            if (pInit->firstFrameHeader.bitsPerSample == 0) {
+                return DRFLAC_FALSE;    /* Failed to initialize because the first frame depends on the STREAMINFO block, which does not exist. */
+            }
+
+            pInit->sampleRate              = pInit->firstFrameHeader.sampleRate;
+            pInit->channels                = drflac__get_channel_count_from_channel_assignment(pInit->firstFrameHeader.channelAssignment);
+            pInit->bitsPerSample           = pInit->firstFrameHeader.bitsPerSample;
+            pInit->maxBlockSizeInPCMFrames = 65535;   /* <-- See notes here: https://xiph.org/flac/format.html#metadata_block_streaminfo */
+            return DRFLAC_TRUE;
+        }
+    } else {
+        drflac_streaminfo streaminfo;
+        if (!drflac__read_streaminfo(onRead, pUserData, &streaminfo)) {
+            return DRFLAC_FALSE;
+        }
+
+        pInit->hasStreamInfoBlock      = DRFLAC_TRUE;
+        pInit->sampleRate              = streaminfo.sampleRate;
+        pInit->channels                = streaminfo.channels;
+        pInit->bitsPerSample           = streaminfo.bitsPerSample;
+        pInit->totalPCMFrameCount      = streaminfo.totalPCMFrameCount;
+        pInit->maxBlockSizeInPCMFrames = streaminfo.maxBlockSizeInPCMFrames;    /* Don't care about the min block size - only the max (used for determining the size of the memory allocation). */
+        pInit->hasMetadataBlocks       = !isLastBlock;
+
+        if (onMeta) {
+            drflac_metadata metadata;
+            metadata.type = DRFLAC_METADATA_BLOCK_TYPE_STREAMINFO;
+            metadata.pRawData = NULL;
+            metadata.rawDataSize = 0;
+            metadata.data.streaminfo = streaminfo;
+            onMeta(pUserDataMD, &metadata);
+        }
+
+        return DRFLAC_TRUE;
+    }
+}
+
+#ifndef DR_FLAC_NO_OGG
+#define DRFLAC_OGG_MAX_PAGE_SIZE            65307
+#define DRFLAC_OGG_CAPTURE_PATTERN_CRC32    1605413199  /* CRC-32 of "OggS". */
+
+typedef enum
+{
+    drflac_ogg_recover_on_crc_mismatch,
+    drflac_ogg_fail_on_crc_mismatch
+} drflac_ogg_crc_mismatch_recovery;
+
+#ifndef DR_FLAC_NO_CRC
+static drflac_uint32 drflac__crc32_table[] = {
+    0x00000000L, 0x04C11DB7L, 0x09823B6EL, 0x0D4326D9L,
+    0x130476DCL, 0x17C56B6BL, 0x1A864DB2L, 0x1E475005L,
+    0x2608EDB8L, 0x22C9F00FL, 0x2F8AD6D6L, 0x2B4BCB61L,
+    0x350C9B64L, 0x31CD86D3L, 0x3C8EA00AL, 0x384FBDBDL,
+    0x4C11DB70L, 0x48D0C6C7L, 0x4593E01EL, 0x4152FDA9L,
+    0x5F15ADACL, 0x5BD4B01BL, 0x569796C2L, 0x52568B75L,
+    0x6A1936C8L, 0x6ED82B7FL, 0x639B0DA6L, 0x675A1011L,
+    0x791D4014L, 0x7DDC5DA3L, 0x709F7B7AL, 0x745E66CDL,
+    0x9823B6E0L, 0x9CE2AB57L, 0x91A18D8EL, 0x95609039L,
+    0x8B27C03CL, 0x8FE6DD8BL, 0x82A5FB52L, 0x8664E6E5L,
+    0xBE2B5B58L, 0xBAEA46EFL, 0xB7A96036L, 0xB3687D81L,
+    0xAD2F2D84L, 0xA9EE3033L, 0xA4AD16EAL, 0xA06C0B5DL,
+    0xD4326D90L, 0xD0F37027L, 0xDDB056FEL, 0xD9714B49L,
+    0xC7361B4CL, 0xC3F706FBL, 0xCEB42022L, 0xCA753D95L,
+    0xF23A8028L, 0xF6FB9D9FL, 0xFBB8BB46L, 0xFF79A6F1L,
+    0xE13EF6F4L, 0xE5FFEB43L, 0xE8BCCD9AL, 0xEC7DD02DL,
+    0x34867077L, 0x30476DC0L, 0x3D044B19L, 0x39C556AEL,
+    0x278206ABL, 0x23431B1CL, 0x2E003DC5L, 0x2AC12072L,
+    0x128E9DCFL, 0x164F8078L, 0x1B0CA6A1L, 0x1FCDBB16L,
+    0x018AEB13L, 0x054BF6A4L, 0x0808D07DL, 0x0CC9CDCAL,
+    0x7897AB07L, 0x7C56B6B0L, 0x71159069L, 0x75D48DDEL,
+    0x6B93DDDBL, 0x6F52C06CL, 0x6211E6B5L, 0x66D0FB02L,
+    0x5E9F46BFL, 0x5A5E5B08L, 0x571D7DD1L, 0x53DC6066L,
+    0x4D9B3063L, 0x495A2DD4L, 0x44190B0DL, 0x40D816BAL,
+    0xACA5C697L, 0xA864DB20L, 0xA527FDF9L, 0xA1E6E04EL,
+    0xBFA1B04BL, 0xBB60ADFCL, 0xB6238B25L, 0xB2E29692L,
+    0x8AAD2B2FL, 0x8E6C3698L, 0x832F1041L, 0x87EE0DF6L,
+    0x99A95DF3L, 0x9D684044L, 0x902B669DL, 0x94EA7B2AL,
+    0xE0B41DE7L, 0xE4750050L, 0xE9362689L, 0xEDF73B3EL,
+    0xF3B06B3BL, 0xF771768CL, 0xFA325055L, 0xFEF34DE2L,
+    0xC6BCF05FL, 0xC27DEDE8L, 0xCF3ECB31L, 0xCBFFD686L,
+    0xD5B88683L, 0xD1799B34L, 0xDC3ABDEDL, 0xD8FBA05AL,
+    0x690CE0EEL, 0x6DCDFD59L, 0x608EDB80L, 0x644FC637L,
+    0x7A089632L, 0x7EC98B85L, 0x738AAD5CL, 0x774BB0EBL,
+    0x4F040D56L, 0x4BC510E1L, 0x46863638L, 0x42472B8FL,
+    0x5C007B8AL, 0x58C1663DL, 0x558240E4L, 0x51435D53L,
+    0x251D3B9EL, 0x21DC2629L, 0x2C9F00F0L, 0x285E1D47L,
+    0x36194D42L, 0x32D850F5L, 0x3F9B762CL, 0x3B5A6B9BL,
+    0x0315D626L, 0x07D4CB91L, 0x0A97ED48L, 0x0E56F0FFL,
+    0x1011A0FAL, 0x14D0BD4DL, 0x19939B94L, 0x1D528623L,
+    0xF12F560EL, 0xF5EE4BB9L, 0xF8AD6D60L, 0xFC6C70D7L,
+    0xE22B20D2L, 0xE6EA3D65L, 0xEBA91BBCL, 0xEF68060BL,
+    0xD727BBB6L, 0xD3E6A601L, 0xDEA580D8L, 0xDA649D6FL,
+    0xC423CD6AL, 0xC0E2D0DDL, 0xCDA1F604L, 0xC960EBB3L,
+    0xBD3E8D7EL, 0xB9FF90C9L, 0xB4BCB610L, 0xB07DABA7L,
+    0xAE3AFBA2L, 0xAAFBE615L, 0xA7B8C0CCL, 0xA379DD7BL,
+    0x9B3660C6L, 0x9FF77D71L, 0x92B45BA8L, 0x9675461FL,
+    0x8832161AL, 0x8CF30BADL, 0x81B02D74L, 0x857130C3L,
+    0x5D8A9099L, 0x594B8D2EL, 0x5408ABF7L, 0x50C9B640L,
+    0x4E8EE645L, 0x4A4FFBF2L, 0x470CDD2BL, 0x43CDC09CL,
+    0x7B827D21L, 0x7F436096L, 0x7200464FL, 0x76C15BF8L,
+    0x68860BFDL, 0x6C47164AL, 0x61043093L, 0x65C52D24L,
+    0x119B4BE9L, 0x155A565EL, 0x18197087L, 0x1CD86D30L,
+    0x029F3D35L, 0x065E2082L, 0x0B1D065BL, 0x0FDC1BECL,
+    0x3793A651L, 0x3352BBE6L, 0x3E119D3FL, 0x3AD08088L,
+    0x2497D08DL, 0x2056CD3AL, 0x2D15EBE3L, 0x29D4F654L,
+    0xC5A92679L, 0xC1683BCEL, 0xCC2B1D17L, 0xC8EA00A0L,
+    0xD6AD50A5L, 0xD26C4D12L, 0xDF2F6BCBL, 0xDBEE767CL,
+    0xE3A1CBC1L, 0xE760D676L, 0xEA23F0AFL, 0xEEE2ED18L,
+    0xF0A5BD1DL, 0xF464A0AAL, 0xF9278673L, 0xFDE69BC4L,
+    0x89B8FD09L, 0x8D79E0BEL, 0x803AC667L, 0x84FBDBD0L,
+    0x9ABC8BD5L, 0x9E7D9662L, 0x933EB0BBL, 0x97FFAD0CL,
+    0xAFB010B1L, 0xAB710D06L, 0xA6322BDFL, 0xA2F33668L,
+    0xBCB4666DL, 0xB8757BDAL, 0xB5365D03L, 0xB1F740B4L
+};
+#endif
+
+static DRFLAC_INLINE drflac_uint32 drflac_crc32_byte(drflac_uint32 crc32, drflac_uint8 data)
+{
+#ifndef DR_FLAC_NO_CRC
+    return (crc32 << 8) ^ drflac__crc32_table[(drflac_uint8)((crc32 >> 24) & 0xFF) ^ data];
+#else
+    (void)data;
+    return crc32;
+#endif
+}
+
+#if 0
+static DRFLAC_INLINE drflac_uint32 drflac_crc32_uint32(drflac_uint32 crc32, drflac_uint32 data)
+{
+    crc32 = drflac_crc32_byte(crc32, (drflac_uint8)((data >> 24) & 0xFF));
+    crc32 = drflac_crc32_byte(crc32, (drflac_uint8)((data >> 16) & 0xFF));
+    crc32 = drflac_crc32_byte(crc32, (drflac_uint8)((data >>  8) & 0xFF));
+    crc32 = drflac_crc32_byte(crc32, (drflac_uint8)((data >>  0) & 0xFF));
+    return crc32;
+}
+
+static DRFLAC_INLINE drflac_uint32 drflac_crc32_uint64(drflac_uint32 crc32, drflac_uint64 data)
+{
+    crc32 = drflac_crc32_uint32(crc32, (drflac_uint32)((data >> 32) & 0xFFFFFFFF));
+    crc32 = drflac_crc32_uint32(crc32, (drflac_uint32)((data >>  0) & 0xFFFFFFFF));
+    return crc32;
+}
+#endif
+
+static DRFLAC_INLINE drflac_uint32 drflac_crc32_buffer(drflac_uint32 crc32, drflac_uint8* pData, drflac_uint32 dataSize)
+{
+    /* This can be optimized. */
+    drflac_uint32 i;
+    for (i = 0; i < dataSize; ++i) {
+        crc32 = drflac_crc32_byte(crc32, pData[i]);
+    }
+    return crc32;
+}
+
+
+static DRFLAC_INLINE drflac_bool32 drflac_ogg__is_capture_pattern(drflac_uint8 pattern[4])
+{
+    return pattern[0] == 'O' && pattern[1] == 'g' && pattern[2] == 'g' && pattern[3] == 'S';
+}
+
+static DRFLAC_INLINE drflac_uint32 drflac_ogg__get_page_header_size(drflac_ogg_page_header* pHeader)
+{
+    return 27 + pHeader->segmentCount;
+}
+
+static DRFLAC_INLINE drflac_uint32 drflac_ogg__get_page_body_size(drflac_ogg_page_header* pHeader)
+{
+    drflac_uint32 pageBodySize = 0;
+    int i;
+
+    for (i = 0; i < pHeader->segmentCount; ++i) {
+        pageBodySize += pHeader->segmentTable[i];
+    }
+
+    return pageBodySize;
+}
+
+static drflac_result drflac_ogg__read_page_header_after_capture_pattern(drflac_read_proc onRead, void* pUserData, drflac_ogg_page_header* pHeader, drflac_uint32* pBytesRead, drflac_uint32* pCRC32)
+{
+    drflac_uint8 data[23];
+    drflac_uint32 i;
+
+    DRFLAC_ASSERT(*pCRC32 == DRFLAC_OGG_CAPTURE_PATTERN_CRC32);
+
+    if (onRead(pUserData, data, 23) != 23) {
+        return DRFLAC_AT_END;
+    }
+    *pBytesRead += 23;
+
+    /*
+    It's not actually used, but set the capture pattern to 'OggS' for completeness. Not doing this will cause static analysers to complain about
+    us trying to access uninitialized data. We could alternatively just comment out this member of the drflac_ogg_page_header structure, but I
+    like to have it map to the structure of the underlying data.
+    */
+    pHeader->capturePattern[0] = 'O';
+    pHeader->capturePattern[1] = 'g';
+    pHeader->capturePattern[2] = 'g';
+    pHeader->capturePattern[3] = 'S';
+
+    pHeader->structureVersion = data[0];
+    pHeader->headerType       = data[1];
+    DRFLAC_COPY_MEMORY(&pHeader->granulePosition, &data[ 2], 8);
+    DRFLAC_COPY_MEMORY(&pHeader->serialNumber,    &data[10], 4);
+    DRFLAC_COPY_MEMORY(&pHeader->sequenceNumber,  &data[14], 4);
+    DRFLAC_COPY_MEMORY(&pHeader->checksum,        &data[18], 4);
+    pHeader->segmentCount     = data[22];
+
+    /* Calculate the CRC. Note that for the calculation the checksum part of the page needs to be set to 0. */
+    data[18] = 0;
+    data[19] = 0;
+    data[20] = 0;
+    data[21] = 0;
+
+    for (i = 0; i < 23; ++i) {
+        *pCRC32 = drflac_crc32_byte(*pCRC32, data[i]);
+    }
+
+
+    if (onRead(pUserData, pHeader->segmentTable, pHeader->segmentCount) != pHeader->segmentCount) {
+        return DRFLAC_AT_END;
+    }
+    *pBytesRead += pHeader->segmentCount;
+
+    for (i = 0; i < pHeader->segmentCount; ++i) {
+        *pCRC32 = drflac_crc32_byte(*pCRC32, pHeader->segmentTable[i]);
+    }
+
+    return DRFLAC_SUCCESS;
+}
+
+static drflac_result drflac_ogg__read_page_header(drflac_read_proc onRead, void* pUserData, drflac_ogg_page_header* pHeader, drflac_uint32* pBytesRead, drflac_uint32* pCRC32)
+{
+    drflac_uint8 id[4];
+
+    *pBytesRead = 0;
+
+    if (onRead(pUserData, id, 4) != 4) {
+        return DRFLAC_AT_END;
+    }
+    *pBytesRead += 4;
+
+    /* We need to read byte-by-byte until we find the OggS capture pattern. */
+    for (;;) {
+        if (drflac_ogg__is_capture_pattern(id)) {
+            drflac_result result;
+
+            *pCRC32 = DRFLAC_OGG_CAPTURE_PATTERN_CRC32;
+
+            result = drflac_ogg__read_page_header_after_capture_pattern(onRead, pUserData, pHeader, pBytesRead, pCRC32);
+            if (result == DRFLAC_SUCCESS) {
+                return DRFLAC_SUCCESS;
+            } else {
+                if (result == DRFLAC_CRC_MISMATCH) {
+                    continue;
+                } else {
+                    return result;
+                }
+            }
+        } else {
+            /* The first 4 bytes did not equal the capture pattern. Read the next byte and try again. */
+            id[0] = id[1];
+            id[1] = id[2];
+            id[2] = id[3];
+            if (onRead(pUserData, &id[3], 1) != 1) {
+                return DRFLAC_AT_END;
+            }
+            *pBytesRead += 1;
+        }
+    }
+}
+
+
+/*
+The main part of the Ogg encapsulation is the conversion from the physical Ogg bitstream to the native FLAC bitstream. It works
+in three general stages: Ogg Physical Bitstream -> Ogg/FLAC Logical Bitstream -> FLAC Native Bitstream. dr_flac is designed
+in such a way that the core sections assume everything is delivered in native format. Therefore, for each encapsulation type
+dr_flac is supporting there needs to be a layer sitting on top of the onRead and onSeek callbacks that ensures the bits read from
+the physical Ogg bitstream are converted and delivered in native FLAC format.
+*/
+typedef struct
+{
+    drflac_read_proc onRead;                /* The original onRead callback from drflac_open() and family. */
+    drflac_seek_proc onSeek;                /* The original onSeek callback from drflac_open() and family. */
+    void* pUserData;                        /* The user data passed on onRead and onSeek. This is the user data that was passed on drflac_open() and family. */
+    drflac_uint64 currentBytePos;           /* The position of the byte we are sitting on in the physical byte stream. Used for efficient seeking. */
+    drflac_uint64 firstBytePos;             /* The position of the first byte in the physical bitstream. Points to the start of the "OggS" identifier of the FLAC bos page. */
+    drflac_uint32 serialNumber;             /* The serial number of the FLAC audio pages. This is determined by the initial header page that was read during initialization. */
+    drflac_ogg_page_header bosPageHeader;   /* Used for seeking. */
+    drflac_ogg_page_header currentPageHeader;
+    drflac_uint32 bytesRemainingInPage;
+    drflac_uint32 pageDataSize;
+    drflac_uint8 pageData[DRFLAC_OGG_MAX_PAGE_SIZE];
+} drflac_oggbs; /* oggbs = Ogg Bitstream */
+
+static size_t drflac_oggbs__read_physical(drflac_oggbs* oggbs, void* bufferOut, size_t bytesToRead)
+{
+    size_t bytesActuallyRead = oggbs->onRead(oggbs->pUserData, bufferOut, bytesToRead);
+    oggbs->currentBytePos += bytesActuallyRead;
+
+    return bytesActuallyRead;
+}
+
+static drflac_bool32 drflac_oggbs__seek_physical(drflac_oggbs* oggbs, drflac_uint64 offset, drflac_seek_origin origin)
+{
+    if (origin == drflac_seek_origin_start) {
+        if (offset <= 0x7FFFFFFF) {
+            if (!oggbs->onSeek(oggbs->pUserData, (int)offset, drflac_seek_origin_start)) {
+                return DRFLAC_FALSE;
+            }
+            oggbs->currentBytePos = offset;
+
+            return DRFLAC_TRUE;
+        } else {
+            if (!oggbs->onSeek(oggbs->pUserData, 0x7FFFFFFF, drflac_seek_origin_start)) {
+                return DRFLAC_FALSE;
+            }
+            oggbs->currentBytePos = offset;
+
+            return drflac_oggbs__seek_physical(oggbs, offset - 0x7FFFFFFF, drflac_seek_origin_current);
+        }
+    } else {
+        while (offset > 0x7FFFFFFF) {
+            if (!oggbs->onSeek(oggbs->pUserData, 0x7FFFFFFF, drflac_seek_origin_current)) {
+                return DRFLAC_FALSE;
+            }
+            oggbs->currentBytePos += 0x7FFFFFFF;
+            offset -= 0x7FFFFFFF;
+        }
+
+        if (!oggbs->onSeek(oggbs->pUserData, (int)offset, drflac_seek_origin_current)) {    /* <-- Safe cast thanks to the loop above. */
+            return DRFLAC_FALSE;
+        }
+        oggbs->currentBytePos += offset;
+
+        return DRFLAC_TRUE;
+    }
+}
+
+static drflac_bool32 drflac_oggbs__goto_next_page(drflac_oggbs* oggbs, drflac_ogg_crc_mismatch_recovery recoveryMethod)
+{
+    drflac_ogg_page_header header;
+    for (;;) {
+        drflac_uint32 crc32 = 0;
+        drflac_uint32 bytesRead;
+        drflac_uint32 pageBodySize;
+#ifndef DR_FLAC_NO_CRC
+        drflac_uint32 actualCRC32;
+#endif
+
+        if (drflac_ogg__read_page_header(oggbs->onRead, oggbs->pUserData, &header, &bytesRead, &crc32) != DRFLAC_SUCCESS) {
+            return DRFLAC_FALSE;
+        }
+        oggbs->currentBytePos += bytesRead;
+
+        pageBodySize = drflac_ogg__get_page_body_size(&header);
+        if (pageBodySize > DRFLAC_OGG_MAX_PAGE_SIZE) {
+            continue;   /* Invalid page size. Assume it's corrupted and just move to the next page. */
+        }
+
+        if (header.serialNumber != oggbs->serialNumber) {
+            /* It's not a FLAC page. Skip it. */
+            if (pageBodySize > 0 && !drflac_oggbs__seek_physical(oggbs, pageBodySize, drflac_seek_origin_current)) {
+                return DRFLAC_FALSE;
+            }
+            continue;
+        }
+
+
+        /* We need to read the entire page and then do a CRC check on it. If there's a CRC mismatch we need to skip this page. */
+        if (drflac_oggbs__read_physical(oggbs, oggbs->pageData, pageBodySize) != pageBodySize) {
+            return DRFLAC_FALSE;
+        }
+        oggbs->pageDataSize = pageBodySize;
+
+#ifndef DR_FLAC_NO_CRC
+        actualCRC32 = drflac_crc32_buffer(crc32, oggbs->pageData, oggbs->pageDataSize);
+        if (actualCRC32 != header.checksum) {
+            if (recoveryMethod == drflac_ogg_recover_on_crc_mismatch) {
+                continue;   /* CRC mismatch. Skip this page. */
+            } else {
+                /*
+                Even though we are failing on a CRC mismatch, we still want our stream to be in a good state. Therefore we
+                go to the next valid page to ensure we're in a good state, but return false to let the caller know that the
+                seek did not fully complete.
+                */
+                drflac_oggbs__goto_next_page(oggbs, drflac_ogg_recover_on_crc_mismatch);
+                return DRFLAC_FALSE;
+            }
+        }
+#else
+        (void)recoveryMethod;   /* <-- Silence a warning. */
+#endif
+
+        oggbs->currentPageHeader = header;
+        oggbs->bytesRemainingInPage = pageBodySize;
+        return DRFLAC_TRUE;
+    }
+}
+
+/* Function below is unused at the moment, but I might be re-adding it later. */
+#if 0
+static drflac_uint8 drflac_oggbs__get_current_segment_index(drflac_oggbs* oggbs, drflac_uint8* pBytesRemainingInSeg)
+{
+    drflac_uint32 bytesConsumedInPage = drflac_ogg__get_page_body_size(&oggbs->currentPageHeader) - oggbs->bytesRemainingInPage;
+    drflac_uint8 iSeg = 0;
+    drflac_uint32 iByte = 0;
+    while (iByte < bytesConsumedInPage) {
+        drflac_uint8 segmentSize = oggbs->currentPageHeader.segmentTable[iSeg];
+        if (iByte + segmentSize > bytesConsumedInPage) {
+            break;
+        } else {
+            iSeg += 1;
+            iByte += segmentSize;
+        }
+    }
+
+    *pBytesRemainingInSeg = oggbs->currentPageHeader.segmentTable[iSeg] - (drflac_uint8)(bytesConsumedInPage - iByte);
+    return iSeg;
+}
+
+static drflac_bool32 drflac_oggbs__seek_to_next_packet(drflac_oggbs* oggbs)
+{
+    /* The current packet ends when we get to the segment with a lacing value of < 255 which is not at the end of a page. */
+    for (;;) {
+        drflac_bool32 atEndOfPage = DRFLAC_FALSE;
+
+        drflac_uint8 bytesRemainingInSeg;
+        drflac_uint8 iFirstSeg = drflac_oggbs__get_current_segment_index(oggbs, &bytesRemainingInSeg);
+
+        drflac_uint32 bytesToEndOfPacketOrPage = bytesRemainingInSeg;
+        for (drflac_uint8 iSeg = iFirstSeg; iSeg < oggbs->currentPageHeader.segmentCount; ++iSeg) {
+            drflac_uint8 segmentSize = oggbs->currentPageHeader.segmentTable[iSeg];
+            if (segmentSize < 255) {
+                if (iSeg == oggbs->currentPageHeader.segmentCount-1) {
+                    atEndOfPage = DRFLAC_TRUE;
+                }
+
+                break;
+            }
+
+            bytesToEndOfPacketOrPage += segmentSize;
+        }
+
+        /*
+        At this point we will have found either the packet or the end of the page. If were at the end of the page we'll
+        want to load the next page and keep searching for the end of the packet.
+        */
+        drflac_oggbs__seek_physical(oggbs, bytesToEndOfPacketOrPage, drflac_seek_origin_current);
+        oggbs->bytesRemainingInPage -= bytesToEndOfPacketOrPage;
+
+        if (atEndOfPage) {
+            /*
+            We're potentially at the next packet, but we need to check the next page first to be sure because the packet may
+            straddle pages.
+            */
+            if (!drflac_oggbs__goto_next_page(oggbs)) {
+                return DRFLAC_FALSE;
+            }
+
+            /* If it's a fresh packet it most likely means we're at the next packet. */
+            if ((oggbs->currentPageHeader.headerType & 0x01) == 0) {
+                return DRFLAC_TRUE;
+            }
+        } else {
+            /* We're at the next packet. */
+            return DRFLAC_TRUE;
+        }
+    }
+}
+
+static drflac_bool32 drflac_oggbs__seek_to_next_frame(drflac_oggbs* oggbs)
+{
+    /* The bitstream should be sitting on the first byte just after the header of the frame. */
+
+    /* What we're actually doing here is seeking to the start of the next packet. */
+    return drflac_oggbs__seek_to_next_packet(oggbs);
+}
+#endif
+
+static size_t drflac__on_read_ogg(void* pUserData, void* bufferOut, size_t bytesToRead)
+{
+    drflac_oggbs* oggbs = (drflac_oggbs*)pUserData;
+    drflac_uint8* pRunningBufferOut = (drflac_uint8*)bufferOut;
+    size_t bytesRead = 0;
+
+    DRFLAC_ASSERT(oggbs != NULL);
+    DRFLAC_ASSERT(pRunningBufferOut != NULL);
+
+    /* Reading is done page-by-page. If we've run out of bytes in the page we need to move to the next one. */
+    while (bytesRead < bytesToRead) {
+        size_t bytesRemainingToRead = bytesToRead - bytesRead;
+
+        if (oggbs->bytesRemainingInPage >= bytesRemainingToRead) {
+            DRFLAC_COPY_MEMORY(pRunningBufferOut, oggbs->pageData + (oggbs->pageDataSize - oggbs->bytesRemainingInPage), bytesRemainingToRead);
+            bytesRead += bytesRemainingToRead;
+            oggbs->bytesRemainingInPage -= (drflac_uint32)bytesRemainingToRead;
+            break;
+        }
+
+        /* If we get here it means some of the requested data is contained in the next pages. */
+        if (oggbs->bytesRemainingInPage > 0) {
+            DRFLAC_COPY_MEMORY(pRunningBufferOut, oggbs->pageData + (oggbs->pageDataSize - oggbs->bytesRemainingInPage), oggbs->bytesRemainingInPage);
+            bytesRead += oggbs->bytesRemainingInPage;
+            pRunningBufferOut += oggbs->bytesRemainingInPage;
+            oggbs->bytesRemainingInPage = 0;
+        }
+
+        DRFLAC_ASSERT(bytesRemainingToRead > 0);
+        if (!drflac_oggbs__goto_next_page(oggbs, drflac_ogg_recover_on_crc_mismatch)) {
+            break;  /* Failed to go to the next page. Might have simply hit the end of the stream. */
+        }
+    }
+
+    return bytesRead;
+}
+
+static drflac_bool32 drflac__on_seek_ogg(void* pUserData, int offset, drflac_seek_origin origin)
+{
+    drflac_oggbs* oggbs = (drflac_oggbs*)pUserData;
+    int bytesSeeked = 0;
+
+    DRFLAC_ASSERT(oggbs != NULL);
+    DRFLAC_ASSERT(offset >= 0);  /* <-- Never seek backwards. */
+
+    /* Seeking is always forward which makes things a lot simpler. */
+    if (origin == drflac_seek_origin_start) {
+        if (!drflac_oggbs__seek_physical(oggbs, (int)oggbs->firstBytePos, drflac_seek_origin_start)) {
+            return DRFLAC_FALSE;
+        }
+
+        if (!drflac_oggbs__goto_next_page(oggbs, drflac_ogg_fail_on_crc_mismatch)) {
+            return DRFLAC_FALSE;
+        }
+
+        return drflac__on_seek_ogg(pUserData, offset, drflac_seek_origin_current);
+    }
+
+    DRFLAC_ASSERT(origin == drflac_seek_origin_current);
+
+    while (bytesSeeked < offset) {
+        int bytesRemainingToSeek = offset - bytesSeeked;
+        DRFLAC_ASSERT(bytesRemainingToSeek >= 0);
+
+        if (oggbs->bytesRemainingInPage >= (size_t)bytesRemainingToSeek) {
+            bytesSeeked += bytesRemainingToSeek;
+            (void)bytesSeeked;  /* <-- Silence a dead store warning emitted by Clang Static Analyzer. */
+            oggbs->bytesRemainingInPage -= bytesRemainingToSeek;
+            break;
+        }
+
+        /* If we get here it means some of the requested data is contained in the next pages. */
+        if (oggbs->bytesRemainingInPage > 0) {
+            bytesSeeked += (int)oggbs->bytesRemainingInPage;
+            oggbs->bytesRemainingInPage = 0;
+        }
+
+        DRFLAC_ASSERT(bytesRemainingToSeek > 0);
+        if (!drflac_oggbs__goto_next_page(oggbs, drflac_ogg_fail_on_crc_mismatch)) {
+            /* Failed to go to the next page. We either hit the end of the stream or had a CRC mismatch. */
+            return DRFLAC_FALSE;
+        }
+    }
+
+    return DRFLAC_TRUE;
+}
+
+
+static drflac_bool32 drflac_ogg__seek_to_pcm_frame(drflac* pFlac, drflac_uint64 pcmFrameIndex)
+{
+    drflac_oggbs* oggbs = (drflac_oggbs*)pFlac->_oggbs;
+    drflac_uint64 originalBytePos;
+    drflac_uint64 runningGranulePosition;
+    drflac_uint64 runningFrameBytePos;
+    drflac_uint64 runningPCMFrameCount;
+
+    DRFLAC_ASSERT(oggbs != NULL);
+
+    originalBytePos = oggbs->currentBytePos;   /* For recovery. Points to the OggS identifier. */
+
+    /* First seek to the first frame. */
+    if (!drflac__seek_to_byte(&pFlac->bs, pFlac->firstFLACFramePosInBytes)) {
+        return DRFLAC_FALSE;
+    }
+    oggbs->bytesRemainingInPage = 0;
+
+    runningGranulePosition = 0;
+    for (;;) {
+        if (!drflac_oggbs__goto_next_page(oggbs, drflac_ogg_recover_on_crc_mismatch)) {
+            drflac_oggbs__seek_physical(oggbs, originalBytePos, drflac_seek_origin_start);
+            return DRFLAC_FALSE;   /* Never did find that sample... */
+        }
+
+        runningFrameBytePos = oggbs->currentBytePos - drflac_ogg__get_page_header_size(&oggbs->currentPageHeader) - oggbs->pageDataSize;
+        if (oggbs->currentPageHeader.granulePosition >= pcmFrameIndex) {
+            break; /* The sample is somewhere in the previous page. */
+        }
+
+        /*
+        At this point we know the sample is not in the previous page. It could possibly be in this page. For simplicity we
+        disregard any pages that do not begin a fresh packet.
+        */
+        if ((oggbs->currentPageHeader.headerType & 0x01) == 0) {    /* <-- Is it a fresh page? */
+            if (oggbs->currentPageHeader.segmentTable[0] >= 2) {
+                drflac_uint8 firstBytesInPage[2];
+                firstBytesInPage[0] = oggbs->pageData[0];
+                firstBytesInPage[1] = oggbs->pageData[1];
+
+                if ((firstBytesInPage[0] == 0xFF) && (firstBytesInPage[1] & 0xFC) == 0xF8) {    /* <-- Does the page begin with a frame's sync code? */
+                    runningGranulePosition = oggbs->currentPageHeader.granulePosition;
+                }
+
+                continue;
+            }
+        }
+    }
+
+    /*
+    We found the page that that is closest to the sample, so now we need to find it. The first thing to do is seek to the
+    start of that page. In the loop above we checked that it was a fresh page which means this page is also the start of
+    a new frame. This property means that after we've seeked to the page we can immediately start looping over frames until
+    we find the one containing the target sample.
+    */
+    if (!drflac_oggbs__seek_physical(oggbs, runningFrameBytePos, drflac_seek_origin_start)) {
+        return DRFLAC_FALSE;
+    }
+    if (!drflac_oggbs__goto_next_page(oggbs, drflac_ogg_recover_on_crc_mismatch)) {
+        return DRFLAC_FALSE;
+    }
+
+    /*
+    At this point we'll be sitting on the first byte of the frame header of the first frame in the page. We just keep
+    looping over these frames until we find the one containing the sample we're after.
+    */
+    runningPCMFrameCount = runningGranulePosition;
+    for (;;) {
+        /*
+        There are two ways to find the sample and seek past irrelevant frames:
+          1) Use the native FLAC decoder.
+          2) Use Ogg's framing system.
+
+        Both of these options have their own pros and cons. Using the native FLAC decoder is slower because it needs to
+        do a full decode of the frame. Using Ogg's framing system is faster, but more complicated and involves some code
+        duplication for the decoding of frame headers.
+
+        Another thing to consider is that using the Ogg framing system will perform direct seeking of the physical Ogg
+        bitstream. This is important to consider because it means we cannot read data from the drflac_bs object using the
+        standard drflac__*() APIs because that will read in extra data for its own internal caching which in turn breaks
+        the positioning of the read pointer of the physical Ogg bitstream. Therefore, anything that would normally be read
+        using the native FLAC decoding APIs, such as drflac__read_next_flac_frame_header(), need to be re-implemented so as to
+        avoid the use of the drflac_bs object.
+
+        Considering these issues, I have decided to use the slower native FLAC decoding method for the following reasons:
+          1) Seeking is already partially accelerated using Ogg's paging system in the code block above.
+          2) Seeking in an Ogg encapsulated FLAC stream is probably quite uncommon.
+          3) Simplicity.
+        */
+        drflac_uint64 firstPCMFrameInFLACFrame = 0;
+        drflac_uint64 lastPCMFrameInFLACFrame = 0;
+        drflac_uint64 pcmFrameCountInThisFrame;
+
+        if (!drflac__read_next_flac_frame_header(&pFlac->bs, pFlac->bitsPerSample, &pFlac->currentFLACFrame.header)) {
+            return DRFLAC_FALSE;
+        }
+
+        drflac__get_pcm_frame_range_of_current_flac_frame(pFlac, &firstPCMFrameInFLACFrame, &lastPCMFrameInFLACFrame);
+
+        pcmFrameCountInThisFrame = (lastPCMFrameInFLACFrame - firstPCMFrameInFLACFrame) + 1;
+
+        /* If we are seeking to the end of the file and we've just hit it, we're done. */
+        if (pcmFrameIndex == pFlac->totalPCMFrameCount && (runningPCMFrameCount + pcmFrameCountInThisFrame) == pFlac->totalPCMFrameCount) {
+            drflac_result result = drflac__decode_flac_frame(pFlac);
+            if (result == DRFLAC_SUCCESS) {
+                pFlac->currentPCMFrame = pcmFrameIndex;
+                pFlac->currentFLACFrame.pcmFramesRemaining = 0;
+                return DRFLAC_TRUE;
+            } else {
+                return DRFLAC_FALSE;
+            }
+        }
+
+        if (pcmFrameIndex < (runningPCMFrameCount + pcmFrameCountInThisFrame)) {
+            /*
+            The sample should be in this FLAC frame. We need to fully decode it, however if it's an invalid frame (a CRC mismatch), we need to pretend
+            it never existed and keep iterating.
+            */
+            drflac_result result = drflac__decode_flac_frame(pFlac);
+            if (result == DRFLAC_SUCCESS) {
+                /* The frame is valid. We just need to skip over some samples to ensure it's sample-exact. */
+                drflac_uint64 pcmFramesToDecode = (size_t)(pcmFrameIndex - runningPCMFrameCount);    /* <-- Safe cast because the maximum number of samples in a frame is 65535. */
+                if (pcmFramesToDecode == 0) {
+                    return DRFLAC_TRUE;
+                }
+
+                pFlac->currentPCMFrame = runningPCMFrameCount;
+
+                return drflac__seek_forward_by_pcm_frames(pFlac, pcmFramesToDecode) == pcmFramesToDecode;  /* <-- If this fails, something bad has happened (it should never fail). */
+            } else {
+                if (result == DRFLAC_CRC_MISMATCH) {
+                    continue;   /* CRC mismatch. Pretend this frame never existed. */
+                } else {
+                    return DRFLAC_FALSE;
+                }
+            }
+        } else {
+            /*
+            It's not in this frame. We need to seek past the frame, but check if there was a CRC mismatch. If so, we pretend this
+            frame never existed and leave the running sample count untouched.
+            */
+            drflac_result result = drflac__seek_to_next_flac_frame(pFlac);
+            if (result == DRFLAC_SUCCESS) {
+                runningPCMFrameCount += pcmFrameCountInThisFrame;
+            } else {
+                if (result == DRFLAC_CRC_MISMATCH) {
+                    continue;   /* CRC mismatch. Pretend this frame never existed. */
+                } else {
+                    return DRFLAC_FALSE;
+                }
+            }
+        }
+    }
+}
+
+
+
+static drflac_bool32 drflac__init_private__ogg(drflac_init_info* pInit, drflac_read_proc onRead, drflac_seek_proc onSeek, drflac_meta_proc onMeta, void* pUserData, void* pUserDataMD, drflac_bool32 relaxed)
+{
+    drflac_ogg_page_header header;
+    drflac_uint32 crc32 = DRFLAC_OGG_CAPTURE_PATTERN_CRC32;
+    drflac_uint32 bytesRead = 0;
+
+    /* Pre Condition: The bit stream should be sitting just past the 4-byte OggS capture pattern. */
+    (void)relaxed;
+
+    pInit->container = drflac_container_ogg;
+    pInit->oggFirstBytePos = 0;
+
+    /*
+    We'll get here if the first 4 bytes of the stream were the OggS capture pattern, however it doesn't necessarily mean the
+    stream includes FLAC encoded audio. To check for this we need to scan the beginning-of-stream page markers and check if
+    any match the FLAC specification. Important to keep in mind that the stream may be multiplexed.
+    */
+    if (drflac_ogg__read_page_header_after_capture_pattern(onRead, pUserData, &header, &bytesRead, &crc32) != DRFLAC_SUCCESS) {
+        return DRFLAC_FALSE;
+    }
+    pInit->runningFilePos += bytesRead;
+
+    for (;;) {
+        int pageBodySize;
+
+        /* Break if we're past the beginning of stream page. */
+        if ((header.headerType & 0x02) == 0) {
+            return DRFLAC_FALSE;
+        }
+
+        /* Check if it's a FLAC header. */
+        pageBodySize = drflac_ogg__get_page_body_size(&header);
+        if (pageBodySize == 51) {   /* 51 = the lacing value of the FLAC header packet. */
+            /* It could be a FLAC page... */
+            drflac_uint32 bytesRemainingInPage = pageBodySize;
+            drflac_uint8 packetType;
+
+            if (onRead(pUserData, &packetType, 1) != 1) {
+                return DRFLAC_FALSE;
+            }
+
+            bytesRemainingInPage -= 1;
+            if (packetType == 0x7F) {
+                /* Increasingly more likely to be a FLAC page... */
+                drflac_uint8 sig[4];
+                if (onRead(pUserData, sig, 4) != 4) {
+                    return DRFLAC_FALSE;
+                }
+
+                bytesRemainingInPage -= 4;
+                if (sig[0] == 'F' && sig[1] == 'L' && sig[2] == 'A' && sig[3] == 'C') {
+                    /* Almost certainly a FLAC page... */
+                    drflac_uint8 mappingVersion[2];
+                    if (onRead(pUserData, mappingVersion, 2) != 2) {
+                        return DRFLAC_FALSE;
+                    }
+
+                    if (mappingVersion[0] != 1) {
+                        return DRFLAC_FALSE;   /* Only supporting version 1.x of the Ogg mapping. */
+                    }
+
+                    /*
+                    The next 2 bytes are the non-audio packets, not including this one. We don't care about this because we're going to
+                    be handling it in a generic way based on the serial number and packet types.
+                    */
+                    if (!onSeek(pUserData, 2, drflac_seek_origin_current)) {
+                        return DRFLAC_FALSE;
+                    }
+
+                    /* Expecting the native FLAC signature "fLaC". */
+                    if (onRead(pUserData, sig, 4) != 4) {
+                        return DRFLAC_FALSE;
+                    }
+
+                    if (sig[0] == 'f' && sig[1] == 'L' && sig[2] == 'a' && sig[3] == 'C') {
+                        /* The remaining data in the page should be the STREAMINFO block. */
+                        drflac_streaminfo streaminfo;
+                        drflac_uint8 isLastBlock;
+                        drflac_uint8 blockType;
+                        drflac_uint32 blockSize;
+                        if (!drflac__read_and_decode_block_header(onRead, pUserData, &isLastBlock, &blockType, &blockSize)) {
+                            return DRFLAC_FALSE;
+                        }
+
+                        if (blockType != DRFLAC_METADATA_BLOCK_TYPE_STREAMINFO || blockSize != 34) {
+                            return DRFLAC_FALSE;    /* Invalid block type. First block must be the STREAMINFO block. */
+                        }
+
+                        if (drflac__read_streaminfo(onRead, pUserData, &streaminfo)) {
+                            /* Success! */
+                            pInit->hasStreamInfoBlock      = DRFLAC_TRUE;
+                            pInit->sampleRate              = streaminfo.sampleRate;
+                            pInit->channels                = streaminfo.channels;
+                            pInit->bitsPerSample           = streaminfo.bitsPerSample;
+                            pInit->totalPCMFrameCount      = streaminfo.totalPCMFrameCount;
+                            pInit->maxBlockSizeInPCMFrames = streaminfo.maxBlockSizeInPCMFrames;
+                            pInit->hasMetadataBlocks       = !isLastBlock;
+
+                            if (onMeta) {
+                                drflac_metadata metadata;
+                                metadata.type = DRFLAC_METADATA_BLOCK_TYPE_STREAMINFO;
+                                metadata.pRawData = NULL;
+                                metadata.rawDataSize = 0;
+                                metadata.data.streaminfo = streaminfo;
+                                onMeta(pUserDataMD, &metadata);
+                            }
+
+                            pInit->runningFilePos  += pageBodySize;
+                            pInit->oggFirstBytePos  = pInit->runningFilePos - 79;   /* Subtracting 79 will place us right on top of the "OggS" identifier of the FLAC bos page. */
+                            pInit->oggSerial        = header.serialNumber;
+                            pInit->oggBosHeader     = header;
+                            break;
+                        } else {
+                            /* Failed to read STREAMINFO block. Aww, so close... */
+                            return DRFLAC_FALSE;
+                        }
+                    } else {
+                        /* Invalid file. */
+                        return DRFLAC_FALSE;
+                    }
+                } else {
+                    /* Not a FLAC header. Skip it. */
+                    if (!onSeek(pUserData, bytesRemainingInPage, drflac_seek_origin_current)) {
+                        return DRFLAC_FALSE;
+                    }
+                }
+            } else {
+                /* Not a FLAC header. Seek past the entire page and move on to the next. */
+                if (!onSeek(pUserData, bytesRemainingInPage, drflac_seek_origin_current)) {
+                    return DRFLAC_FALSE;
+                }
+            }
+        } else {
+            if (!onSeek(pUserData, pageBodySize, drflac_seek_origin_current)) {
+                return DRFLAC_FALSE;
+            }
+        }
+
+        pInit->runningFilePos += pageBodySize;
+
+
+        /* Read the header of the next page. */
+        if (drflac_ogg__read_page_header(onRead, pUserData, &header, &bytesRead, &crc32) != DRFLAC_SUCCESS) {
+            return DRFLAC_FALSE;
+        }
+        pInit->runningFilePos += bytesRead;
+    }
+
+    /*
+    If we get here it means we found a FLAC audio stream. We should be sitting on the first byte of the header of the next page. The next
+    packets in the FLAC logical stream contain the metadata. The only thing left to do in the initialization phase for Ogg is to create the
+    Ogg bistream object.
+    */
+    pInit->hasMetadataBlocks = DRFLAC_TRUE;    /* <-- Always have at least VORBIS_COMMENT metadata block. */
+    return DRFLAC_TRUE;
+}
+#endif
+
+static drflac_bool32 drflac__init_private(drflac_init_info* pInit, drflac_read_proc onRead, drflac_seek_proc onSeek, drflac_meta_proc onMeta, drflac_container container, void* pUserData, void* pUserDataMD)
+{
+    drflac_bool32 relaxed;
+    drflac_uint8 id[4];
+
+    if (pInit == NULL || onRead == NULL || onSeek == NULL) {
+        return DRFLAC_FALSE;
+    }
+
+    DRFLAC_ZERO_MEMORY(pInit, sizeof(*pInit));
+    pInit->onRead       = onRead;
+    pInit->onSeek       = onSeek;
+    pInit->onMeta       = onMeta;
+    pInit->container    = container;
+    pInit->pUserData    = pUserData;
+    pInit->pUserDataMD  = pUserDataMD;
+
+    pInit->bs.onRead    = onRead;
+    pInit->bs.onSeek    = onSeek;
+    pInit->bs.pUserData = pUserData;
+    drflac__reset_cache(&pInit->bs);
+
+
+    /* If the container is explicitly defined then we can try opening in relaxed mode. */
+    relaxed = container != drflac_container_unknown;
+
+    /* Skip over any ID3 tags. */
+    for (;;) {
+        if (onRead(pUserData, id, 4) != 4) {
+            return DRFLAC_FALSE;    /* Ran out of data. */
+        }
+        pInit->runningFilePos += 4;
+
+        if (id[0] == 'I' && id[1] == 'D' && id[2] == '3') {
+            drflac_uint8 header[6];
+            drflac_uint8 flags;
+            drflac_uint32 headerSize;
+
+            if (onRead(pUserData, header, 6) != 6) {
+                return DRFLAC_FALSE;    /* Ran out of data. */
+            }
+            pInit->runningFilePos += 6;
+
+            flags = header[1];
+
+            DRFLAC_COPY_MEMORY(&headerSize, header+2, 4);
+            headerSize = drflac__unsynchsafe_32(drflac__be2host_32(headerSize));
+            if (flags & 0x10) {
+                headerSize += 10;
+            }
+
+            if (!onSeek(pUserData, headerSize, drflac_seek_origin_current)) {
+                return DRFLAC_FALSE;    /* Failed to seek past the tag. */
+            }
+            pInit->runningFilePos += headerSize;
+        } else {
+            break;
+        }
+    }
+
+    if (id[0] == 'f' && id[1] == 'L' && id[2] == 'a' && id[3] == 'C') {
+        return drflac__init_private__native(pInit, onRead, onSeek, onMeta, pUserData, pUserDataMD, relaxed);
+    }
+#ifndef DR_FLAC_NO_OGG
+    if (id[0] == 'O' && id[1] == 'g' && id[2] == 'g' && id[3] == 'S') {
+        return drflac__init_private__ogg(pInit, onRead, onSeek, onMeta, pUserData, pUserDataMD, relaxed);
+    }
+#endif
+
+    /* If we get here it means we likely don't have a header. Try opening in relaxed mode, if applicable. */
+    if (relaxed) {
+        if (container == drflac_container_native) {
+            return drflac__init_private__native(pInit, onRead, onSeek, onMeta, pUserData, pUserDataMD, relaxed);
+        }
+#ifndef DR_FLAC_NO_OGG
+        if (container == drflac_container_ogg) {
+            return drflac__init_private__ogg(pInit, onRead, onSeek, onMeta, pUserData, pUserDataMD, relaxed);
+        }
+#endif
+    }
+
+    /* Unsupported container. */
+    return DRFLAC_FALSE;
+}
+
+static void drflac__init_from_info(drflac* pFlac, const drflac_init_info* pInit)
+{
+    DRFLAC_ASSERT(pFlac != NULL);
+    DRFLAC_ASSERT(pInit != NULL);
+
+    DRFLAC_ZERO_MEMORY(pFlac, sizeof(*pFlac));
+    pFlac->bs                      = pInit->bs;
+    pFlac->onMeta                  = pInit->onMeta;
+    pFlac->pUserDataMD             = pInit->pUserDataMD;
+    pFlac->maxBlockSizeInPCMFrames = pInit->maxBlockSizeInPCMFrames;
+    pFlac->sampleRate              = pInit->sampleRate;
+    pFlac->channels                = (drflac_uint8)pInit->channels;
+    pFlac->bitsPerSample           = (drflac_uint8)pInit->bitsPerSample;
+    pFlac->totalPCMFrameCount      = pInit->totalPCMFrameCount;
+    pFlac->container               = pInit->container;
+}
+
+
+static drflac* drflac_open_with_metadata_private(drflac_read_proc onRead, drflac_seek_proc onSeek, drflac_meta_proc onMeta, drflac_container container, void* pUserData, void* pUserDataMD, const drflac_allocation_callbacks* pAllocationCallbacks)
+{
+    drflac_init_info init;
+    drflac_uint32 allocationSize;
+    drflac_uint32 wholeSIMDVectorCountPerChannel;
+    drflac_uint32 decodedSamplesAllocationSize;
+#ifndef DR_FLAC_NO_OGG
+    drflac_oggbs* pOggbs = NULL;
+#endif
+    drflac_uint64 firstFramePos;
+    drflac_uint64 seektablePos;
+    drflac_uint32 seekpointCount;
+    drflac_allocation_callbacks allocationCallbacks;
+    drflac* pFlac;
+
+    /* CPU support first. */
+    drflac__init_cpu_caps();
+
+    if (!drflac__init_private(&init, onRead, onSeek, onMeta, container, pUserData, pUserDataMD)) {
+        return NULL;
+    }
+
+    if (pAllocationCallbacks != NULL) {
+        allocationCallbacks = *pAllocationCallbacks;
+        if (allocationCallbacks.onFree == NULL || (allocationCallbacks.onMalloc == NULL && allocationCallbacks.onRealloc == NULL)) {
+            return NULL;    /* Invalid allocation callbacks. */
+        }
+    } else {
+        allocationCallbacks.pUserData = NULL;
+        allocationCallbacks.onMalloc  = drflac__malloc_default;
+        allocationCallbacks.onRealloc = drflac__realloc_default;
+        allocationCallbacks.onFree    = drflac__free_default;
+    }
+
+
+    /*
+    The size of the allocation for the drflac object needs to be large enough to fit the following:
+      1) The main members of the drflac structure
+      2) A block of memory large enough to store the decoded samples of the largest frame in the stream
+      3) If the container is Ogg, a drflac_oggbs object
+
+    The complicated part of the allocation is making sure there's enough room the decoded samples, taking into consideration
+    the different SIMD instruction sets.
+    */
+    allocationSize = sizeof(drflac);
+
+    /*
+    The allocation size for decoded frames depends on the number of 32-bit integers that fit inside the largest SIMD vector
+    we are supporting.
+    */
+    if ((init.maxBlockSizeInPCMFrames % (DRFLAC_MAX_SIMD_VECTOR_SIZE / sizeof(drflac_int32))) == 0) {
+        wholeSIMDVectorCountPerChannel = (init.maxBlockSizeInPCMFrames / (DRFLAC_MAX_SIMD_VECTOR_SIZE / sizeof(drflac_int32)));
+    } else {
+        wholeSIMDVectorCountPerChannel = (init.maxBlockSizeInPCMFrames / (DRFLAC_MAX_SIMD_VECTOR_SIZE / sizeof(drflac_int32))) + 1;
+    }
+
+    decodedSamplesAllocationSize = wholeSIMDVectorCountPerChannel * DRFLAC_MAX_SIMD_VECTOR_SIZE * init.channels;
+
+    allocationSize += decodedSamplesAllocationSize;
+    allocationSize += DRFLAC_MAX_SIMD_VECTOR_SIZE;  /* Allocate extra bytes to ensure we have enough for alignment. */
+
+#ifndef DR_FLAC_NO_OGG
+    /* There's additional data required for Ogg streams. */
+    if (init.container == drflac_container_ogg) {
+        allocationSize += sizeof(drflac_oggbs);
+
+        pOggbs = (drflac_oggbs*)drflac__malloc_from_callbacks(sizeof(*pOggbs), &allocationCallbacks);
+        if (pOggbs == NULL) {
+            return NULL; /*DRFLAC_OUT_OF_MEMORY;*/
+        }
+
+        DRFLAC_ZERO_MEMORY(pOggbs, sizeof(*pOggbs));
+        pOggbs->onRead = onRead;
+        pOggbs->onSeek = onSeek;
+        pOggbs->pUserData = pUserData;
+        pOggbs->currentBytePos = init.oggFirstBytePos;
+        pOggbs->firstBytePos = init.oggFirstBytePos;
+        pOggbs->serialNumber = init.oggSerial;
+        pOggbs->bosPageHeader = init.oggBosHeader;
+        pOggbs->bytesRemainingInPage = 0;
+    }
+#endif
+
+    /*
+    This part is a bit awkward. We need to load the seektable so that it can be referenced in-memory, but I want the drflac object to
+    consist of only a single heap allocation. To this, the size of the seek table needs to be known, which we determine when reading
+    and decoding the metadata.
+    */
+    firstFramePos  = 42;   /* <-- We know we are at byte 42 at this point. */
+    seektablePos   = 0;
+    seekpointCount = 0;
+    if (init.hasMetadataBlocks) {
+        drflac_read_proc onReadOverride = onRead;
+        drflac_seek_proc onSeekOverride = onSeek;
+        void* pUserDataOverride = pUserData;
+
+#ifndef DR_FLAC_NO_OGG
+        if (init.container == drflac_container_ogg) {
+            onReadOverride = drflac__on_read_ogg;
+            onSeekOverride = drflac__on_seek_ogg;
+            pUserDataOverride = (void*)pOggbs;
+        }
+#endif
+
+        if (!drflac__read_and_decode_metadata(onReadOverride, onSeekOverride, onMeta, pUserDataOverride, pUserDataMD, &firstFramePos, &seektablePos, &seekpointCount, &allocationCallbacks)) {
+        #ifndef DR_FLAC_NO_OGG
+            drflac__free_from_callbacks(pOggbs, &allocationCallbacks);
+        #endif
+            return NULL;
+        }
+
+        allocationSize += seekpointCount * sizeof(drflac_seekpoint);
+    }
+
+
+    pFlac = (drflac*)drflac__malloc_from_callbacks(allocationSize, &allocationCallbacks);
+    if (pFlac == NULL) {
+    #ifndef DR_FLAC_NO_OGG
+        drflac__free_from_callbacks(pOggbs, &allocationCallbacks);
+    #endif
+        return NULL;
+    }
+
+    drflac__init_from_info(pFlac, &init);
+    pFlac->allocationCallbacks = allocationCallbacks;
+    pFlac->pDecodedSamples = (drflac_int32*)drflac_align((size_t)pFlac->pExtraData, DRFLAC_MAX_SIMD_VECTOR_SIZE);
+
+#ifndef DR_FLAC_NO_OGG
+    if (init.container == drflac_container_ogg) {
+        drflac_oggbs* pInternalOggbs = (drflac_oggbs*)((drflac_uint8*)pFlac->pDecodedSamples + decodedSamplesAllocationSize + (seekpointCount * sizeof(drflac_seekpoint)));
+        DRFLAC_COPY_MEMORY(pInternalOggbs, pOggbs, sizeof(*pOggbs));
+
+        /* At this point the pOggbs object has been handed over to pInternalOggbs and can be freed. */
+        drflac__free_from_callbacks(pOggbs, &allocationCallbacks);
+        pOggbs = NULL;
+
+        /* The Ogg bistream needs to be layered on top of the original bitstream. */
+        pFlac->bs.onRead = drflac__on_read_ogg;
+        pFlac->bs.onSeek = drflac__on_seek_ogg;
+        pFlac->bs.pUserData = (void*)pInternalOggbs;
+        pFlac->_oggbs = (void*)pInternalOggbs;
+    }
+#endif
+
+    pFlac->firstFLACFramePosInBytes = firstFramePos;
+
+    /* NOTE: Seektables are not currently compatible with Ogg encapsulation (Ogg has its own accelerated seeking system). I may change this later, so I'm leaving this here for now. */
+#ifndef DR_FLAC_NO_OGG
+    if (init.container == drflac_container_ogg)
+    {
+        pFlac->pSeekpoints = NULL;
+        pFlac->seekpointCount = 0;
+    }
+    else
+#endif
+    {
+        /* If we have a seektable we need to load it now, making sure we move back to where we were previously. */
+        if (seektablePos != 0) {
+            pFlac->seekpointCount = seekpointCount;
+            pFlac->pSeekpoints = (drflac_seekpoint*)((drflac_uint8*)pFlac->pDecodedSamples + decodedSamplesAllocationSize);
+
+            DRFLAC_ASSERT(pFlac->bs.onSeek != NULL);
+            DRFLAC_ASSERT(pFlac->bs.onRead != NULL);
+
+            /* Seek to the seektable, then just read directly into our seektable buffer. */
+            if (pFlac->bs.onSeek(pFlac->bs.pUserData, (int)seektablePos, drflac_seek_origin_start)) {
+                drflac_uint32 iSeekpoint;
+
+                for (iSeekpoint = 0; iSeekpoint < seekpointCount; iSeekpoint += 1) {
+                    if (pFlac->bs.onRead(pFlac->bs.pUserData, pFlac->pSeekpoints + iSeekpoint, DRFLAC_SEEKPOINT_SIZE_IN_BYTES) == DRFLAC_SEEKPOINT_SIZE_IN_BYTES) {
+                        /* Endian swap. */
+                        pFlac->pSeekpoints[iSeekpoint].firstPCMFrame   = drflac__be2host_64(pFlac->pSeekpoints[iSeekpoint].firstPCMFrame);
+                        pFlac->pSeekpoints[iSeekpoint].flacFrameOffset = drflac__be2host_64(pFlac->pSeekpoints[iSeekpoint].flacFrameOffset);
+                        pFlac->pSeekpoints[iSeekpoint].pcmFrameCount   = drflac__be2host_16(pFlac->pSeekpoints[iSeekpoint].pcmFrameCount);
+                    } else {
+                        /* Failed to read the seektable. Pretend we don't have one. */
+                        pFlac->pSeekpoints = NULL;
+                        pFlac->seekpointCount = 0;
+                        break;
+                    }
+                }
+
+                /* We need to seek back to where we were. If this fails it's a critical error. */
+                if (!pFlac->bs.onSeek(pFlac->bs.pUserData, (int)pFlac->firstFLACFramePosInBytes, drflac_seek_origin_start)) {
+                    drflac__free_from_callbacks(pFlac, &allocationCallbacks);
+                    return NULL;
+                }
+            } else {
+                /* Failed to seek to the seektable. Ominous sign, but for now we can just pretend we don't have one. */
+                pFlac->pSeekpoints = NULL;
+                pFlac->seekpointCount = 0;
+            }
+        }
+    }
+
+
+    /*
+    If we get here, but don't have a STREAMINFO block, it means we've opened the stream in relaxed mode and need to decode
+    the first frame.
+    */
+    if (!init.hasStreamInfoBlock) {
+        pFlac->currentFLACFrame.header = init.firstFrameHeader;
+        for (;;) {
+            drflac_result result = drflac__decode_flac_frame(pFlac);
+            if (result == DRFLAC_SUCCESS) {
+                break;
+            } else {
+                if (result == DRFLAC_CRC_MISMATCH) {
+                    if (!drflac__read_next_flac_frame_header(&pFlac->bs, pFlac->bitsPerSample, &pFlac->currentFLACFrame.header)) {
+                        drflac__free_from_callbacks(pFlac, &allocationCallbacks);
+                        return NULL;
+                    }
+                    continue;
+                } else {
+                    drflac__free_from_callbacks(pFlac, &allocationCallbacks);
+                    return NULL;
+                }
+            }
+        }
+    }
+
+    return pFlac;
+}
+
+
+
+#ifndef DR_FLAC_NO_STDIO
+#include <stdio.h>
+#ifndef DR_FLAC_NO_WCHAR
+#include <wchar.h>      /* For wcslen(), wcsrtombs() */
+#endif
+
+/* Errno */
+/* drflac_result_from_errno() is only used for fopen() and wfopen() so putting it inside DR_WAV_NO_STDIO for now. If something else needs this later we can move it out. */
+#include <errno.h>
+static drflac_result drflac_result_from_errno(int e)
+{
+    switch (e)
+    {
+        case 0: return DRFLAC_SUCCESS;
+    #ifdef EPERM
+        case EPERM: return DRFLAC_INVALID_OPERATION;
+    #endif
+    #ifdef ENOENT
+        case ENOENT: return DRFLAC_DOES_NOT_EXIST;
+    #endif
+    #ifdef ESRCH
+        case ESRCH: return DRFLAC_DOES_NOT_EXIST;
+    #endif
+    #ifdef EINTR
+        case EINTR: return DRFLAC_INTERRUPT;
+    #endif
+    #ifdef EIO
+        case EIO: return DRFLAC_IO_ERROR;
+    #endif
+    #ifdef ENXIO
+        case ENXIO: return DRFLAC_DOES_NOT_EXIST;
+    #endif
+    #ifdef E2BIG
+        case E2BIG: return DRFLAC_INVALID_ARGS;
+    #endif
+    #ifdef ENOEXEC
+        case ENOEXEC: return DRFLAC_INVALID_FILE;
+    #endif
+    #ifdef EBADF
+        case EBADF: return DRFLAC_INVALID_FILE;
+    #endif
+    #ifdef ECHILD
+        case ECHILD: return DRFLAC_ERROR;
+    #endif
+    #ifdef EAGAIN
+        case EAGAIN: return DRFLAC_UNAVAILABLE;
+    #endif
+    #ifdef ENOMEM
+        case ENOMEM: return DRFLAC_OUT_OF_MEMORY;
+    #endif
+    #ifdef EACCES
+        case EACCES: return DRFLAC_ACCESS_DENIED;
+    #endif
+    #ifdef EFAULT
+        case EFAULT: return DRFLAC_BAD_ADDRESS;
+    #endif
+    #ifdef ENOTBLK
+        case ENOTBLK: return DRFLAC_ERROR;
+    #endif
+    #ifdef EBUSY
+        case EBUSY: return DRFLAC_BUSY;
+    #endif
+    #ifdef EEXIST
+        case EEXIST: return DRFLAC_ALREADY_EXISTS;
+    #endif
+    #ifdef EXDEV
+        case EXDEV: return DRFLAC_ERROR;
+    #endif
+    #ifdef ENODEV
+        case ENODEV: return DRFLAC_DOES_NOT_EXIST;
+    #endif
+    #ifdef ENOTDIR
+        case ENOTDIR: return DRFLAC_NOT_DIRECTORY;
+    #endif
+    #ifdef EISDIR
+        case EISDIR: return DRFLAC_IS_DIRECTORY;
+    #endif
+    #ifdef EINVAL
+        case EINVAL: return DRFLAC_INVALID_ARGS;
+    #endif
+    #ifdef ENFILE
+        case ENFILE: return DRFLAC_TOO_MANY_OPEN_FILES;
+    #endif
+    #ifdef EMFILE
+        case EMFILE: return DRFLAC_TOO_MANY_OPEN_FILES;
+    #endif
+    #ifdef ENOTTY
+        case ENOTTY: return DRFLAC_INVALID_OPERATION;
+    #endif
+    #ifdef ETXTBSY
+        case ETXTBSY: return DRFLAC_BUSY;
+    #endif
+    #ifdef EFBIG
+        case EFBIG: return DRFLAC_TOO_BIG;
+    #endif
+    #ifdef ENOSPC
+        case ENOSPC: return DRFLAC_NO_SPACE;
+    #endif
+    #ifdef ESPIPE
+        case ESPIPE: return DRFLAC_BAD_SEEK;
+    #endif
+    #ifdef EROFS
+        case EROFS: return DRFLAC_ACCESS_DENIED;
+    #endif
+    #ifdef EMLINK
+        case EMLINK: return DRFLAC_TOO_MANY_LINKS;
+    #endif
+    #ifdef EPIPE
+        case EPIPE: return DRFLAC_BAD_PIPE;
+    #endif
+    #ifdef EDOM
+        case EDOM: return DRFLAC_OUT_OF_RANGE;
+    #endif
+    #ifdef ERANGE
+        case ERANGE: return DRFLAC_OUT_OF_RANGE;
+    #endif
+    #ifdef EDEADLK
+        case EDEADLK: return DRFLAC_DEADLOCK;
+    #endif
+    #ifdef ENAMETOOLONG
+        case ENAMETOOLONG: return DRFLAC_PATH_TOO_LONG;
+    #endif
+    #ifdef ENOLCK
+        case ENOLCK: return DRFLAC_ERROR;
+    #endif
+    #ifdef ENOSYS
+        case ENOSYS: return DRFLAC_NOT_IMPLEMENTED;
+    #endif
+    #ifdef ENOTEMPTY
+        case ENOTEMPTY: return DRFLAC_DIRECTORY_NOT_EMPTY;
+    #endif
+    #ifdef ELOOP
+        case ELOOP: return DRFLAC_TOO_MANY_LINKS;
+    #endif
+    #ifdef ENOMSG
+        case ENOMSG: return DRFLAC_NO_MESSAGE;
+    #endif
+    #ifdef EIDRM
+        case EIDRM: return DRFLAC_ERROR;
+    #endif
+    #ifdef ECHRNG
+        case ECHRNG: return DRFLAC_ERROR;
+    #endif
+    #ifdef EL2NSYNC
+        case EL2NSYNC: return DRFLAC_ERROR;
+    #endif
+    #ifdef EL3HLT
+        case EL3HLT: return DRFLAC_ERROR;
+    #endif
+    #ifdef EL3RST
+        case EL3RST: return DRFLAC_ERROR;
+    #endif
+    #ifdef ELNRNG
+        case ELNRNG: return DRFLAC_OUT_OF_RANGE;
+    #endif
+    #ifdef EUNATCH
+        case EUNATCH: return DRFLAC_ERROR;
+    #endif
+    #ifdef ENOCSI
+        case ENOCSI: return DRFLAC_ERROR;
+    #endif
+    #ifdef EL2HLT
+        case EL2HLT: return DRFLAC_ERROR;
+    #endif
+    #ifdef EBADE
+        case EBADE: return DRFLAC_ERROR;
+    #endif
+    #ifdef EBADR
+        case EBADR: return DRFLAC_ERROR;
+    #endif
+    #ifdef EXFULL
+        case EXFULL: return DRFLAC_ERROR;
+    #endif
+    #ifdef ENOANO
+        case ENOANO: return DRFLAC_ERROR;
+    #endif
+    #ifdef EBADRQC
+        case EBADRQC: return DRFLAC_ERROR;
+    #endif
+    #ifdef EBADSLT
+        case EBADSLT: return DRFLAC_ERROR;
+    #endif
+    #ifdef EBFONT
+        case EBFONT: return DRFLAC_INVALID_FILE;
+    #endif
+    #ifdef ENOSTR
+        case ENOSTR: return DRFLAC_ERROR;
+    #endif
+    #ifdef ENODATA
+        case ENODATA: return DRFLAC_NO_DATA_AVAILABLE;
+    #endif
+    #ifdef ETIME
+        case ETIME: return DRFLAC_TIMEOUT;
+    #endif
+    #ifdef ENOSR
+        case ENOSR: return DRFLAC_NO_DATA_AVAILABLE;
+    #endif
+    #ifdef ENONET
+        case ENONET: return DRFLAC_NO_NETWORK;
+    #endif
+    #ifdef ENOPKG
+        case ENOPKG: return DRFLAC_ERROR;
+    #endif
+    #ifdef EREMOTE
+        case EREMOTE: return DRFLAC_ERROR;
+    #endif
+    #ifdef ENOLINK
+        case ENOLINK: return DRFLAC_ERROR;
+    #endif
+    #ifdef EADV
+        case EADV: return DRFLAC_ERROR;
+    #endif
+    #ifdef ESRMNT
+        case ESRMNT: return DRFLAC_ERROR;
+    #endif
+    #ifdef ECOMM
+        case ECOMM: return DRFLAC_ERROR;
+    #endif
+    #ifdef EPROTO
+        case EPROTO: return DRFLAC_ERROR;
+    #endif
+    #ifdef EMULTIHOP
+        case EMULTIHOP: return DRFLAC_ERROR;
+    #endif
+    #ifdef EDOTDOT
+        case EDOTDOT: return DRFLAC_ERROR;
+    #endif
+    #ifdef EBADMSG
+        case EBADMSG: return DRFLAC_BAD_MESSAGE;
+    #endif
+    #ifdef EOVERFLOW
+        case EOVERFLOW: return DRFLAC_TOO_BIG;
+    #endif
+    #ifdef ENOTUNIQ
+        case ENOTUNIQ: return DRFLAC_NOT_UNIQUE;
+    #endif
+    #ifdef EBADFD
+        case EBADFD: return DRFLAC_ERROR;
+    #endif
+    #ifdef EREMCHG
+        case EREMCHG: return DRFLAC_ERROR;
+    #endif
+    #ifdef ELIBACC
+        case ELIBACC: return DRFLAC_ACCESS_DENIED;
+    #endif
+    #ifdef ELIBBAD
+        case ELIBBAD: return DRFLAC_INVALID_FILE;
+    #endif
+    #ifdef ELIBSCN
+        case ELIBSCN: return DRFLAC_INVALID_FILE;
+    #endif
+    #ifdef ELIBMAX
+        case ELIBMAX: return DRFLAC_ERROR;
+    #endif
+    #ifdef ELIBEXEC
+        case ELIBEXEC: return DRFLAC_ERROR;
+    #endif
+    #ifdef EILSEQ
+        case EILSEQ: return DRFLAC_INVALID_DATA;
+    #endif
+    #ifdef ERESTART
+        case ERESTART: return DRFLAC_ERROR;
+    #endif
+    #ifdef ESTRPIPE
+        case ESTRPIPE: return DRFLAC_ERROR;
+    #endif
+    #ifdef EUSERS
+        case EUSERS: return DRFLAC_ERROR;
+    #endif
+    #ifdef ENOTSOCK
+        case ENOTSOCK: return DRFLAC_NOT_SOCKET;
+    #endif
+    #ifdef EDESTADDRREQ
+        case EDESTADDRREQ: return DRFLAC_NO_ADDRESS;
+    #endif
+    #ifdef EMSGSIZE
+        case EMSGSIZE: return DRFLAC_TOO_BIG;
+    #endif
+    #ifdef EPROTOTYPE
+        case EPROTOTYPE: return DRFLAC_BAD_PROTOCOL;
+    #endif
+    #ifdef ENOPROTOOPT
+        case ENOPROTOOPT: return DRFLAC_PROTOCOL_UNAVAILABLE;
+    #endif
+    #ifdef EPROTONOSUPPORT
+        case EPROTONOSUPPORT: return DRFLAC_PROTOCOL_NOT_SUPPORTED;
+    #endif
+    #ifdef ESOCKTNOSUPPORT
+        case ESOCKTNOSUPPORT: return DRFLAC_SOCKET_NOT_SUPPORTED;
+    #endif
+    #ifdef EOPNOTSUPP
+        case EOPNOTSUPP: return DRFLAC_INVALID_OPERATION;
+    #endif
+    #ifdef EPFNOSUPPORT
+        case EPFNOSUPPORT: return DRFLAC_PROTOCOL_FAMILY_NOT_SUPPORTED;
+    #endif
+    #ifdef EAFNOSUPPORT
+        case EAFNOSUPPORT: return DRFLAC_ADDRESS_FAMILY_NOT_SUPPORTED;
+    #endif
+    #ifdef EADDRINUSE
+        case EADDRINUSE: return DRFLAC_ALREADY_IN_USE;
+    #endif
+    #ifdef EADDRNOTAVAIL
+        case EADDRNOTAVAIL: return DRFLAC_ERROR;
+    #endif
+    #ifdef ENETDOWN
+        case ENETDOWN: return DRFLAC_NO_NETWORK;
+    #endif
+    #ifdef ENETUNREACH
+        case ENETUNREACH: return DRFLAC_NO_NETWORK;
+    #endif
+    #ifdef ENETRESET
+        case ENETRESET: return DRFLAC_NO_NETWORK;
+    #endif
+    #ifdef ECONNABORTED
+        case ECONNABORTED: return DRFLAC_NO_NETWORK;
+    #endif
+    #ifdef ECONNRESET
+        case ECONNRESET: return DRFLAC_CONNECTION_RESET;
+    #endif
+    #ifdef ENOBUFS
+        case ENOBUFS: return DRFLAC_NO_SPACE;
+    #endif
+    #ifdef EISCONN
+        case EISCONN: return DRFLAC_ALREADY_CONNECTED;
+    #endif
+    #ifdef ENOTCONN
+        case ENOTCONN: return DRFLAC_NOT_CONNECTED;
+    #endif
+    #ifdef ESHUTDOWN
+        case ESHUTDOWN: return DRFLAC_ERROR;
+    #endif
+    #ifdef ETOOMANYREFS
+        case ETOOMANYREFS: return DRFLAC_ERROR;
+    #endif
+    #ifdef ETIMEDOUT
+        case ETIMEDOUT: return DRFLAC_TIMEOUT;
+    #endif
+    #ifdef ECONNREFUSED
+        case ECONNREFUSED: return DRFLAC_CONNECTION_REFUSED;
+    #endif
+    #ifdef EHOSTDOWN
+        case EHOSTDOWN: return DRFLAC_NO_HOST;
+    #endif
+    #ifdef EHOSTUNREACH
+        case EHOSTUNREACH: return DRFLAC_NO_HOST;
+    #endif
+    #ifdef EALREADY
+        case EALREADY: return DRFLAC_IN_PROGRESS;
+    #endif
+    #ifdef EINPROGRESS
+        case EINPROGRESS: return DRFLAC_IN_PROGRESS;
+    #endif
+    #ifdef ESTALE
+        case ESTALE: return DRFLAC_INVALID_FILE;
+    #endif
+    #ifdef EUCLEAN
+        case EUCLEAN: return DRFLAC_ERROR;
+    #endif
+    #ifdef ENOTNAM
+        case ENOTNAM: return DRFLAC_ERROR;
+    #endif
+    #ifdef ENAVAIL
+        case ENAVAIL: return DRFLAC_ERROR;
+    #endif
+    #ifdef EISNAM
+        case EISNAM: return DRFLAC_ERROR;
+    #endif
+    #ifdef EREMOTEIO
+        case EREMOTEIO: return DRFLAC_IO_ERROR;
+    #endif
+    #ifdef EDQUOT
+        case EDQUOT: return DRFLAC_NO_SPACE;
+    #endif
+    #ifdef ENOMEDIUM
+        case ENOMEDIUM: return DRFLAC_DOES_NOT_EXIST;
+    #endif
+    #ifdef EMEDIUMTYPE
+        case EMEDIUMTYPE: return DRFLAC_ERROR;
+    #endif
+    #ifdef ECANCELED
+        case ECANCELED: return DRFLAC_CANCELLED;
+    #endif
+    #ifdef ENOKEY
+        case ENOKEY: return DRFLAC_ERROR;
+    #endif
+    #ifdef EKEYEXPIRED
+        case EKEYEXPIRED: return DRFLAC_ERROR;
+    #endif
+    #ifdef EKEYREVOKED
+        case EKEYREVOKED: return DRFLAC_ERROR;
+    #endif
+    #ifdef EKEYREJECTED
+        case EKEYREJECTED: return DRFLAC_ERROR;
+    #endif
+    #ifdef EOWNERDEAD
+        case EOWNERDEAD: return DRFLAC_ERROR;
+    #endif
+    #ifdef ENOTRECOVERABLE
+        case ENOTRECOVERABLE: return DRFLAC_ERROR;
+    #endif
+    #ifdef ERFKILL
+        case ERFKILL: return DRFLAC_ERROR;
+    #endif
+    #ifdef EHWPOISON
+        case EHWPOISON: return DRFLAC_ERROR;
+    #endif
+        default: return DRFLAC_ERROR;
+    }
+}
+/* End Errno */
+
+/* fopen */
+static drflac_result drflac_fopen(FILE** ppFile, const char* pFilePath, const char* pOpenMode)
+{
+#if defined(_MSC_VER) && _MSC_VER >= 1400
+    errno_t err;
+#endif
+
+    if (ppFile != NULL) {
+        *ppFile = NULL;  /* Safety. */
+    }
+
+    if (pFilePath == NULL || pOpenMode == NULL || ppFile == NULL) {
+        return DRFLAC_INVALID_ARGS;
+    }
+
+#if defined(_MSC_VER) && _MSC_VER >= 1400
+    err = fopen_s(ppFile, pFilePath, pOpenMode);
+    if (err != 0) {
+        return drflac_result_from_errno(err);
+    }
+#else
+#if defined(_WIN32) || defined(__APPLE__)
+    *ppFile = fopen(pFilePath, pOpenMode);
+#else
+    #if defined(_FILE_OFFSET_BITS) && _FILE_OFFSET_BITS == 64 && defined(_LARGEFILE64_SOURCE)
+        *ppFile = fopen64(pFilePath, pOpenMode);
+    #else
+        *ppFile = fopen(pFilePath, pOpenMode);
+    #endif
+#endif
+    if (*ppFile == NULL) {
+        drflac_result result = drflac_result_from_errno(errno);
+        if (result == DRFLAC_SUCCESS) {
+            result = DRFLAC_ERROR;   /* Just a safety check to make sure we never ever return success when pFile == NULL. */
+        }
+
+        return result;
+    }
+#endif
+
+    return DRFLAC_SUCCESS;
+}
+
+/*
+_wfopen() isn't always available in all compilation environments.
+
+    * Windows only.
+    * MSVC seems to support it universally as far back as VC6 from what I can tell (haven't checked further back).
+    * MinGW-64 (both 32- and 64-bit) seems to support it.
+    * MinGW wraps it in !defined(__STRICT_ANSI__).
+    * OpenWatcom wraps it in !defined(_NO_EXT_KEYS).
+
+This can be reviewed as compatibility issues arise. The preference is to use _wfopen_s() and _wfopen() as opposed to the wcsrtombs()
+fallback, so if you notice your compiler not detecting this properly I'm happy to look at adding support.
+*/
+#if defined(_WIN32)
+    #if defined(_MSC_VER) || defined(__MINGW64__) || (!defined(__STRICT_ANSI__) && !defined(_NO_EXT_KEYS))
+        #define DRFLAC_HAS_WFOPEN
+    #endif
+#endif
+
+#ifndef DR_FLAC_NO_WCHAR
+static drflac_result drflac_wfopen(FILE** ppFile, const wchar_t* pFilePath, const wchar_t* pOpenMode, const drflac_allocation_callbacks* pAllocationCallbacks)
+{
+    if (ppFile != NULL) {
+        *ppFile = NULL;  /* Safety. */
+    }
+
+    if (pFilePath == NULL || pOpenMode == NULL || ppFile == NULL) {
+        return DRFLAC_INVALID_ARGS;
+    }
+
+#if defined(DRFLAC_HAS_WFOPEN)
+    {
+        /* Use _wfopen() on Windows. */
+    #if defined(_MSC_VER) && _MSC_VER >= 1400
+        errno_t err = _wfopen_s(ppFile, pFilePath, pOpenMode);
+        if (err != 0) {
+            return drflac_result_from_errno(err);
+        }
+    #else
+        *ppFile = _wfopen(pFilePath, pOpenMode);
+        if (*ppFile == NULL) {
+            return drflac_result_from_errno(errno);
+        }
+    #endif
+        (void)pAllocationCallbacks;
+    }
+#else
+    /*
+    Use fopen() on anything other than Windows. Requires a conversion. This is annoying because
+	fopen() is locale specific. The only real way I can think of to do this is with wcsrtombs(). Note
+	that wcstombs() is apparently not thread-safe because it uses a static global mbstate_t object for
+    maintaining state. I've checked this with -std=c89 and it works, but if somebody get's a compiler
+	error I'll look into improving compatibility.
+    */
+
+	/*
+	Some compilers don't support wchar_t or wcsrtombs() which we're using below. In this case we just
+	need to abort with an error. If you encounter a compiler lacking such support, add it to this list
+	and submit a bug report and it'll be added to the library upstream.
+	*/
+	#if defined(__DJGPP__)
+	{
+		/* Nothing to do here. This will fall through to the error check below. */
+	}
+	#else
+    {
+        mbstate_t mbs;
+        size_t lenMB;
+        const wchar_t* pFilePathTemp = pFilePath;
+        char* pFilePathMB = NULL;
+        char pOpenModeMB[32] = {0};
+
+        /* Get the length first. */
+        DRFLAC_ZERO_OBJECT(&mbs);
+        lenMB = wcsrtombs(NULL, &pFilePathTemp, 0, &mbs);
+        if (lenMB == (size_t)-1) {
+            return drflac_result_from_errno(errno);
+        }
+
+        pFilePathMB = (char*)drflac__malloc_from_callbacks(lenMB + 1, pAllocationCallbacks);
+        if (pFilePathMB == NULL) {
+            return DRFLAC_OUT_OF_MEMORY;
+        }
+
+        pFilePathTemp = pFilePath;
+        DRFLAC_ZERO_OBJECT(&mbs);
+        wcsrtombs(pFilePathMB, &pFilePathTemp, lenMB + 1, &mbs);
+
+        /* The open mode should always consist of ASCII characters so we should be able to do a trivial conversion. */
+        {
+            size_t i = 0;
+            for (;;) {
+                if (pOpenMode[i] == 0) {
+                    pOpenModeMB[i] = '\0';
+                    break;
+                }
+
+                pOpenModeMB[i] = (char)pOpenMode[i];
+                i += 1;
+            }
+        }
+
+        *ppFile = fopen(pFilePathMB, pOpenModeMB);
+
+        drflac__free_from_callbacks(pFilePathMB, pAllocationCallbacks);
+    }
+	#endif
+
+    if (*ppFile == NULL) {
+        return DRFLAC_ERROR;
+    }
+#endif
+
+    return DRFLAC_SUCCESS;
+}
+#endif
+/* End fopen */
+
+static size_t drflac__on_read_stdio(void* pUserData, void* bufferOut, size_t bytesToRead)
+{
+    return fread(bufferOut, 1, bytesToRead, (FILE*)pUserData);
+}
+
+static drflac_bool32 drflac__on_seek_stdio(void* pUserData, int offset, drflac_seek_origin origin)
+{
+    DRFLAC_ASSERT(offset >= 0);  /* <-- Never seek backwards. */
+
+    return fseek((FILE*)pUserData, offset, (origin == drflac_seek_origin_current) ? SEEK_CUR : SEEK_SET) == 0;
+}
+
+
+DRFLAC_API drflac* drflac_open_file(const char* pFileName, const drflac_allocation_callbacks* pAllocationCallbacks)
+{
+    drflac* pFlac;
+    FILE* pFile;
+
+    if (drflac_fopen(&pFile, pFileName, "rb") != DRFLAC_SUCCESS) {
+        return NULL;
+    }
+
+    pFlac = drflac_open(drflac__on_read_stdio, drflac__on_seek_stdio, (void*)pFile, pAllocationCallbacks);
+    if (pFlac == NULL) {
+        fclose(pFile);
+        return NULL;
+    }
+
+    return pFlac;
+}
+
+#ifndef DR_FLAC_NO_WCHAR
+DRFLAC_API drflac* drflac_open_file_w(const wchar_t* pFileName, const drflac_allocation_callbacks* pAllocationCallbacks)
+{
+    drflac* pFlac;
+    FILE* pFile;
+
+    if (drflac_wfopen(&pFile, pFileName, L"rb", pAllocationCallbacks) != DRFLAC_SUCCESS) {
+        return NULL;
+    }
+
+    pFlac = drflac_open(drflac__on_read_stdio, drflac__on_seek_stdio, (void*)pFile, pAllocationCallbacks);
+    if (pFlac == NULL) {
+        fclose(pFile);
+        return NULL;
+    }
+
+    return pFlac;
+}
+#endif
+
+DRFLAC_API drflac* drflac_open_file_with_metadata(const char* pFileName, drflac_meta_proc onMeta, void* pUserData, const drflac_allocation_callbacks* pAllocationCallbacks)
+{
+    drflac* pFlac;
+    FILE* pFile;
+
+    if (drflac_fopen(&pFile, pFileName, "rb") != DRFLAC_SUCCESS) {
+        return NULL;
+    }
+
+    pFlac = drflac_open_with_metadata_private(drflac__on_read_stdio, drflac__on_seek_stdio, onMeta, drflac_container_unknown, (void*)pFile, pUserData, pAllocationCallbacks);
+    if (pFlac == NULL) {
+        fclose(pFile);
+        return pFlac;
+    }
+
+    return pFlac;
+}
+
+#ifndef DR_FLAC_NO_WCHAR
+DRFLAC_API drflac* drflac_open_file_with_metadata_w(const wchar_t* pFileName, drflac_meta_proc onMeta, void* pUserData, const drflac_allocation_callbacks* pAllocationCallbacks)
+{
+    drflac* pFlac;
+    FILE* pFile;
+
+    if (drflac_wfopen(&pFile, pFileName, L"rb", pAllocationCallbacks) != DRFLAC_SUCCESS) {
+        return NULL;
+    }
+
+    pFlac = drflac_open_with_metadata_private(drflac__on_read_stdio, drflac__on_seek_stdio, onMeta, drflac_container_unknown, (void*)pFile, pUserData, pAllocationCallbacks);
+    if (pFlac == NULL) {
+        fclose(pFile);
+        return pFlac;
+    }
+
+    return pFlac;
+}
+#endif
+#endif  /* DR_FLAC_NO_STDIO */
+
+static size_t drflac__on_read_memory(void* pUserData, void* bufferOut, size_t bytesToRead)
+{
+    drflac__memory_stream* memoryStream = (drflac__memory_stream*)pUserData;
+    size_t bytesRemaining;
+
+    DRFLAC_ASSERT(memoryStream != NULL);
+    DRFLAC_ASSERT(memoryStream->dataSize >= memoryStream->currentReadPos);
+
+    bytesRemaining = memoryStream->dataSize - memoryStream->currentReadPos;
+    if (bytesToRead > bytesRemaining) {
+        bytesToRead = bytesRemaining;
+    }
+
+    if (bytesToRead > 0) {
+        DRFLAC_COPY_MEMORY(bufferOut, memoryStream->data + memoryStream->currentReadPos, bytesToRead);
+        memoryStream->currentReadPos += bytesToRead;
+    }
+
+    return bytesToRead;
+}
+
+static drflac_bool32 drflac__on_seek_memory(void* pUserData, int offset, drflac_seek_origin origin)
+{
+    drflac__memory_stream* memoryStream = (drflac__memory_stream*)pUserData;
+
+    DRFLAC_ASSERT(memoryStream != NULL);
+    DRFLAC_ASSERT(offset >= 0); /* <-- Never seek backwards. */
+
+    if (offset > (drflac_int64)memoryStream->dataSize) {
+        return DRFLAC_FALSE;
+    }
+
+    if (origin == drflac_seek_origin_current) {
+        if (memoryStream->currentReadPos + offset <= memoryStream->dataSize) {
+            memoryStream->currentReadPos += offset;
+        } else {
+            return DRFLAC_FALSE;  /* Trying to seek too far forward. */
+        }
+    } else {
+        if ((drflac_uint32)offset <= memoryStream->dataSize) {
+            memoryStream->currentReadPos = offset;
+        } else {
+            return DRFLAC_FALSE;  /* Trying to seek too far forward. */
+        }
+    }
+
+    return DRFLAC_TRUE;
+}
+
+DRFLAC_API drflac* drflac_open_memory(const void* pData, size_t dataSize, const drflac_allocation_callbacks* pAllocationCallbacks)
+{
+    drflac__memory_stream memoryStream;
+    drflac* pFlac;
+
+    memoryStream.data = (const drflac_uint8*)pData;
+    memoryStream.dataSize = dataSize;
+    memoryStream.currentReadPos = 0;
+    pFlac = drflac_open(drflac__on_read_memory, drflac__on_seek_memory, &memoryStream, pAllocationCallbacks);
+    if (pFlac == NULL) {
+        return NULL;
+    }
+
+    pFlac->memoryStream = memoryStream;
+
+    /* This is an awful hack... */
+#ifndef DR_FLAC_NO_OGG
+    if (pFlac->container == drflac_container_ogg)
+    {
+        drflac_oggbs* oggbs = (drflac_oggbs*)pFlac->_oggbs;
+        oggbs->pUserData = &pFlac->memoryStream;
+    }
+    else
+#endif
+    {
+        pFlac->bs.pUserData = &pFlac->memoryStream;
+    }
+
+    return pFlac;
+}
+
+DRFLAC_API drflac* drflac_open_memory_with_metadata(const void* pData, size_t dataSize, drflac_meta_proc onMeta, void* pUserData, const drflac_allocation_callbacks* pAllocationCallbacks)
+{
+    drflac__memory_stream memoryStream;
+    drflac* pFlac;
+
+    memoryStream.data = (const drflac_uint8*)pData;
+    memoryStream.dataSize = dataSize;
+    memoryStream.currentReadPos = 0;
+    pFlac = drflac_open_with_metadata_private(drflac__on_read_memory, drflac__on_seek_memory, onMeta, drflac_container_unknown, &memoryStream, pUserData, pAllocationCallbacks);
+    if (pFlac == NULL) {
+        return NULL;
+    }
+
+    pFlac->memoryStream = memoryStream;
+
+    /* This is an awful hack... */
+#ifndef DR_FLAC_NO_OGG
+    if (pFlac->container == drflac_container_ogg)
+    {
+        drflac_oggbs* oggbs = (drflac_oggbs*)pFlac->_oggbs;
+        oggbs->pUserData = &pFlac->memoryStream;
+    }
+    else
+#endif
+    {
+        pFlac->bs.pUserData = &pFlac->memoryStream;
+    }
+
+    return pFlac;
+}
+
+
+
+DRFLAC_API drflac* drflac_open(drflac_read_proc onRead, drflac_seek_proc onSeek, void* pUserData, const drflac_allocation_callbacks* pAllocationCallbacks)
+{
+    return drflac_open_with_metadata_private(onRead, onSeek, NULL, drflac_container_unknown, pUserData, pUserData, pAllocationCallbacks);
+}
+DRFLAC_API drflac* drflac_open_relaxed(drflac_read_proc onRead, drflac_seek_proc onSeek, drflac_container container, void* pUserData, const drflac_allocation_callbacks* pAllocationCallbacks)
+{
+    return drflac_open_with_metadata_private(onRead, onSeek, NULL, container, pUserData, pUserData, pAllocationCallbacks);
+}
+
+DRFLAC_API drflac* drflac_open_with_metadata(drflac_read_proc onRead, drflac_seek_proc onSeek, drflac_meta_proc onMeta, void* pUserData, const drflac_allocation_callbacks* pAllocationCallbacks)
+{
+    return drflac_open_with_metadata_private(onRead, onSeek, onMeta, drflac_container_unknown, pUserData, pUserData, pAllocationCallbacks);
+}
+DRFLAC_API drflac* drflac_open_with_metadata_relaxed(drflac_read_proc onRead, drflac_seek_proc onSeek, drflac_meta_proc onMeta, drflac_container container, void* pUserData, const drflac_allocation_callbacks* pAllocationCallbacks)
+{
+    return drflac_open_with_metadata_private(onRead, onSeek, onMeta, container, pUserData, pUserData, pAllocationCallbacks);
+}
+
+DRFLAC_API void drflac_close(drflac* pFlac)
+{
+    if (pFlac == NULL) {
+        return;
+    }
+
+#ifndef DR_FLAC_NO_STDIO
+    /*
+    If we opened the file with drflac_open_file() we will want to close the file handle. We can know whether or not drflac_open_file()
+    was used by looking at the callbacks.
+    */
+    if (pFlac->bs.onRead == drflac__on_read_stdio) {
+        fclose((FILE*)pFlac->bs.pUserData);
+    }
+
+#ifndef DR_FLAC_NO_OGG
+    /* Need to clean up Ogg streams a bit differently due to the way the bit streaming is chained. */
+    if (pFlac->container == drflac_container_ogg) {
+        drflac_oggbs* oggbs = (drflac_oggbs*)pFlac->_oggbs;
+        DRFLAC_ASSERT(pFlac->bs.onRead == drflac__on_read_ogg);
+
+        if (oggbs->onRead == drflac__on_read_stdio) {
+            fclose((FILE*)oggbs->pUserData);
+        }
+    }
+#endif
+#endif
+
+    drflac__free_from_callbacks(pFlac, &pFlac->allocationCallbacks);
+}
+
+
+#if 0
+static DRFLAC_INLINE void drflac_read_pcm_frames_s32__decode_left_side__reference(drflac* pFlac, drflac_uint64 frameCount, drflac_uint32 unusedBitsPerSample, const drflac_int32* pInputSamples0, const drflac_int32* pInputSamples1, drflac_int32* pOutputSamples)
+{
+    drflac_uint64 i;
+    for (i = 0; i < frameCount; ++i) {
+        drflac_uint32 left  = (drflac_uint32)pInputSamples0[i] << (unusedBitsPerSample + pFlac->currentFLACFrame.subframes[0].wastedBitsPerSample);
+        drflac_uint32 side  = (drflac_uint32)pInputSamples1[i] << (unusedBitsPerSample + pFlac->currentFLACFrame.subframes[1].wastedBitsPerSample);
+        drflac_uint32 right = left - side;
+
+        pOutputSamples[i*2+0] = (drflac_int32)left;
+        pOutputSamples[i*2+1] = (drflac_int32)right;
+    }
+}
+#endif
+
+static DRFLAC_INLINE void drflac_read_pcm_frames_s32__decode_left_side__scalar(drflac* pFlac, drflac_uint64 frameCount, drflac_uint32 unusedBitsPerSample, const drflac_int32* pInputSamples0, const drflac_int32* pInputSamples1, drflac_int32* pOutputSamples)
+{
+    drflac_uint64 i;
+    drflac_uint64 frameCount4 = frameCount >> 2;
+    const drflac_uint32* pInputSamples0U32 = (const drflac_uint32*)pInputSamples0;
+    const drflac_uint32* pInputSamples1U32 = (const drflac_uint32*)pInputSamples1;
+    drflac_uint32 shift0 = unusedBitsPerSample + pFlac->currentFLACFrame.subframes[0].wastedBitsPerSample;
+    drflac_uint32 shift1 = unusedBitsPerSample + pFlac->currentFLACFrame.subframes[1].wastedBitsPerSample;
+
+    for (i = 0; i < frameCount4; ++i) {
+        drflac_uint32 left0 = pInputSamples0U32[i*4+0] << shift0;
+        drflac_uint32 left1 = pInputSamples0U32[i*4+1] << shift0;
+        drflac_uint32 left2 = pInputSamples0U32[i*4+2] << shift0;
+        drflac_uint32 left3 = pInputSamples0U32[i*4+3] << shift0;
+
+        drflac_uint32 side0 = pInputSamples1U32[i*4+0] << shift1;
+        drflac_uint32 side1 = pInputSamples1U32[i*4+1] << shift1;
+        drflac_uint32 side2 = pInputSamples1U32[i*4+2] << shift1;
+        drflac_uint32 side3 = pInputSamples1U32[i*4+3] << shift1;
+
+        drflac_uint32 right0 = left0 - side0;
+        drflac_uint32 right1 = left1 - side1;
+        drflac_uint32 right2 = left2 - side2;
+        drflac_uint32 right3 = left3 - side3;
+
+        pOutputSamples[i*8+0] = (drflac_int32)left0;
+        pOutputSamples[i*8+1] = (drflac_int32)right0;
+        pOutputSamples[i*8+2] = (drflac_int32)left1;
+        pOutputSamples[i*8+3] = (drflac_int32)right1;
+        pOutputSamples[i*8+4] = (drflac_int32)left2;
+        pOutputSamples[i*8+5] = (drflac_int32)right2;
+        pOutputSamples[i*8+6] = (drflac_int32)left3;
+        pOutputSamples[i*8+7] = (drflac_int32)right3;
+    }
+
+    for (i = (frameCount4 << 2); i < frameCount; ++i) {
+        drflac_uint32 left  = pInputSamples0U32[i] << shift0;
+        drflac_uint32 side  = pInputSamples1U32[i] << shift1;
+        drflac_uint32 right = left - side;
+
+        pOutputSamples[i*2+0] = (drflac_int32)left;
+        pOutputSamples[i*2+1] = (drflac_int32)right;
+    }
+}
+
+#if defined(DRFLAC_SUPPORT_SSE2)
+static DRFLAC_INLINE void drflac_read_pcm_frames_s32__decode_left_side__sse2(drflac* pFlac, drflac_uint64 frameCount, drflac_uint32 unusedBitsPerSample, const drflac_int32* pInputSamples0, const drflac_int32* pInputSamples1, drflac_int32* pOutputSamples)
+{
+    drflac_uint64 i;
+    drflac_uint64 frameCount4 = frameCount >> 2;
+    const drflac_uint32* pInputSamples0U32 = (const drflac_uint32*)pInputSamples0;
+    const drflac_uint32* pInputSamples1U32 = (const drflac_uint32*)pInputSamples1;
+    drflac_uint32 shift0 = unusedBitsPerSample + pFlac->currentFLACFrame.subframes[0].wastedBitsPerSample;
+    drflac_uint32 shift1 = unusedBitsPerSample + pFlac->currentFLACFrame.subframes[1].wastedBitsPerSample;
+
+    DRFLAC_ASSERT(pFlac->bitsPerSample <= 24);
+
+    for (i = 0; i < frameCount4; ++i) {
+        __m128i left  = _mm_slli_epi32(_mm_loadu_si128((const __m128i*)pInputSamples0 + i), shift0);
+        __m128i side  = _mm_slli_epi32(_mm_loadu_si128((const __m128i*)pInputSamples1 + i), shift1);
+        __m128i right = _mm_sub_epi32(left, side);
+
+        _mm_storeu_si128((__m128i*)(pOutputSamples + i*8 + 0), _mm_unpacklo_epi32(left, right));
+        _mm_storeu_si128((__m128i*)(pOutputSamples + i*8 + 4), _mm_unpackhi_epi32(left, right));
+    }
+
+    for (i = (frameCount4 << 2); i < frameCount; ++i) {
+        drflac_uint32 left  = pInputSamples0U32[i] << shift0;
+        drflac_uint32 side  = pInputSamples1U32[i] << shift1;
+        drflac_uint32 right = left - side;
+
+        pOutputSamples[i*2+0] = (drflac_int32)left;
+        pOutputSamples[i*2+1] = (drflac_int32)right;
+    }
+}
+#endif
+
+#if defined(DRFLAC_SUPPORT_NEON)
+static DRFLAC_INLINE void drflac_read_pcm_frames_s32__decode_left_side__neon(drflac* pFlac, drflac_uint64 frameCount, drflac_uint32 unusedBitsPerSample, const drflac_int32* pInputSamples0, const drflac_int32* pInputSamples1, drflac_int32* pOutputSamples)
+{
+    drflac_uint64 i;
+    drflac_uint64 frameCount4 = frameCount >> 2;
+    const drflac_uint32* pInputSamples0U32 = (const drflac_uint32*)pInputSamples0;
+    const drflac_uint32* pInputSamples1U32 = (const drflac_uint32*)pInputSamples1;
+    drflac_uint32 shift0 = unusedBitsPerSample + pFlac->currentFLACFrame.subframes[0].wastedBitsPerSample;
+    drflac_uint32 shift1 = unusedBitsPerSample + pFlac->currentFLACFrame.subframes[1].wastedBitsPerSample;
+    int32x4_t shift0_4;
+    int32x4_t shift1_4;
+
+    DRFLAC_ASSERT(pFlac->bitsPerSample <= 24);
+
+    shift0_4 = vdupq_n_s32(shift0);
+    shift1_4 = vdupq_n_s32(shift1);
+
+    for (i = 0; i < frameCount4; ++i) {
+        uint32x4_t left;
+        uint32x4_t side;
+        uint32x4_t right;
+
+        left  = vshlq_u32(vld1q_u32(pInputSamples0U32 + i*4), shift0_4);
+        side  = vshlq_u32(vld1q_u32(pInputSamples1U32 + i*4), shift1_4);
+        right = vsubq_u32(left, side);
+
+        drflac__vst2q_u32((drflac_uint32*)pOutputSamples + i*8, vzipq_u32(left, right));
+    }
+
+    for (i = (frameCount4 << 2); i < frameCount; ++i) {
+        drflac_uint32 left  = pInputSamples0U32[i] << shift0;
+        drflac_uint32 side  = pInputSamples1U32[i] << shift1;
+        drflac_uint32 right = left - side;
+
+        pOutputSamples[i*2+0] = (drflac_int32)left;
+        pOutputSamples[i*2+1] = (drflac_int32)right;
+    }
+}
+#endif
+
+static DRFLAC_INLINE void drflac_read_pcm_frames_s32__decode_left_side(drflac* pFlac, drflac_uint64 frameCount, drflac_uint32 unusedBitsPerSample, const drflac_int32* pInputSamples0, const drflac_int32* pInputSamples1, drflac_int32* pOutputSamples)
+{
+#if defined(DRFLAC_SUPPORT_SSE2)
+    if (drflac__gIsSSE2Supported && pFlac->bitsPerSample <= 24) {
+        drflac_read_pcm_frames_s32__decode_left_side__sse2(pFlac, frameCount, unusedBitsPerSample, pInputSamples0, pInputSamples1, pOutputSamples);
+    } else
+#elif defined(DRFLAC_SUPPORT_NEON)
+    if (drflac__gIsNEONSupported && pFlac->bitsPerSample <= 24) {
+        drflac_read_pcm_frames_s32__decode_left_side__neon(pFlac, frameCount, unusedBitsPerSample, pInputSamples0, pInputSamples1, pOutputSamples);
+    } else
+#endif
+    {
+        /* Scalar fallback. */
+#if 0
+        drflac_read_pcm_frames_s32__decode_left_side__reference(pFlac, frameCount, unusedBitsPerSample, pInputSamples0, pInputSamples1, pOutputSamples);
+#else
+        drflac_read_pcm_frames_s32__decode_left_side__scalar(pFlac, frameCount, unusedBitsPerSample, pInputSamples0, pInputSamples1, pOutputSamples);
+#endif
+    }
+}
+
+
+#if 0
+static DRFLAC_INLINE void drflac_read_pcm_frames_s32__decode_right_side__reference(drflac* pFlac, drflac_uint64 frameCount, drflac_uint32 unusedBitsPerSample, const drflac_int32* pInputSamples0, const drflac_int32* pInputSamples1, drflac_int32* pOutputSamples)
+{
+    drflac_uint64 i;
+    for (i = 0; i < frameCount; ++i) {
+        drflac_uint32 side  = (drflac_uint32)pInputSamples0[i] << (unusedBitsPerSample + pFlac->currentFLACFrame.subframes[0].wastedBitsPerSample);
+        drflac_uint32 right = (drflac_uint32)pInputSamples1[i] << (unusedBitsPerSample + pFlac->currentFLACFrame.subframes[1].wastedBitsPerSample);
+        drflac_uint32 left  = right + side;
+
+        pOutputSamples[i*2+0] = (drflac_int32)left;
+        pOutputSamples[i*2+1] = (drflac_int32)right;
+    }
+}
+#endif
+
+static DRFLAC_INLINE void drflac_read_pcm_frames_s32__decode_right_side__scalar(drflac* pFlac, drflac_uint64 frameCount, drflac_uint32 unusedBitsPerSample, const drflac_int32* pInputSamples0, const drflac_int32* pInputSamples1, drflac_int32* pOutputSamples)
+{
+    drflac_uint64 i;
+    drflac_uint64 frameCount4 = frameCount >> 2;
+    const drflac_uint32* pInputSamples0U32 = (const drflac_uint32*)pInputSamples0;
+    const drflac_uint32* pInputSamples1U32 = (const drflac_uint32*)pInputSamples1;
+    drflac_uint32 shift0 = unusedBitsPerSample + pFlac->currentFLACFrame.subframes[0].wastedBitsPerSample;
+    drflac_uint32 shift1 = unusedBitsPerSample + pFlac->currentFLACFrame.subframes[1].wastedBitsPerSample;
+
+    for (i = 0; i < frameCount4; ++i) {
+        drflac_uint32 side0  = pInputSamples0U32[i*4+0] << shift0;
+        drflac_uint32 side1  = pInputSamples0U32[i*4+1] << shift0;
+        drflac_uint32 side2  = pInputSamples0U32[i*4+2] << shift0;
+        drflac_uint32 side3  = pInputSamples0U32[i*4+3] << shift0;
+
+        drflac_uint32 right0 = pInputSamples1U32[i*4+0] << shift1;
+        drflac_uint32 right1 = pInputSamples1U32[i*4+1] << shift1;
+        drflac_uint32 right2 = pInputSamples1U32[i*4+2] << shift1;
+        drflac_uint32 right3 = pInputSamples1U32[i*4+3] << shift1;
+
+        drflac_uint32 left0 = right0 + side0;
+        drflac_uint32 left1 = right1 + side1;
+        drflac_uint32 left2 = right2 + side2;
+        drflac_uint32 left3 = right3 + side3;
+
+        pOutputSamples[i*8+0] = (drflac_int32)left0;
+        pOutputSamples[i*8+1] = (drflac_int32)right0;
+        pOutputSamples[i*8+2] = (drflac_int32)left1;
+        pOutputSamples[i*8+3] = (drflac_int32)right1;
+        pOutputSamples[i*8+4] = (drflac_int32)left2;
+        pOutputSamples[i*8+5] = (drflac_int32)right2;
+        pOutputSamples[i*8+6] = (drflac_int32)left3;
+        pOutputSamples[i*8+7] = (drflac_int32)right3;
+    }
+
+    for (i = (frameCount4 << 2); i < frameCount; ++i) {
+        drflac_uint32 side  = pInputSamples0U32[i] << shift0;
+        drflac_uint32 right = pInputSamples1U32[i] << shift1;
+        drflac_uint32 left  = right + side;
+
+        pOutputSamples[i*2+0] = (drflac_int32)left;
+        pOutputSamples[i*2+1] = (drflac_int32)right;
+    }
+}
+
+#if defined(DRFLAC_SUPPORT_SSE2)
+static DRFLAC_INLINE void drflac_read_pcm_frames_s32__decode_right_side__sse2(drflac* pFlac, drflac_uint64 frameCount, drflac_uint32 unusedBitsPerSample, const drflac_int32* pInputSamples0, const drflac_int32* pInputSamples1, drflac_int32* pOutputSamples)
+{
+    drflac_uint64 i;
+    drflac_uint64 frameCount4 = frameCount >> 2;
+    const drflac_uint32* pInputSamples0U32 = (const drflac_uint32*)pInputSamples0;
+    const drflac_uint32* pInputSamples1U32 = (const drflac_uint32*)pInputSamples1;
+    drflac_uint32 shift0 = unusedBitsPerSample + pFlac->currentFLACFrame.subframes[0].wastedBitsPerSample;
+    drflac_uint32 shift1 = unusedBitsPerSample + pFlac->currentFLACFrame.subframes[1].wastedBitsPerSample;
+
+    DRFLAC_ASSERT(pFlac->bitsPerSample <= 24);
+
+    for (i = 0; i < frameCount4; ++i) {
+        __m128i side  = _mm_slli_epi32(_mm_loadu_si128((const __m128i*)pInputSamples0 + i), shift0);
+        __m128i right = _mm_slli_epi32(_mm_loadu_si128((const __m128i*)pInputSamples1 + i), shift1);
+        __m128i left  = _mm_add_epi32(right, side);
+
+        _mm_storeu_si128((__m128i*)(pOutputSamples + i*8 + 0), _mm_unpacklo_epi32(left, right));
+        _mm_storeu_si128((__m128i*)(pOutputSamples + i*8 + 4), _mm_unpackhi_epi32(left, right));
+    }
+
+    for (i = (frameCount4 << 2); i < frameCount; ++i) {
+        drflac_uint32 side  = pInputSamples0U32[i] << shift0;
+        drflac_uint32 right = pInputSamples1U32[i] << shift1;
+        drflac_uint32 left  = right + side;
+
+        pOutputSamples[i*2+0] = (drflac_int32)left;
+        pOutputSamples[i*2+1] = (drflac_int32)right;
+    }
+}
+#endif
+
+#if defined(DRFLAC_SUPPORT_NEON)
+static DRFLAC_INLINE void drflac_read_pcm_frames_s32__decode_right_side__neon(drflac* pFlac, drflac_uint64 frameCount, drflac_uint32 unusedBitsPerSample, const drflac_int32* pInputSamples0, const drflac_int32* pInputSamples1, drflac_int32* pOutputSamples)
+{
+    drflac_uint64 i;
+    drflac_uint64 frameCount4 = frameCount >> 2;
+    const drflac_uint32* pInputSamples0U32 = (const drflac_uint32*)pInputSamples0;
+    const drflac_uint32* pInputSamples1U32 = (const drflac_uint32*)pInputSamples1;
+    drflac_uint32 shift0 = unusedBitsPerSample + pFlac->currentFLACFrame.subframes[0].wastedBitsPerSample;
+    drflac_uint32 shift1 = unusedBitsPerSample + pFlac->currentFLACFrame.subframes[1].wastedBitsPerSample;
+    int32x4_t shift0_4;
+    int32x4_t shift1_4;
+
+    DRFLAC_ASSERT(pFlac->bitsPerSample <= 24);
+
+    shift0_4 = vdupq_n_s32(shift0);
+    shift1_4 = vdupq_n_s32(shift1);
+
+    for (i = 0; i < frameCount4; ++i) {
+        uint32x4_t side;
+        uint32x4_t right;
+        uint32x4_t left;
+
+        side  = vshlq_u32(vld1q_u32(pInputSamples0U32 + i*4), shift0_4);
+        right = vshlq_u32(vld1q_u32(pInputSamples1U32 + i*4), shift1_4);
+        left  = vaddq_u32(right, side);
+
+        drflac__vst2q_u32((drflac_uint32*)pOutputSamples + i*8, vzipq_u32(left, right));
+    }
+
+    for (i = (frameCount4 << 2); i < frameCount; ++i) {
+        drflac_uint32 side  = pInputSamples0U32[i] << shift0;
+        drflac_uint32 right = pInputSamples1U32[i] << shift1;
+        drflac_uint32 left  = right + side;
+
+        pOutputSamples[i*2+0] = (drflac_int32)left;
+        pOutputSamples[i*2+1] = (drflac_int32)right;
+    }
+}
+#endif
+
+static DRFLAC_INLINE void drflac_read_pcm_frames_s32__decode_right_side(drflac* pFlac, drflac_uint64 frameCount, drflac_uint32 unusedBitsPerSample, const drflac_int32* pInputSamples0, const drflac_int32* pInputSamples1, drflac_int32* pOutputSamples)
+{
+#if defined(DRFLAC_SUPPORT_SSE2)
+    if (drflac__gIsSSE2Supported && pFlac->bitsPerSample <= 24) {
+        drflac_read_pcm_frames_s32__decode_right_side__sse2(pFlac, frameCount, unusedBitsPerSample, pInputSamples0, pInputSamples1, pOutputSamples);
+    } else
+#elif defined(DRFLAC_SUPPORT_NEON)
+    if (drflac__gIsNEONSupported && pFlac->bitsPerSample <= 24) {
+        drflac_read_pcm_frames_s32__decode_right_side__neon(pFlac, frameCount, unusedBitsPerSample, pInputSamples0, pInputSamples1, pOutputSamples);
+    } else
+#endif
+    {
+        /* Scalar fallback. */
+#if 0
+        drflac_read_pcm_frames_s32__decode_right_side__reference(pFlac, frameCount, unusedBitsPerSample, pInputSamples0, pInputSamples1, pOutputSamples);
+#else
+        drflac_read_pcm_frames_s32__decode_right_side__scalar(pFlac, frameCount, unusedBitsPerSample, pInputSamples0, pInputSamples1, pOutputSamples);
+#endif
+    }
+}
+
+
+#if 0
+static DRFLAC_INLINE void drflac_read_pcm_frames_s32__decode_mid_side__reference(drflac* pFlac, drflac_uint64 frameCount, drflac_uint32 unusedBitsPerSample, const drflac_int32* pInputSamples0, const drflac_int32* pInputSamples1, drflac_int32* pOutputSamples)
+{
+    for (drflac_uint64 i = 0; i < frameCount; ++i) {
+        drflac_uint32 mid  = pInputSamples0U32[i] << pFlac->currentFLACFrame.subframes[0].wastedBitsPerSample;
+        drflac_uint32 side = pInputSamples1U32[i] << pFlac->currentFLACFrame.subframes[1].wastedBitsPerSample;
+
+        mid = (mid << 1) | (side & 0x01);
+
+        pOutputSamples[i*2+0] = (drflac_int32)((drflac_uint32)((drflac_int32)(mid + side) >> 1) << unusedBitsPerSample);
+        pOutputSamples[i*2+1] = (drflac_int32)((drflac_uint32)((drflac_int32)(mid - side) >> 1) << unusedBitsPerSample);
+    }
+}
+#endif
+
+static DRFLAC_INLINE void drflac_read_pcm_frames_s32__decode_mid_side__scalar(drflac* pFlac, drflac_uint64 frameCount, drflac_uint32 unusedBitsPerSample, const drflac_int32* pInputSamples0, const drflac_int32* pInputSamples1, drflac_int32* pOutputSamples)
+{
+    drflac_uint64 i;
+    drflac_uint64 frameCount4 = frameCount >> 2;
+    const drflac_uint32* pInputSamples0U32 = (const drflac_uint32*)pInputSamples0;
+    const drflac_uint32* pInputSamples1U32 = (const drflac_uint32*)pInputSamples1;
+    drflac_int32 shift = unusedBitsPerSample;
+
+    if (shift > 0) {
+        shift -= 1;
+        for (i = 0; i < frameCount4; ++i) {
+            drflac_uint32 temp0L;
+            drflac_uint32 temp1L;
+            drflac_uint32 temp2L;
+            drflac_uint32 temp3L;
+            drflac_uint32 temp0R;
+            drflac_uint32 temp1R;
+            drflac_uint32 temp2R;
+            drflac_uint32 temp3R;
+
+            drflac_uint32 mid0  = pInputSamples0U32[i*4+0] << pFlac->currentFLACFrame.subframes[0].wastedBitsPerSample;
+            drflac_uint32 mid1  = pInputSamples0U32[i*4+1] << pFlac->currentFLACFrame.subframes[0].wastedBitsPerSample;
+            drflac_uint32 mid2  = pInputSamples0U32[i*4+2] << pFlac->currentFLACFrame.subframes[0].wastedBitsPerSample;
+            drflac_uint32 mid3  = pInputSamples0U32[i*4+3] << pFlac->currentFLACFrame.subframes[0].wastedBitsPerSample;
+
+            drflac_uint32 side0 = pInputSamples1U32[i*4+0] << pFlac->currentFLACFrame.subframes[1].wastedBitsPerSample;
+            drflac_uint32 side1 = pInputSamples1U32[i*4+1] << pFlac->currentFLACFrame.subframes[1].wastedBitsPerSample;
+            drflac_uint32 side2 = pInputSamples1U32[i*4+2] << pFlac->currentFLACFrame.subframes[1].wastedBitsPerSample;
+            drflac_uint32 side3 = pInputSamples1U32[i*4+3] << pFlac->currentFLACFrame.subframes[1].wastedBitsPerSample;
+
+            mid0 = (mid0 << 1) | (side0 & 0x01);
+            mid1 = (mid1 << 1) | (side1 & 0x01);
+            mid2 = (mid2 << 1) | (side2 & 0x01);
+            mid3 = (mid3 << 1) | (side3 & 0x01);
+
+            temp0L = (mid0 + side0) << shift;
+            temp1L = (mid1 + side1) << shift;
+            temp2L = (mid2 + side2) << shift;
+            temp3L = (mid3 + side3) << shift;
+
+            temp0R = (mid0 - side0) << shift;
+            temp1R = (mid1 - side1) << shift;
+            temp2R = (mid2 - side2) << shift;
+            temp3R = (mid3 - side3) << shift;
+
+            pOutputSamples[i*8+0] = (drflac_int32)temp0L;
+            pOutputSamples[i*8+1] = (drflac_int32)temp0R;
+            pOutputSamples[i*8+2] = (drflac_int32)temp1L;
+            pOutputSamples[i*8+3] = (drflac_int32)temp1R;
+            pOutputSamples[i*8+4] = (drflac_int32)temp2L;
+            pOutputSamples[i*8+5] = (drflac_int32)temp2R;
+            pOutputSamples[i*8+6] = (drflac_int32)temp3L;
+            pOutputSamples[i*8+7] = (drflac_int32)temp3R;
+        }
+    } else {
+        for (i = 0; i < frameCount4; ++i) {
+            drflac_uint32 temp0L;
+            drflac_uint32 temp1L;
+            drflac_uint32 temp2L;
+            drflac_uint32 temp3L;
+            drflac_uint32 temp0R;
+            drflac_uint32 temp1R;
+            drflac_uint32 temp2R;
+            drflac_uint32 temp3R;
+
+            drflac_uint32 mid0  = pInputSamples0U32[i*4+0] << pFlac->currentFLACFrame.subframes[0].wastedBitsPerSample;
+            drflac_uint32 mid1  = pInputSamples0U32[i*4+1] << pFlac->currentFLACFrame.subframes[0].wastedBitsPerSample;
+            drflac_uint32 mid2  = pInputSamples0U32[i*4+2] << pFlac->currentFLACFrame.subframes[0].wastedBitsPerSample;
+            drflac_uint32 mid3  = pInputSamples0U32[i*4+3] << pFlac->currentFLACFrame.subframes[0].wastedBitsPerSample;
+
+            drflac_uint32 side0 = pInputSamples1U32[i*4+0] << pFlac->currentFLACFrame.subframes[1].wastedBitsPerSample;
+            drflac_uint32 side1 = pInputSamples1U32[i*4+1] << pFlac->currentFLACFrame.subframes[1].wastedBitsPerSample;
+            drflac_uint32 side2 = pInputSamples1U32[i*4+2] << pFlac->currentFLACFrame.subframes[1].wastedBitsPerSample;
+            drflac_uint32 side3 = pInputSamples1U32[i*4+3] << pFlac->currentFLACFrame.subframes[1].wastedBitsPerSample;
+
+            mid0 = (mid0 << 1) | (side0 & 0x01);
+            mid1 = (mid1 << 1) | (side1 & 0x01);
+            mid2 = (mid2 << 1) | (side2 & 0x01);
+            mid3 = (mid3 << 1) | (side3 & 0x01);
+
+            temp0L = (drflac_uint32)((drflac_int32)(mid0 + side0) >> 1);
+            temp1L = (drflac_uint32)((drflac_int32)(mid1 + side1) >> 1);
+            temp2L = (drflac_uint32)((drflac_int32)(mid2 + side2) >> 1);
+            temp3L = (drflac_uint32)((drflac_int32)(mid3 + side3) >> 1);
+
+            temp0R = (drflac_uint32)((drflac_int32)(mid0 - side0) >> 1);
+            temp1R = (drflac_uint32)((drflac_int32)(mid1 - side1) >> 1);
+            temp2R = (drflac_uint32)((drflac_int32)(mid2 - side2) >> 1);
+            temp3R = (drflac_uint32)((drflac_int32)(mid3 - side3) >> 1);
+
+            pOutputSamples[i*8+0] = (drflac_int32)temp0L;
+            pOutputSamples[i*8+1] = (drflac_int32)temp0R;
+            pOutputSamples[i*8+2] = (drflac_int32)temp1L;
+            pOutputSamples[i*8+3] = (drflac_int32)temp1R;
+            pOutputSamples[i*8+4] = (drflac_int32)temp2L;
+            pOutputSamples[i*8+5] = (drflac_int32)temp2R;
+            pOutputSamples[i*8+6] = (drflac_int32)temp3L;
+            pOutputSamples[i*8+7] = (drflac_int32)temp3R;
+        }
+    }
+
+    for (i = (frameCount4 << 2); i < frameCount; ++i) {
+        drflac_uint32 mid  = pInputSamples0U32[i] << pFlac->currentFLACFrame.subframes[0].wastedBitsPerSample;
+        drflac_uint32 side = pInputSamples1U32[i] << pFlac->currentFLACFrame.subframes[1].wastedBitsPerSample;
+
+        mid = (mid << 1) | (side & 0x01);
+
+        pOutputSamples[i*2+0] = (drflac_int32)((drflac_uint32)((drflac_int32)(mid + side) >> 1) << unusedBitsPerSample);
+        pOutputSamples[i*2+1] = (drflac_int32)((drflac_uint32)((drflac_int32)(mid - side) >> 1) << unusedBitsPerSample);
+    }
+}
+
+#if defined(DRFLAC_SUPPORT_SSE2)
+static DRFLAC_INLINE void drflac_read_pcm_frames_s32__decode_mid_side__sse2(drflac* pFlac, drflac_uint64 frameCount, drflac_uint32 unusedBitsPerSample, const drflac_int32* pInputSamples0, const drflac_int32* pInputSamples1, drflac_int32* pOutputSamples)
+{
+    drflac_uint64 i;
+    drflac_uint64 frameCount4 = frameCount >> 2;
+    const drflac_uint32* pInputSamples0U32 = (const drflac_uint32*)pInputSamples0;
+    const drflac_uint32* pInputSamples1U32 = (const drflac_uint32*)pInputSamples1;
+    drflac_int32 shift = unusedBitsPerSample;
+
+    DRFLAC_ASSERT(pFlac->bitsPerSample <= 24);
+
+    if (shift == 0) {
+        for (i = 0; i < frameCount4; ++i) {
+            __m128i mid;
+            __m128i side;
+            __m128i left;
+            __m128i right;
+
+            mid   = _mm_slli_epi32(_mm_loadu_si128((const __m128i*)pInputSamples0 + i), pFlac->currentFLACFrame.subframes[0].wastedBitsPerSample);
+            side  = _mm_slli_epi32(_mm_loadu_si128((const __m128i*)pInputSamples1 + i), pFlac->currentFLACFrame.subframes[1].wastedBitsPerSample);
+
+            mid   = _mm_or_si128(_mm_slli_epi32(mid, 1), _mm_and_si128(side, _mm_set1_epi32(0x01)));
+
+            left  = _mm_srai_epi32(_mm_add_epi32(mid, side), 1);
+            right = _mm_srai_epi32(_mm_sub_epi32(mid, side), 1);
+
+            _mm_storeu_si128((__m128i*)(pOutputSamples + i*8 + 0), _mm_unpacklo_epi32(left, right));
+            _mm_storeu_si128((__m128i*)(pOutputSamples + i*8 + 4), _mm_unpackhi_epi32(left, right));
+        }
+
+        for (i = (frameCount4 << 2); i < frameCount; ++i) {
+            drflac_uint32 mid  = pInputSamples0U32[i] << pFlac->currentFLACFrame.subframes[0].wastedBitsPerSample;
+            drflac_uint32 side = pInputSamples1U32[i] << pFlac->currentFLACFrame.subframes[1].wastedBitsPerSample;
+
+            mid = (mid << 1) | (side & 0x01);
+
+            pOutputSamples[i*2+0] = (drflac_int32)(mid + side) >> 1;
+            pOutputSamples[i*2+1] = (drflac_int32)(mid - side) >> 1;
+        }
+    } else {
+        shift -= 1;
+        for (i = 0; i < frameCount4; ++i) {
+            __m128i mid;
+            __m128i side;
+            __m128i left;
+            __m128i right;
+
+            mid   = _mm_slli_epi32(_mm_loadu_si128((const __m128i*)pInputSamples0 + i), pFlac->currentFLACFrame.subframes[0].wastedBitsPerSample);
+            side  = _mm_slli_epi32(_mm_loadu_si128((const __m128i*)pInputSamples1 + i), pFlac->currentFLACFrame.subframes[1].wastedBitsPerSample);
+
+            mid   = _mm_or_si128(_mm_slli_epi32(mid, 1), _mm_and_si128(side, _mm_set1_epi32(0x01)));
+
+            left  = _mm_slli_epi32(_mm_add_epi32(mid, side), shift);
+            right = _mm_slli_epi32(_mm_sub_epi32(mid, side), shift);
+
+            _mm_storeu_si128((__m128i*)(pOutputSamples + i*8 + 0), _mm_unpacklo_epi32(left, right));
+            _mm_storeu_si128((__m128i*)(pOutputSamples + i*8 + 4), _mm_unpackhi_epi32(left, right));
+        }
+
+        for (i = (frameCount4 << 2); i < frameCount; ++i) {
+            drflac_uint32 mid  = pInputSamples0U32[i] << pFlac->currentFLACFrame.subframes[0].wastedBitsPerSample;
+            drflac_uint32 side = pInputSamples1U32[i] << pFlac->currentFLACFrame.subframes[1].wastedBitsPerSample;
+
+            mid = (mid << 1) | (side & 0x01);
+
+            pOutputSamples[i*2+0] = (drflac_int32)((mid + side) << shift);
+            pOutputSamples[i*2+1] = (drflac_int32)((mid - side) << shift);
+        }
+    }
+}
+#endif
+
+#if defined(DRFLAC_SUPPORT_NEON)
+static DRFLAC_INLINE void drflac_read_pcm_frames_s32__decode_mid_side__neon(drflac* pFlac, drflac_uint64 frameCount, drflac_uint32 unusedBitsPerSample, const drflac_int32* pInputSamples0, const drflac_int32* pInputSamples1, drflac_int32* pOutputSamples)
+{
+    drflac_uint64 i;
+    drflac_uint64 frameCount4 = frameCount >> 2;
+    const drflac_uint32* pInputSamples0U32 = (const drflac_uint32*)pInputSamples0;
+    const drflac_uint32* pInputSamples1U32 = (const drflac_uint32*)pInputSamples1;
+    drflac_int32 shift = unusedBitsPerSample;
+    int32x4_t  wbpsShift0_4; /* wbps = Wasted Bits Per Sample */
+    int32x4_t  wbpsShift1_4; /* wbps = Wasted Bits Per Sample */
+    uint32x4_t one4;
+
+    DRFLAC_ASSERT(pFlac->bitsPerSample <= 24);
+
+    wbpsShift0_4 = vdupq_n_s32(pFlac->currentFLACFrame.subframes[0].wastedBitsPerSample);
+    wbpsShift1_4 = vdupq_n_s32(pFlac->currentFLACFrame.subframes[1].wastedBitsPerSample);
+    one4         = vdupq_n_u32(1);
+
+    if (shift == 0) {
+        for (i = 0; i < frameCount4; ++i) {
+            uint32x4_t mid;
+            uint32x4_t side;
+            int32x4_t left;
+            int32x4_t right;
+
+            mid   = vshlq_u32(vld1q_u32(pInputSamples0U32 + i*4), wbpsShift0_4);
+            side  = vshlq_u32(vld1q_u32(pInputSamples1U32 + i*4), wbpsShift1_4);
+
+            mid   = vorrq_u32(vshlq_n_u32(mid, 1), vandq_u32(side, one4));
+
+            left  = vshrq_n_s32(vreinterpretq_s32_u32(vaddq_u32(mid, side)), 1);
+            right = vshrq_n_s32(vreinterpretq_s32_u32(vsubq_u32(mid, side)), 1);
+
+            drflac__vst2q_s32(pOutputSamples + i*8, vzipq_s32(left, right));
+        }
+
+        for (i = (frameCount4 << 2); i < frameCount; ++i) {
+            drflac_uint32 mid  = pInputSamples0U32[i] << pFlac->currentFLACFrame.subframes[0].wastedBitsPerSample;
+            drflac_uint32 side = pInputSamples1U32[i] << pFlac->currentFLACFrame.subframes[1].wastedBitsPerSample;
+
+            mid = (mid << 1) | (side & 0x01);
+
+            pOutputSamples[i*2+0] = (drflac_int32)(mid + side) >> 1;
+            pOutputSamples[i*2+1] = (drflac_int32)(mid - side) >> 1;
+        }
+    } else {
+        int32x4_t shift4;
+
+        shift -= 1;
+        shift4 = vdupq_n_s32(shift);
+
+        for (i = 0; i < frameCount4; ++i) {
+            uint32x4_t mid;
+            uint32x4_t side;
+            int32x4_t left;
+            int32x4_t right;
+
+            mid   = vshlq_u32(vld1q_u32(pInputSamples0U32 + i*4), wbpsShift0_4);
+            side  = vshlq_u32(vld1q_u32(pInputSamples1U32 + i*4), wbpsShift1_4);
+
+            mid   = vorrq_u32(vshlq_n_u32(mid, 1), vandq_u32(side, one4));
+
+            left  = vreinterpretq_s32_u32(vshlq_u32(vaddq_u32(mid, side), shift4));
+            right = vreinterpretq_s32_u32(vshlq_u32(vsubq_u32(mid, side), shift4));
+
+            drflac__vst2q_s32(pOutputSamples + i*8, vzipq_s32(left, right));
+        }
+
+        for (i = (frameCount4 << 2); i < frameCount; ++i) {
+            drflac_uint32 mid  = pInputSamples0U32[i] << pFlac->currentFLACFrame.subframes[0].wastedBitsPerSample;
+            drflac_uint32 side = pInputSamples1U32[i] << pFlac->currentFLACFrame.subframes[1].wastedBitsPerSample;
+
+            mid = (mid << 1) | (side & 0x01);
+
+            pOutputSamples[i*2+0] = (drflac_int32)((mid + side) << shift);
+            pOutputSamples[i*2+1] = (drflac_int32)((mid - side) << shift);
+        }
+    }
+}
+#endif
+
+static DRFLAC_INLINE void drflac_read_pcm_frames_s32__decode_mid_side(drflac* pFlac, drflac_uint64 frameCount, drflac_uint32 unusedBitsPerSample, const drflac_int32* pInputSamples0, const drflac_int32* pInputSamples1, drflac_int32* pOutputSamples)
+{
+#if defined(DRFLAC_SUPPORT_SSE2)
+    if (drflac__gIsSSE2Supported && pFlac->bitsPerSample <= 24) {
+        drflac_read_pcm_frames_s32__decode_mid_side__sse2(pFlac, frameCount, unusedBitsPerSample, pInputSamples0, pInputSamples1, pOutputSamples);
+    } else
+#elif defined(DRFLAC_SUPPORT_NEON)
+    if (drflac__gIsNEONSupported && pFlac->bitsPerSample <= 24) {
+        drflac_read_pcm_frames_s32__decode_mid_side__neon(pFlac, frameCount, unusedBitsPerSample, pInputSamples0, pInputSamples1, pOutputSamples);
+    } else
+#endif
+    {
+        /* Scalar fallback. */
+#if 0
+        drflac_read_pcm_frames_s32__decode_mid_side__reference(pFlac, frameCount, unusedBitsPerSample, pInputSamples0, pInputSamples1, pOutputSamples);
+#else
+        drflac_read_pcm_frames_s32__decode_mid_side__scalar(pFlac, frameCount, unusedBitsPerSample, pInputSamples0, pInputSamples1, pOutputSamples);
+#endif
+    }
+}
+
+
+#if 0
+static DRFLAC_INLINE void drflac_read_pcm_frames_s32__decode_independent_stereo__reference(drflac* pFlac, drflac_uint64 frameCount, drflac_uint32 unusedBitsPerSample, const drflac_int32* pInputSamples0, const drflac_int32* pInputSamples1, drflac_int32* pOutputSamples)
+{
+    for (drflac_uint64 i = 0; i < frameCount; ++i) {
+        pOutputSamples[i*2+0] = (drflac_int32)((drflac_uint32)pInputSamples0[i] << (unusedBitsPerSample + pFlac->currentFLACFrame.subframes[0].wastedBitsPerSample));
+        pOutputSamples[i*2+1] = (drflac_int32)((drflac_uint32)pInputSamples1[i] << (unusedBitsPerSample + pFlac->currentFLACFrame.subframes[1].wastedBitsPerSample));
+    }
+}
+#endif
+
+static DRFLAC_INLINE void drflac_read_pcm_frames_s32__decode_independent_stereo__scalar(drflac* pFlac, drflac_uint64 frameCount, drflac_uint32 unusedBitsPerSample, const drflac_int32* pInputSamples0, const drflac_int32* pInputSamples1, drflac_int32* pOutputSamples)
+{
+    drflac_uint64 i;
+    drflac_uint64 frameCount4 = frameCount >> 2;
+    const drflac_uint32* pInputSamples0U32 = (const drflac_uint32*)pInputSamples0;
+    const drflac_uint32* pInputSamples1U32 = (const drflac_uint32*)pInputSamples1;
+    drflac_uint32 shift0 = unusedBitsPerSample + pFlac->currentFLACFrame.subframes[0].wastedBitsPerSample;
+    drflac_uint32 shift1 = unusedBitsPerSample + pFlac->currentFLACFrame.subframes[1].wastedBitsPerSample;
+
+    for (i = 0; i < frameCount4; ++i) {
+        drflac_uint32 tempL0 = pInputSamples0U32[i*4+0] << shift0;
+        drflac_uint32 tempL1 = pInputSamples0U32[i*4+1] << shift0;
+        drflac_uint32 tempL2 = pInputSamples0U32[i*4+2] << shift0;
+        drflac_uint32 tempL3 = pInputSamples0U32[i*4+3] << shift0;
+
+        drflac_uint32 tempR0 = pInputSamples1U32[i*4+0] << shift1;
+        drflac_uint32 tempR1 = pInputSamples1U32[i*4+1] << shift1;
+        drflac_uint32 tempR2 = pInputSamples1U32[i*4+2] << shift1;
+        drflac_uint32 tempR3 = pInputSamples1U32[i*4+3] << shift1;
+
+        pOutputSamples[i*8+0] = (drflac_int32)tempL0;
+        pOutputSamples[i*8+1] = (drflac_int32)tempR0;
+        pOutputSamples[i*8+2] = (drflac_int32)tempL1;
+        pOutputSamples[i*8+3] = (drflac_int32)tempR1;
+        pOutputSamples[i*8+4] = (drflac_int32)tempL2;
+        pOutputSamples[i*8+5] = (drflac_int32)tempR2;
+        pOutputSamples[i*8+6] = (drflac_int32)tempL3;
+        pOutputSamples[i*8+7] = (drflac_int32)tempR3;
+    }
+
+    for (i = (frameCount4 << 2); i < frameCount; ++i) {
+        pOutputSamples[i*2+0] = (drflac_int32)(pInputSamples0U32[i] << shift0);
+        pOutputSamples[i*2+1] = (drflac_int32)(pInputSamples1U32[i] << shift1);
+    }
+}
+
+#if defined(DRFLAC_SUPPORT_SSE2)
+static DRFLAC_INLINE void drflac_read_pcm_frames_s32__decode_independent_stereo__sse2(drflac* pFlac, drflac_uint64 frameCount, drflac_uint32 unusedBitsPerSample, const drflac_int32* pInputSamples0, const drflac_int32* pInputSamples1, drflac_int32* pOutputSamples)
+{
+    drflac_uint64 i;
+    drflac_uint64 frameCount4 = frameCount >> 2;
+    const drflac_uint32* pInputSamples0U32 = (const drflac_uint32*)pInputSamples0;
+    const drflac_uint32* pInputSamples1U32 = (const drflac_uint32*)pInputSamples1;
+    drflac_uint32 shift0 = unusedBitsPerSample + pFlac->currentFLACFrame.subframes[0].wastedBitsPerSample;
+    drflac_uint32 shift1 = unusedBitsPerSample + pFlac->currentFLACFrame.subframes[1].wastedBitsPerSample;
+
+    for (i = 0; i < frameCount4; ++i) {
+        __m128i left  = _mm_slli_epi32(_mm_loadu_si128((const __m128i*)pInputSamples0 + i), shift0);
+        __m128i right = _mm_slli_epi32(_mm_loadu_si128((const __m128i*)pInputSamples1 + i), shift1);
+
+        _mm_storeu_si128((__m128i*)(pOutputSamples + i*8 + 0), _mm_unpacklo_epi32(left, right));
+        _mm_storeu_si128((__m128i*)(pOutputSamples + i*8 + 4), _mm_unpackhi_epi32(left, right));
+    }
+
+    for (i = (frameCount4 << 2); i < frameCount; ++i) {
+        pOutputSamples[i*2+0] = (drflac_int32)(pInputSamples0U32[i] << shift0);
+        pOutputSamples[i*2+1] = (drflac_int32)(pInputSamples1U32[i] << shift1);
+    }
+}
+#endif
+
+#if defined(DRFLAC_SUPPORT_NEON)
+static DRFLAC_INLINE void drflac_read_pcm_frames_s32__decode_independent_stereo__neon(drflac* pFlac, drflac_uint64 frameCount, drflac_uint32 unusedBitsPerSample, const drflac_int32* pInputSamples0, const drflac_int32* pInputSamples1, drflac_int32* pOutputSamples)
+{
+    drflac_uint64 i;
+    drflac_uint64 frameCount4 = frameCount >> 2;
+    const drflac_uint32* pInputSamples0U32 = (const drflac_uint32*)pInputSamples0;
+    const drflac_uint32* pInputSamples1U32 = (const drflac_uint32*)pInputSamples1;
+    drflac_uint32 shift0 = unusedBitsPerSample + pFlac->currentFLACFrame.subframes[0].wastedBitsPerSample;
+    drflac_uint32 shift1 = unusedBitsPerSample + pFlac->currentFLACFrame.subframes[1].wastedBitsPerSample;
+
+    int32x4_t shift4_0 = vdupq_n_s32(shift0);
+    int32x4_t shift4_1 = vdupq_n_s32(shift1);
+
+    for (i = 0; i < frameCount4; ++i) {
+        int32x4_t left;
+        int32x4_t right;
+
+        left  = vreinterpretq_s32_u32(vshlq_u32(vld1q_u32(pInputSamples0U32 + i*4), shift4_0));
+        right = vreinterpretq_s32_u32(vshlq_u32(vld1q_u32(pInputSamples1U32 + i*4), shift4_1));
+
+        drflac__vst2q_s32(pOutputSamples + i*8, vzipq_s32(left, right));
+    }
+
+    for (i = (frameCount4 << 2); i < frameCount; ++i) {
+        pOutputSamples[i*2+0] = (drflac_int32)(pInputSamples0U32[i] << shift0);
+        pOutputSamples[i*2+1] = (drflac_int32)(pInputSamples1U32[i] << shift1);
+    }
+}
+#endif
+
+static DRFLAC_INLINE void drflac_read_pcm_frames_s32__decode_independent_stereo(drflac* pFlac, drflac_uint64 frameCount, drflac_uint32 unusedBitsPerSample, const drflac_int32* pInputSamples0, const drflac_int32* pInputSamples1, drflac_int32* pOutputSamples)
+{
+#if defined(DRFLAC_SUPPORT_SSE2)
+    if (drflac__gIsSSE2Supported && pFlac->bitsPerSample <= 24) {
+        drflac_read_pcm_frames_s32__decode_independent_stereo__sse2(pFlac, frameCount, unusedBitsPerSample, pInputSamples0, pInputSamples1, pOutputSamples);
+    } else
+#elif defined(DRFLAC_SUPPORT_NEON)
+    if (drflac__gIsNEONSupported && pFlac->bitsPerSample <= 24) {
+        drflac_read_pcm_frames_s32__decode_independent_stereo__neon(pFlac, frameCount, unusedBitsPerSample, pInputSamples0, pInputSamples1, pOutputSamples);
+    } else
+#endif
+    {
+        /* Scalar fallback. */
+#if 0
+        drflac_read_pcm_frames_s32__decode_independent_stereo__reference(pFlac, frameCount, unusedBitsPerSample, pInputSamples0, pInputSamples1, pOutputSamples);
+#else
+        drflac_read_pcm_frames_s32__decode_independent_stereo__scalar(pFlac, frameCount, unusedBitsPerSample, pInputSamples0, pInputSamples1, pOutputSamples);
+#endif
+    }
+}
+
+
+DRFLAC_API drflac_uint64 drflac_read_pcm_frames_s32(drflac* pFlac, drflac_uint64 framesToRead, drflac_int32* pBufferOut)
+{
+    drflac_uint64 framesRead;
+    drflac_uint32 unusedBitsPerSample;
+
+    if (pFlac == NULL || framesToRead == 0) {
+        return 0;
+    }
+
+    if (pBufferOut == NULL) {
+        return drflac__seek_forward_by_pcm_frames(pFlac, framesToRead);
+    }
+
+    DRFLAC_ASSERT(pFlac->bitsPerSample <= 32);
+    unusedBitsPerSample = 32 - pFlac->bitsPerSample;
+
+    framesRead = 0;
+    while (framesToRead > 0) {
+        /* If we've run out of samples in this frame, go to the next. */
+        if (pFlac->currentFLACFrame.pcmFramesRemaining == 0) {
+            if (!drflac__read_and_decode_next_flac_frame(pFlac)) {
+                break;  /* Couldn't read the next frame, so just break from the loop and return. */
+            }
+        } else {
+            unsigned int channelCount = drflac__get_channel_count_from_channel_assignment(pFlac->currentFLACFrame.header.channelAssignment);
+            drflac_uint64 iFirstPCMFrame = pFlac->currentFLACFrame.header.blockSizeInPCMFrames - pFlac->currentFLACFrame.pcmFramesRemaining;
+            drflac_uint64 frameCountThisIteration = framesToRead;
+
+            if (frameCountThisIteration > pFlac->currentFLACFrame.pcmFramesRemaining) {
+                frameCountThisIteration = pFlac->currentFLACFrame.pcmFramesRemaining;
+            }
+
+            if (channelCount == 2) {
+                const drflac_int32* pDecodedSamples0 = pFlac->currentFLACFrame.subframes[0].pSamplesS32 + iFirstPCMFrame;
+                const drflac_int32* pDecodedSamples1 = pFlac->currentFLACFrame.subframes[1].pSamplesS32 + iFirstPCMFrame;
+
+                switch (pFlac->currentFLACFrame.header.channelAssignment)
+                {
+                    case DRFLAC_CHANNEL_ASSIGNMENT_LEFT_SIDE:
+                    {
+                        drflac_read_pcm_frames_s32__decode_left_side(pFlac, frameCountThisIteration, unusedBitsPerSample, pDecodedSamples0, pDecodedSamples1, pBufferOut);
+                    } break;
+
+                    case DRFLAC_CHANNEL_ASSIGNMENT_RIGHT_SIDE:
+                    {
+                        drflac_read_pcm_frames_s32__decode_right_side(pFlac, frameCountThisIteration, unusedBitsPerSample, pDecodedSamples0, pDecodedSamples1, pBufferOut);
+                    } break;
+
+                    case DRFLAC_CHANNEL_ASSIGNMENT_MID_SIDE:
+                    {
+                        drflac_read_pcm_frames_s32__decode_mid_side(pFlac, frameCountThisIteration, unusedBitsPerSample, pDecodedSamples0, pDecodedSamples1, pBufferOut);
+                    } break;
+
+                    case DRFLAC_CHANNEL_ASSIGNMENT_INDEPENDENT:
+                    default:
+                    {
+                        drflac_read_pcm_frames_s32__decode_independent_stereo(pFlac, frameCountThisIteration, unusedBitsPerSample, pDecodedSamples0, pDecodedSamples1, pBufferOut);
+                    } break;
+                }
+            } else {
+                /* Generic interleaving. */
+                drflac_uint64 i;
+                for (i = 0; i < frameCountThisIteration; ++i) {
+                    unsigned int j;
+                    for (j = 0; j < channelCount; ++j) {
+                        pBufferOut[(i*channelCount)+j] = (drflac_int32)((drflac_uint32)(pFlac->currentFLACFrame.subframes[j].pSamplesS32[iFirstPCMFrame + i]) << (unusedBitsPerSample + pFlac->currentFLACFrame.subframes[j].wastedBitsPerSample));
+                    }
+                }
+            }
+
+            framesRead                += frameCountThisIteration;
+            pBufferOut                += frameCountThisIteration * channelCount;
+            framesToRead              -= frameCountThisIteration;
+            pFlac->currentPCMFrame    += frameCountThisIteration;
+            pFlac->currentFLACFrame.pcmFramesRemaining -= (drflac_uint32)frameCountThisIteration;
+        }
+    }
+
+    return framesRead;
+}
+
+
+#if 0
+static DRFLAC_INLINE void drflac_read_pcm_frames_s16__decode_left_side__reference(drflac* pFlac, drflac_uint64 frameCount, drflac_uint32 unusedBitsPerSample, const drflac_int32* pInputSamples0, const drflac_int32* pInputSamples1, drflac_int16* pOutputSamples)
+{
+    drflac_uint64 i;
+    for (i = 0; i < frameCount; ++i) {
+        drflac_uint32 left  = (drflac_uint32)pInputSamples0[i] << (unusedBitsPerSample + pFlac->currentFLACFrame.subframes[0].wastedBitsPerSample);
+        drflac_uint32 side  = (drflac_uint32)pInputSamples1[i] << (unusedBitsPerSample + pFlac->currentFLACFrame.subframes[1].wastedBitsPerSample);
+        drflac_uint32 right = left - side;
+
+        left  >>= 16;
+        right >>= 16;
+
+        pOutputSamples[i*2+0] = (drflac_int16)left;
+        pOutputSamples[i*2+1] = (drflac_int16)right;
+    }
+}
+#endif
+
+static DRFLAC_INLINE void drflac_read_pcm_frames_s16__decode_left_side__scalar(drflac* pFlac, drflac_uint64 frameCount, drflac_uint32 unusedBitsPerSample, const drflac_int32* pInputSamples0, const drflac_int32* pInputSamples1, drflac_int16* pOutputSamples)
+{
+    drflac_uint64 i;
+    drflac_uint64 frameCount4 = frameCount >> 2;
+    const drflac_uint32* pInputSamples0U32 = (const drflac_uint32*)pInputSamples0;
+    const drflac_uint32* pInputSamples1U32 = (const drflac_uint32*)pInputSamples1;
+    drflac_uint32 shift0 = unusedBitsPerSample + pFlac->currentFLACFrame.subframes[0].wastedBitsPerSample;
+    drflac_uint32 shift1 = unusedBitsPerSample + pFlac->currentFLACFrame.subframes[1].wastedBitsPerSample;
+
+    for (i = 0; i < frameCount4; ++i) {
+        drflac_uint32 left0 = pInputSamples0U32[i*4+0] << shift0;
+        drflac_uint32 left1 = pInputSamples0U32[i*4+1] << shift0;
+        drflac_uint32 left2 = pInputSamples0U32[i*4+2] << shift0;
+        drflac_uint32 left3 = pInputSamples0U32[i*4+3] << shift0;
+
+        drflac_uint32 side0 = pInputSamples1U32[i*4+0] << shift1;
+        drflac_uint32 side1 = pInputSamples1U32[i*4+1] << shift1;
+        drflac_uint32 side2 = pInputSamples1U32[i*4+2] << shift1;
+        drflac_uint32 side3 = pInputSamples1U32[i*4+3] << shift1;
+
+        drflac_uint32 right0 = left0 - side0;
+        drflac_uint32 right1 = left1 - side1;
+        drflac_uint32 right2 = left2 - side2;
+        drflac_uint32 right3 = left3 - side3;
+
+        left0  >>= 16;
+        left1  >>= 16;
+        left2  >>= 16;
+        left3  >>= 16;
+
+        right0 >>= 16;
+        right1 >>= 16;
+        right2 >>= 16;
+        right3 >>= 16;
+
+        pOutputSamples[i*8+0] = (drflac_int16)left0;
+        pOutputSamples[i*8+1] = (drflac_int16)right0;
+        pOutputSamples[i*8+2] = (drflac_int16)left1;
+        pOutputSamples[i*8+3] = (drflac_int16)right1;
+        pOutputSamples[i*8+4] = (drflac_int16)left2;
+        pOutputSamples[i*8+5] = (drflac_int16)right2;
+        pOutputSamples[i*8+6] = (drflac_int16)left3;
+        pOutputSamples[i*8+7] = (drflac_int16)right3;
+    }
+
+    for (i = (frameCount4 << 2); i < frameCount; ++i) {
+        drflac_uint32 left  = pInputSamples0U32[i] << shift0;
+        drflac_uint32 side  = pInputSamples1U32[i] << shift1;
+        drflac_uint32 right = left - side;
+
+        left  >>= 16;
+        right >>= 16;
+
+        pOutputSamples[i*2+0] = (drflac_int16)left;
+        pOutputSamples[i*2+1] = (drflac_int16)right;
+    }
+}
+
+#if defined(DRFLAC_SUPPORT_SSE2)
+static DRFLAC_INLINE void drflac_read_pcm_frames_s16__decode_left_side__sse2(drflac* pFlac, drflac_uint64 frameCount, drflac_uint32 unusedBitsPerSample, const drflac_int32* pInputSamples0, const drflac_int32* pInputSamples1, drflac_int16* pOutputSamples)
+{
+    drflac_uint64 i;
+    drflac_uint64 frameCount4 = frameCount >> 2;
+    const drflac_uint32* pInputSamples0U32 = (const drflac_uint32*)pInputSamples0;
+    const drflac_uint32* pInputSamples1U32 = (const drflac_uint32*)pInputSamples1;
+    drflac_uint32 shift0 = unusedBitsPerSample + pFlac->currentFLACFrame.subframes[0].wastedBitsPerSample;
+    drflac_uint32 shift1 = unusedBitsPerSample + pFlac->currentFLACFrame.subframes[1].wastedBitsPerSample;
+
+    DRFLAC_ASSERT(pFlac->bitsPerSample <= 24);
+
+    for (i = 0; i < frameCount4; ++i) {
+        __m128i left  = _mm_slli_epi32(_mm_loadu_si128((const __m128i*)pInputSamples0 + i), shift0);
+        __m128i side  = _mm_slli_epi32(_mm_loadu_si128((const __m128i*)pInputSamples1 + i), shift1);
+        __m128i right = _mm_sub_epi32(left, side);
+
+        left  = _mm_srai_epi32(left,  16);
+        right = _mm_srai_epi32(right, 16);
+
+        _mm_storeu_si128((__m128i*)(pOutputSamples + i*8), drflac__mm_packs_interleaved_epi32(left, right));
+    }
+
+    for (i = (frameCount4 << 2); i < frameCount; ++i) {
+        drflac_uint32 left  = pInputSamples0U32[i] << shift0;
+        drflac_uint32 side  = pInputSamples1U32[i] << shift1;
+        drflac_uint32 right = left - side;
+
+        left  >>= 16;
+        right >>= 16;
+
+        pOutputSamples[i*2+0] = (drflac_int16)left;
+        pOutputSamples[i*2+1] = (drflac_int16)right;
+    }
+}
+#endif
+
+#if defined(DRFLAC_SUPPORT_NEON)
+static DRFLAC_INLINE void drflac_read_pcm_frames_s16__decode_left_side__neon(drflac* pFlac, drflac_uint64 frameCount, drflac_uint32 unusedBitsPerSample, const drflac_int32* pInputSamples0, const drflac_int32* pInputSamples1, drflac_int16* pOutputSamples)
+{
+    drflac_uint64 i;
+    drflac_uint64 frameCount4 = frameCount >> 2;
+    const drflac_uint32* pInputSamples0U32 = (const drflac_uint32*)pInputSamples0;
+    const drflac_uint32* pInputSamples1U32 = (const drflac_uint32*)pInputSamples1;
+    drflac_uint32 shift0 = unusedBitsPerSample + pFlac->currentFLACFrame.subframes[0].wastedBitsPerSample;
+    drflac_uint32 shift1 = unusedBitsPerSample + pFlac->currentFLACFrame.subframes[1].wastedBitsPerSample;
+    int32x4_t shift0_4;
+    int32x4_t shift1_4;
+
+    DRFLAC_ASSERT(pFlac->bitsPerSample <= 24);
+
+    shift0_4 = vdupq_n_s32(shift0);
+    shift1_4 = vdupq_n_s32(shift1);
+
+    for (i = 0; i < frameCount4; ++i) {
+        uint32x4_t left;
+        uint32x4_t side;
+        uint32x4_t right;
+
+        left  = vshlq_u32(vld1q_u32(pInputSamples0U32 + i*4), shift0_4);
+        side  = vshlq_u32(vld1q_u32(pInputSamples1U32 + i*4), shift1_4);
+        right = vsubq_u32(left, side);
+
+        left  = vshrq_n_u32(left,  16);
+        right = vshrq_n_u32(right, 16);
+
+        drflac__vst2q_u16((drflac_uint16*)pOutputSamples + i*8, vzip_u16(vmovn_u32(left), vmovn_u32(right)));
+    }
+
+    for (i = (frameCount4 << 2); i < frameCount; ++i) {
+        drflac_uint32 left  = pInputSamples0U32[i] << shift0;
+        drflac_uint32 side  = pInputSamples1U32[i] << shift1;
+        drflac_uint32 right = left - side;
+
+        left  >>= 16;
+        right >>= 16;
+
+        pOutputSamples[i*2+0] = (drflac_int16)left;
+        pOutputSamples[i*2+1] = (drflac_int16)right;
+    }
+}
+#endif
+
+static DRFLAC_INLINE void drflac_read_pcm_frames_s16__decode_left_side(drflac* pFlac, drflac_uint64 frameCount, drflac_uint32 unusedBitsPerSample, const drflac_int32* pInputSamples0, const drflac_int32* pInputSamples1, drflac_int16* pOutputSamples)
+{
+#if defined(DRFLAC_SUPPORT_SSE2)
+    if (drflac__gIsSSE2Supported && pFlac->bitsPerSample <= 24) {
+        drflac_read_pcm_frames_s16__decode_left_side__sse2(pFlac, frameCount, unusedBitsPerSample, pInputSamples0, pInputSamples1, pOutputSamples);
+    } else
+#elif defined(DRFLAC_SUPPORT_NEON)
+    if (drflac__gIsNEONSupported && pFlac->bitsPerSample <= 24) {
+        drflac_read_pcm_frames_s16__decode_left_side__neon(pFlac, frameCount, unusedBitsPerSample, pInputSamples0, pInputSamples1, pOutputSamples);
+    } else
+#endif
+    {
+        /* Scalar fallback. */
+#if 0
+        drflac_read_pcm_frames_s16__decode_left_side__reference(pFlac, frameCount, unusedBitsPerSample, pInputSamples0, pInputSamples1, pOutputSamples);
+#else
+        drflac_read_pcm_frames_s16__decode_left_side__scalar(pFlac, frameCount, unusedBitsPerSample, pInputSamples0, pInputSamples1, pOutputSamples);
+#endif
+    }
+}
+
+
+#if 0
+static DRFLAC_INLINE void drflac_read_pcm_frames_s16__decode_right_side__reference(drflac* pFlac, drflac_uint64 frameCount, drflac_uint32 unusedBitsPerSample, const drflac_int32* pInputSamples0, const drflac_int32* pInputSamples1, drflac_int16* pOutputSamples)
+{
+    drflac_uint64 i;
+    for (i = 0; i < frameCount; ++i) {
+        drflac_uint32 side  = (drflac_uint32)pInputSamples0[i] << (unusedBitsPerSample + pFlac->currentFLACFrame.subframes[0].wastedBitsPerSample);
+        drflac_uint32 right = (drflac_uint32)pInputSamples1[i] << (unusedBitsPerSample + pFlac->currentFLACFrame.subframes[1].wastedBitsPerSample);
+        drflac_uint32 left  = right + side;
+
+        left  >>= 16;
+        right >>= 16;
+
+        pOutputSamples[i*2+0] = (drflac_int16)left;
+        pOutputSamples[i*2+1] = (drflac_int16)right;
+    }
+}
+#endif
+
+static DRFLAC_INLINE void drflac_read_pcm_frames_s16__decode_right_side__scalar(drflac* pFlac, drflac_uint64 frameCount, drflac_uint32 unusedBitsPerSample, const drflac_int32* pInputSamples0, const drflac_int32* pInputSamples1, drflac_int16* pOutputSamples)
+{
+    drflac_uint64 i;
+    drflac_uint64 frameCount4 = frameCount >> 2;
+    const drflac_uint32* pInputSamples0U32 = (const drflac_uint32*)pInputSamples0;
+    const drflac_uint32* pInputSamples1U32 = (const drflac_uint32*)pInputSamples1;
+    drflac_uint32 shift0 = unusedBitsPerSample + pFlac->currentFLACFrame.subframes[0].wastedBitsPerSample;
+    drflac_uint32 shift1 = unusedBitsPerSample + pFlac->currentFLACFrame.subframes[1].wastedBitsPerSample;
+
+    for (i = 0; i < frameCount4; ++i) {
+        drflac_uint32 side0  = pInputSamples0U32[i*4+0] << shift0;
+        drflac_uint32 side1  = pInputSamples0U32[i*4+1] << shift0;
+        drflac_uint32 side2  = pInputSamples0U32[i*4+2] << shift0;
+        drflac_uint32 side3  = pInputSamples0U32[i*4+3] << shift0;
+
+        drflac_uint32 right0 = pInputSamples1U32[i*4+0] << shift1;
+        drflac_uint32 right1 = pInputSamples1U32[i*4+1] << shift1;
+        drflac_uint32 right2 = pInputSamples1U32[i*4+2] << shift1;
+        drflac_uint32 right3 = pInputSamples1U32[i*4+3] << shift1;
+
+        drflac_uint32 left0 = right0 + side0;
+        drflac_uint32 left1 = right1 + side1;
+        drflac_uint32 left2 = right2 + side2;
+        drflac_uint32 left3 = right3 + side3;
+
+        left0  >>= 16;
+        left1  >>= 16;
+        left2  >>= 16;
+        left3  >>= 16;
+
+        right0 >>= 16;
+        right1 >>= 16;
+        right2 >>= 16;
+        right3 >>= 16;
+
+        pOutputSamples[i*8+0] = (drflac_int16)left0;
+        pOutputSamples[i*8+1] = (drflac_int16)right0;
+        pOutputSamples[i*8+2] = (drflac_int16)left1;
+        pOutputSamples[i*8+3] = (drflac_int16)right1;
+        pOutputSamples[i*8+4] = (drflac_int16)left2;
+        pOutputSamples[i*8+5] = (drflac_int16)right2;
+        pOutputSamples[i*8+6] = (drflac_int16)left3;
+        pOutputSamples[i*8+7] = (drflac_int16)right3;
+    }
+
+    for (i = (frameCount4 << 2); i < frameCount; ++i) {
+        drflac_uint32 side  = pInputSamples0U32[i] << shift0;
+        drflac_uint32 right = pInputSamples1U32[i] << shift1;
+        drflac_uint32 left  = right + side;
+
+        left  >>= 16;
+        right >>= 16;
+
+        pOutputSamples[i*2+0] = (drflac_int16)left;
+        pOutputSamples[i*2+1] = (drflac_int16)right;
+    }
+}
+
+#if defined(DRFLAC_SUPPORT_SSE2)
+static DRFLAC_INLINE void drflac_read_pcm_frames_s16__decode_right_side__sse2(drflac* pFlac, drflac_uint64 frameCount, drflac_uint32 unusedBitsPerSample, const drflac_int32* pInputSamples0, const drflac_int32* pInputSamples1, drflac_int16* pOutputSamples)
+{
+    drflac_uint64 i;
+    drflac_uint64 frameCount4 = frameCount >> 2;
+    const drflac_uint32* pInputSamples0U32 = (const drflac_uint32*)pInputSamples0;
+    const drflac_uint32* pInputSamples1U32 = (const drflac_uint32*)pInputSamples1;
+    drflac_uint32 shift0 = unusedBitsPerSample + pFlac->currentFLACFrame.subframes[0].wastedBitsPerSample;
+    drflac_uint32 shift1 = unusedBitsPerSample + pFlac->currentFLACFrame.subframes[1].wastedBitsPerSample;
+
+    DRFLAC_ASSERT(pFlac->bitsPerSample <= 24);
+
+    for (i = 0; i < frameCount4; ++i) {
+        __m128i side  = _mm_slli_epi32(_mm_loadu_si128((const __m128i*)pInputSamples0 + i), shift0);
+        __m128i right = _mm_slli_epi32(_mm_loadu_si128((const __m128i*)pInputSamples1 + i), shift1);
+        __m128i left  = _mm_add_epi32(right, side);
+
+        left  = _mm_srai_epi32(left,  16);
+        right = _mm_srai_epi32(right, 16);
+
+        _mm_storeu_si128((__m128i*)(pOutputSamples + i*8), drflac__mm_packs_interleaved_epi32(left, right));
+    }
+
+    for (i = (frameCount4 << 2); i < frameCount; ++i) {
+        drflac_uint32 side  = pInputSamples0U32[i] << shift0;
+        drflac_uint32 right = pInputSamples1U32[i] << shift1;
+        drflac_uint32 left  = right + side;
+
+        left  >>= 16;
+        right >>= 16;
+
+        pOutputSamples[i*2+0] = (drflac_int16)left;
+        pOutputSamples[i*2+1] = (drflac_int16)right;
+    }
+}
+#endif
+
+#if defined(DRFLAC_SUPPORT_NEON)
+static DRFLAC_INLINE void drflac_read_pcm_frames_s16__decode_right_side__neon(drflac* pFlac, drflac_uint64 frameCount, drflac_uint32 unusedBitsPerSample, const drflac_int32* pInputSamples0, const drflac_int32* pInputSamples1, drflac_int16* pOutputSamples)
+{
+    drflac_uint64 i;
+    drflac_uint64 frameCount4 = frameCount >> 2;
+    const drflac_uint32* pInputSamples0U32 = (const drflac_uint32*)pInputSamples0;
+    const drflac_uint32* pInputSamples1U32 = (const drflac_uint32*)pInputSamples1;
+    drflac_uint32 shift0 = unusedBitsPerSample + pFlac->currentFLACFrame.subframes[0].wastedBitsPerSample;
+    drflac_uint32 shift1 = unusedBitsPerSample + pFlac->currentFLACFrame.subframes[1].wastedBitsPerSample;
+    int32x4_t shift0_4;
+    int32x4_t shift1_4;
+
+    DRFLAC_ASSERT(pFlac->bitsPerSample <= 24);
+
+    shift0_4 = vdupq_n_s32(shift0);
+    shift1_4 = vdupq_n_s32(shift1);
+
+    for (i = 0; i < frameCount4; ++i) {
+        uint32x4_t side;
+        uint32x4_t right;
+        uint32x4_t left;
+
+        side  = vshlq_u32(vld1q_u32(pInputSamples0U32 + i*4), shift0_4);
+        right = vshlq_u32(vld1q_u32(pInputSamples1U32 + i*4), shift1_4);
+        left  = vaddq_u32(right, side);
+
+        left  = vshrq_n_u32(left,  16);
+        right = vshrq_n_u32(right, 16);
+
+        drflac__vst2q_u16((drflac_uint16*)pOutputSamples + i*8, vzip_u16(vmovn_u32(left), vmovn_u32(right)));
+    }
+
+    for (i = (frameCount4 << 2); i < frameCount; ++i) {
+        drflac_uint32 side  = pInputSamples0U32[i] << shift0;
+        drflac_uint32 right = pInputSamples1U32[i] << shift1;
+        drflac_uint32 left  = right + side;
+
+        left  >>= 16;
+        right >>= 16;
+
+        pOutputSamples[i*2+0] = (drflac_int16)left;
+        pOutputSamples[i*2+1] = (drflac_int16)right;
+    }
+}
+#endif
+
+static DRFLAC_INLINE void drflac_read_pcm_frames_s16__decode_right_side(drflac* pFlac, drflac_uint64 frameCount, drflac_uint32 unusedBitsPerSample, const drflac_int32* pInputSamples0, const drflac_int32* pInputSamples1, drflac_int16* pOutputSamples)
+{
+#if defined(DRFLAC_SUPPORT_SSE2)
+    if (drflac__gIsSSE2Supported && pFlac->bitsPerSample <= 24) {
+        drflac_read_pcm_frames_s16__decode_right_side__sse2(pFlac, frameCount, unusedBitsPerSample, pInputSamples0, pInputSamples1, pOutputSamples);
+    } else
+#elif defined(DRFLAC_SUPPORT_NEON)
+    if (drflac__gIsNEONSupported && pFlac->bitsPerSample <= 24) {
+        drflac_read_pcm_frames_s16__decode_right_side__neon(pFlac, frameCount, unusedBitsPerSample, pInputSamples0, pInputSamples1, pOutputSamples);
+    } else
+#endif
+    {
+        /* Scalar fallback. */
+#if 0
+        drflac_read_pcm_frames_s16__decode_right_side__reference(pFlac, frameCount, unusedBitsPerSample, pInputSamples0, pInputSamples1, pOutputSamples);
+#else
+        drflac_read_pcm_frames_s16__decode_right_side__scalar(pFlac, frameCount, unusedBitsPerSample, pInputSamples0, pInputSamples1, pOutputSamples);
+#endif
+    }
+}
+
+
+#if 0
+static DRFLAC_INLINE void drflac_read_pcm_frames_s16__decode_mid_side__reference(drflac* pFlac, drflac_uint64 frameCount, drflac_uint32 unusedBitsPerSample, const drflac_int32* pInputSamples0, const drflac_int32* pInputSamples1, drflac_int16* pOutputSamples)
+{
+    for (drflac_uint64 i = 0; i < frameCount; ++i) {
+        drflac_uint32 mid  = (drflac_uint32)pInputSamples0[i] << pFlac->currentFLACFrame.subframes[0].wastedBitsPerSample;
+        drflac_uint32 side = (drflac_uint32)pInputSamples1[i] << pFlac->currentFLACFrame.subframes[1].wastedBitsPerSample;
+
+        mid = (mid << 1) | (side & 0x01);
+
+        pOutputSamples[i*2+0] = (drflac_int16)(((drflac_uint32)((drflac_int32)(mid + side) >> 1) << unusedBitsPerSample) >> 16);
+        pOutputSamples[i*2+1] = (drflac_int16)(((drflac_uint32)((drflac_int32)(mid - side) >> 1) << unusedBitsPerSample) >> 16);
+    }
+}
+#endif
+
+static DRFLAC_INLINE void drflac_read_pcm_frames_s16__decode_mid_side__scalar(drflac* pFlac, drflac_uint64 frameCount, drflac_uint32 unusedBitsPerSample, const drflac_int32* pInputSamples0, const drflac_int32* pInputSamples1, drflac_int16* pOutputSamples)
+{
+    drflac_uint64 i;
+    drflac_uint64 frameCount4 = frameCount >> 2;
+    const drflac_uint32* pInputSamples0U32 = (const drflac_uint32*)pInputSamples0;
+    const drflac_uint32* pInputSamples1U32 = (const drflac_uint32*)pInputSamples1;
+    drflac_uint32 shift = unusedBitsPerSample;
+
+    if (shift > 0) {
+        shift -= 1;
+        for (i = 0; i < frameCount4; ++i) {
+            drflac_uint32 temp0L;
+            drflac_uint32 temp1L;
+            drflac_uint32 temp2L;
+            drflac_uint32 temp3L;
+            drflac_uint32 temp0R;
+            drflac_uint32 temp1R;
+            drflac_uint32 temp2R;
+            drflac_uint32 temp3R;
+
+            drflac_uint32 mid0  = pInputSamples0U32[i*4+0] << pFlac->currentFLACFrame.subframes[0].wastedBitsPerSample;
+            drflac_uint32 mid1  = pInputSamples0U32[i*4+1] << pFlac->currentFLACFrame.subframes[0].wastedBitsPerSample;
+            drflac_uint32 mid2  = pInputSamples0U32[i*4+2] << pFlac->currentFLACFrame.subframes[0].wastedBitsPerSample;
+            drflac_uint32 mid3  = pInputSamples0U32[i*4+3] << pFlac->currentFLACFrame.subframes[0].wastedBitsPerSample;
+
+            drflac_uint32 side0 = pInputSamples1U32[i*4+0] << pFlac->currentFLACFrame.subframes[1].wastedBitsPerSample;
+            drflac_uint32 side1 = pInputSamples1U32[i*4+1] << pFlac->currentFLACFrame.subframes[1].wastedBitsPerSample;
+            drflac_uint32 side2 = pInputSamples1U32[i*4+2] << pFlac->currentFLACFrame.subframes[1].wastedBitsPerSample;
+            drflac_uint32 side3 = pInputSamples1U32[i*4+3] << pFlac->currentFLACFrame.subframes[1].wastedBitsPerSample;
+
+            mid0 = (mid0 << 1) | (side0 & 0x01);
+            mid1 = (mid1 << 1) | (side1 & 0x01);
+            mid2 = (mid2 << 1) | (side2 & 0x01);
+            mid3 = (mid3 << 1) | (side3 & 0x01);
+
+            temp0L = (mid0 + side0) << shift;
+            temp1L = (mid1 + side1) << shift;
+            temp2L = (mid2 + side2) << shift;
+            temp3L = (mid3 + side3) << shift;
+
+            temp0R = (mid0 - side0) << shift;
+            temp1R = (mid1 - side1) << shift;
+            temp2R = (mid2 - side2) << shift;
+            temp3R = (mid3 - side3) << shift;
+
+            temp0L >>= 16;
+            temp1L >>= 16;
+            temp2L >>= 16;
+            temp3L >>= 16;
+
+            temp0R >>= 16;
+            temp1R >>= 16;
+            temp2R >>= 16;
+            temp3R >>= 16;
+
+            pOutputSamples[i*8+0] = (drflac_int16)temp0L;
+            pOutputSamples[i*8+1] = (drflac_int16)temp0R;
+            pOutputSamples[i*8+2] = (drflac_int16)temp1L;
+            pOutputSamples[i*8+3] = (drflac_int16)temp1R;
+            pOutputSamples[i*8+4] = (drflac_int16)temp2L;
+            pOutputSamples[i*8+5] = (drflac_int16)temp2R;
+            pOutputSamples[i*8+6] = (drflac_int16)temp3L;
+            pOutputSamples[i*8+7] = (drflac_int16)temp3R;
+        }
+    } else {
+        for (i = 0; i < frameCount4; ++i) {
+            drflac_uint32 temp0L;
+            drflac_uint32 temp1L;
+            drflac_uint32 temp2L;
+            drflac_uint32 temp3L;
+            drflac_uint32 temp0R;
+            drflac_uint32 temp1R;
+            drflac_uint32 temp2R;
+            drflac_uint32 temp3R;
+
+            drflac_uint32 mid0  = pInputSamples0U32[i*4+0] << pFlac->currentFLACFrame.subframes[0].wastedBitsPerSample;
+            drflac_uint32 mid1  = pInputSamples0U32[i*4+1] << pFlac->currentFLACFrame.subframes[0].wastedBitsPerSample;
+            drflac_uint32 mid2  = pInputSamples0U32[i*4+2] << pFlac->currentFLACFrame.subframes[0].wastedBitsPerSample;
+            drflac_uint32 mid3  = pInputSamples0U32[i*4+3] << pFlac->currentFLACFrame.subframes[0].wastedBitsPerSample;
+
+            drflac_uint32 side0 = pInputSamples1U32[i*4+0] << pFlac->currentFLACFrame.subframes[1].wastedBitsPerSample;
+            drflac_uint32 side1 = pInputSamples1U32[i*4+1] << pFlac->currentFLACFrame.subframes[1].wastedBitsPerSample;
+            drflac_uint32 side2 = pInputSamples1U32[i*4+2] << pFlac->currentFLACFrame.subframes[1].wastedBitsPerSample;
+            drflac_uint32 side3 = pInputSamples1U32[i*4+3] << pFlac->currentFLACFrame.subframes[1].wastedBitsPerSample;
+
+            mid0 = (mid0 << 1) | (side0 & 0x01);
+            mid1 = (mid1 << 1) | (side1 & 0x01);
+            mid2 = (mid2 << 1) | (side2 & 0x01);
+            mid3 = (mid3 << 1) | (side3 & 0x01);
+
+            temp0L = ((drflac_int32)(mid0 + side0) >> 1);
+            temp1L = ((drflac_int32)(mid1 + side1) >> 1);
+            temp2L = ((drflac_int32)(mid2 + side2) >> 1);
+            temp3L = ((drflac_int32)(mid3 + side3) >> 1);
+
+            temp0R = ((drflac_int32)(mid0 - side0) >> 1);
+            temp1R = ((drflac_int32)(mid1 - side1) >> 1);
+            temp2R = ((drflac_int32)(mid2 - side2) >> 1);
+            temp3R = ((drflac_int32)(mid3 - side3) >> 1);
+
+            temp0L >>= 16;
+            temp1L >>= 16;
+            temp2L >>= 16;
+            temp3L >>= 16;
+
+            temp0R >>= 16;
+            temp1R >>= 16;
+            temp2R >>= 16;
+            temp3R >>= 16;
+
+            pOutputSamples[i*8+0] = (drflac_int16)temp0L;
+            pOutputSamples[i*8+1] = (drflac_int16)temp0R;
+            pOutputSamples[i*8+2] = (drflac_int16)temp1L;
+            pOutputSamples[i*8+3] = (drflac_int16)temp1R;
+            pOutputSamples[i*8+4] = (drflac_int16)temp2L;
+            pOutputSamples[i*8+5] = (drflac_int16)temp2R;
+            pOutputSamples[i*8+6] = (drflac_int16)temp3L;
+            pOutputSamples[i*8+7] = (drflac_int16)temp3R;
+        }
+    }
+
+    for (i = (frameCount4 << 2); i < frameCount; ++i) {
+        drflac_uint32 mid  = pInputSamples0U32[i] << pFlac->currentFLACFrame.subframes[0].wastedBitsPerSample;
+        drflac_uint32 side = pInputSamples1U32[i] << pFlac->currentFLACFrame.subframes[1].wastedBitsPerSample;
+
+        mid = (mid << 1) | (side & 0x01);
+
+        pOutputSamples[i*2+0] = (drflac_int16)(((drflac_uint32)((drflac_int32)(mid + side) >> 1) << unusedBitsPerSample) >> 16);
+        pOutputSamples[i*2+1] = (drflac_int16)(((drflac_uint32)((drflac_int32)(mid - side) >> 1) << unusedBitsPerSample) >> 16);
+    }
+}
+
+#if defined(DRFLAC_SUPPORT_SSE2)
+static DRFLAC_INLINE void drflac_read_pcm_frames_s16__decode_mid_side__sse2(drflac* pFlac, drflac_uint64 frameCount, drflac_uint32 unusedBitsPerSample, const drflac_int32* pInputSamples0, const drflac_int32* pInputSamples1, drflac_int16* pOutputSamples)
+{
+    drflac_uint64 i;
+    drflac_uint64 frameCount4 = frameCount >> 2;
+    const drflac_uint32* pInputSamples0U32 = (const drflac_uint32*)pInputSamples0;
+    const drflac_uint32* pInputSamples1U32 = (const drflac_uint32*)pInputSamples1;
+    drflac_uint32 shift = unusedBitsPerSample;
+
+    DRFLAC_ASSERT(pFlac->bitsPerSample <= 24);
+
+    if (shift == 0) {
+        for (i = 0; i < frameCount4; ++i) {
+            __m128i mid;
+            __m128i side;
+            __m128i left;
+            __m128i right;
+
+            mid   = _mm_slli_epi32(_mm_loadu_si128((const __m128i*)pInputSamples0 + i), pFlac->currentFLACFrame.subframes[0].wastedBitsPerSample);
+            side  = _mm_slli_epi32(_mm_loadu_si128((const __m128i*)pInputSamples1 + i), pFlac->currentFLACFrame.subframes[1].wastedBitsPerSample);
+
+            mid   = _mm_or_si128(_mm_slli_epi32(mid, 1), _mm_and_si128(side, _mm_set1_epi32(0x01)));
+
+            left  = _mm_srai_epi32(_mm_add_epi32(mid, side), 1);
+            right = _mm_srai_epi32(_mm_sub_epi32(mid, side), 1);
+
+            left  = _mm_srai_epi32(left,  16);
+            right = _mm_srai_epi32(right, 16);
+
+            _mm_storeu_si128((__m128i*)(pOutputSamples + i*8), drflac__mm_packs_interleaved_epi32(left, right));
+        }
+
+        for (i = (frameCount4 << 2); i < frameCount; ++i) {
+            drflac_uint32 mid  = pInputSamples0U32[i] << pFlac->currentFLACFrame.subframes[0].wastedBitsPerSample;
+            drflac_uint32 side = pInputSamples1U32[i] << pFlac->currentFLACFrame.subframes[1].wastedBitsPerSample;
+
+            mid = (mid << 1) | (side & 0x01);
+
+            pOutputSamples[i*2+0] = (drflac_int16)(((drflac_int32)(mid + side) >> 1) >> 16);
+            pOutputSamples[i*2+1] = (drflac_int16)(((drflac_int32)(mid - side) >> 1) >> 16);
+        }
+    } else {
+        shift -= 1;
+        for (i = 0; i < frameCount4; ++i) {
+            __m128i mid;
+            __m128i side;
+            __m128i left;
+            __m128i right;
+
+            mid   = _mm_slli_epi32(_mm_loadu_si128((const __m128i*)pInputSamples0 + i), pFlac->currentFLACFrame.subframes[0].wastedBitsPerSample);
+            side  = _mm_slli_epi32(_mm_loadu_si128((const __m128i*)pInputSamples1 + i), pFlac->currentFLACFrame.subframes[1].wastedBitsPerSample);
+
+            mid   = _mm_or_si128(_mm_slli_epi32(mid, 1), _mm_and_si128(side, _mm_set1_epi32(0x01)));
+
+            left  = _mm_slli_epi32(_mm_add_epi32(mid, side), shift);
+            right = _mm_slli_epi32(_mm_sub_epi32(mid, side), shift);
+
+            left  = _mm_srai_epi32(left,  16);
+            right = _mm_srai_epi32(right, 16);
+
+            _mm_storeu_si128((__m128i*)(pOutputSamples + i*8), drflac__mm_packs_interleaved_epi32(left, right));
+        }
+
+        for (i = (frameCount4 << 2); i < frameCount; ++i) {
+            drflac_uint32 mid  = pInputSamples0U32[i] << pFlac->currentFLACFrame.subframes[0].wastedBitsPerSample;
+            drflac_uint32 side = pInputSamples1U32[i] << pFlac->currentFLACFrame.subframes[1].wastedBitsPerSample;
+
+            mid = (mid << 1) | (side & 0x01);
+
+            pOutputSamples[i*2+0] = (drflac_int16)(((mid + side) << shift) >> 16);
+            pOutputSamples[i*2+1] = (drflac_int16)(((mid - side) << shift) >> 16);
+        }
+    }
+}
+#endif
+
+#if defined(DRFLAC_SUPPORT_NEON)
+static DRFLAC_INLINE void drflac_read_pcm_frames_s16__decode_mid_side__neon(drflac* pFlac, drflac_uint64 frameCount, drflac_uint32 unusedBitsPerSample, const drflac_int32* pInputSamples0, const drflac_int32* pInputSamples1, drflac_int16* pOutputSamples)
+{
+    drflac_uint64 i;
+    drflac_uint64 frameCount4 = frameCount >> 2;
+    const drflac_uint32* pInputSamples0U32 = (const drflac_uint32*)pInputSamples0;
+    const drflac_uint32* pInputSamples1U32 = (const drflac_uint32*)pInputSamples1;
+    drflac_uint32 shift = unusedBitsPerSample;
+    int32x4_t wbpsShift0_4; /* wbps = Wasted Bits Per Sample */
+    int32x4_t wbpsShift1_4; /* wbps = Wasted Bits Per Sample */
+
+    DRFLAC_ASSERT(pFlac->bitsPerSample <= 24);
+
+    wbpsShift0_4 = vdupq_n_s32(pFlac->currentFLACFrame.subframes[0].wastedBitsPerSample);
+    wbpsShift1_4 = vdupq_n_s32(pFlac->currentFLACFrame.subframes[1].wastedBitsPerSample);
+
+    if (shift == 0) {
+        for (i = 0; i < frameCount4; ++i) {
+            uint32x4_t mid;
+            uint32x4_t side;
+            int32x4_t left;
+            int32x4_t right;
+
+            mid   = vshlq_u32(vld1q_u32(pInputSamples0U32 + i*4), wbpsShift0_4);
+            side  = vshlq_u32(vld1q_u32(pInputSamples1U32 + i*4), wbpsShift1_4);
+
+            mid   = vorrq_u32(vshlq_n_u32(mid, 1), vandq_u32(side, vdupq_n_u32(1)));
+
+            left  = vshrq_n_s32(vreinterpretq_s32_u32(vaddq_u32(mid, side)), 1);
+            right = vshrq_n_s32(vreinterpretq_s32_u32(vsubq_u32(mid, side)), 1);
+
+            left  = vshrq_n_s32(left,  16);
+            right = vshrq_n_s32(right, 16);
+
+            drflac__vst2q_s16(pOutputSamples + i*8, vzip_s16(vmovn_s32(left), vmovn_s32(right)));
+        }
+
+        for (i = (frameCount4 << 2); i < frameCount; ++i) {
+            drflac_uint32 mid  = pInputSamples0U32[i] << pFlac->currentFLACFrame.subframes[0].wastedBitsPerSample;
+            drflac_uint32 side = pInputSamples1U32[i] << pFlac->currentFLACFrame.subframes[1].wastedBitsPerSample;
+
+            mid = (mid << 1) | (side & 0x01);
+
+            pOutputSamples[i*2+0] = (drflac_int16)(((drflac_int32)(mid + side) >> 1) >> 16);
+            pOutputSamples[i*2+1] = (drflac_int16)(((drflac_int32)(mid - side) >> 1) >> 16);
+        }
+    } else {
+        int32x4_t shift4;
+
+        shift -= 1;
+        shift4 = vdupq_n_s32(shift);
+
+        for (i = 0; i < frameCount4; ++i) {
+            uint32x4_t mid;
+            uint32x4_t side;
+            int32x4_t left;
+            int32x4_t right;
+
+            mid   = vshlq_u32(vld1q_u32(pInputSamples0U32 + i*4), wbpsShift0_4);
+            side  = vshlq_u32(vld1q_u32(pInputSamples1U32 + i*4), wbpsShift1_4);
+
+            mid   = vorrq_u32(vshlq_n_u32(mid, 1), vandq_u32(side, vdupq_n_u32(1)));
+
+            left  = vreinterpretq_s32_u32(vshlq_u32(vaddq_u32(mid, side), shift4));
+            right = vreinterpretq_s32_u32(vshlq_u32(vsubq_u32(mid, side), shift4));
+
+            left  = vshrq_n_s32(left,  16);
+            right = vshrq_n_s32(right, 16);
+
+            drflac__vst2q_s16(pOutputSamples + i*8, vzip_s16(vmovn_s32(left), vmovn_s32(right)));
+        }
+
+        for (i = (frameCount4 << 2); i < frameCount; ++i) {
+            drflac_uint32 mid  = pInputSamples0U32[i] << pFlac->currentFLACFrame.subframes[0].wastedBitsPerSample;
+            drflac_uint32 side = pInputSamples1U32[i] << pFlac->currentFLACFrame.subframes[1].wastedBitsPerSample;
+
+            mid = (mid << 1) | (side & 0x01);
+
+            pOutputSamples[i*2+0] = (drflac_int16)(((mid + side) << shift) >> 16);
+            pOutputSamples[i*2+1] = (drflac_int16)(((mid - side) << shift) >> 16);
+        }
+    }
+}
+#endif
+
+static DRFLAC_INLINE void drflac_read_pcm_frames_s16__decode_mid_side(drflac* pFlac, drflac_uint64 frameCount, drflac_uint32 unusedBitsPerSample, const drflac_int32* pInputSamples0, const drflac_int32* pInputSamples1, drflac_int16* pOutputSamples)
+{
+#if defined(DRFLAC_SUPPORT_SSE2)
+    if (drflac__gIsSSE2Supported && pFlac->bitsPerSample <= 24) {
+        drflac_read_pcm_frames_s16__decode_mid_side__sse2(pFlac, frameCount, unusedBitsPerSample, pInputSamples0, pInputSamples1, pOutputSamples);
+    } else
+#elif defined(DRFLAC_SUPPORT_NEON)
+    if (drflac__gIsNEONSupported && pFlac->bitsPerSample <= 24) {
+        drflac_read_pcm_frames_s16__decode_mid_side__neon(pFlac, frameCount, unusedBitsPerSample, pInputSamples0, pInputSamples1, pOutputSamples);
+    } else
+#endif
+    {
+        /* Scalar fallback. */
+#if 0
+        drflac_read_pcm_frames_s16__decode_mid_side__reference(pFlac, frameCount, unusedBitsPerSample, pInputSamples0, pInputSamples1, pOutputSamples);
+#else
+        drflac_read_pcm_frames_s16__decode_mid_side__scalar(pFlac, frameCount, unusedBitsPerSample, pInputSamples0, pInputSamples1, pOutputSamples);
+#endif
+    }
+}
+
+
+#if 0
+static DRFLAC_INLINE void drflac_read_pcm_frames_s16__decode_independent_stereo__reference(drflac* pFlac, drflac_uint64 frameCount, drflac_uint32 unusedBitsPerSample, const drflac_int32* pInputSamples0, const drflac_int32* pInputSamples1, drflac_int16* pOutputSamples)
+{
+    for (drflac_uint64 i = 0; i < frameCount; ++i) {
+        pOutputSamples[i*2+0] = (drflac_int16)((drflac_int32)((drflac_uint32)pInputSamples0[i] << (unusedBitsPerSample + pFlac->currentFLACFrame.subframes[0].wastedBitsPerSample)) >> 16);
+        pOutputSamples[i*2+1] = (drflac_int16)((drflac_int32)((drflac_uint32)pInputSamples1[i] << (unusedBitsPerSample + pFlac->currentFLACFrame.subframes[1].wastedBitsPerSample)) >> 16);
+    }
+}
+#endif
+
+static DRFLAC_INLINE void drflac_read_pcm_frames_s16__decode_independent_stereo__scalar(drflac* pFlac, drflac_uint64 frameCount, drflac_uint32 unusedBitsPerSample, const drflac_int32* pInputSamples0, const drflac_int32* pInputSamples1, drflac_int16* pOutputSamples)
+{
+    drflac_uint64 i;
+    drflac_uint64 frameCount4 = frameCount >> 2;
+    const drflac_uint32* pInputSamples0U32 = (const drflac_uint32*)pInputSamples0;
+    const drflac_uint32* pInputSamples1U32 = (const drflac_uint32*)pInputSamples1;
+    drflac_uint32 shift0 = unusedBitsPerSample + pFlac->currentFLACFrame.subframes[0].wastedBitsPerSample;
+    drflac_uint32 shift1 = unusedBitsPerSample + pFlac->currentFLACFrame.subframes[1].wastedBitsPerSample;
+
+    for (i = 0; i < frameCount4; ++i) {
+        drflac_uint32 tempL0 = pInputSamples0U32[i*4+0] << shift0;
+        drflac_uint32 tempL1 = pInputSamples0U32[i*4+1] << shift0;
+        drflac_uint32 tempL2 = pInputSamples0U32[i*4+2] << shift0;
+        drflac_uint32 tempL3 = pInputSamples0U32[i*4+3] << shift0;
+
+        drflac_uint32 tempR0 = pInputSamples1U32[i*4+0] << shift1;
+        drflac_uint32 tempR1 = pInputSamples1U32[i*4+1] << shift1;
+        drflac_uint32 tempR2 = pInputSamples1U32[i*4+2] << shift1;
+        drflac_uint32 tempR3 = pInputSamples1U32[i*4+3] << shift1;
+
+        tempL0 >>= 16;
+        tempL1 >>= 16;
+        tempL2 >>= 16;
+        tempL3 >>= 16;
+
+        tempR0 >>= 16;
+        tempR1 >>= 16;
+        tempR2 >>= 16;
+        tempR3 >>= 16;
+
+        pOutputSamples[i*8+0] = (drflac_int16)tempL0;
+        pOutputSamples[i*8+1] = (drflac_int16)tempR0;
+        pOutputSamples[i*8+2] = (drflac_int16)tempL1;
+        pOutputSamples[i*8+3] = (drflac_int16)tempR1;
+        pOutputSamples[i*8+4] = (drflac_int16)tempL2;
+        pOutputSamples[i*8+5] = (drflac_int16)tempR2;
+        pOutputSamples[i*8+6] = (drflac_int16)tempL3;
+        pOutputSamples[i*8+7] = (drflac_int16)tempR3;
+    }
+
+    for (i = (frameCount4 << 2); i < frameCount; ++i) {
+        pOutputSamples[i*2+0] = (drflac_int16)((pInputSamples0U32[i] << shift0) >> 16);
+        pOutputSamples[i*2+1] = (drflac_int16)((pInputSamples1U32[i] << shift1) >> 16);
+    }
+}
+
+#if defined(DRFLAC_SUPPORT_SSE2)
+static DRFLAC_INLINE void drflac_read_pcm_frames_s16__decode_independent_stereo__sse2(drflac* pFlac, drflac_uint64 frameCount, drflac_uint32 unusedBitsPerSample, const drflac_int32* pInputSamples0, const drflac_int32* pInputSamples1, drflac_int16* pOutputSamples)
+{
+    drflac_uint64 i;
+    drflac_uint64 frameCount4 = frameCount >> 2;
+    const drflac_uint32* pInputSamples0U32 = (const drflac_uint32*)pInputSamples0;
+    const drflac_uint32* pInputSamples1U32 = (const drflac_uint32*)pInputSamples1;
+    drflac_uint32 shift0 = unusedBitsPerSample + pFlac->currentFLACFrame.subframes[0].wastedBitsPerSample;
+    drflac_uint32 shift1 = unusedBitsPerSample + pFlac->currentFLACFrame.subframes[1].wastedBitsPerSample;
+
+    for (i = 0; i < frameCount4; ++i) {
+        __m128i left  = _mm_slli_epi32(_mm_loadu_si128((const __m128i*)pInputSamples0 + i), shift0);
+        __m128i right = _mm_slli_epi32(_mm_loadu_si128((const __m128i*)pInputSamples1 + i), shift1);
+
+        left  = _mm_srai_epi32(left,  16);
+        right = _mm_srai_epi32(right, 16);
+
+        /* At this point we have results. We can now pack and interleave these into a single __m128i object and then store the in the output buffer. */
+        _mm_storeu_si128((__m128i*)(pOutputSamples + i*8), drflac__mm_packs_interleaved_epi32(left, right));
+    }
+
+    for (i = (frameCount4 << 2); i < frameCount; ++i) {
+        pOutputSamples[i*2+0] = (drflac_int16)((pInputSamples0U32[i] << shift0) >> 16);
+        pOutputSamples[i*2+1] = (drflac_int16)((pInputSamples1U32[i] << shift1) >> 16);
+    }
+}
+#endif
+
+#if defined(DRFLAC_SUPPORT_NEON)
+static DRFLAC_INLINE void drflac_read_pcm_frames_s16__decode_independent_stereo__neon(drflac* pFlac, drflac_uint64 frameCount, drflac_uint32 unusedBitsPerSample, const drflac_int32* pInputSamples0, const drflac_int32* pInputSamples1, drflac_int16* pOutputSamples)
+{
+    drflac_uint64 i;
+    drflac_uint64 frameCount4 = frameCount >> 2;
+    const drflac_uint32* pInputSamples0U32 = (const drflac_uint32*)pInputSamples0;
+    const drflac_uint32* pInputSamples1U32 = (const drflac_uint32*)pInputSamples1;
+    drflac_uint32 shift0 = unusedBitsPerSample + pFlac->currentFLACFrame.subframes[0].wastedBitsPerSample;
+    drflac_uint32 shift1 = unusedBitsPerSample + pFlac->currentFLACFrame.subframes[1].wastedBitsPerSample;
+
+    int32x4_t shift0_4 = vdupq_n_s32(shift0);
+    int32x4_t shift1_4 = vdupq_n_s32(shift1);
+
+    for (i = 0; i < frameCount4; ++i) {
+        int32x4_t left;
+        int32x4_t right;
+
+        left  = vreinterpretq_s32_u32(vshlq_u32(vld1q_u32(pInputSamples0U32 + i*4), shift0_4));
+        right = vreinterpretq_s32_u32(vshlq_u32(vld1q_u32(pInputSamples1U32 + i*4), shift1_4));
+
+        left  = vshrq_n_s32(left,  16);
+        right = vshrq_n_s32(right, 16);
+
+        drflac__vst2q_s16(pOutputSamples + i*8, vzip_s16(vmovn_s32(left), vmovn_s32(right)));
+    }
+
+    for (i = (frameCount4 << 2); i < frameCount; ++i) {
+        pOutputSamples[i*2+0] = (drflac_int16)((pInputSamples0U32[i] << shift0) >> 16);
+        pOutputSamples[i*2+1] = (drflac_int16)((pInputSamples1U32[i] << shift1) >> 16);
+    }
+}
+#endif
+
+static DRFLAC_INLINE void drflac_read_pcm_frames_s16__decode_independent_stereo(drflac* pFlac, drflac_uint64 frameCount, drflac_uint32 unusedBitsPerSample, const drflac_int32* pInputSamples0, const drflac_int32* pInputSamples1, drflac_int16* pOutputSamples)
+{
+#if defined(DRFLAC_SUPPORT_SSE2)
+    if (drflac__gIsSSE2Supported && pFlac->bitsPerSample <= 24) {
+        drflac_read_pcm_frames_s16__decode_independent_stereo__sse2(pFlac, frameCount, unusedBitsPerSample, pInputSamples0, pInputSamples1, pOutputSamples);
+    } else
+#elif defined(DRFLAC_SUPPORT_NEON)
+    if (drflac__gIsNEONSupported && pFlac->bitsPerSample <= 24) {
+        drflac_read_pcm_frames_s16__decode_independent_stereo__neon(pFlac, frameCount, unusedBitsPerSample, pInputSamples0, pInputSamples1, pOutputSamples);
+    } else
+#endif
+    {
+        /* Scalar fallback. */
+#if 0
+        drflac_read_pcm_frames_s16__decode_independent_stereo__reference(pFlac, frameCount, unusedBitsPerSample, pInputSamples0, pInputSamples1, pOutputSamples);
+#else
+        drflac_read_pcm_frames_s16__decode_independent_stereo__scalar(pFlac, frameCount, unusedBitsPerSample, pInputSamples0, pInputSamples1, pOutputSamples);
+#endif
+    }
+}
+
+DRFLAC_API drflac_uint64 drflac_read_pcm_frames_s16(drflac* pFlac, drflac_uint64 framesToRead, drflac_int16* pBufferOut)
+{
+    drflac_uint64 framesRead;
+    drflac_uint32 unusedBitsPerSample;
+
+    if (pFlac == NULL || framesToRead == 0) {
+        return 0;
+    }
+
+    if (pBufferOut == NULL) {
+        return drflac__seek_forward_by_pcm_frames(pFlac, framesToRead);
+    }
+
+    DRFLAC_ASSERT(pFlac->bitsPerSample <= 32);
+    unusedBitsPerSample = 32 - pFlac->bitsPerSample;
+
+    framesRead = 0;
+    while (framesToRead > 0) {
+        /* If we've run out of samples in this frame, go to the next. */
+        if (pFlac->currentFLACFrame.pcmFramesRemaining == 0) {
+            if (!drflac__read_and_decode_next_flac_frame(pFlac)) {
+                break;  /* Couldn't read the next frame, so just break from the loop and return. */
+            }
+        } else {
+            unsigned int channelCount = drflac__get_channel_count_from_channel_assignment(pFlac->currentFLACFrame.header.channelAssignment);
+            drflac_uint64 iFirstPCMFrame = pFlac->currentFLACFrame.header.blockSizeInPCMFrames - pFlac->currentFLACFrame.pcmFramesRemaining;
+            drflac_uint64 frameCountThisIteration = framesToRead;
+
+            if (frameCountThisIteration > pFlac->currentFLACFrame.pcmFramesRemaining) {
+                frameCountThisIteration = pFlac->currentFLACFrame.pcmFramesRemaining;
+            }
+
+            if (channelCount == 2) {
+                const drflac_int32* pDecodedSamples0 = pFlac->currentFLACFrame.subframes[0].pSamplesS32 + iFirstPCMFrame;
+                const drflac_int32* pDecodedSamples1 = pFlac->currentFLACFrame.subframes[1].pSamplesS32 + iFirstPCMFrame;
+
+                switch (pFlac->currentFLACFrame.header.channelAssignment)
+                {
+                    case DRFLAC_CHANNEL_ASSIGNMENT_LEFT_SIDE:
+                    {
+                        drflac_read_pcm_frames_s16__decode_left_side(pFlac, frameCountThisIteration, unusedBitsPerSample, pDecodedSamples0, pDecodedSamples1, pBufferOut);
+                    } break;
+
+                    case DRFLAC_CHANNEL_ASSIGNMENT_RIGHT_SIDE:
+                    {
+                        drflac_read_pcm_frames_s16__decode_right_side(pFlac, frameCountThisIteration, unusedBitsPerSample, pDecodedSamples0, pDecodedSamples1, pBufferOut);
+                    } break;
+
+                    case DRFLAC_CHANNEL_ASSIGNMENT_MID_SIDE:
+                    {
+                        drflac_read_pcm_frames_s16__decode_mid_side(pFlac, frameCountThisIteration, unusedBitsPerSample, pDecodedSamples0, pDecodedSamples1, pBufferOut);
+                    } break;
+
+                    case DRFLAC_CHANNEL_ASSIGNMENT_INDEPENDENT:
+                    default:
+                    {
+                        drflac_read_pcm_frames_s16__decode_independent_stereo(pFlac, frameCountThisIteration, unusedBitsPerSample, pDecodedSamples0, pDecodedSamples1, pBufferOut);
+                    } break;
+                }
+            } else {
+                /* Generic interleaving. */
+                drflac_uint64 i;
+                for (i = 0; i < frameCountThisIteration; ++i) {
+                    unsigned int j;
+                    for (j = 0; j < channelCount; ++j) {
+                        drflac_int32 sampleS32 = (drflac_int32)((drflac_uint32)(pFlac->currentFLACFrame.subframes[j].pSamplesS32[iFirstPCMFrame + i]) << (unusedBitsPerSample + pFlac->currentFLACFrame.subframes[j].wastedBitsPerSample));
+                        pBufferOut[(i*channelCount)+j] = (drflac_int16)(sampleS32 >> 16);
+                    }
+                }
+            }
+
+            framesRead                += frameCountThisIteration;
+            pBufferOut                += frameCountThisIteration * channelCount;
+            framesToRead              -= frameCountThisIteration;
+            pFlac->currentPCMFrame    += frameCountThisIteration;
+            pFlac->currentFLACFrame.pcmFramesRemaining -= (drflac_uint32)frameCountThisIteration;
+        }
+    }
+
+    return framesRead;
+}
+
+
+#if 0
+static DRFLAC_INLINE void drflac_read_pcm_frames_f32__decode_left_side__reference(drflac* pFlac, drflac_uint64 frameCount, drflac_uint32 unusedBitsPerSample, const drflac_int32* pInputSamples0, const drflac_int32* pInputSamples1, float* pOutputSamples)
+{
+    drflac_uint64 i;
+    for (i = 0; i < frameCount; ++i) {
+        drflac_uint32 left  = (drflac_uint32)pInputSamples0[i] << (unusedBitsPerSample + pFlac->currentFLACFrame.subframes[0].wastedBitsPerSample);
+        drflac_uint32 side  = (drflac_uint32)pInputSamples1[i] << (unusedBitsPerSample + pFlac->currentFLACFrame.subframes[1].wastedBitsPerSample);
+        drflac_uint32 right = left - side;
+
+        pOutputSamples[i*2+0] = (float)((drflac_int32)left  / 2147483648.0);
+        pOutputSamples[i*2+1] = (float)((drflac_int32)right / 2147483648.0);
+    }
+}
+#endif
+
+static DRFLAC_INLINE void drflac_read_pcm_frames_f32__decode_left_side__scalar(drflac* pFlac, drflac_uint64 frameCount, drflac_uint32 unusedBitsPerSample, const drflac_int32* pInputSamples0, const drflac_int32* pInputSamples1, float* pOutputSamples)
+{
+    drflac_uint64 i;
+    drflac_uint64 frameCount4 = frameCount >> 2;
+    const drflac_uint32* pInputSamples0U32 = (const drflac_uint32*)pInputSamples0;
+    const drflac_uint32* pInputSamples1U32 = (const drflac_uint32*)pInputSamples1;
+    drflac_uint32 shift0 = unusedBitsPerSample + pFlac->currentFLACFrame.subframes[0].wastedBitsPerSample;
+    drflac_uint32 shift1 = unusedBitsPerSample + pFlac->currentFLACFrame.subframes[1].wastedBitsPerSample;
+
+    float factor = 1 / 2147483648.0;
+
+    for (i = 0; i < frameCount4; ++i) {
+        drflac_uint32 left0 = pInputSamples0U32[i*4+0] << shift0;
+        drflac_uint32 left1 = pInputSamples0U32[i*4+1] << shift0;
+        drflac_uint32 left2 = pInputSamples0U32[i*4+2] << shift0;
+        drflac_uint32 left3 = pInputSamples0U32[i*4+3] << shift0;
+
+        drflac_uint32 side0 = pInputSamples1U32[i*4+0] << shift1;
+        drflac_uint32 side1 = pInputSamples1U32[i*4+1] << shift1;
+        drflac_uint32 side2 = pInputSamples1U32[i*4+2] << shift1;
+        drflac_uint32 side3 = pInputSamples1U32[i*4+3] << shift1;
+
+        drflac_uint32 right0 = left0 - side0;
+        drflac_uint32 right1 = left1 - side1;
+        drflac_uint32 right2 = left2 - side2;
+        drflac_uint32 right3 = left3 - side3;
+
+        pOutputSamples[i*8+0] = (drflac_int32)left0  * factor;
+        pOutputSamples[i*8+1] = (drflac_int32)right0 * factor;
+        pOutputSamples[i*8+2] = (drflac_int32)left1  * factor;
+        pOutputSamples[i*8+3] = (drflac_int32)right1 * factor;
+        pOutputSamples[i*8+4] = (drflac_int32)left2  * factor;
+        pOutputSamples[i*8+5] = (drflac_int32)right2 * factor;
+        pOutputSamples[i*8+6] = (drflac_int32)left3  * factor;
+        pOutputSamples[i*8+7] = (drflac_int32)right3 * factor;
+    }
+
+    for (i = (frameCount4 << 2); i < frameCount; ++i) {
+        drflac_uint32 left  = pInputSamples0U32[i] << shift0;
+        drflac_uint32 side  = pInputSamples1U32[i] << shift1;
+        drflac_uint32 right = left - side;
+
+        pOutputSamples[i*2+0] = (drflac_int32)left  * factor;
+        pOutputSamples[i*2+1] = (drflac_int32)right * factor;
+    }
+}
+
+#if defined(DRFLAC_SUPPORT_SSE2)
+static DRFLAC_INLINE void drflac_read_pcm_frames_f32__decode_left_side__sse2(drflac* pFlac, drflac_uint64 frameCount, drflac_uint32 unusedBitsPerSample, const drflac_int32* pInputSamples0, const drflac_int32* pInputSamples1, float* pOutputSamples)
+{
+    drflac_uint64 i;
+    drflac_uint64 frameCount4 = frameCount >> 2;
+    const drflac_uint32* pInputSamples0U32 = (const drflac_uint32*)pInputSamples0;
+    const drflac_uint32* pInputSamples1U32 = (const drflac_uint32*)pInputSamples1;
+    drflac_uint32 shift0 = (unusedBitsPerSample + pFlac->currentFLACFrame.subframes[0].wastedBitsPerSample) - 8;
+    drflac_uint32 shift1 = (unusedBitsPerSample + pFlac->currentFLACFrame.subframes[1].wastedBitsPerSample) - 8;
+    __m128 factor;
+
+    DRFLAC_ASSERT(pFlac->bitsPerSample <= 24);
+
+    factor = _mm_set1_ps(1.0f / 8388608.0f);
+
+    for (i = 0; i < frameCount4; ++i) {
+        __m128i left  = _mm_slli_epi32(_mm_loadu_si128((const __m128i*)pInputSamples0 + i), shift0);
+        __m128i side  = _mm_slli_epi32(_mm_loadu_si128((const __m128i*)pInputSamples1 + i), shift1);
+        __m128i right = _mm_sub_epi32(left, side);
+        __m128 leftf  = _mm_mul_ps(_mm_cvtepi32_ps(left),  factor);
+        __m128 rightf = _mm_mul_ps(_mm_cvtepi32_ps(right), factor);
+
+        _mm_storeu_ps(pOutputSamples + i*8 + 0, _mm_unpacklo_ps(leftf, rightf));
+        _mm_storeu_ps(pOutputSamples + i*8 + 4, _mm_unpackhi_ps(leftf, rightf));
+    }
+
+    for (i = (frameCount4 << 2); i < frameCount; ++i) {
+        drflac_uint32 left  = pInputSamples0U32[i] << shift0;
+        drflac_uint32 side  = pInputSamples1U32[i] << shift1;
+        drflac_uint32 right = left - side;
+
+        pOutputSamples[i*2+0] = (drflac_int32)left  / 8388608.0f;
+        pOutputSamples[i*2+1] = (drflac_int32)right / 8388608.0f;
+    }
+}
+#endif
+
+#if defined(DRFLAC_SUPPORT_NEON)
+static DRFLAC_INLINE void drflac_read_pcm_frames_f32__decode_left_side__neon(drflac* pFlac, drflac_uint64 frameCount, drflac_uint32 unusedBitsPerSample, const drflac_int32* pInputSamples0, const drflac_int32* pInputSamples1, float* pOutputSamples)
+{
+    drflac_uint64 i;
+    drflac_uint64 frameCount4 = frameCount >> 2;
+    const drflac_uint32* pInputSamples0U32 = (const drflac_uint32*)pInputSamples0;
+    const drflac_uint32* pInputSamples1U32 = (const drflac_uint32*)pInputSamples1;
+    drflac_uint32 shift0 = (unusedBitsPerSample + pFlac->currentFLACFrame.subframes[0].wastedBitsPerSample) - 8;
+    drflac_uint32 shift1 = (unusedBitsPerSample + pFlac->currentFLACFrame.subframes[1].wastedBitsPerSample) - 8;
+    float32x4_t factor4;
+    int32x4_t shift0_4;
+    int32x4_t shift1_4;
+
+    DRFLAC_ASSERT(pFlac->bitsPerSample <= 24);
+
+    factor4  = vdupq_n_f32(1.0f / 8388608.0f);
+    shift0_4 = vdupq_n_s32(shift0);
+    shift1_4 = vdupq_n_s32(shift1);
+
+    for (i = 0; i < frameCount4; ++i) {
+        uint32x4_t left;
+        uint32x4_t side;
+        uint32x4_t right;
+        float32x4_t leftf;
+        float32x4_t rightf;
+
+        left   = vshlq_u32(vld1q_u32(pInputSamples0U32 + i*4), shift0_4);
+        side   = vshlq_u32(vld1q_u32(pInputSamples1U32 + i*4), shift1_4);
+        right  = vsubq_u32(left, side);
+        leftf  = vmulq_f32(vcvtq_f32_s32(vreinterpretq_s32_u32(left)),  factor4);
+        rightf = vmulq_f32(vcvtq_f32_s32(vreinterpretq_s32_u32(right)), factor4);
+
+        drflac__vst2q_f32(pOutputSamples + i*8, vzipq_f32(leftf, rightf));
+    }
+
+    for (i = (frameCount4 << 2); i < frameCount; ++i) {
+        drflac_uint32 left  = pInputSamples0U32[i] << shift0;
+        drflac_uint32 side  = pInputSamples1U32[i] << shift1;
+        drflac_uint32 right = left - side;
+
+        pOutputSamples[i*2+0] = (drflac_int32)left  / 8388608.0f;
+        pOutputSamples[i*2+1] = (drflac_int32)right / 8388608.0f;
+    }
+}
+#endif
+
+static DRFLAC_INLINE void drflac_read_pcm_frames_f32__decode_left_side(drflac* pFlac, drflac_uint64 frameCount, drflac_uint32 unusedBitsPerSample, const drflac_int32* pInputSamples0, const drflac_int32* pInputSamples1, float* pOutputSamples)
+{
+#if defined(DRFLAC_SUPPORT_SSE2)
+    if (drflac__gIsSSE2Supported && pFlac->bitsPerSample <= 24) {
+        drflac_read_pcm_frames_f32__decode_left_side__sse2(pFlac, frameCount, unusedBitsPerSample, pInputSamples0, pInputSamples1, pOutputSamples);
+    } else
+#elif defined(DRFLAC_SUPPORT_NEON)
+    if (drflac__gIsNEONSupported && pFlac->bitsPerSample <= 24) {
+        drflac_read_pcm_frames_f32__decode_left_side__neon(pFlac, frameCount, unusedBitsPerSample, pInputSamples0, pInputSamples1, pOutputSamples);
+    } else
+#endif
+    {
+        /* Scalar fallback. */
+#if 0
+        drflac_read_pcm_frames_f32__decode_left_side__reference(pFlac, frameCount, unusedBitsPerSample, pInputSamples0, pInputSamples1, pOutputSamples);
+#else
+        drflac_read_pcm_frames_f32__decode_left_side__scalar(pFlac, frameCount, unusedBitsPerSample, pInputSamples0, pInputSamples1, pOutputSamples);
+#endif
+    }
+}
+
+
+#if 0
+static DRFLAC_INLINE void drflac_read_pcm_frames_f32__decode_right_side__reference(drflac* pFlac, drflac_uint64 frameCount, drflac_uint32 unusedBitsPerSample, const drflac_int32* pInputSamples0, const drflac_int32* pInputSamples1, float* pOutputSamples)
+{
+    drflac_uint64 i;
+    for (i = 0; i < frameCount; ++i) {
+        drflac_uint32 side  = (drflac_uint32)pInputSamples0[i] << (unusedBitsPerSample + pFlac->currentFLACFrame.subframes[0].wastedBitsPerSample);
+        drflac_uint32 right = (drflac_uint32)pInputSamples1[i] << (unusedBitsPerSample + pFlac->currentFLACFrame.subframes[1].wastedBitsPerSample);
+        drflac_uint32 left  = right + side;
+
+        pOutputSamples[i*2+0] = (float)((drflac_int32)left  / 2147483648.0);
+        pOutputSamples[i*2+1] = (float)((drflac_int32)right / 2147483648.0);
+    }
+}
+#endif
+
+static DRFLAC_INLINE void drflac_read_pcm_frames_f32__decode_right_side__scalar(drflac* pFlac, drflac_uint64 frameCount, drflac_uint32 unusedBitsPerSample, const drflac_int32* pInputSamples0, const drflac_int32* pInputSamples1, float* pOutputSamples)
+{
+    drflac_uint64 i;
+    drflac_uint64 frameCount4 = frameCount >> 2;
+    const drflac_uint32* pInputSamples0U32 = (const drflac_uint32*)pInputSamples0;
+    const drflac_uint32* pInputSamples1U32 = (const drflac_uint32*)pInputSamples1;
+    drflac_uint32 shift0 = unusedBitsPerSample + pFlac->currentFLACFrame.subframes[0].wastedBitsPerSample;
+    drflac_uint32 shift1 = unusedBitsPerSample + pFlac->currentFLACFrame.subframes[1].wastedBitsPerSample;
+    float factor = 1 / 2147483648.0;
+
+    for (i = 0; i < frameCount4; ++i) {
+        drflac_uint32 side0  = pInputSamples0U32[i*4+0] << shift0;
+        drflac_uint32 side1  = pInputSamples0U32[i*4+1] << shift0;
+        drflac_uint32 side2  = pInputSamples0U32[i*4+2] << shift0;
+        drflac_uint32 side3  = pInputSamples0U32[i*4+3] << shift0;
+
+        drflac_uint32 right0 = pInputSamples1U32[i*4+0] << shift1;
+        drflac_uint32 right1 = pInputSamples1U32[i*4+1] << shift1;
+        drflac_uint32 right2 = pInputSamples1U32[i*4+2] << shift1;
+        drflac_uint32 right3 = pInputSamples1U32[i*4+3] << shift1;
+
+        drflac_uint32 left0 = right0 + side0;
+        drflac_uint32 left1 = right1 + side1;
+        drflac_uint32 left2 = right2 + side2;
+        drflac_uint32 left3 = right3 + side3;
+
+        pOutputSamples[i*8+0] = (drflac_int32)left0  * factor;
+        pOutputSamples[i*8+1] = (drflac_int32)right0 * factor;
+        pOutputSamples[i*8+2] = (drflac_int32)left1  * factor;
+        pOutputSamples[i*8+3] = (drflac_int32)right1 * factor;
+        pOutputSamples[i*8+4] = (drflac_int32)left2  * factor;
+        pOutputSamples[i*8+5] = (drflac_int32)right2 * factor;
+        pOutputSamples[i*8+6] = (drflac_int32)left3  * factor;
+        pOutputSamples[i*8+7] = (drflac_int32)right3 * factor;
+    }
+
+    for (i = (frameCount4 << 2); i < frameCount; ++i) {
+        drflac_uint32 side  = pInputSamples0U32[i] << shift0;
+        drflac_uint32 right = pInputSamples1U32[i] << shift1;
+        drflac_uint32 left  = right + side;
+
+        pOutputSamples[i*2+0] = (drflac_int32)left  * factor;
+        pOutputSamples[i*2+1] = (drflac_int32)right * factor;
+    }
+}
+
+#if defined(DRFLAC_SUPPORT_SSE2)
+static DRFLAC_INLINE void drflac_read_pcm_frames_f32__decode_right_side__sse2(drflac* pFlac, drflac_uint64 frameCount, drflac_uint32 unusedBitsPerSample, const drflac_int32* pInputSamples0, const drflac_int32* pInputSamples1, float* pOutputSamples)
+{
+    drflac_uint64 i;
+    drflac_uint64 frameCount4 = frameCount >> 2;
+    const drflac_uint32* pInputSamples0U32 = (const drflac_uint32*)pInputSamples0;
+    const drflac_uint32* pInputSamples1U32 = (const drflac_uint32*)pInputSamples1;
+    drflac_uint32 shift0 = (unusedBitsPerSample + pFlac->currentFLACFrame.subframes[0].wastedBitsPerSample) - 8;
+    drflac_uint32 shift1 = (unusedBitsPerSample + pFlac->currentFLACFrame.subframes[1].wastedBitsPerSample) - 8;
+    __m128 factor;
+
+    DRFLAC_ASSERT(pFlac->bitsPerSample <= 24);
+
+    factor = _mm_set1_ps(1.0f / 8388608.0f);
+
+    for (i = 0; i < frameCount4; ++i) {
+        __m128i side  = _mm_slli_epi32(_mm_loadu_si128((const __m128i*)pInputSamples0 + i), shift0);
+        __m128i right = _mm_slli_epi32(_mm_loadu_si128((const __m128i*)pInputSamples1 + i), shift1);
+        __m128i left  = _mm_add_epi32(right, side);
+        __m128 leftf  = _mm_mul_ps(_mm_cvtepi32_ps(left),  factor);
+        __m128 rightf = _mm_mul_ps(_mm_cvtepi32_ps(right), factor);
+
+        _mm_storeu_ps(pOutputSamples + i*8 + 0, _mm_unpacklo_ps(leftf, rightf));
+        _mm_storeu_ps(pOutputSamples + i*8 + 4, _mm_unpackhi_ps(leftf, rightf));
+    }
+
+    for (i = (frameCount4 << 2); i < frameCount; ++i) {
+        drflac_uint32 side  = pInputSamples0U32[i] << shift0;
+        drflac_uint32 right = pInputSamples1U32[i] << shift1;
+        drflac_uint32 left  = right + side;
+
+        pOutputSamples[i*2+0] = (drflac_int32)left  / 8388608.0f;
+        pOutputSamples[i*2+1] = (drflac_int32)right / 8388608.0f;
+    }
+}
+#endif
+
+#if defined(DRFLAC_SUPPORT_NEON)
+static DRFLAC_INLINE void drflac_read_pcm_frames_f32__decode_right_side__neon(drflac* pFlac, drflac_uint64 frameCount, drflac_uint32 unusedBitsPerSample, const drflac_int32* pInputSamples0, const drflac_int32* pInputSamples1, float* pOutputSamples)
+{
+    drflac_uint64 i;
+    drflac_uint64 frameCount4 = frameCount >> 2;
+    const drflac_uint32* pInputSamples0U32 = (const drflac_uint32*)pInputSamples0;
+    const drflac_uint32* pInputSamples1U32 = (const drflac_uint32*)pInputSamples1;
+    drflac_uint32 shift0 = (unusedBitsPerSample + pFlac->currentFLACFrame.subframes[0].wastedBitsPerSample) - 8;
+    drflac_uint32 shift1 = (unusedBitsPerSample + pFlac->currentFLACFrame.subframes[1].wastedBitsPerSample) - 8;
+    float32x4_t factor4;
+    int32x4_t shift0_4;
+    int32x4_t shift1_4;
+
+    DRFLAC_ASSERT(pFlac->bitsPerSample <= 24);
+
+    factor4  = vdupq_n_f32(1.0f / 8388608.0f);
+    shift0_4 = vdupq_n_s32(shift0);
+    shift1_4 = vdupq_n_s32(shift1);
+
+    for (i = 0; i < frameCount4; ++i) {
+        uint32x4_t side;
+        uint32x4_t right;
+        uint32x4_t left;
+        float32x4_t leftf;
+        float32x4_t rightf;
+
+        side   = vshlq_u32(vld1q_u32(pInputSamples0U32 + i*4), shift0_4);
+        right  = vshlq_u32(vld1q_u32(pInputSamples1U32 + i*4), shift1_4);
+        left   = vaddq_u32(right, side);
+        leftf  = vmulq_f32(vcvtq_f32_s32(vreinterpretq_s32_u32(left)),  factor4);
+        rightf = vmulq_f32(vcvtq_f32_s32(vreinterpretq_s32_u32(right)), factor4);
+
+        drflac__vst2q_f32(pOutputSamples + i*8, vzipq_f32(leftf, rightf));
+    }
+
+    for (i = (frameCount4 << 2); i < frameCount; ++i) {
+        drflac_uint32 side  = pInputSamples0U32[i] << shift0;
+        drflac_uint32 right = pInputSamples1U32[i] << shift1;
+        drflac_uint32 left  = right + side;
+
+        pOutputSamples[i*2+0] = (drflac_int32)left  / 8388608.0f;
+        pOutputSamples[i*2+1] = (drflac_int32)right / 8388608.0f;
+    }
+}
+#endif
+
+static DRFLAC_INLINE void drflac_read_pcm_frames_f32__decode_right_side(drflac* pFlac, drflac_uint64 frameCount, drflac_uint32 unusedBitsPerSample, const drflac_int32* pInputSamples0, const drflac_int32* pInputSamples1, float* pOutputSamples)
+{
+#if defined(DRFLAC_SUPPORT_SSE2)
+    if (drflac__gIsSSE2Supported && pFlac->bitsPerSample <= 24) {
+        drflac_read_pcm_frames_f32__decode_right_side__sse2(pFlac, frameCount, unusedBitsPerSample, pInputSamples0, pInputSamples1, pOutputSamples);
+    } else
+#elif defined(DRFLAC_SUPPORT_NEON)
+    if (drflac__gIsNEONSupported && pFlac->bitsPerSample <= 24) {
+        drflac_read_pcm_frames_f32__decode_right_side__neon(pFlac, frameCount, unusedBitsPerSample, pInputSamples0, pInputSamples1, pOutputSamples);
+    } else
+#endif
+    {
+        /* Scalar fallback. */
+#if 0
+        drflac_read_pcm_frames_f32__decode_right_side__reference(pFlac, frameCount, unusedBitsPerSample, pInputSamples0, pInputSamples1, pOutputSamples);
+#else
+        drflac_read_pcm_frames_f32__decode_right_side__scalar(pFlac, frameCount, unusedBitsPerSample, pInputSamples0, pInputSamples1, pOutputSamples);
+#endif
+    }
+}
+
+
+#if 0
+static DRFLAC_INLINE void drflac_read_pcm_frames_f32__decode_mid_side__reference(drflac* pFlac, drflac_uint64 frameCount, drflac_uint32 unusedBitsPerSample, const drflac_int32* pInputSamples0, const drflac_int32* pInputSamples1, float* pOutputSamples)
+{
+    for (drflac_uint64 i = 0; i < frameCount; ++i) {
+        drflac_uint32 mid  = (drflac_uint32)pInputSamples0[i] << pFlac->currentFLACFrame.subframes[0].wastedBitsPerSample;
+        drflac_uint32 side = (drflac_uint32)pInputSamples1[i] << pFlac->currentFLACFrame.subframes[1].wastedBitsPerSample;
+
+        mid = (mid << 1) | (side & 0x01);
+
+        pOutputSamples[i*2+0] = (float)((((drflac_int32)(mid + side) >> 1) << (unusedBitsPerSample)) / 2147483648.0);
+        pOutputSamples[i*2+1] = (float)((((drflac_int32)(mid - side) >> 1) << (unusedBitsPerSample)) / 2147483648.0);
+    }
+}
+#endif
+
+static DRFLAC_INLINE void drflac_read_pcm_frames_f32__decode_mid_side__scalar(drflac* pFlac, drflac_uint64 frameCount, drflac_uint32 unusedBitsPerSample, const drflac_int32* pInputSamples0, const drflac_int32* pInputSamples1, float* pOutputSamples)
+{
+    drflac_uint64 i;
+    drflac_uint64 frameCount4 = frameCount >> 2;
+    const drflac_uint32* pInputSamples0U32 = (const drflac_uint32*)pInputSamples0;
+    const drflac_uint32* pInputSamples1U32 = (const drflac_uint32*)pInputSamples1;
+    drflac_uint32 shift = unusedBitsPerSample;
+    float factor = 1 / 2147483648.0;
+
+    if (shift > 0) {
+        shift -= 1;
+        for (i = 0; i < frameCount4; ++i) {
+            drflac_uint32 temp0L;
+            drflac_uint32 temp1L;
+            drflac_uint32 temp2L;
+            drflac_uint32 temp3L;
+            drflac_uint32 temp0R;
+            drflac_uint32 temp1R;
+            drflac_uint32 temp2R;
+            drflac_uint32 temp3R;
+
+            drflac_uint32 mid0  = pInputSamples0U32[i*4+0] << pFlac->currentFLACFrame.subframes[0].wastedBitsPerSample;
+            drflac_uint32 mid1  = pInputSamples0U32[i*4+1] << pFlac->currentFLACFrame.subframes[0].wastedBitsPerSample;
+            drflac_uint32 mid2  = pInputSamples0U32[i*4+2] << pFlac->currentFLACFrame.subframes[0].wastedBitsPerSample;
+            drflac_uint32 mid3  = pInputSamples0U32[i*4+3] << pFlac->currentFLACFrame.subframes[0].wastedBitsPerSample;
+
+            drflac_uint32 side0 = pInputSamples1U32[i*4+0] << pFlac->currentFLACFrame.subframes[1].wastedBitsPerSample;
+            drflac_uint32 side1 = pInputSamples1U32[i*4+1] << pFlac->currentFLACFrame.subframes[1].wastedBitsPerSample;
+            drflac_uint32 side2 = pInputSamples1U32[i*4+2] << pFlac->currentFLACFrame.subframes[1].wastedBitsPerSample;
+            drflac_uint32 side3 = pInputSamples1U32[i*4+3] << pFlac->currentFLACFrame.subframes[1].wastedBitsPerSample;
+
+            mid0 = (mid0 << 1) | (side0 & 0x01);
+            mid1 = (mid1 << 1) | (side1 & 0x01);
+            mid2 = (mid2 << 1) | (side2 & 0x01);
+            mid3 = (mid3 << 1) | (side3 & 0x01);
+
+            temp0L = (mid0 + side0) << shift;
+            temp1L = (mid1 + side1) << shift;
+            temp2L = (mid2 + side2) << shift;
+            temp3L = (mid3 + side3) << shift;
+
+            temp0R = (mid0 - side0) << shift;
+            temp1R = (mid1 - side1) << shift;
+            temp2R = (mid2 - side2) << shift;
+            temp3R = (mid3 - side3) << shift;
+
+            pOutputSamples[i*8+0] = (drflac_int32)temp0L * factor;
+            pOutputSamples[i*8+1] = (drflac_int32)temp0R * factor;
+            pOutputSamples[i*8+2] = (drflac_int32)temp1L * factor;
+            pOutputSamples[i*8+3] = (drflac_int32)temp1R * factor;
+            pOutputSamples[i*8+4] = (drflac_int32)temp2L * factor;
+            pOutputSamples[i*8+5] = (drflac_int32)temp2R * factor;
+            pOutputSamples[i*8+6] = (drflac_int32)temp3L * factor;
+            pOutputSamples[i*8+7] = (drflac_int32)temp3R * factor;
+        }
+    } else {
+        for (i = 0; i < frameCount4; ++i) {
+            drflac_uint32 temp0L;
+            drflac_uint32 temp1L;
+            drflac_uint32 temp2L;
+            drflac_uint32 temp3L;
+            drflac_uint32 temp0R;
+            drflac_uint32 temp1R;
+            drflac_uint32 temp2R;
+            drflac_uint32 temp3R;
+
+            drflac_uint32 mid0  = pInputSamples0U32[i*4+0] << pFlac->currentFLACFrame.subframes[0].wastedBitsPerSample;
+            drflac_uint32 mid1  = pInputSamples0U32[i*4+1] << pFlac->currentFLACFrame.subframes[0].wastedBitsPerSample;
+            drflac_uint32 mid2  = pInputSamples0U32[i*4+2] << pFlac->currentFLACFrame.subframes[0].wastedBitsPerSample;
+            drflac_uint32 mid3  = pInputSamples0U32[i*4+3] << pFlac->currentFLACFrame.subframes[0].wastedBitsPerSample;
+
+            drflac_uint32 side0 = pInputSamples1U32[i*4+0] << pFlac->currentFLACFrame.subframes[1].wastedBitsPerSample;
+            drflac_uint32 side1 = pInputSamples1U32[i*4+1] << pFlac->currentFLACFrame.subframes[1].wastedBitsPerSample;
+            drflac_uint32 side2 = pInputSamples1U32[i*4+2] << pFlac->currentFLACFrame.subframes[1].wastedBitsPerSample;
+            drflac_uint32 side3 = pInputSamples1U32[i*4+3] << pFlac->currentFLACFrame.subframes[1].wastedBitsPerSample;
+
+            mid0 = (mid0 << 1) | (side0 & 0x01);
+            mid1 = (mid1 << 1) | (side1 & 0x01);
+            mid2 = (mid2 << 1) | (side2 & 0x01);
+            mid3 = (mid3 << 1) | (side3 & 0x01);
+
+            temp0L = (drflac_uint32)((drflac_int32)(mid0 + side0) >> 1);
+            temp1L = (drflac_uint32)((drflac_int32)(mid1 + side1) >> 1);
+            temp2L = (drflac_uint32)((drflac_int32)(mid2 + side2) >> 1);
+            temp3L = (drflac_uint32)((drflac_int32)(mid3 + side3) >> 1);
+
+            temp0R = (drflac_uint32)((drflac_int32)(mid0 - side0) >> 1);
+            temp1R = (drflac_uint32)((drflac_int32)(mid1 - side1) >> 1);
+            temp2R = (drflac_uint32)((drflac_int32)(mid2 - side2) >> 1);
+            temp3R = (drflac_uint32)((drflac_int32)(mid3 - side3) >> 1);
+
+            pOutputSamples[i*8+0] = (drflac_int32)temp0L * factor;
+            pOutputSamples[i*8+1] = (drflac_int32)temp0R * factor;
+            pOutputSamples[i*8+2] = (drflac_int32)temp1L * factor;
+            pOutputSamples[i*8+3] = (drflac_int32)temp1R * factor;
+            pOutputSamples[i*8+4] = (drflac_int32)temp2L * factor;
+            pOutputSamples[i*8+5] = (drflac_int32)temp2R * factor;
+            pOutputSamples[i*8+6] = (drflac_int32)temp3L * factor;
+            pOutputSamples[i*8+7] = (drflac_int32)temp3R * factor;
+        }
+    }
+
+    for (i = (frameCount4 << 2); i < frameCount; ++i) {
+        drflac_uint32 mid  = pInputSamples0U32[i] << pFlac->currentFLACFrame.subframes[0].wastedBitsPerSample;
+        drflac_uint32 side = pInputSamples1U32[i] << pFlac->currentFLACFrame.subframes[1].wastedBitsPerSample;
+
+        mid = (mid << 1) | (side & 0x01);
+
+        pOutputSamples[i*2+0] = (drflac_int32)((drflac_uint32)((drflac_int32)(mid + side) >> 1) << unusedBitsPerSample) * factor;
+        pOutputSamples[i*2+1] = (drflac_int32)((drflac_uint32)((drflac_int32)(mid - side) >> 1) << unusedBitsPerSample) * factor;
+    }
+}
+
+#if defined(DRFLAC_SUPPORT_SSE2)
+static DRFLAC_INLINE void drflac_read_pcm_frames_f32__decode_mid_side__sse2(drflac* pFlac, drflac_uint64 frameCount, drflac_uint32 unusedBitsPerSample, const drflac_int32* pInputSamples0, const drflac_int32* pInputSamples1, float* pOutputSamples)
+{
+    drflac_uint64 i;
+    drflac_uint64 frameCount4 = frameCount >> 2;
+    const drflac_uint32* pInputSamples0U32 = (const drflac_uint32*)pInputSamples0;
+    const drflac_uint32* pInputSamples1U32 = (const drflac_uint32*)pInputSamples1;
+    drflac_uint32 shift = unusedBitsPerSample - 8;
+    float factor;
+    __m128 factor128;
+
+    DRFLAC_ASSERT(pFlac->bitsPerSample <= 24);
+
+    factor = 1.0f / 8388608.0f;
+    factor128 = _mm_set1_ps(factor);
+
+    if (shift == 0) {
+        for (i = 0; i < frameCount4; ++i) {
+            __m128i mid;
+            __m128i side;
+            __m128i tempL;
+            __m128i tempR;
+            __m128  leftf;
+            __m128  rightf;
+
+            mid    = _mm_slli_epi32(_mm_loadu_si128((const __m128i*)pInputSamples0 + i), pFlac->currentFLACFrame.subframes[0].wastedBitsPerSample);
+            side   = _mm_slli_epi32(_mm_loadu_si128((const __m128i*)pInputSamples1 + i), pFlac->currentFLACFrame.subframes[1].wastedBitsPerSample);
+
+            mid    = _mm_or_si128(_mm_slli_epi32(mid, 1), _mm_and_si128(side, _mm_set1_epi32(0x01)));
+
+            tempL  = _mm_srai_epi32(_mm_add_epi32(mid, side), 1);
+            tempR  = _mm_srai_epi32(_mm_sub_epi32(mid, side), 1);
+
+            leftf  = _mm_mul_ps(_mm_cvtepi32_ps(tempL), factor128);
+            rightf = _mm_mul_ps(_mm_cvtepi32_ps(tempR), factor128);
+
+            _mm_storeu_ps(pOutputSamples + i*8 + 0, _mm_unpacklo_ps(leftf, rightf));
+            _mm_storeu_ps(pOutputSamples + i*8 + 4, _mm_unpackhi_ps(leftf, rightf));
+        }
+
+        for (i = (frameCount4 << 2); i < frameCount; ++i) {
+            drflac_uint32 mid  = pInputSamples0U32[i] << pFlac->currentFLACFrame.subframes[0].wastedBitsPerSample;
+            drflac_uint32 side = pInputSamples1U32[i] << pFlac->currentFLACFrame.subframes[1].wastedBitsPerSample;
+
+            mid = (mid << 1) | (side & 0x01);
+
+            pOutputSamples[i*2+0] = ((drflac_int32)(mid + side) >> 1) * factor;
+            pOutputSamples[i*2+1] = ((drflac_int32)(mid - side) >> 1) * factor;
+        }
+    } else {
+        shift -= 1;
+        for (i = 0; i < frameCount4; ++i) {
+            __m128i mid;
+            __m128i side;
+            __m128i tempL;
+            __m128i tempR;
+            __m128 leftf;
+            __m128 rightf;
+
+            mid    = _mm_slli_epi32(_mm_loadu_si128((const __m128i*)pInputSamples0 + i), pFlac->currentFLACFrame.subframes[0].wastedBitsPerSample);
+            side   = _mm_slli_epi32(_mm_loadu_si128((const __m128i*)pInputSamples1 + i), pFlac->currentFLACFrame.subframes[1].wastedBitsPerSample);
+
+            mid    = _mm_or_si128(_mm_slli_epi32(mid, 1), _mm_and_si128(side, _mm_set1_epi32(0x01)));
+
+            tempL  = _mm_slli_epi32(_mm_add_epi32(mid, side), shift);
+            tempR  = _mm_slli_epi32(_mm_sub_epi32(mid, side), shift);
+
+            leftf  = _mm_mul_ps(_mm_cvtepi32_ps(tempL), factor128);
+            rightf = _mm_mul_ps(_mm_cvtepi32_ps(tempR), factor128);
+
+            _mm_storeu_ps(pOutputSamples + i*8 + 0, _mm_unpacklo_ps(leftf, rightf));
+            _mm_storeu_ps(pOutputSamples + i*8 + 4, _mm_unpackhi_ps(leftf, rightf));
+        }
+
+        for (i = (frameCount4 << 2); i < frameCount; ++i) {
+            drflac_uint32 mid  = pInputSamples0U32[i] << pFlac->currentFLACFrame.subframes[0].wastedBitsPerSample;
+            drflac_uint32 side = pInputSamples1U32[i] << pFlac->currentFLACFrame.subframes[1].wastedBitsPerSample;
+
+            mid = (mid << 1) | (side & 0x01);
+
+            pOutputSamples[i*2+0] = (drflac_int32)((mid + side) << shift) * factor;
+            pOutputSamples[i*2+1] = (drflac_int32)((mid - side) << shift) * factor;
+        }
+    }
+}
+#endif
+
+#if defined(DRFLAC_SUPPORT_NEON)
+static DRFLAC_INLINE void drflac_read_pcm_frames_f32__decode_mid_side__neon(drflac* pFlac, drflac_uint64 frameCount, drflac_uint32 unusedBitsPerSample, const drflac_int32* pInputSamples0, const drflac_int32* pInputSamples1, float* pOutputSamples)
+{
+    drflac_uint64 i;
+    drflac_uint64 frameCount4 = frameCount >> 2;
+    const drflac_uint32* pInputSamples0U32 = (const drflac_uint32*)pInputSamples0;
+    const drflac_uint32* pInputSamples1U32 = (const drflac_uint32*)pInputSamples1;
+    drflac_uint32 shift = unusedBitsPerSample - 8;
+    float factor;
+    float32x4_t factor4;
+    int32x4_t shift4;
+    int32x4_t wbps0_4;  /* Wasted Bits Per Sample */
+    int32x4_t wbps1_4;  /* Wasted Bits Per Sample */
+
+    DRFLAC_ASSERT(pFlac->bitsPerSample <= 24);
+
+    factor  = 1.0f / 8388608.0f;
+    factor4 = vdupq_n_f32(factor);
+    wbps0_4 = vdupq_n_s32(pFlac->currentFLACFrame.subframes[0].wastedBitsPerSample);
+    wbps1_4 = vdupq_n_s32(pFlac->currentFLACFrame.subframes[1].wastedBitsPerSample);
+
+    if (shift == 0) {
+        for (i = 0; i < frameCount4; ++i) {
+            int32x4_t lefti;
+            int32x4_t righti;
+            float32x4_t leftf;
+            float32x4_t rightf;
+
+            uint32x4_t mid  = vshlq_u32(vld1q_u32(pInputSamples0U32 + i*4), wbps0_4);
+            uint32x4_t side = vshlq_u32(vld1q_u32(pInputSamples1U32 + i*4), wbps1_4);
+
+            mid    = vorrq_u32(vshlq_n_u32(mid, 1), vandq_u32(side, vdupq_n_u32(1)));
+
+            lefti  = vshrq_n_s32(vreinterpretq_s32_u32(vaddq_u32(mid, side)), 1);
+            righti = vshrq_n_s32(vreinterpretq_s32_u32(vsubq_u32(mid, side)), 1);
+
+            leftf  = vmulq_f32(vcvtq_f32_s32(lefti),  factor4);
+            rightf = vmulq_f32(vcvtq_f32_s32(righti), factor4);
+
+            drflac__vst2q_f32(pOutputSamples + i*8, vzipq_f32(leftf, rightf));
+        }
+
+        for (i = (frameCount4 << 2); i < frameCount; ++i) {
+            drflac_uint32 mid  = pInputSamples0U32[i] << pFlac->currentFLACFrame.subframes[0].wastedBitsPerSample;
+            drflac_uint32 side = pInputSamples1U32[i] << pFlac->currentFLACFrame.subframes[1].wastedBitsPerSample;
+
+            mid = (mid << 1) | (side & 0x01);
+
+            pOutputSamples[i*2+0] = ((drflac_int32)(mid + side) >> 1) * factor;
+            pOutputSamples[i*2+1] = ((drflac_int32)(mid - side) >> 1) * factor;
+        }
+    } else {
+        shift -= 1;
+        shift4 = vdupq_n_s32(shift);
+        for (i = 0; i < frameCount4; ++i) {
+            uint32x4_t mid;
+            uint32x4_t side;
+            int32x4_t lefti;
+            int32x4_t righti;
+            float32x4_t leftf;
+            float32x4_t rightf;
+
+            mid    = vshlq_u32(vld1q_u32(pInputSamples0U32 + i*4), wbps0_4);
+            side   = vshlq_u32(vld1q_u32(pInputSamples1U32 + i*4), wbps1_4);
+
+            mid    = vorrq_u32(vshlq_n_u32(mid, 1), vandq_u32(side, vdupq_n_u32(1)));
+
+            lefti  = vreinterpretq_s32_u32(vshlq_u32(vaddq_u32(mid, side), shift4));
+            righti = vreinterpretq_s32_u32(vshlq_u32(vsubq_u32(mid, side), shift4));
+
+            leftf  = vmulq_f32(vcvtq_f32_s32(lefti),  factor4);
+            rightf = vmulq_f32(vcvtq_f32_s32(righti), factor4);
+
+            drflac__vst2q_f32(pOutputSamples + i*8, vzipq_f32(leftf, rightf));
+        }
+
+        for (i = (frameCount4 << 2); i < frameCount; ++i) {
+            drflac_uint32 mid  = pInputSamples0U32[i] << pFlac->currentFLACFrame.subframes[0].wastedBitsPerSample;
+            drflac_uint32 side = pInputSamples1U32[i] << pFlac->currentFLACFrame.subframes[1].wastedBitsPerSample;
+
+            mid = (mid << 1) | (side & 0x01);
+
+            pOutputSamples[i*2+0] = (drflac_int32)((mid + side) << shift) * factor;
+            pOutputSamples[i*2+1] = (drflac_int32)((mid - side) << shift) * factor;
+        }
+    }
+}
+#endif
+
+static DRFLAC_INLINE void drflac_read_pcm_frames_f32__decode_mid_side(drflac* pFlac, drflac_uint64 frameCount, drflac_uint32 unusedBitsPerSample, const drflac_int32* pInputSamples0, const drflac_int32* pInputSamples1, float* pOutputSamples)
+{
+#if defined(DRFLAC_SUPPORT_SSE2)
+    if (drflac__gIsSSE2Supported && pFlac->bitsPerSample <= 24) {
+        drflac_read_pcm_frames_f32__decode_mid_side__sse2(pFlac, frameCount, unusedBitsPerSample, pInputSamples0, pInputSamples1, pOutputSamples);
+    } else
+#elif defined(DRFLAC_SUPPORT_NEON)
+    if (drflac__gIsNEONSupported && pFlac->bitsPerSample <= 24) {
+        drflac_read_pcm_frames_f32__decode_mid_side__neon(pFlac, frameCount, unusedBitsPerSample, pInputSamples0, pInputSamples1, pOutputSamples);
+    } else
+#endif
+    {
+        /* Scalar fallback. */
+#if 0
+        drflac_read_pcm_frames_f32__decode_mid_side__reference(pFlac, frameCount, unusedBitsPerSample, pInputSamples0, pInputSamples1, pOutputSamples);
+#else
+        drflac_read_pcm_frames_f32__decode_mid_side__scalar(pFlac, frameCount, unusedBitsPerSample, pInputSamples0, pInputSamples1, pOutputSamples);
+#endif
+    }
+}
+
+#if 0
+static DRFLAC_INLINE void drflac_read_pcm_frames_f32__decode_independent_stereo__reference(drflac* pFlac, drflac_uint64 frameCount, drflac_uint32 unusedBitsPerSample, const drflac_int32* pInputSamples0, const drflac_int32* pInputSamples1, float* pOutputSamples)
+{
+    for (drflac_uint64 i = 0; i < frameCount; ++i) {
+        pOutputSamples[i*2+0] = (float)((drflac_int32)((drflac_uint32)pInputSamples0[i] << (unusedBitsPerSample + pFlac->currentFLACFrame.subframes[0].wastedBitsPerSample)) / 2147483648.0);
+        pOutputSamples[i*2+1] = (float)((drflac_int32)((drflac_uint32)pInputSamples1[i] << (unusedBitsPerSample + pFlac->currentFLACFrame.subframes[1].wastedBitsPerSample)) / 2147483648.0);
+    }
+}
+#endif
+
+static DRFLAC_INLINE void drflac_read_pcm_frames_f32__decode_independent_stereo__scalar(drflac* pFlac, drflac_uint64 frameCount, drflac_uint32 unusedBitsPerSample, const drflac_int32* pInputSamples0, const drflac_int32* pInputSamples1, float* pOutputSamples)
+{
+    drflac_uint64 i;
+    drflac_uint64 frameCount4 = frameCount >> 2;
+    const drflac_uint32* pInputSamples0U32 = (const drflac_uint32*)pInputSamples0;
+    const drflac_uint32* pInputSamples1U32 = (const drflac_uint32*)pInputSamples1;
+    drflac_uint32 shift0 = unusedBitsPerSample + pFlac->currentFLACFrame.subframes[0].wastedBitsPerSample;
+    drflac_uint32 shift1 = unusedBitsPerSample + pFlac->currentFLACFrame.subframes[1].wastedBitsPerSample;
+    float factor = 1 / 2147483648.0;
+
+    for (i = 0; i < frameCount4; ++i) {
+        drflac_uint32 tempL0 = pInputSamples0U32[i*4+0] << shift0;
+        drflac_uint32 tempL1 = pInputSamples0U32[i*4+1] << shift0;
+        drflac_uint32 tempL2 = pInputSamples0U32[i*4+2] << shift0;
+        drflac_uint32 tempL3 = pInputSamples0U32[i*4+3] << shift0;
+
+        drflac_uint32 tempR0 = pInputSamples1U32[i*4+0] << shift1;
+        drflac_uint32 tempR1 = pInputSamples1U32[i*4+1] << shift1;
+        drflac_uint32 tempR2 = pInputSamples1U32[i*4+2] << shift1;
+        drflac_uint32 tempR3 = pInputSamples1U32[i*4+3] << shift1;
+
+        pOutputSamples[i*8+0] = (drflac_int32)tempL0 * factor;
+        pOutputSamples[i*8+1] = (drflac_int32)tempR0 * factor;
+        pOutputSamples[i*8+2] = (drflac_int32)tempL1 * factor;
+        pOutputSamples[i*8+3] = (drflac_int32)tempR1 * factor;
+        pOutputSamples[i*8+4] = (drflac_int32)tempL2 * factor;
+        pOutputSamples[i*8+5] = (drflac_int32)tempR2 * factor;
+        pOutputSamples[i*8+6] = (drflac_int32)tempL3 * factor;
+        pOutputSamples[i*8+7] = (drflac_int32)tempR3 * factor;
+    }
+
+    for (i = (frameCount4 << 2); i < frameCount; ++i) {
+        pOutputSamples[i*2+0] = (drflac_int32)(pInputSamples0U32[i] << shift0) * factor;
+        pOutputSamples[i*2+1] = (drflac_int32)(pInputSamples1U32[i] << shift1) * factor;
+    }
+}
+
+#if defined(DRFLAC_SUPPORT_SSE2)
+static DRFLAC_INLINE void drflac_read_pcm_frames_f32__decode_independent_stereo__sse2(drflac* pFlac, drflac_uint64 frameCount, drflac_uint32 unusedBitsPerSample, const drflac_int32* pInputSamples0, const drflac_int32* pInputSamples1, float* pOutputSamples)
+{
+    drflac_uint64 i;
+    drflac_uint64 frameCount4 = frameCount >> 2;
+    const drflac_uint32* pInputSamples0U32 = (const drflac_uint32*)pInputSamples0;
+    const drflac_uint32* pInputSamples1U32 = (const drflac_uint32*)pInputSamples1;
+    drflac_uint32 shift0 = (unusedBitsPerSample + pFlac->currentFLACFrame.subframes[0].wastedBitsPerSample) - 8;
+    drflac_uint32 shift1 = (unusedBitsPerSample + pFlac->currentFLACFrame.subframes[1].wastedBitsPerSample) - 8;
+
+    float factor = 1.0f / 8388608.0f;
+    __m128 factor128 = _mm_set1_ps(factor);
+
+    for (i = 0; i < frameCount4; ++i) {
+        __m128i lefti;
+        __m128i righti;
+        __m128 leftf;
+        __m128 rightf;
+
+        lefti  = _mm_slli_epi32(_mm_loadu_si128((const __m128i*)pInputSamples0 + i), shift0);
+        righti = _mm_slli_epi32(_mm_loadu_si128((const __m128i*)pInputSamples1 + i), shift1);
+
+        leftf  = _mm_mul_ps(_mm_cvtepi32_ps(lefti),  factor128);
+        rightf = _mm_mul_ps(_mm_cvtepi32_ps(righti), factor128);
+
+        _mm_storeu_ps(pOutputSamples + i*8 + 0, _mm_unpacklo_ps(leftf, rightf));
+        _mm_storeu_ps(pOutputSamples + i*8 + 4, _mm_unpackhi_ps(leftf, rightf));
+    }
+
+    for (i = (frameCount4 << 2); i < frameCount; ++i) {
+        pOutputSamples[i*2+0] = (drflac_int32)(pInputSamples0U32[i] << shift0) * factor;
+        pOutputSamples[i*2+1] = (drflac_int32)(pInputSamples1U32[i] << shift1) * factor;
+    }
+}
+#endif
+
+#if defined(DRFLAC_SUPPORT_NEON)
+static DRFLAC_INLINE void drflac_read_pcm_frames_f32__decode_independent_stereo__neon(drflac* pFlac, drflac_uint64 frameCount, drflac_uint32 unusedBitsPerSample, const drflac_int32* pInputSamples0, const drflac_int32* pInputSamples1, float* pOutputSamples)
+{
+    drflac_uint64 i;
+    drflac_uint64 frameCount4 = frameCount >> 2;
+    const drflac_uint32* pInputSamples0U32 = (const drflac_uint32*)pInputSamples0;
+    const drflac_uint32* pInputSamples1U32 = (const drflac_uint32*)pInputSamples1;
+    drflac_uint32 shift0 = (unusedBitsPerSample + pFlac->currentFLACFrame.subframes[0].wastedBitsPerSample) - 8;
+    drflac_uint32 shift1 = (unusedBitsPerSample + pFlac->currentFLACFrame.subframes[1].wastedBitsPerSample) - 8;
+
+    float factor = 1.0f / 8388608.0f;
+    float32x4_t factor4 = vdupq_n_f32(factor);
+    int32x4_t shift0_4  = vdupq_n_s32(shift0);
+    int32x4_t shift1_4  = vdupq_n_s32(shift1);
+
+    for (i = 0; i < frameCount4; ++i) {
+        int32x4_t lefti;
+        int32x4_t righti;
+        float32x4_t leftf;
+        float32x4_t rightf;
+
+        lefti  = vreinterpretq_s32_u32(vshlq_u32(vld1q_u32(pInputSamples0U32 + i*4), shift0_4));
+        righti = vreinterpretq_s32_u32(vshlq_u32(vld1q_u32(pInputSamples1U32 + i*4), shift1_4));
+
+        leftf  = vmulq_f32(vcvtq_f32_s32(lefti),  factor4);
+        rightf = vmulq_f32(vcvtq_f32_s32(righti), factor4);
+
+        drflac__vst2q_f32(pOutputSamples + i*8, vzipq_f32(leftf, rightf));
+    }
+
+    for (i = (frameCount4 << 2); i < frameCount; ++i) {
+        pOutputSamples[i*2+0] = (drflac_int32)(pInputSamples0U32[i] << shift0) * factor;
+        pOutputSamples[i*2+1] = (drflac_int32)(pInputSamples1U32[i] << shift1) * factor;
+    }
+}
+#endif
+
+static DRFLAC_INLINE void drflac_read_pcm_frames_f32__decode_independent_stereo(drflac* pFlac, drflac_uint64 frameCount, drflac_uint32 unusedBitsPerSample, const drflac_int32* pInputSamples0, const drflac_int32* pInputSamples1, float* pOutputSamples)
+{
+#if defined(DRFLAC_SUPPORT_SSE2)
+    if (drflac__gIsSSE2Supported && pFlac->bitsPerSample <= 24) {
+        drflac_read_pcm_frames_f32__decode_independent_stereo__sse2(pFlac, frameCount, unusedBitsPerSample, pInputSamples0, pInputSamples1, pOutputSamples);
+    } else
+#elif defined(DRFLAC_SUPPORT_NEON)
+    if (drflac__gIsNEONSupported && pFlac->bitsPerSample <= 24) {
+        drflac_read_pcm_frames_f32__decode_independent_stereo__neon(pFlac, frameCount, unusedBitsPerSample, pInputSamples0, pInputSamples1, pOutputSamples);
+    } else
+#endif
+    {
+        /* Scalar fallback. */
+#if 0
+        drflac_read_pcm_frames_f32__decode_independent_stereo__reference(pFlac, frameCount, unusedBitsPerSample, pInputSamples0, pInputSamples1, pOutputSamples);
+#else
+        drflac_read_pcm_frames_f32__decode_independent_stereo__scalar(pFlac, frameCount, unusedBitsPerSample, pInputSamples0, pInputSamples1, pOutputSamples);
+#endif
+    }
+}
+
+DRFLAC_API drflac_uint64 drflac_read_pcm_frames_f32(drflac* pFlac, drflac_uint64 framesToRead, float* pBufferOut)
+{
+    drflac_uint64 framesRead;
+    drflac_uint32 unusedBitsPerSample;
+
+    if (pFlac == NULL || framesToRead == 0) {
+        return 0;
+    }
+
+    if (pBufferOut == NULL) {
+        return drflac__seek_forward_by_pcm_frames(pFlac, framesToRead);
+    }
+
+    DRFLAC_ASSERT(pFlac->bitsPerSample <= 32);
+    unusedBitsPerSample = 32 - pFlac->bitsPerSample;
+
+    framesRead = 0;
+    while (framesToRead > 0) {
+        /* If we've run out of samples in this frame, go to the next. */
+        if (pFlac->currentFLACFrame.pcmFramesRemaining == 0) {
+            if (!drflac__read_and_decode_next_flac_frame(pFlac)) {
+                break;  /* Couldn't read the next frame, so just break from the loop and return. */
+            }
+        } else {
+            unsigned int channelCount = drflac__get_channel_count_from_channel_assignment(pFlac->currentFLACFrame.header.channelAssignment);
+            drflac_uint64 iFirstPCMFrame = pFlac->currentFLACFrame.header.blockSizeInPCMFrames - pFlac->currentFLACFrame.pcmFramesRemaining;
+            drflac_uint64 frameCountThisIteration = framesToRead;
+
+            if (frameCountThisIteration > pFlac->currentFLACFrame.pcmFramesRemaining) {
+                frameCountThisIteration = pFlac->currentFLACFrame.pcmFramesRemaining;
+            }
+
+            if (channelCount == 2) {
+                const drflac_int32* pDecodedSamples0 = pFlac->currentFLACFrame.subframes[0].pSamplesS32 + iFirstPCMFrame;
+                const drflac_int32* pDecodedSamples1 = pFlac->currentFLACFrame.subframes[1].pSamplesS32 + iFirstPCMFrame;
+
+                switch (pFlac->currentFLACFrame.header.channelAssignment)
+                {
+                    case DRFLAC_CHANNEL_ASSIGNMENT_LEFT_SIDE:
+                    {
+                        drflac_read_pcm_frames_f32__decode_left_side(pFlac, frameCountThisIteration, unusedBitsPerSample, pDecodedSamples0, pDecodedSamples1, pBufferOut);
+                    } break;
+
+                    case DRFLAC_CHANNEL_ASSIGNMENT_RIGHT_SIDE:
+                    {
+                        drflac_read_pcm_frames_f32__decode_right_side(pFlac, frameCountThisIteration, unusedBitsPerSample, pDecodedSamples0, pDecodedSamples1, pBufferOut);
+                    } break;
+
+                    case DRFLAC_CHANNEL_ASSIGNMENT_MID_SIDE:
+                    {
+                        drflac_read_pcm_frames_f32__decode_mid_side(pFlac, frameCountThisIteration, unusedBitsPerSample, pDecodedSamples0, pDecodedSamples1, pBufferOut);
+                    } break;
+
+                    case DRFLAC_CHANNEL_ASSIGNMENT_INDEPENDENT:
+                    default:
+                    {
+                        drflac_read_pcm_frames_f32__decode_independent_stereo(pFlac, frameCountThisIteration, unusedBitsPerSample, pDecodedSamples0, pDecodedSamples1, pBufferOut);
+                    } break;
+                }
+            } else {
+                /* Generic interleaving. */
+                drflac_uint64 i;
+                for (i = 0; i < frameCountThisIteration; ++i) {
+                    unsigned int j;
+                    for (j = 0; j < channelCount; ++j) {
+                        drflac_int32 sampleS32 = (drflac_int32)((drflac_uint32)(pFlac->currentFLACFrame.subframes[j].pSamplesS32[iFirstPCMFrame + i]) << (unusedBitsPerSample + pFlac->currentFLACFrame.subframes[j].wastedBitsPerSample));
+                        pBufferOut[(i*channelCount)+j] = (float)(sampleS32 / 2147483648.0);
+                    }
+                }
+            }
+
+            framesRead                += frameCountThisIteration;
+            pBufferOut                += frameCountThisIteration * channelCount;
+            framesToRead              -= frameCountThisIteration;
+            pFlac->currentPCMFrame    += frameCountThisIteration;
+            pFlac->currentFLACFrame.pcmFramesRemaining -= (unsigned int)frameCountThisIteration;
+        }
+    }
+
+    return framesRead;
+}
+
+
+DRFLAC_API drflac_bool32 drflac_seek_to_pcm_frame(drflac* pFlac, drflac_uint64 pcmFrameIndex)
+{
+    if (pFlac == NULL) {
+        return DRFLAC_FALSE;
+    }
+
+    /* Don't do anything if we're already on the seek point. */
+    if (pFlac->currentPCMFrame == pcmFrameIndex) {
+        return DRFLAC_TRUE;
+    }
+
+    /*
+    If we don't know where the first frame begins then we can't seek. This will happen when the STREAMINFO block was not present
+    when the decoder was opened.
+    */
+    if (pFlac->firstFLACFramePosInBytes == 0) {
+        return DRFLAC_FALSE;
+    }
+
+    if (pcmFrameIndex == 0) {
+        pFlac->currentPCMFrame = 0;
+        return drflac__seek_to_first_frame(pFlac);
+    } else {
+        drflac_bool32 wasSuccessful = DRFLAC_FALSE;
+        drflac_uint64 originalPCMFrame = pFlac->currentPCMFrame;
+
+        /* Clamp the sample to the end. */
+        if (pcmFrameIndex > pFlac->totalPCMFrameCount) {
+            pcmFrameIndex = pFlac->totalPCMFrameCount;
+        }
+
+        /* If the target sample and the current sample are in the same frame we just move the position forward. */
+        if (pcmFrameIndex > pFlac->currentPCMFrame) {
+            /* Forward. */
+            drflac_uint32 offset = (drflac_uint32)(pcmFrameIndex - pFlac->currentPCMFrame);
+            if (pFlac->currentFLACFrame.pcmFramesRemaining >  offset) {
+                pFlac->currentFLACFrame.pcmFramesRemaining -= offset;
+                pFlac->currentPCMFrame = pcmFrameIndex;
+                return DRFLAC_TRUE;
+            }
+        } else {
+            /* Backward. */
+            drflac_uint32 offsetAbs = (drflac_uint32)(pFlac->currentPCMFrame - pcmFrameIndex);
+            drflac_uint32 currentFLACFramePCMFrameCount = pFlac->currentFLACFrame.header.blockSizeInPCMFrames;
+            drflac_uint32 currentFLACFramePCMFramesConsumed = currentFLACFramePCMFrameCount - pFlac->currentFLACFrame.pcmFramesRemaining;
+            if (currentFLACFramePCMFramesConsumed > offsetAbs) {
+                pFlac->currentFLACFrame.pcmFramesRemaining += offsetAbs;
+                pFlac->currentPCMFrame = pcmFrameIndex;
+                return DRFLAC_TRUE;
+            }
+        }
+
+        /*
+        Different techniques depending on encapsulation. Using the native FLAC seektable with Ogg encapsulation is a bit awkward so
+        we'll instead use Ogg's natural seeking facility.
+        */
+#ifndef DR_FLAC_NO_OGG
+        if (pFlac->container == drflac_container_ogg)
+        {
+            wasSuccessful = drflac_ogg__seek_to_pcm_frame(pFlac, pcmFrameIndex);
+        }
+        else
+#endif
+        {
+            /* First try seeking via the seek table. If this fails, fall back to a brute force seek which is much slower. */
+            if (/*!wasSuccessful && */!pFlac->_noSeekTableSeek) {
+                wasSuccessful = drflac__seek_to_pcm_frame__seek_table(pFlac, pcmFrameIndex);
+            }
+
+#if !defined(DR_FLAC_NO_CRC)
+            /* Fall back to binary search if seek table seeking fails. This requires the length of the stream to be known. */
+            if (!wasSuccessful && !pFlac->_noBinarySearchSeek && pFlac->totalPCMFrameCount > 0) {
+                wasSuccessful = drflac__seek_to_pcm_frame__binary_search(pFlac, pcmFrameIndex);
+            }
+#endif
+
+            /* Fall back to brute force if all else fails. */
+            if (!wasSuccessful && !pFlac->_noBruteForceSeek) {
+                wasSuccessful = drflac__seek_to_pcm_frame__brute_force(pFlac, pcmFrameIndex);
+            }
+        }
+
+        if (wasSuccessful) {
+            pFlac->currentPCMFrame = pcmFrameIndex;
+        } else {
+            /* Seek failed. Try putting the decoder back to it's original state. */
+            if (drflac_seek_to_pcm_frame(pFlac, originalPCMFrame) == DRFLAC_FALSE) {
+                /* Failed to seek back to the original PCM frame. Fall back to 0. */
+                drflac_seek_to_pcm_frame(pFlac, 0);
+            }
+        }
+
+        return wasSuccessful;
+    }
+}
+
+
+
+/* High Level APIs */
+
+/* SIZE_MAX */
+#if defined(SIZE_MAX)
+    #define DRFLAC_SIZE_MAX  SIZE_MAX
+#else
+    #if defined(DRFLAC_64BIT)
+        #define DRFLAC_SIZE_MAX  ((drflac_uint64)0xFFFFFFFFFFFFFFFF)
+    #else
+        #define DRFLAC_SIZE_MAX  0xFFFFFFFF
+    #endif
+#endif
+/* End SIZE_MAX */
+
+
+/* Using a macro as the definition of the drflac__full_decode_and_close_*() API family. Sue me. */
+#define DRFLAC_DEFINE_FULL_READ_AND_CLOSE(extension, type) \
+static type* drflac__full_read_and_close_ ## extension (drflac* pFlac, unsigned int* channelsOut, unsigned int* sampleRateOut, drflac_uint64* totalPCMFrameCountOut)\
+{                                                                                                                                                                   \
+    type* pSampleData = NULL;                                                                                                                                       \
+    drflac_uint64 totalPCMFrameCount;                                                                                                                               \
+                                                                                                                                                                    \
+    DRFLAC_ASSERT(pFlac != NULL);                                                                                                                                   \
+                                                                                                                                                                    \
+    totalPCMFrameCount = pFlac->totalPCMFrameCount;                                                                                                                 \
+                                                                                                                                                                    \
+    if (totalPCMFrameCount == 0) {                                                                                                                                  \
+        type buffer[4096];                                                                                                                                          \
+        drflac_uint64 pcmFramesRead;                                                                                                                                \
+        size_t sampleDataBufferSize = sizeof(buffer);                                                                                                               \
+                                                                                                                                                                    \
+        pSampleData = (type*)drflac__malloc_from_callbacks(sampleDataBufferSize, &pFlac->allocationCallbacks);                                                      \
+        if (pSampleData == NULL) {                                                                                                                                  \
+            goto on_error;                                                                                                                                          \
+        }                                                                                                                                                           \
+                                                                                                                                                                    \
+        while ((pcmFramesRead = (drflac_uint64)drflac_read_pcm_frames_##extension(pFlac, sizeof(buffer)/sizeof(buffer[0])/pFlac->channels, buffer)) > 0) {          \
+            if (((totalPCMFrameCount + pcmFramesRead) * pFlac->channels * sizeof(type)) > sampleDataBufferSize) {                                                   \
+                type* pNewSampleData;                                                                                                                               \
+                size_t newSampleDataBufferSize;                                                                                                                     \
+                                                                                                                                                                    \
+                newSampleDataBufferSize = sampleDataBufferSize * 2;                                                                                                 \
+                pNewSampleData = (type*)drflac__realloc_from_callbacks(pSampleData, newSampleDataBufferSize, sampleDataBufferSize, &pFlac->allocationCallbacks);    \
+                if (pNewSampleData == NULL) {                                                                                                                       \
+                    drflac__free_from_callbacks(pSampleData, &pFlac->allocationCallbacks);                                                                          \
+                    goto on_error;                                                                                                                                  \
+                }                                                                                                                                                   \
+                                                                                                                                                                    \
+                sampleDataBufferSize = newSampleDataBufferSize;                                                                                                     \
+                pSampleData = pNewSampleData;                                                                                                                       \
+            }                                                                                                                                                       \
+                                                                                                                                                                    \
+            DRFLAC_COPY_MEMORY(pSampleData + (totalPCMFrameCount*pFlac->channels), buffer, (size_t)(pcmFramesRead*pFlac->channels*sizeof(type)));                   \
+            totalPCMFrameCount += pcmFramesRead;                                                                                                                    \
+        }                                                                                                                                                           \
+                                                                                                                                                                    \
+        /* At this point everything should be decoded, but we just want to fill the unused part buffer with silence - need to                                       \
+           protect those ears from random noise! */                                                                                                                 \
+        DRFLAC_ZERO_MEMORY(pSampleData + (totalPCMFrameCount*pFlac->channels), (size_t)(sampleDataBufferSize - totalPCMFrameCount*pFlac->channels*sizeof(type)));   \
+    } else {                                                                                                                                                        \
+        drflac_uint64 dataSize = totalPCMFrameCount*pFlac->channels*sizeof(type);                                                                                   \
+        if (dataSize > (drflac_uint64)DRFLAC_SIZE_MAX) {                                                                                                            \
+            goto on_error;  /* The decoded data is too big. */                                                                                                      \
+        }                                                                                                                                                           \
+                                                                                                                                                                    \
+        pSampleData = (type*)drflac__malloc_from_callbacks((size_t)dataSize, &pFlac->allocationCallbacks);    /* <-- Safe cast as per the check above. */           \
+        if (pSampleData == NULL) {                                                                                                                                  \
+            goto on_error;                                                                                                                                          \
+        }                                                                                                                                                           \
+                                                                                                                                                                    \
+        totalPCMFrameCount = drflac_read_pcm_frames_##extension(pFlac, pFlac->totalPCMFrameCount, pSampleData);                                                     \
+    }                                                                                                                                                               \
+                                                                                                                                                                    \
+    if (sampleRateOut) *sampleRateOut = pFlac->sampleRate;                                                                                                          \
+    if (channelsOut) *channelsOut = pFlac->channels;                                                                                                                \
+    if (totalPCMFrameCountOut) *totalPCMFrameCountOut = totalPCMFrameCount;                                                                                         \
+                                                                                                                                                                    \
+    drflac_close(pFlac);                                                                                                                                            \
+    return pSampleData;                                                                                                                                             \
+                                                                                                                                                                    \
+on_error:                                                                                                                                                           \
+    drflac_close(pFlac);                                                                                                                                            \
+    return NULL;                                                                                                                                                    \
+}
+
+DRFLAC_DEFINE_FULL_READ_AND_CLOSE(s32, drflac_int32)
+DRFLAC_DEFINE_FULL_READ_AND_CLOSE(s16, drflac_int16)
+DRFLAC_DEFINE_FULL_READ_AND_CLOSE(f32, float)
+
+DRFLAC_API drflac_int32* drflac_open_and_read_pcm_frames_s32(drflac_read_proc onRead, drflac_seek_proc onSeek, void* pUserData, unsigned int* channelsOut, unsigned int* sampleRateOut, drflac_uint64* totalPCMFrameCountOut, const drflac_allocation_callbacks* pAllocationCallbacks)
+{
+    drflac* pFlac;
+
+    if (channelsOut) {
+        *channelsOut = 0;
+    }
+    if (sampleRateOut) {
+        *sampleRateOut = 0;
+    }
+    if (totalPCMFrameCountOut) {
+        *totalPCMFrameCountOut = 0;
+    }
+
+    pFlac = drflac_open(onRead, onSeek, pUserData, pAllocationCallbacks);
+    if (pFlac == NULL) {
+        return NULL;
+    }
+
+    return drflac__full_read_and_close_s32(pFlac, channelsOut, sampleRateOut, totalPCMFrameCountOut);
+}
+
+DRFLAC_API drflac_int16* drflac_open_and_read_pcm_frames_s16(drflac_read_proc onRead, drflac_seek_proc onSeek, void* pUserData, unsigned int* channelsOut, unsigned int* sampleRateOut, drflac_uint64* totalPCMFrameCountOut, const drflac_allocation_callbacks* pAllocationCallbacks)
+{
+    drflac* pFlac;
+
+    if (channelsOut) {
+        *channelsOut = 0;
+    }
+    if (sampleRateOut) {
+        *sampleRateOut = 0;
+    }
+    if (totalPCMFrameCountOut) {
+        *totalPCMFrameCountOut = 0;
+    }
+
+    pFlac = drflac_open(onRead, onSeek, pUserData, pAllocationCallbacks);
+    if (pFlac == NULL) {
+        return NULL;
+    }
+
+    return drflac__full_read_and_close_s16(pFlac, channelsOut, sampleRateOut, totalPCMFrameCountOut);
+}
+
+DRFLAC_API float* drflac_open_and_read_pcm_frames_f32(drflac_read_proc onRead, drflac_seek_proc onSeek, void* pUserData, unsigned int* channelsOut, unsigned int* sampleRateOut, drflac_uint64* totalPCMFrameCountOut, const drflac_allocation_callbacks* pAllocationCallbacks)
+{
+    drflac* pFlac;
+
+    if (channelsOut) {
+        *channelsOut = 0;
+    }
+    if (sampleRateOut) {
+        *sampleRateOut = 0;
+    }
+    if (totalPCMFrameCountOut) {
+        *totalPCMFrameCountOut = 0;
+    }
+
+    pFlac = drflac_open(onRead, onSeek, pUserData, pAllocationCallbacks);
+    if (pFlac == NULL) {
+        return NULL;
+    }
+
+    return drflac__full_read_and_close_f32(pFlac, channelsOut, sampleRateOut, totalPCMFrameCountOut);
+}
+
+#ifndef DR_FLAC_NO_STDIO
+DRFLAC_API drflac_int32* drflac_open_file_and_read_pcm_frames_s32(const char* filename, unsigned int* channels, unsigned int* sampleRate, drflac_uint64* totalPCMFrameCount, const drflac_allocation_callbacks* pAllocationCallbacks)
+{
+    drflac* pFlac;
+
+    if (sampleRate) {
+        *sampleRate = 0;
+    }
+    if (channels) {
+        *channels = 0;
+    }
+    if (totalPCMFrameCount) {
+        *totalPCMFrameCount = 0;
+    }
+
+    pFlac = drflac_open_file(filename, pAllocationCallbacks);
+    if (pFlac == NULL) {
+        return NULL;
+    }
+
+    return drflac__full_read_and_close_s32(pFlac, channels, sampleRate, totalPCMFrameCount);
+}
+
+DRFLAC_API drflac_int16* drflac_open_file_and_read_pcm_frames_s16(const char* filename, unsigned int* channels, unsigned int* sampleRate, drflac_uint64* totalPCMFrameCount, const drflac_allocation_callbacks* pAllocationCallbacks)
+{
+    drflac* pFlac;
+
+    if (sampleRate) {
+        *sampleRate = 0;
+    }
+    if (channels) {
+        *channels = 0;
+    }
+    if (totalPCMFrameCount) {
+        *totalPCMFrameCount = 0;
+    }
+
+    pFlac = drflac_open_file(filename, pAllocationCallbacks);
+    if (pFlac == NULL) {
+        return NULL;
+    }
+
+    return drflac__full_read_and_close_s16(pFlac, channels, sampleRate, totalPCMFrameCount);
+}
+
+DRFLAC_API float* drflac_open_file_and_read_pcm_frames_f32(const char* filename, unsigned int* channels, unsigned int* sampleRate, drflac_uint64* totalPCMFrameCount, const drflac_allocation_callbacks* pAllocationCallbacks)
+{
+    drflac* pFlac;
+
+    if (sampleRate) {
+        *sampleRate = 0;
+    }
+    if (channels) {
+        *channels = 0;
+    }
+    if (totalPCMFrameCount) {
+        *totalPCMFrameCount = 0;
+    }
+
+    pFlac = drflac_open_file(filename, pAllocationCallbacks);
+    if (pFlac == NULL) {
+        return NULL;
+    }
+
+    return drflac__full_read_and_close_f32(pFlac, channels, sampleRate, totalPCMFrameCount);
+}
+#endif
+
+DRFLAC_API drflac_int32* drflac_open_memory_and_read_pcm_frames_s32(const void* data, size_t dataSize, unsigned int* channels, unsigned int* sampleRate, drflac_uint64* totalPCMFrameCount, const drflac_allocation_callbacks* pAllocationCallbacks)
+{
+    drflac* pFlac;
+
+    if (sampleRate) {
+        *sampleRate = 0;
+    }
+    if (channels) {
+        *channels = 0;
+    }
+    if (totalPCMFrameCount) {
+        *totalPCMFrameCount = 0;
+    }
+
+    pFlac = drflac_open_memory(data, dataSize, pAllocationCallbacks);
+    if (pFlac == NULL) {
+        return NULL;
+    }
+
+    return drflac__full_read_and_close_s32(pFlac, channels, sampleRate, totalPCMFrameCount);
+}
+
+DRFLAC_API drflac_int16* drflac_open_memory_and_read_pcm_frames_s16(const void* data, size_t dataSize, unsigned int* channels, unsigned int* sampleRate, drflac_uint64* totalPCMFrameCount, const drflac_allocation_callbacks* pAllocationCallbacks)
+{
+    drflac* pFlac;
+
+    if (sampleRate) {
+        *sampleRate = 0;
+    }
+    if (channels) {
+        *channels = 0;
+    }
+    if (totalPCMFrameCount) {
+        *totalPCMFrameCount = 0;
+    }
+
+    pFlac = drflac_open_memory(data, dataSize, pAllocationCallbacks);
+    if (pFlac == NULL) {
+        return NULL;
+    }
+
+    return drflac__full_read_and_close_s16(pFlac, channels, sampleRate, totalPCMFrameCount);
+}
+
+DRFLAC_API float* drflac_open_memory_and_read_pcm_frames_f32(const void* data, size_t dataSize, unsigned int* channels, unsigned int* sampleRate, drflac_uint64* totalPCMFrameCount, const drflac_allocation_callbacks* pAllocationCallbacks)
+{
+    drflac* pFlac;
+
+    if (sampleRate) {
+        *sampleRate = 0;
+    }
+    if (channels) {
+        *channels = 0;
+    }
+    if (totalPCMFrameCount) {
+        *totalPCMFrameCount = 0;
+    }
+
+    pFlac = drflac_open_memory(data, dataSize, pAllocationCallbacks);
+    if (pFlac == NULL) {
+        return NULL;
+    }
+
+    return drflac__full_read_and_close_f32(pFlac, channels, sampleRate, totalPCMFrameCount);
+}
+
+
+DRFLAC_API void drflac_free(void* p, const drflac_allocation_callbacks* pAllocationCallbacks)
+{
+    if (pAllocationCallbacks != NULL) {
+        drflac__free_from_callbacks(p, pAllocationCallbacks);
+    } else {
+        drflac__free_default(p, NULL);
+    }
+}
+
+
+
+
+DRFLAC_API void drflac_init_vorbis_comment_iterator(drflac_vorbis_comment_iterator* pIter, drflac_uint32 commentCount, const void* pComments)
+{
+    if (pIter == NULL) {
+        return;
+    }
+
+    pIter->countRemaining = commentCount;
+    pIter->pRunningData   = (const char*)pComments;
+}
+
+DRFLAC_API const char* drflac_next_vorbis_comment(drflac_vorbis_comment_iterator* pIter, drflac_uint32* pCommentLengthOut)
+{
+    drflac_int32 length;
+    const char* pComment;
+
+    /* Safety. */
+    if (pCommentLengthOut) {
+        *pCommentLengthOut = 0;
+    }
+
+    if (pIter == NULL || pIter->countRemaining == 0 || pIter->pRunningData == NULL) {
+        return NULL;
+    }
+
+    length = drflac__le2host_32_ptr_unaligned(pIter->pRunningData);
+    pIter->pRunningData += 4;
+
+    pComment = pIter->pRunningData;
+    pIter->pRunningData += length;
+    pIter->countRemaining -= 1;
+
+    if (pCommentLengthOut) {
+        *pCommentLengthOut = length;
+    }
+
+    return pComment;
+}
+
+
+
+
+DRFLAC_API void drflac_init_cuesheet_track_iterator(drflac_cuesheet_track_iterator* pIter, drflac_uint32 trackCount, const void* pTrackData)
+{
+    if (pIter == NULL) {
+        return;
+    }
+
+    pIter->countRemaining = trackCount;
+    pIter->pRunningData   = (const char*)pTrackData;
+}
+
+DRFLAC_API drflac_bool32 drflac_next_cuesheet_track(drflac_cuesheet_track_iterator* pIter, drflac_cuesheet_track* pCuesheetTrack)
+{
+    drflac_cuesheet_track cuesheetTrack;
+    const char* pRunningData;
+    drflac_uint64 offsetHi;
+    drflac_uint64 offsetLo;
+
+    if (pIter == NULL || pIter->countRemaining == 0 || pIter->pRunningData == NULL) {
+        return DRFLAC_FALSE;
+    }
+
+    pRunningData = pIter->pRunningData;
+
+    offsetHi                   = drflac__be2host_32(*(const drflac_uint32*)pRunningData); pRunningData += 4;
+    offsetLo                   = drflac__be2host_32(*(const drflac_uint32*)pRunningData); pRunningData += 4;
+    cuesheetTrack.offset       = offsetLo | (offsetHi << 32);
+    cuesheetTrack.trackNumber  = pRunningData[0];                                         pRunningData += 1;
+    DRFLAC_COPY_MEMORY(cuesheetTrack.ISRC, pRunningData, sizeof(cuesheetTrack.ISRC));     pRunningData += 12;
+    cuesheetTrack.isAudio      = (pRunningData[0] & 0x80) != 0;
+    cuesheetTrack.preEmphasis  = (pRunningData[0] & 0x40) != 0;                           pRunningData += 14;
+    cuesheetTrack.indexCount   = pRunningData[0];                                         pRunningData += 1;
+    cuesheetTrack.pIndexPoints = (const drflac_cuesheet_track_index*)pRunningData;        pRunningData += cuesheetTrack.indexCount * sizeof(drflac_cuesheet_track_index);
+
+    pIter->pRunningData = pRunningData;
+    pIter->countRemaining -= 1;
+
+    if (pCuesheetTrack) {
+        *pCuesheetTrack = cuesheetTrack;
+    }
+
+    return DRFLAC_TRUE;
+}
+
+#if defined(__clang__) || (defined(__GNUC__) && (__GNUC__ > 4 || (__GNUC__ == 4 && __GNUC_MINOR__ >= 6)))
+    #pragma GCC diagnostic pop
+#endif
+#endif  /* dr_flac_c */
+#endif  /* DR_FLAC_IMPLEMENTATION */
+
+
+/*
+REVISION HISTORY
+================
+v0.12.42 - 2023-11-02
+  - Fix build for ARMv6-M.
+  - Fix a compilation warning with GCC.
+
+v0.12.41 - 2023-06-17
+  - Fix an incorrect date in revision history. No functional change.
+
+v0.12.40 - 2023-05-22
+  - Minor code restructure. No functional change.
+
+v0.12.39 - 2022-09-17
+  - Fix compilation with DJGPP.
+  - Fix compilation error with Visual Studio 2019 and the ARM build.
+  - Fix an error with SSE 4.1 detection.
+  - Add support for disabling wchar_t with DR_WAV_NO_WCHAR.
+  - Improve compatibility with compilers which lack support for explicit struct packing.
+  - Improve compatibility with low-end and embedded hardware by reducing the amount of stack
+    allocation when loading an Ogg encapsulated file.
+
+v0.12.38 - 2022-04-10
+  - Fix compilation error on older versions of GCC.
+
+v0.12.37 - 2022-02-12
+  - Improve ARM detection.
+
+v0.12.36 - 2022-02-07
+  - Fix a compilation error with the ARM build.
+
+v0.12.35 - 2022-02-06
+  - Fix a bug due to underestimating the amount of precision required for the prediction stage.
+  - Fix some bugs found from fuzz testing.
+
+v0.12.34 - 2022-01-07
+  - Fix some misalignment bugs when reading metadata.
+
+v0.12.33 - 2021-12-22
+  - Fix a bug with seeking when the seek table does not start at PCM frame 0.
+
+v0.12.32 - 2021-12-11
+  - Fix a warning with Clang.
+
+v0.12.31 - 2021-08-16
+  - Silence some warnings.
+
+v0.12.30 - 2021-07-31
+  - Fix platform detection for ARM64.
+
+v0.12.29 - 2021-04-02
+  - Fix a bug where the running PCM frame index is set to an invalid value when over-seeking.
+  - Fix a decoding error due to an incorrect validation check.
+
+v0.12.28 - 2021-02-21
+  - Fix a warning due to referencing _MSC_VER when it is undefined.
+
+v0.12.27 - 2021-01-31
+  - Fix a static analysis warning.
+
+v0.12.26 - 2021-01-17
+  - Fix a compilation warning due to _BSD_SOURCE being deprecated.
+
+v0.12.25 - 2020-12-26
+  - Update documentation.
+
+v0.12.24 - 2020-11-29
+  - Fix ARM64/NEON detection when compiling with MSVC.
+
+v0.12.23 - 2020-11-21
+  - Fix compilation with OpenWatcom.
+
+v0.12.22 - 2020-11-01
+  - Fix an error with the previous release.
+
+v0.12.21 - 2020-11-01
+  - Fix a possible deadlock when seeking.
+  - Improve compiler support for older versions of GCC.
+
+v0.12.20 - 2020-09-08
+  - Fix a compilation error on older compilers.
+
+v0.12.19 - 2020-08-30
+  - Fix a bug due to an undefined 32-bit shift.
+
+v0.12.18 - 2020-08-14
+  - Fix a crash when compiling with clang-cl.
+
+v0.12.17 - 2020-08-02
+  - Simplify sized types.
+
+v0.12.16 - 2020-07-25
+  - Fix a compilation warning.
+
+v0.12.15 - 2020-07-06
+  - Check for negative LPC shifts and return an error.
+
+v0.12.14 - 2020-06-23
+  - Add include guard for the implementation section.
+
+v0.12.13 - 2020-05-16
+  - Add compile-time and run-time version querying.
+    - DRFLAC_VERSION_MINOR
+    - DRFLAC_VERSION_MAJOR
+    - DRFLAC_VERSION_REVISION
+    - DRFLAC_VERSION_STRING
+    - drflac_version()
+    - drflac_version_string()
+
+v0.12.12 - 2020-04-30
+  - Fix compilation errors with VC6.
+
+v0.12.11 - 2020-04-19
+  - Fix some pedantic warnings.
+  - Fix some undefined behaviour warnings.
+
+v0.12.10 - 2020-04-10
+  - Fix some bugs when trying to seek with an invalid seek table.
+
+v0.12.9 - 2020-04-05
+  - Fix warnings.
+
+v0.12.8 - 2020-04-04
+  - Add drflac_open_file_w() and drflac_open_file_with_metadata_w().
+  - Fix some static analysis warnings.
+  - Minor documentation updates.
+
+v0.12.7 - 2020-03-14
+  - Fix compilation errors with VC6.
+
+v0.12.6 - 2020-03-07
+  - Fix compilation error with Visual Studio .NET 2003.
+
+v0.12.5 - 2020-01-30
+  - Silence some static analysis warnings.
+
+v0.12.4 - 2020-01-29
+  - Silence some static analysis warnings.
+
+v0.12.3 - 2019-12-02
+  - Fix some warnings when compiling with GCC and the -Og flag.
+  - Fix a crash in out-of-memory situations.
+  - Fix potential integer overflow bug.
+  - Fix some static analysis warnings.
+  - Fix a possible crash when using custom memory allocators without a custom realloc() implementation.
+  - Fix a bug with binary search seeking where the bits per sample is not a multiple of 8.
+
+v0.12.2 - 2019-10-07
+  - Internal code clean up.
+
+v0.12.1 - 2019-09-29
+  - Fix some Clang Static Analyzer warnings.
+  - Fix an unused variable warning.
+
+v0.12.0 - 2019-09-23
+  - API CHANGE: Add support for user defined memory allocation routines. This system allows the program to specify their own memory allocation
+    routines with a user data pointer for client-specific contextual data. This adds an extra parameter to the end of the following APIs:
+    - drflac_open()
+    - drflac_open_relaxed()
+    - drflac_open_with_metadata()
+    - drflac_open_with_metadata_relaxed()
+    - drflac_open_file()
+    - drflac_open_file_with_metadata()
+    - drflac_open_memory()
+    - drflac_open_memory_with_metadata()
+    - drflac_open_and_read_pcm_frames_s32()
+    - drflac_open_and_read_pcm_frames_s16()
+    - drflac_open_and_read_pcm_frames_f32()
+    - drflac_open_file_and_read_pcm_frames_s32()
+    - drflac_open_file_and_read_pcm_frames_s16()
+    - drflac_open_file_and_read_pcm_frames_f32()
+    - drflac_open_memory_and_read_pcm_frames_s32()
+    - drflac_open_memory_and_read_pcm_frames_s16()
+    - drflac_open_memory_and_read_pcm_frames_f32()
+    Set this extra parameter to NULL to use defaults which is the same as the previous behaviour. Setting this NULL will use
+    DRFLAC_MALLOC, DRFLAC_REALLOC and DRFLAC_FREE.
+  - Remove deprecated APIs:
+    - drflac_read_s32()
+    - drflac_read_s16()
+    - drflac_read_f32()
+    - drflac_seek_to_sample()
+    - drflac_open_and_decode_s32()
+    - drflac_open_and_decode_s16()
+    - drflac_open_and_decode_f32()
+    - drflac_open_and_decode_file_s32()
+    - drflac_open_and_decode_file_s16()
+    - drflac_open_and_decode_file_f32()
+    - drflac_open_and_decode_memory_s32()
+    - drflac_open_and_decode_memory_s16()
+    - drflac_open_and_decode_memory_f32()
+  - Remove drflac.totalSampleCount which is now replaced with drflac.totalPCMFrameCount. You can emulate drflac.totalSampleCount
+    by doing pFlac->totalPCMFrameCount*pFlac->channels.
+  - Rename drflac.currentFrame to drflac.currentFLACFrame to remove ambiguity with PCM frames.
+  - Fix errors when seeking to the end of a stream.
+  - Optimizations to seeking.
+  - SSE improvements and optimizations.
+  - ARM NEON optimizations.
+  - Optimizations to drflac_read_pcm_frames_s16().
+  - Optimizations to drflac_read_pcm_frames_s32().
+
+v0.11.10 - 2019-06-26
+  - Fix a compiler error.
+
+v0.11.9 - 2019-06-16
+  - Silence some ThreadSanitizer warnings.
+
+v0.11.8 - 2019-05-21
+  - Fix warnings.
+
+v0.11.7 - 2019-05-06
+  - C89 fixes.
+
+v0.11.6 - 2019-05-05
+  - Add support for C89.
+  - Fix a compiler warning when CRC is disabled.
+  - Change license to choice of public domain or MIT-0.
+
+v0.11.5 - 2019-04-19
+  - Fix a compiler error with GCC.
+
+v0.11.4 - 2019-04-17
+  - Fix some warnings with GCC when compiling with -std=c99.
+
+v0.11.3 - 2019-04-07
+  - Silence warnings with GCC.
+
+v0.11.2 - 2019-03-10
+  - Fix a warning.
+
+v0.11.1 - 2019-02-17
+  - Fix a potential bug with seeking.
+
+v0.11.0 - 2018-12-16
+  - API CHANGE: Deprecated drflac_read_s32(), drflac_read_s16() and drflac_read_f32() and replaced them with
+    drflac_read_pcm_frames_s32(), drflac_read_pcm_frames_s16() and drflac_read_pcm_frames_f32(). The new APIs take
+    and return PCM frame counts instead of sample counts. To upgrade you will need to change the input count by
+    dividing it by the channel count, and then do the same with the return value.
+  - API_CHANGE: Deprecated drflac_seek_to_sample() and replaced with drflac_seek_to_pcm_frame(). Same rules as
+    the changes to drflac_read_*() apply.
+  - API CHANGE: Deprecated drflac_open_and_decode_*() and replaced with drflac_open_*_and_read_*(). Same rules as
+    the changes to drflac_read_*() apply.
+  - Optimizations.
+
+v0.10.0 - 2018-09-11
+  - Remove the DR_FLAC_NO_WIN32_IO option and the Win32 file IO functionality. If you need to use Win32 file IO you
+    need to do it yourself via the callback API.
+  - Fix the clang build.
+  - Fix undefined behavior.
+  - Fix errors with CUESHEET metdata blocks.
+  - Add an API for iterating over each cuesheet track in the CUESHEET metadata block. This works the same way as the
+    Vorbis comment API.
+  - Other miscellaneous bug fixes, mostly relating to invalid FLAC streams.
+  - Minor optimizations.
+
+v0.9.11 - 2018-08-29
+  - Fix a bug with sample reconstruction.
+
+v0.9.10 - 2018-08-07
+  - Improve 64-bit detection.
+
+v0.9.9 - 2018-08-05
+  - Fix C++ build on older versions of GCC.
+
+v0.9.8 - 2018-07-24
+  - Fix compilation errors.
+
+v0.9.7 - 2018-07-05
+  - Fix a warning.
+
+v0.9.6 - 2018-06-29
+  - Fix some typos.
+
+v0.9.5 - 2018-06-23
+  - Fix some warnings.
+
+v0.9.4 - 2018-06-14
+  - Optimizations to seeking.
+  - Clean up.
+
+v0.9.3 - 2018-05-22
+  - Bug fix.
+
+v0.9.2 - 2018-05-12
+  - Fix a compilation error due to a missing break statement.
+
+v0.9.1 - 2018-04-29
+  - Fix compilation error with Clang.
+
+v0.9 - 2018-04-24
+  - Fix Clang build.
+  - Start using major.minor.revision versioning.
+
+v0.8g - 2018-04-19
+  - Fix build on non-x86/x64 architectures.
+
+v0.8f - 2018-02-02
+  - Stop pretending to support changing rate/channels mid stream.
+
+v0.8e - 2018-02-01
+  - Fix a crash when the block size of a frame is larger than the maximum block size defined by the FLAC stream.
+  - Fix a crash the the Rice partition order is invalid.
+
+v0.8d - 2017-09-22
+  - Add support for decoding streams with ID3 tags. ID3 tags are just skipped.
+
+v0.8c - 2017-09-07
+  - Fix warning on non-x86/x64 architectures.
+
+v0.8b - 2017-08-19
+  - Fix build on non-x86/x64 architectures.
+
+v0.8a - 2017-08-13
+  - A small optimization for the Clang build.
+
+v0.8 - 2017-08-12
+  - API CHANGE: Rename dr_* types to drflac_*.
+  - Optimizations. This brings dr_flac back to about the same class of efficiency as the reference implementation.
+  - Add support for custom implementations of malloc(), realloc(), etc.
+  - Add CRC checking to Ogg encapsulated streams.
+  - Fix VC++ 6 build. This is only for the C++ compiler. The C compiler is not currently supported.
+  - Bug fixes.
+
+v0.7 - 2017-07-23
+  - Add support for opening a stream without a header block. To do this, use drflac_open_relaxed() / drflac_open_with_metadata_relaxed().
+
+v0.6 - 2017-07-22
+  - Add support for recovering from invalid frames. With this change, dr_flac will simply skip over invalid frames as if they
+    never existed. Frames are checked against their sync code, the CRC-8 of the frame header and the CRC-16 of the whole frame.
+
+v0.5 - 2017-07-16
+  - Fix typos.
+  - Change drflac_bool* types to unsigned.
+  - Add CRC checking. This makes dr_flac slower, but can be disabled with #define DR_FLAC_NO_CRC.
+
+v0.4f - 2017-03-10
+  - Fix a couple of bugs with the bitstreaming code.
+
+v0.4e - 2017-02-17
+  - Fix some warnings.
+
+v0.4d - 2016-12-26
+  - Add support for 32-bit floating-point PCM decoding.
+  - Use drflac_int* and drflac_uint* sized types to improve compiler support.
+  - Minor improvements to documentation.
+
+v0.4c - 2016-12-26
+  - Add support for signed 16-bit integer PCM decoding.
+
+v0.4b - 2016-10-23
+  - A minor change to drflac_bool8 and drflac_bool32 types.
+
+v0.4a - 2016-10-11
+  - Rename drBool32 to drflac_bool32 for styling consistency.
+
+v0.4 - 2016-09-29
+  - API/ABI CHANGE: Use fixed size 32-bit booleans instead of the built-in bool type.
+  - API CHANGE: Rename drflac_open_and_decode*() to drflac_open_and_decode*_s32().
+  - API CHANGE: Swap the order of "channels" and "sampleRate" parameters in drflac_open_and_decode*(). Rationale for this is to
+    keep it consistent with drflac_audio.
+
+v0.3f - 2016-09-21
+  - Fix a warning with GCC.
+
+v0.3e - 2016-09-18
+  - Fixed a bug where GCC 4.3+ was not getting properly identified.
+  - Fixed a few typos.
+  - Changed date formats to ISO 8601 (YYYY-MM-DD).
+
+v0.3d - 2016-06-11
+  - Minor clean up.
+
+v0.3c - 2016-05-28
+  - Fixed compilation error.
+
+v0.3b - 2016-05-16
+  - Fixed Linux/GCC build.
+  - Updated documentation.
+
+v0.3a - 2016-05-15
+  - Minor fixes to documentation.
+
+v0.3 - 2016-05-11
+  - Optimizations. Now at about parity with the reference implementation on 32-bit builds.
+  - Lots of clean up.
+
+v0.2b - 2016-05-10
+  - Bug fixes.
+
+v0.2a - 2016-05-10
+  - Made drflac_open_and_decode() more robust.
+  - Removed an unused debugging variable
+
+v0.2 - 2016-05-09
+  - Added support for Ogg encapsulation.
+  - API CHANGE. Have the onSeek callback take a third argument which specifies whether or not the seek
+    should be relative to the start or the current position. Also changes the seeking rules such that
+    seeking offsets will never be negative.
+  - Have drflac_open_and_decode() fail gracefully if the stream has an unknown total sample count.
+
+v0.1b - 2016-05-07
+  - Properly close the file handle in drflac_open_file() and family when the decoder fails to initialize.
+  - Removed a stale comment.
+
+v0.1a - 2016-05-05
+  - Minor formatting changes.
+  - Fixed a warning on the GCC build.
+
+v0.1 - 2016-05-03
+  - Initial versioned release.
+*/
+
+/*
+This software is available as a choice of the following licenses. Choose
+whichever you prefer.
+
+===============================================================================
+ALTERNATIVE 1 - Public Domain (www.unlicense.org)
+===============================================================================
+This is free and unencumbered software released into the public domain.
+
+Anyone is free to copy, modify, publish, use, compile, sell, or distribute this
+software, either in source code form or as a compiled binary, for any purpose,
+commercial or non-commercial, and by any means.
+
+In jurisdictions that recognize copyright laws, the author or authors of this
+software dedicate any and all copyright interest in the software to the public
+domain. We make this dedication for the benefit of the public at large and to
+the detriment of our heirs and successors. We intend this dedication to be an
+overt act of relinquishment in perpetuity of all present and future rights to
+this software under copyright law.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION
+WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
+
+For more information, please refer to <http://unlicense.org/>
+
+===============================================================================
+ALTERNATIVE 2 - MIT No Attribution
+===============================================================================
+Copyright 2023 David Reid
+
+Permission is hereby granted, free of charge, to any person obtaining a copy of
+this software and associated documentation files (the "Software"), to deal in
+the Software without restriction, including without limitation the rights to
+use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies
+of the Software, and to permit persons to whom the Software is furnished to do
+so.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.
+*/
diff --git a/error.h b/error.h
index 96413cb..7ca91ac 100644
--- a/error.h
+++ b/error.h
@@ -2,13 +2,9 @@
 #define error_h
 
 typedef enum {
-	error_none = 0,
-	error_file_not_found,
+	error_none,
 	error_out_of_memory,
-	error_platform_error,
-	error_multithreading,
-	error_font,
-	error_ui
+	error_read_error
 } Error;
 
 #endif
diff --git a/font.c b/font.c
deleted file mode 100644
index ba6dfb6..0000000
--- a/font.c
+++ /dev/null
@@ -1,238 +0,0 @@
-#include "memory.h"
-#include "plat.h"
-#include "render.h"
-#include "stb_rect_pack.h"
-#include "stb_truetype.h"
-#include <math.h>
-
-typedef struct {
-	Bitmap bmp;
-	stbtt_bakedchar glyphs[max_glyphset];
-} Glyph_Set;
-
-struct Font {
-	const void* data;
-	Heap* heap;
-	stbtt_fontinfo info;
-	Glyph_Set* sets[max_glyphset];
-	int size, height;
-};
-
-int font_height(Font* f) {
-	return f->height;
-}
-
-static const char* utf8_to_codepoint(
-	const char* p,
-	unsigned* dst
-) {
-	unsigned res, n;
-	switch (*p & 0xf0) {
-		case 0xf0 : res = *p & 0x07; n = 3; break;
-		case 0xe0 : res = *p & 0x0f; n = 2; break;
-		case 0xd0 :
-		case 0xc0 : res = *p & 0x1f; n = 1; break;
-		default   : res = *p;        n = 0; break;
-	}
-	while (n--) {
-		res = (res << 6) | (*(++p) & 0x3f);
-	}
-	*dst = res;
-	return p + 1;
-}
-
-static Glyph_Set* load_glyph_set(
-	Font* font,
-	int idx
-) {
-	Glyph_Set* gs;
-	Bitmap* bmp;
-	int r, i;
-	int ascent, descent, linegap, scaled_ascent;
-	float s;
-	gs = heap_alloc(font->heap, sizeof *gs);
-	bmp = &gs->bmp;
-	bmp->w = 512;
-	bmp->h = 512;
-retry:
-	bmp->pixels = heap_alloc(
-		font->heap,
-		bmp->w * bmp->h * 4
-	);
-	s =
-		stbtt_ScaleForMappingEmToPixels(&font->info, 1) /
-		stbtt_ScaleForPixelHeight(&font->info, 1);
-	r = stbtt_BakeFontBitmap(
-		font->data,
-		0,
-		font->size * s,
-		(unsigned char*)bmp->pixels,
-		bmp->w,
-		bmp->h,
-		idx * 256,
-		256,
-		gs->glyphs
-	);
-	if (r <= 0) {
-		bmp->w *= 2;
-		bmp->h *= 2;
-		heap_free(font->heap, bmp->pixels);
-		goto retry;
-	}
-
-	stbtt_GetFontVMetrics(&font->info, &ascent, &descent, &linegap);
-	s = stbtt_ScaleForMappingEmToPixels(&font->info, font->size);
-	scaled_ascent = (int)(ascent * s + 0.5f);
-	for (i = 0; i < 256; i++) {
-		gs->glyphs[i].yoff += scaled_ascent;
-		gs->glyphs[i].xadvance = (float)floor(gs->glyphs[i].xadvance);
-	}
-	for (i = bmp->w * bmp->h - 1; i >= 0; i--) {
-		unsigned char n = ((unsigned char*)bmp->pixels)[i];
-		bmp->pixels[i].r = 255;
-		bmp->pixels[i].g = 255;
-		bmp->pixels[i].b = 255;
-		bmp->pixels[i].a = n;
-	}
-	return gs;
-}
-
-static Glyph_Set* get_glyph_set(
-	Font* font,
-	int codepoint
-) {
-	int idx = (codepoint >> 8) % max_glyphset;
-	if (!font->sets[idx]) {
-		font->sets[idx] = load_glyph_set(font, idx);
-	}
-	return font->sets[idx];
-}
-
-void init_font(
-	Font* font,
-	Heap* heap,
-	const unsigned char* raw,
-	int size
-) {
-	int r, i;
-	int ascent, descent, linegap;
-	float scale;
-	font->heap = heap;
-	font->data = raw;
-	font->size = size;
-	stbtt_bakedchar* g;
-	for (i = 0; i < max_glyphset; i++)
-		font->sets[i] = 0;
-	r = stbtt_InitFont(&font->info, font->data, 0);
-	if (!r) {
-		print("Invalid font.\n");
-		pbreak(error_font);
-	}
-	stbtt_GetFontVMetrics(
-		&font->info,
-		&ascent,
-		&descent,
-		&linegap
-	);
-	scale = stbtt_ScaleForMappingEmToPixels(
-		&font->info,
-		size
-	);
-	font->height = (int)(
-		(ascent - descent + linegap) *
-		scale + 0.5f
-	);
-	g = get_glyph_set(font, '\n')->glyphs;
-	g['\t'].x1 = g['\t'].x0;
-	g['\n'].x1 = g['\n'].x0;
-}
-
-Font* new_font(
-	Heap* h,
-	const unsigned char* raw,
-	int size
-) {
-	Font* f;
-	f = heap_alloc(h, sizeof *f);
-	init_font(f, h, raw, size);
-	return f;
-}
-
-Rectangle text_rect(
-	Font* font,
-	const char* text
-) {
-	const char* p = text;
-	unsigned codepoint;
-	int x = 0;
-	Glyph_Set* gs;
-	stbtt_bakedchar* g;
-	Rectangle re = { 0 }, r;
-	while (*p) {
-		p = utf8_to_codepoint(p, &codepoint);
-		gs = get_glyph_set(font, codepoint);
-		g = &gs->glyphs[codepoint & 0xff];
-		r.x = x + (int)g->xoff;
-		r.y = (int)g->yoff;
-		r.w = (int)g->x1 - g->x0;
-		r.h = (int)g->y1 - g->y0;
-		rect_merge(&re, &r);
-		x += (int)g->xadvance;
-	}
-	return re;
-}
-
-#define rf(expr) \
-	const char* p = text; \
-	unsigned codepoint; \
-	Glyph_Set* gs; \
-	stbtt_bakedchar* g; \
-	Rectangle r; \
-	while (*p) { \
-		p = utf8_to_codepoint(p, &codepoint); \
-		gs = get_glyph_set(font, codepoint); \
-		g = &gs->glyphs[codepoint & 0xff]; \
-		r.x = (int)g->x0; \
-		r.y = (int)g->y0; \
-		r.w = (int)g->x1 - g->x0; \
-		r.h = (int)g->y1 - g->y0; \
-		expr; \
-		x += (int)g->xadvance; \
-	}
-
-void rfont_text(
-	Font* font,
-	int x,
-	int y,
-	const char* text
-) {
-	rf(render_bitmap(
-		&gs->bmp,
-		x + (int)g->xoff,
-		y + (int)g->yoff,
-		&r
-	));
-}
-
-void rfont_text_col(
-	Font* font,
-	int x,
-	int y,
-	const char* text,
-	Colour colour
-) {
-	rf(render_bitmap_col(
-		&gs->bmp,
-		x + (int)g->xoff,
-		y + (int)g->yoff,
-		&r,
-		colour
-	));
-}
-
-#undef rf
-
-#define STB_RECT_PACK_IMPLEMENTATION
-#define STB_TRUETYPE_IMPLEMENTATION
-#include "stb_rect_pack.h"
-#include "stb_truetype.h"
diff --git a/library.c b/library.c
index 7459579..4eb4ab0 100644
--- a/library.c
+++ b/library.c
@@ -1,14 +1,9 @@
-#include <string.h>
-
 #include "library.h"
 #include "plat.h"
+#include <string.h>
+#include <stdio.h>
 
-void init_library(Library* l) {
-	int i;
-	l->song_count = 0;
-	for (i = 0; i < max_songs; i++)
-		l->songs[i].file[0] = 0;
-}
+#include "dr_flac.h"
 
 static unsigned hash(const char* s) {
 	const unsigned char* p = (const unsigned char*)s;
@@ -19,30 +14,128 @@ static unsigned hash(const char* s) {
 }
 
 Song* find_song(Library* l, const char* path) {
-	int i = hash(path) % max_songs;
+	int i = hash(path) % l->cap;
 	int c;
 	Song* s;
-	for (c = 0; c < max_songs; c++) {
+	for (c = 0; c < l->cap; c++) {
 		s = &l->songs[i];
-		if (!s->file[0] || !strcmp(path, s->file))
+		if (!s->path[0] || !strcmp(path, s->path))
 			return s;
 		i++;
-		i %= max_songs;
+		i %= l->cap;
 	}
 	return 0;
 }
 
-void add_song(void* l, const char* path) {
+static void cap_iter(
+	void* uptr,
+	const char* path
+) {
+	Library* lib = uptr;
+	(void)path;
+	lib->cap++;
+}
+
+static void adder_iter(
+	void* uptr,
+	const char* path
+) {
 	Song* s;
-	s = find_song(l, path);
-	if (!s) {
-		print_err("Too many songs!");
-		return;
+	Library* lib = uptr;
+	s = find_song(lib, path);
+	if (get_song_meta(path, s)) {
+		strcpy(s->path, path);
+		lib->indices[lib->cnt++] = s - lib->songs;
+	}
+}
+
+void build_library(
+	Arena* a,
+	Library* lib,
+	const char* path
+) {
+	int i;
+	lib->cnt = lib->cap = 0;
+	iter_dir(path, cap_iter, lib);
+	lib->cap += 16;
+	lib->songs = arena_alloc(
+		a,
+		lib->cap * sizeof *lib->songs
+	);
+	lib->indices = arena_alloc(
+		a,
+		lib->cap * sizeof *lib->indices
+	);
+	for (i = 0; i < lib->cap; i++)
+		lib->songs[i].path[0] = 0;
+	iter_dir(path, adder_iter, lib);
+}
+
+void parse_vorbis_comment(
+	Song* s,
+	const char* com,
+	int len
+) {
+	/* this is not safe xDDDD idrc */
+	if (len > 64) {
+		len = 64;
+	}
+	if (!memcmp(com, "ARTIST=", 7)) {
+		len -= 7;
+		memcpy(s->artist, com + 7, len);
+		s->artist[len] = 0;
+	} else if (!memcmp(com, "TITLE=", 6)) {
+		len -= 6;
+		memcpy(s->name, com + 6, len);
+		s->name[len] = 0;
+	} else if (!memcmp(com, "ALBUM=", 6)) {
+		len -= 6;
+		memcpy(s->album, com + 6, len);
+		s->album[len] = 0;
+	}
+}
+
+void flac_meta(
+	void* uptr,
+	drflac_metadata* m
+) {
+	Song* s = uptr;
+	drflac_vorbis_comment_iterator i;
+	drflac_uint32 len;
+	const char* c;
+	int size;
+	switch (m->type) {
+		case DRFLAC_METADATA_BLOCK_TYPE_VORBIS_COMMENT:
+			drflac_init_vorbis_comment_iterator(
+				&i,
+				m->data.vorbis_comment.commentCount,
+				m->data.vorbis_comment.pComments
+			);
+			while ((c = drflac_next_vorbis_comment(
+				&i,
+				&len
+			))) {
+				parse_vorbis_comment(s, c, len);
+			}
+			break;
 	}
-	strcpy(s->file, path);
-	strcpy(s->name, path);
 }
 
-void build_library(Library* l, const char* path) {
-	iter_dir(path, add_song, l);
+int get_song_meta(const char* path, Song* s) {
+	drflac* f;
+	f = drflac_open_file_with_metadata(
+		path,
+		flac_meta,
+		s,
+		0
+	);
+	if (f) {
+		drflac_close(f);
+		return 1;
+	}
+	s->path[0] = 0;
+	return 0;
 }
+
+#define DR_FLAC_IMPLEMENTATION
+#include "dr_flac.h"
diff --git a/library.h b/library.h
index 7c9e000..e94c10a 100644
--- a/library.h
+++ b/library.h
@@ -1,19 +1,27 @@
 #ifndef library_h
 #define library_h
 
-#include "config.h"
+#include "memory.h"
 
 typedef struct {
-	char name[song_name_max];
-	char file[song_filename_max];
+	char path[256];
+	char name[64];
+	char artist[64];
+	char album[64];
 } Song;
 
 typedef struct {
-	Song songs[max_songs];
-	int song_count;
+	Song* songs;
+	int* indices;
+	int cap, cnt;
 } Library;
 
-void init_library(Library* l);
-void build_library(Library* l, const char* path);
+void build_library(
+	Arena* a,
+	Library* lib,
+	const char* path
+);
+
+int get_song_meta(const char* path, Song* s);
 
 #endif
diff --git a/luigi.c b/luigi.c
new file mode 100644
index 0000000..5c371d6
--- /dev/null
+++ b/luigi.c
@@ -0,0 +1,87 @@
+#include <stddef.h>
+
+/*
+todo get this working proper. i don't think it's
+possible in X however.
+*/
+/*
+#include "memory.h"
+#include "config.h"
+
+Heap ui_heap = { 0 };
+
+static void try_init_heap(void) {
+	if (!ui_heap.buffer) {
+		init_heap(
+			&ui_heap,
+			galloc(ui_memory_size),
+			ui_memory_size
+		);
+	}
+}
+
+void* ui_malloc(size_t size) {
+	try_init_heap();
+	return heap_alloc(&ui_heap, size);
+}
+
+void ui_free(void* ptr) {
+	try_init_heap();
+	heap_free(&ui_heap, ptr);
+}
+
+void* ui_realloc(void* ptr, size_t size) {
+	void* old;
+	size_t os, diff;
+	if (!ptr)
+		return ui_malloc(size);
+	if (!size && ptr)
+		ui_free(ptr);
+	os = heap_block_size(ptr);
+	if (size == os) return ptr;
+	diff = size;
+	if (size > os)
+		diff = os;
+	old = ptr;
+	ptr = ui_malloc(size);
+	memcpy(ptr, old, diff);
+	ui_free(old);
+	return ptr;
+}
+
+void* ui_calloc(size_t c, size_t s) {
+	void* p;
+	s *= c;
+	p = ui_malloc(s);
+	memset(p, 0, s);
+	return p;
+}
+
+#ifdef malloc
+#undef malloc
+#endif
+#ifdef free
+#undef free
+#endif
+#ifdef realloc
+#undef realloc
+#endif
+#ifdef calloc
+#undef calloc
+#endif
+
+#define malloc ui_malloc
+#define free ui_free
+#define realloc ui_realloc
+#define calloc ui_calloc
+*/
+
+#define UI_IMPLEMENTATION
+#include "luigi.h"
+
+/*
+#undef malloc
+#undef free
+#undef realloc
+#undef calloc
+*/
diff --git a/luigi.h b/luigi.h
new file mode 100644
index 0000000..ec1711d
--- /dev/null
+++ b/luigi.h
@@ -0,0 +1,5931 @@
+// TODO UITextbox features - mouse input, multi-line, clipboard, undo, IME support, number dragging.
+// TODO New elements - list view, menu bar.
+// TODO Keyboard navigation - menus, dialogs, tables.
+// TODO Easier to use fonts; GDI font support.
+// TODO Formalize the notion of size-stability? See _UIExpandPaneButtonInvoke.
+
+#include <stdint.h>
+#include <stddef.h>
+#include <stdbool.h>
+#include <stdarg.h>
+
+#ifdef UI_LINUX
+#include <X11/Xlib.h>
+#include <X11/Xutil.h>
+#include <X11/Xatom.h>
+#include <X11/cursorfont.h>
+
+#include <xmmintrin.h>
+#endif
+
+#define _UI_TO_STRING_1(x) #x
+#define _UI_TO_STRING_2(x) _UI_TO_STRING_1(x)
+
+#ifdef UI_WINDOWS
+#undef _UNICODE
+#undef UNICODE
+#include <windows.h>
+#include <shellapi.h>
+
+#define UI_ASSERT(x) do { if (!(x)) { ui.assertionFailure = true; \
+	MessageBox(0, "Assertion failure on line " _UI_TO_STRING_2(__LINE__), 0, 0); \
+	ExitProcess(1); } } while (0)
+#define UI_CALLOC(x) HeapAlloc(ui.heap, HEAP_ZERO_MEMORY, (x))
+#define UI_FREE(x) HeapFree(ui.heap, 0, (x))
+#define UI_MALLOC(x) HeapAlloc(ui.heap, 0, (x))
+#define UI_REALLOC _UIHeapReAlloc
+#define UI_CLOCK GetTickCount
+#define UI_CLOCKS_PER_SECOND (1000)
+#define UI_CLOCK_T DWORD
+#endif
+
+#if defined(UI_LINUX)
+#include <stdlib.h>
+#include <string.h>
+#include <assert.h>
+#include <time.h>
+#include <math.h>
+
+#define UI_ASSERT assert
+#define UI_CALLOC(x) calloc(1, (x))
+#define UI_FREE free
+#define UI_MALLOC malloc
+#define UI_REALLOC realloc
+#define UI_CLOCK clock
+#define UI_CLOCKS_PER_SECOND CLOCKS_PER_SEC
+#define UI_CLOCK_T clock_t
+#endif
+
+#if defined(UI_ESSENCE)
+#include <essence.h>
+
+#define UI_ASSERT EsAssert
+#define UI_CALLOC(x) EsHeapAllocate((x), true)
+#define UI_FREE EsHeapFree
+#define UI_MALLOC(x) EsHeapAllocate((x), false)
+#define UI_REALLOC(x, y) EsHeapReallocate((x), (y), false)
+#define UI_CLOCK EsTimeStampMs
+#define UI_CLOCKS_PER_SECOND 1000
+#define UI_CLOCK_T uint64_t
+
+// Callback to allow the application to process messages.
+void _UIMessageProcess(EsMessage *message);
+#endif
+
+#ifdef UI_DEBUG
+#include <stdio.h>
+#endif
+
+#ifdef UI_FREETYPE
+#include <ft2build.h>
+#include FT_FREETYPE_H
+#include <freetype/ftbitmap.h>
+#endif
+
+#define UI_SIZE_BUTTON_MINIMUM_WIDTH (100)
+#define UI_SIZE_BUTTON_PADDING (16)
+#define UI_SIZE_BUTTON_HEIGHT (27)
+#define UI_SIZE_BUTTON_CHECKED_AREA (4)
+
+#define UI_SIZE_CHECKBOX_BOX (14)
+#define UI_SIZE_CHECKBOX_GAP (8)
+
+#define UI_SIZE_MENU_ITEM_HEIGHT (24)
+#define UI_SIZE_MENU_ITEM_MINIMUM_WIDTH (160)
+#define UI_SIZE_MENU_ITEM_MARGIN (9)
+
+#define UI_SIZE_GAUGE_WIDTH (200)
+#define UI_SIZE_GAUGE_HEIGHT (22)
+
+#define UI_SIZE_SLIDER_WIDTH (200)
+#define UI_SIZE_SLIDER_HEIGHT (25)
+#define UI_SIZE_SLIDER_THUMB (15)
+#define UI_SIZE_SLIDER_TRACK (5)
+
+#define UI_SIZE_TEXTBOX_MARGIN (3)
+#define UI_SIZE_TEXTBOX_WIDTH (200)
+#define UI_SIZE_TEXTBOX_HEIGHT (27)
+
+#define UI_SIZE_TAB_PANE_SPACE_TOP (2)
+#define UI_SIZE_TAB_PANE_SPACE_LEFT (4)
+
+#define UI_SIZE_SPLITTER (8)
+
+#define UI_SIZE_SCROLL_BAR (16)
+#define UI_SIZE_SCROLL_MINIMUM_THUMB (20)
+
+#define UI_SIZE_CODE_MARGIN (ui.activeFont->glyphWidth * 5)
+#define UI_SIZE_CODE_MARGIN_GAP (ui.activeFont->glyphWidth * 1)
+
+#define UI_SIZE_TABLE_HEADER (26)
+#define UI_SIZE_TABLE_COLUMN_GAP (20)
+#define UI_SIZE_TABLE_ROW (20)
+
+#define UI_SIZE_PANE_MEDIUM_BORDER (5)
+#define UI_SIZE_PANE_MEDIUM_GAP (5)
+#define UI_SIZE_PANE_SMALL_BORDER (3)
+#define UI_SIZE_PANE_SMALL_GAP (3)
+
+#define UI_SIZE_MDI_CHILD_BORDER (6)
+#define UI_SIZE_MDI_CHILD_TITLE (30)
+#define UI_SIZE_MDI_CHILD_CORNER (12)
+#define UI_SIZE_MDI_CHILD_MINIMUM_WIDTH (100)
+#define UI_SIZE_MDI_CHILD_MINIMUM_HEIGHT (50)
+#define UI_SIZE_MDI_CASCADE (30)
+
+#define UI_UPDATE_HOVERED (1)
+#define UI_UPDATE_PRESSED (2)
+#define UI_UPDATE_FOCUSED (3)
+
+typedef enum UIMessage {
+	UI_MSG_PAINT, // dp = pointer to UIPainter
+	UI_MSG_LAYOUT,
+	UI_MSG_DESTROY,
+	UI_MSG_UPDATE, // di = UI_UPDATE_... constant
+	UI_MSG_ANIMATE,
+	UI_MSG_SCROLLED,
+	UI_MSG_GET_WIDTH, // di = height (if known); return width
+	UI_MSG_GET_HEIGHT, // di = width (if known); return height
+	UI_MSG_FIND_BY_POINT, // dp = pointer to UIFindByPoint; return 1 if handled
+	UI_MSG_CLIENT_PARENT, // dp = pointer to UIElement *, set it to the parent for client elements
+
+	UI_MSG_INPUT_EVENTS_START, // not sent to disabled elements
+	UI_MSG_LEFT_DOWN,
+	UI_MSG_LEFT_UP,
+	UI_MSG_MIDDLE_DOWN,
+	UI_MSG_MIDDLE_UP,
+	UI_MSG_RIGHT_DOWN,
+	UI_MSG_RIGHT_UP,
+	UI_MSG_KEY_TYPED, // dp = pointer to UIKeyTyped; return 1 if handled
+	UI_MSG_MOUSE_MOVE,
+	UI_MSG_MOUSE_DRAG,
+	UI_MSG_MOUSE_WHEEL, // di = delta; return 1 if handled
+	UI_MSG_CLICKED,
+	UI_MSG_GET_CURSOR, // return cursor code
+	UI_MSG_PRESSED_DESCENDENT, // dp = pointer to child that is/contains pressed element
+	UI_MSG_INPUT_EVENTS_END,
+
+	UI_MSG_VALUE_CHANGED, // sent to notify that the element's value has changed
+	UI_MSG_TABLE_GET_ITEM, // dp = pointer to UITableGetItem; return string length
+	UI_MSG_CODE_GET_MARGIN_COLOR, // di = line index (starts at 1); return color
+	UI_MSG_CODE_DECORATE_LINE, // dp = pointer to UICodeDecorateLine
+	UI_MSG_WINDOW_CLOSE, // return 1 to prevent default (process exit for UIWindow; close for UIMDIChild)
+	UI_MSG_TAB_SELECTED, // sent to the tab that was selected (not the tab pane itself)
+	UI_MSG_WINDOW_DROP_FILES, // di = count, dp = char ** of paths
+	UI_MSG_WINDOW_ACTIVATE,
+
+	UI_MSG_USER,
+} UIMessage;
+
+#ifdef UI_ESSENCE
+#define UIRectangle EsRectangle
+#else
+typedef struct UIRectangle {
+	int l, r, t, b;
+} UIRectangle;
+#endif
+
+typedef struct UITheme {
+	uint32_t panel1, panel2, selected, border;
+	uint32_t text, textDisabled, textSelected;
+	uint32_t buttonNormal, buttonHovered, buttonPressed, buttonDisabled;
+	uint32_t textboxNormal, textboxFocused;
+	uint32_t codeFocused, codeBackground, codeDefault, codeComment, codeString, codeNumber, codeOperator, codePreprocessor;
+} UITheme;
+
+typedef struct UIPainter {
+	UIRectangle clip;
+	uint32_t *bits;
+	int width, height;
+#ifdef UI_DEBUG
+	int fillCount;
+#endif
+} UIPainter;
+
+typedef struct UIFont {
+	int glyphWidth, glyphHeight;
+
+#ifdef UI_FREETYPE
+	bool isFreeType;
+	FT_Face font;
+	FT_Bitmap glyphs[128];
+	bool glyphsRendered[128];
+	int glyphOffsetsX[128], glyphOffsetsY[128];
+#endif
+} UIFont;
+
+typedef struct UIShortcut {
+	intptr_t code;
+	bool ctrl, shift, alt;
+	void (*invoke)(void *cp);
+	void *cp;
+} UIShortcut;
+
+typedef struct UIStringSelection {
+	int carets[2];
+	uint32_t colorText, colorBackground;
+} UIStringSelection;
+
+typedef struct UIKeyTyped {
+	char *text;
+	int textBytes;
+	intptr_t code;
+} UIKeyTyped;
+
+typedef struct UITableGetItem {
+	char *buffer;
+	size_t bufferBytes;
+	int index, column;
+	bool isSelected;
+} UITableGetItem;
+
+typedef struct UICodeDecorateLine {
+	UIRectangle bounds;
+	int index; // Starting at 1!
+	int x, y; // Position where additional text can be drawn.
+	UIPainter *painter;
+} UICodeDecorateLine;
+
+typedef struct UIFindByPoint {
+	int x, y;
+	struct UIElement *result;
+} UIFindByPoint;
+
+#define UI_RECT_1(x) ((UIRectangle) { (x), (x), (x), (x) })
+#define UI_RECT_1I(x) ((UIRectangle) { (x), -(x), (x), -(x) })
+#define UI_RECT_2(x, y) ((UIRectangle) { (x), (x), (y), (y) })
+#define UI_RECT_2I(x, y) ((UIRectangle) { (x), -(x), (y), -(y) })
+#define UI_RECT_2S(x, y) ((UIRectangle) { 0, (x), 0, (y) })
+#define UI_RECT_4(x, y, z, w) ((UIRectangle) { (x), (y), (z), (w) })
+#define UI_RECT_WIDTH(_r) ((_r).r - (_r).l)
+#define UI_RECT_HEIGHT(_r) ((_r).b - (_r).t)
+#define UI_RECT_TOTAL_H(_r) ((_r).r + (_r).l)
+#define UI_RECT_TOTAL_V(_r) ((_r).b + (_r).t)
+#define UI_RECT_SIZE(_r) UI_RECT_WIDTH(_r), UI_RECT_HEIGHT(_r)
+#define UI_RECT_TOP_LEFT(_r) (_r).l, (_r).t
+#define UI_RECT_BOTTOM_LEFT(_r) (_r).l, (_r).b
+#define UI_RECT_BOTTOM_RIGHT(_r) (_r).r, (_r).b
+#define UI_RECT_ALL(_r) (_r).l, (_r).r, (_r).t, (_r).b
+#define UI_RECT_VALID(_r) (UI_RECT_WIDTH(_r) > 0 && UI_RECT_HEIGHT(_r) > 0)
+
+#define UI_COLOR_ALPHA_F(x) ((((x) >> 24) & 0xFF) / 255.0f)
+#define UI_COLOR_RED_F(x) ((((x) >> 16) & 0xFF) / 255.0f)
+#define UI_COLOR_GREEN_F(x) ((((x) >> 8) & 0xFF) / 255.0f)
+#define UI_COLOR_BLUE_F(x) ((((x) >> 0) & 0xFF) / 255.0f)
+#define UI_COLOR_ALPHA(x) ((((x) >> 24) & 0xFF))
+#define UI_COLOR_RED(x) ((((x) >> 16) & 0xFF))
+#define UI_COLOR_GREEN(x) ((((x) >> 8) & 0xFF))
+#define UI_COLOR_BLUE(x) ((((x) >> 0) & 0xFF))
+#define UI_COLOR_FROM_FLOAT(r, g, b) (((uint32_t) ((r) * 255.0f) << 16) | ((uint32_t) ((g) * 255.0f) << 8) | ((uint32_t) ((b) * 255.0f) << 0))
+#define UI_COLOR_FROM_RGBA_F(r, g, b, a) (((uint32_t) ((r) * 255.0f) << 16) | ((uint32_t) ((g) * 255.0f) << 8) \
+		| ((uint32_t) ((b) * 255.0f) << 0) | ((uint32_t) ((a) * 255.0f) << 24))
+
+#define UI_SWAP(s, a, b) do { s t = (a); (a) = (b); (b) = t; } while (0)
+
+#define UI_CURSOR_ARROW (0)
+#define UI_CURSOR_TEXT (1)
+#define UI_CURSOR_SPLIT_V (2)
+#define UI_CURSOR_SPLIT_H (3)
+#define UI_CURSOR_FLIPPED_ARROW (4)
+#define UI_CURSOR_CROSS_HAIR (5)
+#define UI_CURSOR_HAND (6)
+#define UI_CURSOR_RESIZE_UP (7)
+#define UI_CURSOR_RESIZE_LEFT (8)
+#define UI_CURSOR_RESIZE_UP_RIGHT (9)
+#define UI_CURSOR_RESIZE_UP_LEFT (10)
+#define UI_CURSOR_RESIZE_DOWN (11)
+#define UI_CURSOR_RESIZE_RIGHT (12)
+#define UI_CURSOR_RESIZE_DOWN_RIGHT (13)
+#define UI_CURSOR_RESIZE_DOWN_LEFT (14)
+#define UI_CURSOR_COUNT (15)
+
+#define UI_ALIGN_LEFT (1)
+#define UI_ALIGN_RIGHT (2)
+#define UI_ALIGN_CENTER (3)
+
+extern const int UI_KEYCODE_A;
+extern const int UI_KEYCODE_BACKSPACE;
+extern const int UI_KEYCODE_DELETE;
+extern const int UI_KEYCODE_DOWN;
+extern const int UI_KEYCODE_END;
+extern const int UI_KEYCODE_ENTER;
+extern const int UI_KEYCODE_ESCAPE;
+extern const int UI_KEYCODE_F1;
+extern const int UI_KEYCODE_HOME;
+extern const int UI_KEYCODE_LEFT;
+extern const int UI_KEYCODE_RIGHT;
+extern const int UI_KEYCODE_SPACE;
+extern const int UI_KEYCODE_TAB;
+extern const int UI_KEYCODE_UP;
+extern const int UI_KEYCODE_INSERT;
+extern const int UI_KEYCODE_0;
+
+#define UI_KEYCODE_LETTER(x) (UI_KEYCODE_A + (x) - 'A')
+#define UI_KEYCODE_DIGIT(x) (UI_KEYCODE_0 + (x) - '0')
+#define UI_KEYCODE_FKEY(x) (UI_KEYCODE_F1 + (x) - 1)
+
+typedef struct UIElement {
+#define UI_ELEMENT_V_FILL (1 << 16)
+#define UI_ELEMENT_H_FILL (1 << 17)
+#define UI_ELEMENT_WINDOW (1 << 18)
+#define UI_ELEMENT_PARENT_PUSH (1 << 19)
+#define UI_ELEMENT_TAB_STOP (1 << 20)
+#define UI_ELEMENT_NON_CLIENT (1 << 21) // Don't destroy in UIElementDestroyDescendents, like scroll bars.
+#define UI_ELEMENT_DISABLED (1 << 22) // Don't receive input events.
+
+#define UI_ELEMENT_HIDE (1 << 29)
+#define UI_ELEMENT_DESTROY (1 << 30)
+#define UI_ELEMENT_DESTROY_DESCENDENT (1 << 31)
+
+	uint32_t flags; // First 16 bits are element specific.
+	uint32_t id;
+
+	struct UIElement *parent;
+	struct UIElement *next;
+	struct UIElement *children;
+	struct UIWindow *window;
+
+	UIRectangle bounds, clip;
+	
+	void *cp; // Context pointer (for user).
+
+	int (*messageClass)(struct UIElement *element, UIMessage message, int di /* data integer */, void *dp /* data pointer */);
+	int (*messageUser)(struct UIElement *element, UIMessage message, int di, void *dp);
+
+	const char *cClassName;
+} UIElement;
+
+#define UI_SHORTCUT(code, ctrl, shift, alt, invoke, cp) ((UIShortcut) { (code), (ctrl), (shift), (alt), (invoke), (cp) })
+
+typedef struct UIWindow {
+#define UI_WINDOW_MENU (1 << 0)
+#define UI_WINDOW_INSPECTOR (1 << 1)
+#define UI_WINDOW_CENTER_IN_OWNER (1 << 2)
+#define UI_WINDOW_MAXIMIZE (1 << 3)
+
+	UIElement e;
+
+	UIElement *dialog;
+
+	UIShortcut *shortcuts;
+	size_t shortcutCount, shortcutAllocated;
+
+	float scale;
+
+	uint32_t *bits;
+	int width, height;
+	struct UIWindow *next;
+
+	UIElement *hovered, *pressed, *focused, *dialogOldFocus;
+	int pressedButton;
+
+	int cursorX, cursorY;
+	int cursorStyle;
+
+	// Set when a textbox is modified. 
+	// Useful for tracking whether changes to the loaded document have been saved.
+	bool textboxModifiedFlag; 
+
+	bool ctrl, shift, alt;
+
+	UIRectangle updateRegion;
+
+#ifdef UI_DEBUG
+	float lastFullFillCount;
+#endif
+
+#ifdef UI_LINUX
+	Window window;
+	XImage *image;
+	XIC xic;
+	unsigned ctrlCode, shiftCode, altCode;
+	Window dragSource;
+#endif
+
+#ifdef UI_WINDOWS
+	HWND hwnd;
+	bool trackingLeave;
+#endif
+
+#ifdef UI_ESSENCE
+	EsWindow *window;
+	EsElement *canvas;
+	int cursor;
+#endif
+} UIWindow;
+
+typedef struct UIPanel {
+#define UI_PANEL_HORIZONTAL (1 << 0)
+#define UI_PANEL_GRAY (1 << 2)
+#define UI_PANEL_WHITE (1 << 3)
+#define UI_PANEL_EXPAND (1 << 4)
+#define UI_PANEL_MEDIUM_SPACING (1 << 5)
+#define UI_PANEL_SMALL_SPACING (1 << 6)
+#define UI_PANEL_SCROLL (1 << 7)
+#define UI_PANEL_BORDER (1 << 8)
+	UIElement e;
+	struct UIScrollBar *scrollBar;
+	UIRectangle border;
+	int gap;
+} UIPanel;
+
+typedef struct UIButton {
+#define UI_BUTTON_SMALL (1 << 0)
+#define UI_BUTTON_MENU_ITEM (1 << 1)
+#define UI_BUTTON_CAN_FOCUS (1 << 2)
+#define UI_BUTTON_DROP_DOWN (1 << 3)
+#define UI_BUTTON_CHECKED (1 << 15)
+	UIElement e;
+	char *label;
+	ptrdiff_t labelBytes;
+	void (*invoke)(void *cp);
+} UIButton;
+
+typedef struct UICheckbox {
+#define UI_CHECKBOX_ALLOW_INDETERMINATE (1 << 0)
+	UIElement e;
+#define UI_CHECK_UNCHECKED (0)
+#define UI_CHECK_CHECKED (1)
+#define UI_CHECK_INDETERMINATE (2)
+	uint8_t check;
+	char *label;
+	ptrdiff_t labelBytes;
+	void (*invoke)(void *cp);
+} UICheckbox;
+
+typedef struct UILabel {
+	UIElement e;
+	char *label;
+	ptrdiff_t labelBytes;
+} UILabel;
+
+typedef struct UISpacer {
+#define UI_SPACER_LINE (1 << 0)
+	UIElement e;
+	int width, height;
+} UISpacer;
+
+typedef struct UISplitPane {
+#define UI_SPLIT_PANE_VERTICAL (1 << 0)
+	UIElement e;
+	float weight;
+} UISplitPane;
+
+typedef struct UITabPane {
+	UIElement e;
+	char *tabs;
+	int active;
+} UITabPane;
+
+typedef struct UIScrollBar {
+#define UI_SCROLL_BAR_HORIZONTAL (1 << 0)
+	UIElement e;
+	int64_t maximum, page;
+	int64_t dragOffset;
+	double position;
+	uint64_t lastAnimateTime;
+	bool inDrag, horizontal;
+} UIScrollBar;
+
+typedef struct UICodeLine {
+	int offset, bytes;
+} UICodeLine;
+
+typedef struct UICode {
+#define UI_CODE_NO_MARGIN (1 << 0)
+	UIElement e;
+	UIScrollBar *vScroll;
+	UICodeLine *lines;
+	UIFont *font;
+	int lineCount, focused;
+	bool moveScrollToFocusNextLayout;
+	char *content;
+	size_t contentBytes;
+	int tabSize;
+} UICode;
+
+typedef struct UIGauge {
+	UIElement e;
+	float position;
+} UIGauge;
+
+typedef struct UITable {
+	UIElement e;
+	UIScrollBar *vScroll;
+	int itemCount;
+	char *columns;
+	int *columnWidths, columnCount, columnHighlight;
+} UITable;
+
+typedef struct UITextbox {
+	UIElement e;
+	char *string;
+	ptrdiff_t bytes;
+	int carets[2];
+	int scroll;
+	bool rejectNextKey;
+} UITextbox;
+
+#define UI_MENU_PLACE_ABOVE (1 << 0)
+#define UI_MENU_NO_SCROLL (1 << 1)
+#ifdef UI_ESSENCE
+typedef EsMenu UIMenu;
+#else
+typedef struct UIMenu {
+	UIElement e;
+	int pointX, pointY;
+	UIScrollBar *vScroll;
+} UIMenu;
+#endif
+
+typedef struct UISlider {
+	UIElement e;
+	float position;
+	int steps;
+} UISlider;
+
+typedef struct UIColorPicker {
+#define UI_COLOR_PICKER_HAS_OPACITY (1 << 0)
+	UIElement e;
+	float hue, saturation, value, opacity;
+} UIColorPicker;
+
+typedef struct UIMDIClient {
+#define UI_MDI_CLIENT_TRANSPARENT (1 << 0)
+	UIElement e;
+	struct UIMDIChild *active;
+	int cascade;
+} UIMDIClient;
+
+typedef struct UIMDIChild {
+#define UI_MDI_CHILD_CLOSE_BUTTON (1 << 0)
+	UIElement e;
+	UIRectangle bounds;
+	char *title;
+	ptrdiff_t titleBytes;
+	int dragHitTest;
+	UIRectangle dragOffset;
+	struct UIMDIChild *previous;
+} UIMDIChild;
+
+typedef struct UIExpandPane {
+	UIElement e;
+	UIButton *button;
+	UIPanel *panel;
+	bool expanded;
+} UIExpandPane;
+
+typedef struct UIImageDisplay {
+#define UI_IMAGE_DISPLAY_INTERACTIVE (1 << 0)
+#define _UI_IMAGE_DISPLAY_ZOOM_FIT (1 << 1)
+
+	UIElement e;
+	uint32_t *bits;
+	int width, height;
+	float panX, panY, zoom;
+	
+	// Internals:
+	int previousWidth, previousHeight;
+	int previousPanPointX, previousPanPointY;
+} UIImageDisplay;
+
+typedef struct UIWrapPanel {
+	UIElement e;
+} UIWrapPanel;
+
+void UIInitialise();
+int UIMessageLoop();
+
+UIElement *UIElementCreate(size_t bytes, UIElement *parent, uint32_t flags, 
+	int (*messageClass)(UIElement *, UIMessage, int, void *), const char *cClassName);
+
+UIButton *UIButtonCreate(UIElement *parent, uint32_t flags, const char *label, ptrdiff_t labelBytes);
+UICheckbox *UICheckboxCreate(UIElement *parent, uint32_t flags, const char *label, ptrdiff_t labelBytes);
+UIColorPicker *UIColorPickerCreate(UIElement *parent, uint32_t flags);
+UIExpandPane *UIExpandPaneCreate(UIElement *parent, uint32_t flags, const char *label, ptrdiff_t labelBytes, uint32_t panelFlags);
+UIGauge *UIGaugeCreate(UIElement *parent, uint32_t flags);
+UIMDIClient *UIMDIClientCreate(UIElement *parent, uint32_t flags);
+UIMDIChild *UIMDIChildCreate(UIElement *parent, uint32_t flags, UIRectangle initialBounds, const char *title, ptrdiff_t titleBytes);
+UIPanel *UIPanelCreate(UIElement *parent, uint32_t flags);
+UIScrollBar *UIScrollBarCreate(UIElement *parent, uint32_t flags);
+UISlider *UISliderCreate(UIElement *parent, uint32_t flags);
+UISpacer *UISpacerCreate(UIElement *parent, uint32_t flags, int width, int height);
+UISplitPane *UISplitPaneCreate(UIElement *parent, uint32_t flags, float weight);
+UITabPane *UITabPaneCreate(UIElement *parent, uint32_t flags, const char *tabs /* separate with \t, terminate with \0 */);
+UIWrapPanel *UIWrapPanelCreate(UIElement *parent, uint32_t flags);
+
+UILabel *UILabelCreate(UIElement *parent, uint32_t flags, const char *label, ptrdiff_t labelBytes);
+void UILabelSetContent(UILabel *code, const char *content, ptrdiff_t byteCount);
+
+UIImageDisplay *UIImageDisplayCreate(UIElement *parent, uint32_t flags, uint32_t *bits, size_t width, size_t height, size_t stride);
+void UIImageDisplaySetContent(UIImageDisplay *display, uint32_t *bits, size_t width, size_t height, size_t stride);
+
+UIWindow *UIWindowCreate(UIWindow *owner, uint32_t flags, const char *cTitle, int width, int height);
+void UIWindowRegisterShortcut(UIWindow *window, UIShortcut shortcut);
+void UIWindowPostMessage(UIWindow *window, UIMessage message, void *dp); // Thread-safe.
+void UIWindowPack(UIWindow *window, int width); // Change the size of the window to best match its contents.
+
+typedef void (*UIDialogUserCallback)(UIElement *);
+const char *UIDialogShow(UIWindow *window, uint32_t flags, const char *format, ...);
+
+UIMenu *UIMenuCreate(UIElement *parent, uint32_t flags);
+void UIMenuAddItem(UIMenu *menu, uint32_t flags, const char *label, ptrdiff_t labelBytes, void (*invoke)(void *cp), void *cp);
+void UIMenuShow(UIMenu *menu);
+
+UITextbox *UITextboxCreate(UIElement *parent, uint32_t flags);
+void UITextboxReplace(UITextbox *textbox, const char *text, ptrdiff_t bytes, bool sendChangedMessage);
+void UITextboxClear(UITextbox *textbox, bool sendChangedMessage);
+void UITextboxMoveCaret(UITextbox *textbox, bool backward, bool word);
+
+UITable *UITableCreate(UIElement *parent, uint32_t flags, const char *columns /* separate with \t, terminate with \0 */);
+int UITableHitTest(UITable *table, int x, int y); // Returns item index. Returns -1 if not on an item.
+int UITableHeaderHitTest(UITable *table, int x, int y); // Returns column index or -1.
+bool UITableEnsureVisible(UITable *table, int index); // Returns false if the item was already visible.
+void UITableResizeColumns(UITable *table);
+
+UICode *UICodeCreate(UIElement *parent, uint32_t flags);
+void UICodeFocusLine(UICode *code, int index); // Line numbers are 1-indexed!!
+int UICodeHitTest(UICode *code, int x, int y); // Returns line number; negates if in margin. Returns 0 if not on a line.
+void UICodeInsertContent(UICode *code, const char *content, ptrdiff_t byteCount, bool replace);
+
+void UIDrawBlock(UIPainter *painter, UIRectangle rectangle, uint32_t color);
+void UIDrawInvert(UIPainter *painter, UIRectangle rectangle);
+bool UIDrawLine(UIPainter *painter, int x0, int y0, int x1, int y1, uint32_t color); // Returns false if the line was not visible.
+void UIDrawTriangle(UIPainter *painter, int x0, int y0, int x1, int y1, int x2, int y2, uint32_t color);
+void UIDrawTriangleOutline(UIPainter *painter, int x0, int y0, int x1, int y1, int x2, int y2, uint32_t color);
+void UIDrawGlyph(UIPainter *painter, int x, int y, int c, uint32_t color);
+void UIDrawRectangle(UIPainter *painter, UIRectangle r, uint32_t mainColor, uint32_t borderColor, UIRectangle borderSize);
+void UIDrawBorder(UIPainter *painter, UIRectangle r, uint32_t borderColor, UIRectangle borderSize);
+void UIDrawString(UIPainter *painter, UIRectangle r, const char *string, ptrdiff_t bytes, uint32_t color, int align, UIStringSelection *selection);
+int UIDrawStringHighlighted(UIPainter *painter, UIRectangle r, const char *string, ptrdiff_t bytes, int tabSize);
+
+int UIMeasureStringWidth(const char *string, ptrdiff_t bytes);
+int UIMeasureStringHeight();
+
+uint64_t UIAnimateClock(); // In ms.
+
+bool UIElementAnimate(UIElement *element, bool stop);
+void UIElementDestroy(UIElement *element);
+void UIElementDestroyDescendents(UIElement *element);
+UIElement *UIElementFindByPoint(UIElement *element, int x, int y);
+void UIElementFocus(UIElement *element);
+UIRectangle UIElementScreenBounds(UIElement *element); // Returns bounds of element in same coordinate system as used by UIWindowCreate.
+void UIElementRefresh(UIElement *element);
+void UIElementRepaint(UIElement *element, UIRectangle *region);
+void UIElementMove(UIElement *element, UIRectangle bounds, bool alwaysLayout);
+int UIElementMessage(UIElement *element, UIMessage message, int di, void *dp);
+void UIElementChangeParent(UIElement *element, UIElement *newParent, UIElement *insertBefore); // Set insertBefore to null to insert at the end.
+
+UIElement *UIParentPush(UIElement *element);
+UIElement *UIParentPop();
+
+UIRectangle UIRectangleIntersection(UIRectangle a, UIRectangle b);
+UIRectangle UIRectangleBounding(UIRectangle a, UIRectangle b);
+UIRectangle UIRectangleAdd(UIRectangle a, UIRectangle b);
+UIRectangle UIRectangleTranslate(UIRectangle a, UIRectangle b);
+bool UIRectangleEquals(UIRectangle a, UIRectangle b);
+bool UIRectangleContains(UIRectangle a, int x, int y);
+
+bool UIColorToHSV(uint32_t rgb, float *hue, float *saturation, float *value);
+void UIColorToRGB(float hue, float saturation, float value, uint32_t *rgb);
+
+char *UIStringCopy(const char *in, ptrdiff_t inBytes);
+
+UIFont *UIFontCreate(const char *cPath, uint32_t size);
+UIFont *UIFontActivate(UIFont *font); // Returns the previously active font.
+
+#ifdef UI_DEBUG
+void UIInspectorLog(const char *cFormat, ...);
+#endif
+
+#ifdef UI_IMPLEMENTATION
+
+struct {
+	UIWindow *windows;
+	UIElement *animating;
+	UITheme theme;
+
+	UIElement *parentStack[16];
+	int parentStackCount;
+
+	bool quit;
+	const char *dialogResult;
+	UIElement *dialogOldFocus;
+
+	UIFont *activeFont;
+
+#ifdef UI_DEBUG
+	UIWindow *inspector;
+	UITable *inspectorTable;
+	UIWindow *inspectorTarget;
+	UICode *inspectorLog;
+#endif
+
+#ifdef UI_LINUX
+	Display *display;
+	Visual *visual;
+	XIM xim;
+	Atom windowClosedID, primaryID, uriListID, plainTextID;
+	Atom dndEnterID, dndPositionID, dndStatusID, dndActionCopyID, dndDropID, dndSelectionID, dndFinishedID, dndAwareID;
+	Atom clipboardID, xSelectionDataID, textID, targetID, incrID;
+	Cursor cursors[UI_CURSOR_COUNT];
+	char *pasteText;
+	XEvent copyEvent;
+#endif
+
+#ifdef UI_WINDOWS
+	HCURSOR cursors[UI_CURSOR_COUNT];
+	HANDLE heap;
+	bool assertionFailure;
+#endif
+
+#ifdef UI_ESSENCE
+	EsInstance *instance;
+
+	void *menuData[256]; // HACK This limits the number of menu items to 128.
+	uintptr_t menuIndex;
+#endif
+
+#ifdef UI_FREETYPE
+	FT_Library ft;
+#endif
+} ui;
+
+UITheme _uiThemeClassic = {
+	.panel1 = 0xFFF0F0F0,
+	.panel2 = 0xFFFFFFFF,
+	.selected = 0xFF94BEFE,
+	.border = 0xFF404040,
+
+	.text = 0xFF000000,
+	.textDisabled = 0xFF404040,
+	.textSelected = 0xFF000000,
+
+	.buttonNormal = 0xFFE0E0E0,
+	.buttonHovered = 0xFFF0F0F0,
+	.buttonPressed = 0xFFA0A0A0,
+	.buttonDisabled = 0xFFF0F0F0,
+
+	.textboxNormal = 0xFFF8F8F8,
+	.textboxFocused = 0xFFFFFFFF,
+
+	.codeFocused = 0xFFE0E0E0,
+	.codeBackground = 0xFFFFFFFF,
+	.codeDefault = 0xFF000000,
+	.codeComment = 0xFFA11F20,
+	.codeString = 0xFF037E01,
+	.codeNumber = 0xFF213EF1,
+	.codeOperator = 0xFF7F0480,
+	.codePreprocessor = 0xFF545D70,
+};
+
+UITheme _uiThemeDark = {
+	.panel1 = 0xFF252B31,
+	.panel2 = 0xFF14181E,
+	.selected = 0xFF94BEFE,
+	.border = 0xFF000000,
+
+	.text = 0xFFFFFFFF,
+	.textDisabled = 0xFF787D81,
+	.textSelected = 0xFF000000,
+
+	.buttonNormal = 0xFF383D41,
+	.buttonHovered = 0xFF4B5874,
+	.buttonPressed = 0xFF0D0D0F,
+	.buttonDisabled = 0xFF1B1F23,
+
+	.textboxNormal = 0xFF31353C,
+	.textboxFocused = 0xFF4D4D59,
+
+	.codeFocused = 0xFF505055,
+	.codeBackground = 0xFF212126,
+	.codeDefault = 0xFFFFFFFF,
+	.codeComment = 0xFFB4B4B4,
+	.codeString = 0xFFF5DDD1,
+	.codeNumber = 0xFFC3F5D3,
+	.codeOperator = 0xFFF5D499,
+	.codePreprocessor = 0xFFF5F3D1,
+};
+
+// Taken from https://commons.wikimedia.org/wiki/File:Codepage-437.png
+// Public domain.
+
+const uint64_t _uiFont[] = {
+	0x0000000000000000UL, 0x0000000000000000UL, 0xBD8181A5817E0000UL, 0x000000007E818199UL, 0xC3FFFFDBFF7E0000UL, 0x000000007EFFFFE7UL, 0x7F7F7F3600000000UL, 0x00000000081C3E7FUL, 
+	0x7F3E1C0800000000UL, 0x0000000000081C3EUL, 0xE7E73C3C18000000UL, 0x000000003C1818E7UL, 0xFFFF7E3C18000000UL, 0x000000003C18187EUL, 0x3C18000000000000UL, 0x000000000000183CUL, 
+	0xC3E7FFFFFFFFFFFFUL, 0xFFFFFFFFFFFFE7C3UL, 0x42663C0000000000UL, 0x00000000003C6642UL, 0xBD99C3FFFFFFFFFFUL, 0xFFFFFFFFFFC399BDUL, 0x331E4C5870780000UL, 0x000000001E333333UL, 
+	0x3C666666663C0000UL, 0x0000000018187E18UL, 0x0C0C0CFCCCFC0000UL, 0x00000000070F0E0CUL, 0xC6C6C6FEC6FE0000UL, 0x0000000367E7E6C6UL, 0xE73CDB1818000000UL, 0x000000001818DB3CUL, 
+	0x1F7F1F0F07030100UL, 0x000000000103070FUL, 0x7C7F7C7870604000UL, 0x0000000040607078UL, 0x1818187E3C180000UL, 0x0000000000183C7EUL, 0x6666666666660000UL, 0x0000000066660066UL, 
+	0xD8DEDBDBDBFE0000UL, 0x00000000D8D8D8D8UL, 0x6363361C06633E00UL, 0x0000003E63301C36UL, 0x0000000000000000UL, 0x000000007F7F7F7FUL, 0x1818187E3C180000UL, 0x000000007E183C7EUL, 
+	0x1818187E3C180000UL, 0x0000000018181818UL, 0x1818181818180000UL, 0x00000000183C7E18UL, 0x7F30180000000000UL, 0x0000000000001830UL, 0x7F060C0000000000UL, 0x0000000000000C06UL, 
+	0x0303000000000000UL, 0x0000000000007F03UL, 0xFF66240000000000UL, 0x0000000000002466UL, 0x3E1C1C0800000000UL, 0x00000000007F7F3EUL, 0x3E3E7F7F00000000UL, 0x0000000000081C1CUL, 
+	0x0000000000000000UL, 0x0000000000000000UL, 0x18183C3C3C180000UL, 0x0000000018180018UL, 0x0000002466666600UL, 0x0000000000000000UL, 0x36367F3636000000UL, 0x0000000036367F36UL, 
+	0x603E0343633E1818UL, 0x000018183E636160UL, 0x1830634300000000UL, 0x000000006163060CUL, 0x3B6E1C36361C0000UL, 0x000000006E333333UL, 0x000000060C0C0C00UL, 0x0000000000000000UL, 
+	0x0C0C0C0C18300000UL, 0x0000000030180C0CUL, 0x30303030180C0000UL, 0x000000000C183030UL, 0xFF3C660000000000UL, 0x000000000000663CUL, 0x7E18180000000000UL, 0x0000000000001818UL, 
+	0x0000000000000000UL, 0x0000000C18181800UL, 0x7F00000000000000UL, 0x0000000000000000UL, 0x0000000000000000UL, 0x0000000018180000UL, 0x1830604000000000UL, 0x000000000103060CUL, 
+	0xDBDBC3C3663C0000UL, 0x000000003C66C3C3UL, 0x1818181E1C180000UL, 0x000000007E181818UL, 0x0C183060633E0000UL, 0x000000007F630306UL, 0x603C6060633E0000UL, 0x000000003E636060UL, 
+	0x7F33363C38300000UL, 0x0000000078303030UL, 0x603F0303037F0000UL, 0x000000003E636060UL, 0x633F0303061C0000UL, 0x000000003E636363UL, 0x18306060637F0000UL, 0x000000000C0C0C0CUL, 
+	0x633E6363633E0000UL, 0x000000003E636363UL, 0x607E6363633E0000UL, 0x000000001E306060UL, 0x0000181800000000UL, 0x0000000000181800UL, 0x0000181800000000UL, 0x000000000C181800UL, 
+	0x060C183060000000UL, 0x000000006030180CUL, 0x00007E0000000000UL, 0x000000000000007EUL, 0x6030180C06000000UL, 0x00000000060C1830UL, 0x18183063633E0000UL, 0x0000000018180018UL, 
+	0x7B7B63633E000000UL, 0x000000003E033B7BUL, 0x7F6363361C080000UL, 0x0000000063636363UL, 0x663E6666663F0000UL, 0x000000003F666666UL, 0x03030343663C0000UL, 0x000000003C664303UL, 
+	0x66666666361F0000UL, 0x000000001F366666UL, 0x161E1646667F0000UL, 0x000000007F664606UL, 0x161E1646667F0000UL, 0x000000000F060606UL, 0x7B030343663C0000UL, 0x000000005C666363UL, 
+	0x637F636363630000UL, 0x0000000063636363UL, 0x18181818183C0000UL, 0x000000003C181818UL, 0x3030303030780000UL, 0x000000001E333333UL, 0x1E1E366666670000UL, 0x0000000067666636UL, 
+	0x06060606060F0000UL, 0x000000007F664606UL, 0xC3DBFFFFE7C30000UL, 0x00000000C3C3C3C3UL, 0x737B7F6F67630000UL, 0x0000000063636363UL, 0x63636363633E0000UL, 0x000000003E636363UL, 
+	0x063E6666663F0000UL, 0x000000000F060606UL, 0x63636363633E0000UL, 0x000070303E7B6B63UL, 0x363E6666663F0000UL, 0x0000000067666666UL, 0x301C0663633E0000UL, 0x000000003E636360UL, 
+	0x18181899DBFF0000UL, 0x000000003C181818UL, 0x6363636363630000UL, 0x000000003E636363UL, 0xC3C3C3C3C3C30000UL, 0x00000000183C66C3UL, 0xDBC3C3C3C3C30000UL, 0x000000006666FFDBUL, 
+	0x18183C66C3C30000UL, 0x00000000C3C3663CUL, 0x183C66C3C3C30000UL, 0x000000003C181818UL, 0x0C183061C3FF0000UL, 0x00000000FFC38306UL, 0x0C0C0C0C0C3C0000UL, 0x000000003C0C0C0CUL, 
+	0x1C0E070301000000UL, 0x0000000040607038UL, 0x30303030303C0000UL, 0x000000003C303030UL, 0x0000000063361C08UL, 0x0000000000000000UL, 0x0000000000000000UL, 0x0000FF0000000000UL, 
+	0x0000000000180C0CUL, 0x0000000000000000UL, 0x3E301E0000000000UL, 0x000000006E333333UL, 0x66361E0606070000UL, 0x000000003E666666UL, 0x03633E0000000000UL, 0x000000003E630303UL, 
+	0x33363C3030380000UL, 0x000000006E333333UL, 0x7F633E0000000000UL, 0x000000003E630303UL, 0x060F0626361C0000UL, 0x000000000F060606UL, 0x33336E0000000000UL, 0x001E33303E333333UL, 
+	0x666E360606070000UL, 0x0000000067666666UL, 0x18181C0018180000UL, 0x000000003C181818UL, 0x6060700060600000UL, 0x003C666660606060UL, 0x1E36660606070000UL, 0x000000006766361EUL, 
+	0x18181818181C0000UL, 0x000000003C181818UL, 0xDBFF670000000000UL, 0x00000000DBDBDBDBUL, 0x66663B0000000000UL, 0x0000000066666666UL, 0x63633E0000000000UL, 0x000000003E636363UL, 
+	0x66663B0000000000UL, 0x000F06063E666666UL, 0x33336E0000000000UL, 0x007830303E333333UL, 0x666E3B0000000000UL, 0x000000000F060606UL, 0x06633E0000000000UL, 0x000000003E63301CUL, 
+	0x0C0C3F0C0C080000UL, 0x00000000386C0C0CUL, 0x3333330000000000UL, 0x000000006E333333UL, 0xC3C3C30000000000UL, 0x00000000183C66C3UL, 0xC3C3C30000000000UL, 0x0000000066FFDBDBUL, 
+	0x3C66C30000000000UL, 0x00000000C3663C18UL, 0x6363630000000000UL, 0x001F30607E636363UL, 0x18337F0000000000UL, 0x000000007F63060CUL, 0x180E181818700000UL, 0x0000000070181818UL, 
+	0x1800181818180000UL, 0x0000000018181818UL, 0x18701818180E0000UL, 0x000000000E181818UL, 0x000000003B6E0000UL, 0x0000000000000000UL, 0x63361C0800000000UL, 0x00000000007F6363UL, 
+};
+
+void _UIWindowEndPaint(UIWindow *window, UIPainter *painter);
+void _UIWindowSetCursor(UIWindow *window, int cursor);
+void _UIWindowGetScreenPosition(UIWindow *window, int *x, int *y);
+void _UIWindowSetPressed(UIWindow *window, UIElement *element, int button);
+void _UIClipboardWriteText(UIWindow *window, char *text);
+char *_UIClipboardReadTextStart(UIWindow *window, size_t *bytes);
+void _UIClipboardReadTextEnd(UIWindow *window, char *text);
+bool _UIMessageLoopSingle(int *result);
+void _UIInspectorRefresh();
+void _UIUpdate();
+
+#ifdef UI_WINDOWS
+void *_UIHeapReAlloc(void *pointer, size_t size);
+#endif
+
+UIRectangle UIRectangleIntersection(UIRectangle a, UIRectangle b) {
+	if (a.l < b.l) a.l = b.l;
+	if (a.t < b.t) a.t = b.t;
+	if (a.r > b.r) a.r = b.r;
+	if (a.b > b.b) a.b = b.b;
+	return a;
+}
+
+UIRectangle UIRectangleBounding(UIRectangle a, UIRectangle b) {
+	if (a.l > b.l) a.l = b.l;
+	if (a.t > b.t) a.t = b.t;
+	if (a.r < b.r) a.r = b.r;
+	if (a.b < b.b) a.b = b.b;
+	return a;
+}
+
+UIRectangle UIRectangleAdd(UIRectangle a, UIRectangle b) {
+	a.l += b.l;
+	a.t += b.t;
+	a.r += b.r;
+	a.b += b.b;
+	return a;
+}
+
+UIRectangle UIRectangleTranslate(UIRectangle a, UIRectangle b) {
+	a.l += b.l;
+	a.t += b.t;
+	a.r += b.l;
+	a.b += b.t;
+	return a;
+}
+
+bool UIRectangleEquals(UIRectangle a, UIRectangle b) {
+	return a.l == b.l && a.r == b.r && a.t == b.t && a.b == b.b;
+}
+
+bool UIRectangleContains(UIRectangle a, int x, int y) {
+	return a.l <= x && a.r > x && a.t <= y && a.b > y;
+}
+
+#include <xmmintrin.h>
+
+typedef union _UIConvertFloatInteger {
+	float f;
+	uint32_t i;
+} _UIConvertFloatInteger;
+
+float _UIFloorFloat(float x) {
+	_UIConvertFloatInteger convert = {x};
+	uint32_t sign = convert.i & 0x80000000;
+	int exponent = (int) ((convert.i >> 23) & 0xFF) - 0x7F;
+
+	if (exponent >= 23) {
+		// There aren't any bits representing a fractional part.
+	} else if (exponent >= 0) {
+		// Positive exponent.
+		uint32_t mask = 0x7FFFFF >> exponent;
+		if (!(mask & convert.i)) return x; // Already an integer.
+		if (sign) convert.i += mask;
+		convert.i &= ~mask; // Mask out the fractional bits.
+	} else if (exponent < 0) {
+		// Negative exponent.
+		return sign ? -1.0 : 0.0;
+	}
+
+	return convert.f;
+}
+
+float _UISquareRootFloat(float x) {
+	float result[4];
+	_mm_storeu_ps(result, _mm_sqrt_ps(_mm_set_ps(0, 0, 0, x)));
+	return result[0];
+}
+
+#define _F(x) (((_UIConvertFloatInteger) { .i = (x) }).f)
+
+float _UIArcTanFloatI(float x) {
+	float x2 = x * x;
+	return x * (_F(0x3F7FFFF8) + x2 * (_F(0xBEAAA53C) + x2 * (_F(0x3E4BC990) + x2 * (_F(0xBE084A60) + x2 * _F(0x3D8864B0)))));
+}
+
+float _UISinFloatI(float x) {
+	float x2 = x * x;
+	return x * (_F(0x3F800000) + x2 * (_F(0xBE2AAAA0) + x2 * (_F(0x3C0882C0) + x2 * _F(0xB94C6000))));
+}
+
+float _UICosFloatI(float x) {
+	float x2 = x * x;
+	return _F(0x3F800000) + x2 * (_F(0xBEFFFFDA) + x2 * (_F(0x3D2A9F60) + x2 * _F(0xBAB22C00)));
+}
+
+#undef _F
+
+float _UISinFloat(float x) {
+	bool negate = false;
+	if (x < 0) { x = -x; negate = true; }
+	x -= 2 * 3.141592654f * _UIFloorFloat(x / (2 * 3.141592654f));
+	if (x < 3.141592654f / 2) {}
+	else if (x < 3.141592654f) { x = 3.141592654f - x; }
+	else if (x < 3 * 3.141592654f / 2) { x = x - 3.141592654f; negate = !negate; }
+	else { x = 3.141592654f * 2 - x; negate = !negate; }
+	float y = x < 3.141592654f / 4 ? _UISinFloatI(x) : _UICosFloatI(3.141592654f / 2 - x);
+	return negate ? -y : y;
+}
+
+float _UICosFloat(float x) {
+	return _UISinFloat(3.141592654f / 2 - x);
+}
+
+float _UIArcTanFloat(float x) {
+	bool negate = false, reciprocalTaken = false;
+	if (x < 0) { x = -x; negate = true; }
+	if (x > 1) { x = 1 / x; reciprocalTaken = true; }
+	float y = x < 0.5f ? _UIArcTanFloatI(x) : (0.463647609f + _UIArcTanFloatI((2 * x - 1) / (2 + x)));
+	if (reciprocalTaken) { y = 3.141592654f / 2 - y; }
+	return negate ? -y : y;
+}
+
+float _UIArcTan2Float(float y, float x) {
+	if (x == 0) return y > 0 ? 3.141592654f / 2 : -3.141592654f / 2;
+	else if (x > 0) return _UIArcTanFloat(y / x);
+	else if (y >= 0) return 3.141592654f + _UIArcTanFloat(y / x);
+	else return -3.141592654f + _UIArcTanFloat(y / x);
+}
+
+float _UILinearMap(float value, float inFrom, float inTo, float outFrom, float outTo) {
+	float inRange = inTo - inFrom, outRange = outTo - outFrom;
+	float normalisedValue = (value - inFrom) / inRange;
+	return normalisedValue * outRange + outFrom;
+}
+
+bool UIColorToHSV(uint32_t rgb, float *hue, float *saturation, float *value) {
+	float r = UI_COLOR_RED_F(rgb);
+	float g = UI_COLOR_GREEN_F(rgb);
+	float b = UI_COLOR_BLUE_F(rgb);
+
+	float maximum = (r > g && r > b) ? r : (g > b ? g : b),
+	      minimum = (r < g && r < b) ? r : (g < b ? g : b),
+	      difference = maximum - minimum;
+	*value = maximum;
+
+	if (!difference) {
+		*saturation = 0;
+		return false;
+	} else {
+		if (r == maximum) *hue = (g - b) / difference + 0;
+		if (g == maximum) *hue = (b - r) / difference + 2;
+		if (b == maximum) *hue = (r - g) / difference + 4;
+		if (*hue < 0) *hue += 6;
+		*saturation = difference / maximum;
+		return true;
+	}
+}
+
+void UIColorToRGB(float h, float s, float v, uint32_t *rgb) {
+	float r, g, b;
+
+	if (!s) {
+		r = g = b = v;
+	} else {
+		int h0 = ((int) h) % 6;
+		float f = h - _UIFloorFloat(h);
+		float x = v * (1 - s), y = v * (1 - s * f), z = v * (1 - s * (1 - f));
+
+		switch (h0) {
+			case 0:  r = v, g = z, b = x; break;
+			case 1:  r = y, g = v, b = x; break;
+			case 2:  r = x, g = v, b = z; break;
+			case 3:  r = x, g = y, b = v; break;
+			case 4:  r = z, g = x, b = v; break;
+			default: r = v, g = x, b = y; break;
+		}
+	}
+
+	*rgb = UI_COLOR_FROM_FLOAT(r, g, b);
+}
+
+void UIElementRefresh(UIElement *element) {
+	UIElementMessage(element, UI_MSG_LAYOUT, 0, 0);
+	UIElementRepaint(element, NULL);
+}
+
+void UIElementRepaint(UIElement *element, UIRectangle *region) {
+	if (!region) {
+		region = &element->bounds;
+	}
+
+	UIRectangle r = UIRectangleIntersection(*region, element->clip);
+
+	if (!UI_RECT_VALID(r)) {
+		return;
+	}
+
+	if (UI_RECT_VALID(element->window->updateRegion)) {
+		element->window->updateRegion = UIRectangleBounding(element->window->updateRegion, r);
+	} else {
+		element->window->updateRegion = r;
+	}
+}
+
+bool UIElementAnimate(UIElement *element, bool stop) {
+	if (stop) {
+		if (ui.animating != element) {
+			return false;
+		}
+
+		ui.animating = NULL;
+	} else {
+		if (ui.animating && ui.animating != element) {
+			return false;
+		}
+
+		ui.animating = element;
+	}
+
+	return true;
+}
+
+uint64_t UIAnimateClock() {
+	return (uint64_t) UI_CLOCK() * 1000 / UI_CLOCKS_PER_SECOND;
+}
+
+void _UIElementDestroyDescendents(UIElement *element, bool topLevel) {
+	UIElement *child = element->children;
+
+	while (child) {
+		if (!topLevel || (~child->flags & UI_ELEMENT_NON_CLIENT)) {
+			UIElementDestroy(child);
+		}
+
+		child = child->next;
+	}
+
+#ifdef UI_DEBUG
+	_UIInspectorRefresh();
+#endif
+}
+
+void UIElementDestroyDescendents(UIElement *element) {
+	_UIElementDestroyDescendents(element, true);
+}
+
+void UIElementDestroy(UIElement *element) {
+	if (element->flags & UI_ELEMENT_DESTROY) {
+		return;
+	}
+
+	element->flags |= UI_ELEMENT_DESTROY | UI_ELEMENT_HIDE;
+
+	UIElement *ancestor = element->parent;
+
+	while (ancestor) {
+		ancestor->flags |= UI_ELEMENT_DESTROY_DESCENDENT;
+		ancestor = ancestor->parent;
+	}
+
+	_UIElementDestroyDescendents(element, false);
+}
+
+void UIDrawBlock(UIPainter *painter, UIRectangle rectangle, uint32_t color) {
+	rectangle = UIRectangleIntersection(painter->clip, rectangle);
+
+	if (!UI_RECT_VALID(rectangle)) {
+		return;
+	}
+
+#ifdef UI_SSE2
+	__m128i color4 = _mm_set_epi32(color, color, color, color);
+#endif
+
+	for (int line = rectangle.t; line < rectangle.b; line++) {
+		uint32_t *bits = painter->bits + line * painter->width + rectangle.l;
+		int count = UI_RECT_WIDTH(rectangle);
+
+#ifdef UI_SSE2
+		while (count >= 4) {
+			_mm_storeu_si128((__m128i *) bits, color4);
+			bits += 4;
+			count -= 4;
+		} 
+#endif
+
+		while (count--) {
+			*bits++ = color;
+		}
+	}
+
+#ifdef UI_DEBUG
+	painter->fillCount += UI_RECT_WIDTH(rectangle) * UI_RECT_HEIGHT(rectangle);
+#endif
+}
+
+bool UIDrawLine(UIPainter *painter, int x0, int y0, int x1, int y1, uint32_t color) {
+	// Apply the clip.
+
+	UIRectangle c = painter->clip;
+	if (!UI_RECT_VALID(c)) return false;
+	int dx = x1 - x0, dy = y1 - y0;
+	const int p[4] = { -dx, dx, -dy, dy };
+	const int q[4] = { x0 - c.l, c.r - 1 - x0, y0 - c.t, c.b - 1 - y0 };
+	float t0 = 0.0f, t1 = 1.0f; // How far along the line the points end up.
+
+	for (int i = 0; i < 4; i++) {
+		if (!p[i] && q[i] < 0) return false;
+		float r = (float) q[i] / p[i];
+		if (p[i] < 0 && r > t1) return false;
+		if (p[i] > 0 && r < t0) return false;
+		if (p[i] < 0 && r > t0) t0 = r;
+		if (p[i] > 0 && r < t1) t1 = r;
+	}
+
+	x1 = x0 + t1 * dx, y1 = y0 + t1 * dy;
+	x0 += t0 * dx, y0 += t0 * dy;
+
+	// Calculate the delta X and delta Y.
+
+	if (y1 < y0) {
+		int t;
+		t = x0, x0 = x1, x1 = t;
+		t = y0, y0 = y1, y1 = t;
+	}
+
+	dx = x1 - x0, dy = y1 - y0;
+	int dxs = dx < 0 ? -1 : 1;
+	if (dx < 0) dx = -dx;
+
+	// Draw the line using Bresenham's line algorithm.
+
+	uint32_t *bits = painter->bits + y0 * painter->width + x0;
+
+	if (dy * dy < dx * dx) {
+		int m = 2 * dy - dx;
+
+		for (int i = 0; i < dx; i++, bits += dxs) {
+			*bits = color;
+			if (m > 0) bits += painter->width, m -= 2 * dx;
+			m += 2 * dy;
+		}
+	} else {
+		int m = 2 * dx - dy;
+
+		for (int i = 0; i < dy; i++, bits += painter->width) {
+			*bits = color;
+			if (m > 0) bits += dxs, m -= 2 * dy;
+			m += 2 * dx;
+		}
+	}
+
+	return true;
+}
+
+void UIDrawTriangle(UIPainter *painter, int x0, int y0, int x1, int y1, int x2, int y2, uint32_t color) {
+	// Step 1: Sort the points by their y-coordinate.
+	if (y1 < y0) { int xt = x0; x0 = x1, x1 = xt; int yt = y0; y0 = y1, y1 = yt; }
+	if (y2 < y1) { int xt = x1; x1 = x2, x2 = xt; int yt = y1; y1 = y2, y2 = yt; }
+	if (y1 < y0) { int xt = x0; x0 = x1, x1 = xt; int yt = y0; y0 = y1, y1 = yt; }
+	if (y2 == y0) return;
+
+	// Step 2: Clip the triangle.
+	if (x0 < painter->clip.l && x1 < painter->clip.l && x2 < painter->clip.l) return;
+	if (x0 >= painter->clip.r && x1 >= painter->clip.r && x2 >= painter->clip.r) return;
+	if (y2 < painter->clip.t || y0 >= painter->clip.b) return;
+	bool needsXClip = x0 < painter->clip.l + 1 || x0 >= painter->clip.r - 1
+		|| x1 < painter->clip.l + 1 || x1 >= painter->clip.r - 1
+		|| x2 < painter->clip.l + 1 || x2 >= painter->clip.r - 1;
+	bool needsYClip = y0 < painter->clip.t + 1 || y2 >= painter->clip.b - 1;
+#define _UI_DRAW_TRIANGLE_APPLY_CLIP(xo, yo) \
+	if (needsYClip && (yi + yo < painter->clip.t || yi + yo >= painter->clip.b)) continue; \
+	if (needsXClip && xf + xo < painter->clip.l) xf = painter->clip.l - xo; \
+	if (needsXClip && xt + xo > painter->clip.r) xt = painter->clip.r - xo;
+
+	// Step 3: Split into 2 triangles with bases aligned with the x-axis.
+	float xm0 = (x2 - x0) * (y1 - y0) / (y2 - y0), xm1 = x1 - x0;
+	if (xm1 < xm0) { float xmt = xm0; xm0 = xm1, xm1 = xmt; }
+	float xe0 = xm0 + x0 - x2, xe1 = xm1 + x0 - x2;
+	int ym = y1 - y0, ye = y2 - y1;
+	float ymr = 1.0f / ym, yer = 1.0f / ye;
+
+	// Step 4: Draw the top part.
+	for (float y = 0; y < ym; y++) {
+		int xf = xm0 * y * ymr, xt = xm1 * y * ymr, yi = (int) y;
+		_UI_DRAW_TRIANGLE_APPLY_CLIP(x0, y0);
+		uint32_t *b = &painter->bits[(yi + y0) * painter->width + x0];
+		for (int x = xf; x < xt; x++) b[x] = color;
+	}
+
+	// Step 5: Draw the bottom part.
+	for (float y = 0; y < ye; y++) {
+		int xf = xe0 * (ye - y) * yer, xt = xe1 * (ye - y) * yer, yi = (int) y;
+		_UI_DRAW_TRIANGLE_APPLY_CLIP(x2, y1);
+		uint32_t *b = &painter->bits[(yi + y1) * painter->width + x2];
+		for (int x = xf; x < xt; x++) b[x] = color;
+	}
+}
+
+void UIDrawTriangleOutline(UIPainter *painter, int x0, int y0, int x1, int y1, int x2, int y2, uint32_t color) {
+	UIDrawLine(painter, x0, y0, x1, y1, color);
+	UIDrawLine(painter, x1, y1, x2, y2, color);
+	UIDrawLine(painter, x2, y2, x0, y0, color);
+}
+
+void UIDrawInvert(UIPainter *painter, UIRectangle rectangle) {
+	rectangle = UIRectangleIntersection(painter->clip, rectangle);
+
+	if (!UI_RECT_VALID(rectangle)) {
+		return;
+	}
+
+	for (int line = rectangle.t; line < rectangle.b; line++) {
+		uint32_t *bits = painter->bits + line * painter->width + rectangle.l;
+		int count = UI_RECT_WIDTH(rectangle);
+
+		while (count--) {
+			uint32_t in = *bits;
+			*bits = in ^ 0xFFFFFF;
+			bits++;
+		}
+	}
+}
+
+void UIDrawGlyph(UIPainter *painter, int x0, int y0, int c, uint32_t color) {
+#ifdef UI_FREETYPE
+	UIFont *font = ui.activeFont;
+
+	if (font->isFreeType) {
+		if (c < 0 || c > 127) c = '?';
+		if (c == '\r') c = ' ';
+
+		if (!font->glyphsRendered[c]) {
+			FT_Load_Char(font->font, c == 24 ? 0x2191 : c == 25 ? 0x2193 : c == 26 ? 0x2192 : c == 27 ? 0x2190 : c, FT_LOAD_DEFAULT);
+#ifdef UI_FREETYPE_SUBPIXEL
+			FT_Render_Glyph(font->font->glyph, FT_RENDER_MODE_LCD);
+#else
+			FT_Render_Glyph(font->font->glyph, FT_RENDER_MODE_NORMAL);
+#endif
+			FT_Bitmap_Copy(ui.ft, &font->font->glyph->bitmap, &font->glyphs[c]);
+			font->glyphOffsetsX[c] = font->font->glyph->bitmap_left;
+			font->glyphOffsetsY[c] = font->font->size->metrics.ascender / 64 - font->font->glyph->bitmap_top;
+			font->glyphsRendered[c] = true;
+		}
+
+		FT_Bitmap *bitmap = &font->glyphs[c];
+		x0 += font->glyphOffsetsX[c], y0 += font->glyphOffsetsY[c];
+
+		for (int y = 0; y < (int) bitmap->rows; y++) {
+			if (y0 + y < painter->clip.t) continue;
+			if (y0 + y >= painter->clip.b) break;
+
+			int width = bitmap->width;
+#ifdef UI_FREETYPE_SUBPIXEL
+			width /= 3;
+#endif
+
+			for (int x = 0; x < width; x++) {
+				if (x0 + x < painter->clip.l) continue;
+				if (x0 + x >= painter->clip.r) break;
+
+				uint32_t *destination = painter->bits + (x0 + x) + (y0 + y) * painter->width;
+				uint32_t original = *destination;
+
+#ifdef UI_FREETYPE_SUBPIXEL
+				uint32_t ra = ((uint8_t *) bitmap->buffer)[x * 3 + y * bitmap->pitch + 0];
+				uint32_t ga = ((uint8_t *) bitmap->buffer)[x * 3 + y * bitmap->pitch + 1];
+				uint32_t ba = ((uint8_t *) bitmap->buffer)[x * 3 + y * bitmap->pitch + 2];
+				ra += (ga - ra) / 2, ba += (ga - ba) / 2;
+#else
+				uint32_t ra = ((uint8_t *) bitmap->buffer)[x + y * bitmap->pitch];
+				uint32_t ga = ra, ba = ra;
+#endif
+				uint32_t r2 = (255 - ra) * ((original & 0x000000FF) >> 0);
+				uint32_t g2 = (255 - ga) * ((original & 0x0000FF00) >> 8);
+				uint32_t b2 = (255 - ba) * ((original & 0x00FF0000) >> 16);
+				uint32_t r1 = ra * ((color & 0x000000FF) >> 0);
+				uint32_t g1 = ga * ((color & 0x0000FF00) >> 8);
+				uint32_t b1 = ba * ((color & 0x00FF0000) >> 16);
+
+				uint32_t result = 0xFF000000 | (0x00FF0000 & ((b1 + b2) << 8)) 
+					| (0x0000FF00 & ((g1 + g2) << 0)) 
+					| (0x000000FF & ((r1 + r2) >> 8));
+				*destination = result;
+			}
+		}
+
+		return;
+	}
+#endif
+
+	if (c < 0 || c > 127) c = '?';
+
+	UIRectangle rectangle = UIRectangleIntersection(painter->clip, UI_RECT_4(x0, x0 + 8, y0, y0 + 16));
+
+	const uint8_t *data = (const uint8_t *) _uiFont + c * 16;
+
+	for (int i = rectangle.t; i < rectangle.b; i++) {
+		uint32_t *bits = painter->bits + i * painter->width + rectangle.l;
+		uint8_t byte = data[i - y0];
+
+		for (int j = rectangle.l; j < rectangle.r; j++) {
+			if (byte & (1 << (j - x0))) {
+				*bits = color;
+			}
+
+			bits++;
+		}
+	}
+}
+
+ptrdiff_t _UIStringLength(const char *cString) {
+	if (!cString) return 0;
+	ptrdiff_t length;
+	for (length = 0; cString[length]; length++);
+	return length;
+}
+
+char *UIStringCopy(const char *in, ptrdiff_t inBytes) {
+	if (inBytes == -1) {
+		inBytes = _UIStringLength(in);
+	}
+
+	char *buffer = (char *) UI_MALLOC(inBytes + 1);
+	
+	for (intptr_t i = 0; i < inBytes; i++) {
+		buffer[i] = in[i];
+	}
+	
+	buffer[inBytes] = 0;
+	return buffer;
+}
+
+int UIMeasureStringWidth(const char *string, ptrdiff_t bytes) {
+	if (bytes == -1) {
+		bytes = _UIStringLength(string);
+	}
+	
+	return bytes * ui.activeFont->glyphWidth;
+}
+
+int UIMeasureStringHeight() {
+	return ui.activeFont->glyphHeight;
+}
+
+void UIDrawString(UIPainter *painter, UIRectangle r, const char *string, ptrdiff_t bytes, uint32_t color, int align, UIStringSelection *selection) {
+	UIRectangle oldClip = painter->clip;
+	painter->clip = UIRectangleIntersection(r, oldClip);
+
+	if (!UI_RECT_VALID(painter->clip)) {
+		painter->clip = oldClip;
+		return;
+	}
+
+	if (bytes == -1) {
+		bytes = _UIStringLength(string);
+	}
+
+	int width = UIMeasureStringWidth(string, bytes);
+	int height = UIMeasureStringHeight();
+	int x = align == UI_ALIGN_CENTER ? ((r.l + r.r - width) / 2) : align == UI_ALIGN_RIGHT ? (r.r - width) : r.l;
+	int y = (r.t + r.b - height) / 2;
+	int i = 0, j = 0;
+
+	int selectFrom = -1, selectTo = -1;
+
+	if (selection) {
+		selectFrom = selection->carets[0];
+		selectTo = selection->carets[1];
+		
+		if (selectFrom > selectTo) {
+			UI_SWAP(int, selectFrom, selectTo);
+		}
+	}
+
+	for (; j < bytes; j++) {
+		char c = *string++;
+		uint32_t colorText = color;
+
+		if (j >= selectFrom && j < selectTo) {
+			UIDrawBlock(painter, UI_RECT_4(x, x + ui.activeFont->glyphWidth, y, y + height), selection->colorBackground);
+			colorText = selection->colorText;
+		}
+
+		if (c != '\t') {
+			UIDrawGlyph(painter, x, y, c, colorText);
+		}
+
+		if (selection && selection->carets[0] == j) {
+			UIDrawInvert(painter, UI_RECT_4(x, x + 1, y, y + height));
+		}
+
+		x += ui.activeFont->glyphWidth, i++;
+
+		if (c == '\t') {
+			while (i & 3) x += ui.activeFont->glyphWidth, i++;
+		}
+	}
+
+	if (selection && selection->carets[0] == j) {
+		UIDrawInvert(painter, UI_RECT_4(x, x + 1, y, y + height));
+	}
+
+	painter->clip = oldClip;
+}
+
+void UIDrawBorder(UIPainter *painter, UIRectangle r, uint32_t borderColor, UIRectangle borderSize) {
+	UIDrawBlock(painter, UI_RECT_4(r.l, r.r, r.t, r.t + borderSize.t), borderColor);
+	UIDrawBlock(painter, UI_RECT_4(r.l, r.l + borderSize.l, r.t + borderSize.t, r.b - borderSize.b), borderColor);
+	UIDrawBlock(painter, UI_RECT_4(r.r - borderSize.r, r.r, r.t + borderSize.t, r.b - borderSize.b), borderColor);
+	UIDrawBlock(painter, UI_RECT_4(r.l, r.r, r.b - borderSize.b, r.b), borderColor);
+}
+
+void UIDrawRectangle(UIPainter *painter, UIRectangle r, uint32_t mainColor, uint32_t borderColor, UIRectangle borderSize) {
+	UIDrawBorder(painter, r, borderColor, borderSize);
+	UIDrawBlock(painter, UI_RECT_4(r.l + borderSize.l, r.r - borderSize.r, r.t + borderSize.t, r.b - borderSize.b), mainColor);
+}
+
+void UIElementMove(UIElement *element, UIRectangle bounds, bool alwaysLayout) {
+	UIRectangle oldClip = element->clip;
+	element->clip = UIRectangleIntersection(element->parent->clip, bounds);
+
+	if (!UIRectangleEquals(element->bounds, bounds) || !UIRectangleEquals(element->clip, oldClip) || alwaysLayout) {
+		element->bounds = bounds;
+		UIElementMessage(element, UI_MSG_LAYOUT, 0, 0);
+	}
+}
+
+int UIElementMessage(UIElement *element, UIMessage message, int di, void *dp) {
+	if (message != UI_MSG_DESTROY && (element->flags & UI_ELEMENT_DESTROY)) {
+		return 0;
+	}
+
+	if (message >= UI_MSG_INPUT_EVENTS_START && message <= UI_MSG_INPUT_EVENTS_END && (element->flags & UI_ELEMENT_DISABLED)) {
+		return 0;
+	}
+
+	if (element->messageUser) {
+		int result = element->messageUser(element, message, di, dp);
+
+		if (result) {
+			return result;
+		}
+	}
+
+	if (element->messageClass) {
+		return element->messageClass(element, message, di, dp);
+	} else {
+		return 0;
+	}
+}
+
+void UIElementChangeParent(UIElement *element, UIElement *newParent, UIElement *insertBefore) {
+	UIElement **link = &element->parent->children;
+
+	while (true) {
+		if (*link == element) {
+			*link = element->next;
+			break;
+		} else {
+			link = &(*link)->next;
+		}
+	}
+
+	link = &newParent->children;
+	element->next = insertBefore;
+
+	while (true) {
+		if ((*link) == insertBefore) {
+			*link = element;
+			break;
+		} else {
+			link = &(*link)->next;
+		}
+	}
+
+	element->parent = newParent;
+	element->window = newParent->window;
+}
+
+UIElement *UIElementCreate(size_t bytes, UIElement *parent, uint32_t flags, int (*message)(UIElement *, UIMessage, int, void *), const char *cClassName) {
+	UI_ASSERT(bytes >= sizeof(UIElement));
+	UIElement *element = (UIElement *) UI_CALLOC(bytes);
+	element->flags = flags;
+	element->messageClass = message;
+
+	if (!parent && (~flags & UI_ELEMENT_WINDOW)) {
+		UI_ASSERT(ui.parentStackCount);
+		parent = ui.parentStack[ui.parentStackCount - 1];
+	}
+
+	if ((~flags & UI_ELEMENT_NON_CLIENT) && parent) {
+		UIElementMessage(parent, UI_MSG_CLIENT_PARENT, 0, &parent);
+	}
+
+	if (parent) {
+		element->window = parent->window;
+		element->parent = parent;
+
+		if (parent->children) {
+			UIElement *sibling = parent->children;
+
+			while (sibling->next) {
+				sibling = sibling->next;
+			}
+
+			sibling->next = element;
+		} else {
+			parent->children = element;
+		}
+
+		UI_ASSERT(~parent->flags & UI_ELEMENT_DESTROY);
+	}
+
+	element->cClassName = cClassName;
+	static uint32_t id = 0;
+	element->id = ++id;
+
+#ifdef UI_DEBUG
+	_UIInspectorRefresh();
+#endif
+
+	if (flags & UI_ELEMENT_PARENT_PUSH) {
+		UIParentPush(element);
+	}
+
+	return element;
+}
+
+UIElement *UIParentPush(UIElement *element) {
+	UI_ASSERT(ui.parentStackCount != sizeof(ui.parentStack) / sizeof(ui.parentStack[0]));
+	ui.parentStack[ui.parentStackCount++] = element;
+	return element;
+}
+
+UIElement *UIParentPop() {
+	UI_ASSERT(ui.parentStackCount);
+	ui.parentStackCount--;
+	return ui.parentStack[ui.parentStackCount];
+}
+
+int _UIPanelMeasure(UIPanel *panel) {
+	bool horizontal = panel->e.flags & UI_PANEL_HORIZONTAL;
+	int size = 0;
+	UIElement *child = panel->e.children;
+
+	while (child) {
+		if (~child->flags & UI_ELEMENT_HIDE) {
+			if (horizontal) {
+				int height = UIElementMessage(child, UI_MSG_GET_HEIGHT, 0, 0);
+
+				if (height > size) {
+					size = height;
+				}
+			} else {
+				int width = UIElementMessage(child, UI_MSG_GET_WIDTH, 0, 0);
+
+				if (width > size) {
+					size = width;
+				}
+			}
+		}
+
+		child = child->next;
+	}
+
+	int border = 0;
+
+	if (horizontal) {
+		border = panel->border.t + panel->border.b;
+	} else {
+		border = panel->border.l + panel->border.r;
+	}
+
+	return size + border * panel->e.window->scale;
+}
+
+int _UIPanelLayout(UIPanel *panel, UIRectangle bounds, bool measure) {
+	bool horizontal = panel->e.flags & UI_PANEL_HORIZONTAL;
+	float scale = panel->e.window->scale;
+	int position = (horizontal ? panel->border.l : panel->border.t) * scale;
+	if (panel->scrollBar && !measure) position -= panel->scrollBar->position;
+	int hSpace = UI_RECT_WIDTH(bounds) - UI_RECT_TOTAL_H(panel->border) * scale;
+	int vSpace = UI_RECT_HEIGHT(bounds) - UI_RECT_TOTAL_V(panel->border) * scale;
+
+	int available = horizontal ? hSpace : vSpace;
+	int fill = 0, count = 0, perFill = 0;
+
+	for (UIElement *child = panel->e.children; child; child = child->next) {
+		if (child->flags & (UI_ELEMENT_HIDE | UI_ELEMENT_NON_CLIENT)) {
+			continue;
+		}
+
+		count++;
+
+		if (horizontal) {
+			if (child->flags & UI_ELEMENT_H_FILL) {
+				fill++;
+			} else if (available > 0) {
+				available -= UIElementMessage(child, UI_MSG_GET_WIDTH, vSpace, 0);
+			}
+		} else {
+			if (child->flags & UI_ELEMENT_V_FILL) {
+				fill++;
+			} else if (available > 0) {
+				available -= UIElementMessage(child, UI_MSG_GET_HEIGHT, hSpace, 0);
+			}
+		}
+	}
+
+	if (count) {
+		available -= (count - 1) * (int) (panel->gap * scale);
+	}
+
+	if (available > 0 && fill) {
+		perFill = available / fill;
+	}
+
+	bool expand = panel->e.flags & UI_PANEL_EXPAND;
+	int scaledBorder2 = (horizontal ? panel->border.t : panel->border.l) * panel->e.window->scale;
+
+	for (UIElement *child = panel->e.children; child; child = child->next) {
+		if (child->flags & (UI_ELEMENT_HIDE | UI_ELEMENT_NON_CLIENT)) {
+			continue;
+		}
+
+		if (horizontal) {
+			int height = ((child->flags & UI_ELEMENT_V_FILL) || expand) ? vSpace : UIElementMessage(child, UI_MSG_GET_HEIGHT, 0, 0);
+			int width = (child->flags & UI_ELEMENT_H_FILL) ? perFill : UIElementMessage(child, UI_MSG_GET_WIDTH, height, 0);
+			UIRectangle relative = UI_RECT_4(position, position + width, 
+					scaledBorder2 + (vSpace - height) / 2, 
+					scaledBorder2 + (vSpace + height) / 2);
+			if (!measure) UIElementMove(child, UIRectangleTranslate(relative, bounds), false);
+			position += width + panel->gap * scale;
+		} else {
+			int width = ((child->flags & UI_ELEMENT_H_FILL) || expand) ? hSpace : UIElementMessage(child, UI_MSG_GET_WIDTH, 0, 0);
+			int height = (child->flags & UI_ELEMENT_V_FILL) ? perFill : UIElementMessage(child, UI_MSG_GET_HEIGHT, width, 0);
+			UIRectangle relative = UI_RECT_4(scaledBorder2 + (hSpace - width) / 2, 
+					scaledBorder2 + (hSpace + width) / 2, position, position + height);
+			if (!measure) UIElementMove(child, UIRectangleTranslate(relative, bounds), false);
+			position += height + panel->gap * scale;
+		}
+	}
+
+	return position - panel->gap * scale + (horizontal ? panel->border.r : panel->border.b) * scale;
+}
+
+int _UIPanelMessage(UIElement *element, UIMessage message, int di, void *dp) {
+	UIPanel *panel = (UIPanel *) element;
+	bool horizontal = element->flags & UI_PANEL_HORIZONTAL;
+
+	if (message == UI_MSG_LAYOUT) {
+		int scrollBarWidth = panel->scrollBar ? (UI_SIZE_SCROLL_BAR * element->window->scale) : 0;
+		UIRectangle bounds = element->bounds;
+		bounds.r -= scrollBarWidth;
+
+		if (panel->scrollBar) {
+			UIRectangle scrollBarBounds = element->bounds;
+			scrollBarBounds.l = scrollBarBounds.r - scrollBarWidth;
+			panel->scrollBar->maximum = _UIPanelLayout(panel, bounds, true);
+			panel->scrollBar->page = UI_RECT_HEIGHT(element->bounds);
+			UIElementMove(&panel->scrollBar->e, scrollBarBounds, true);
+		}
+
+		_UIPanelLayout(panel, bounds, false);
+	} else if (message == UI_MSG_GET_WIDTH) {
+		if (horizontal) {
+			return _UIPanelLayout(panel, UI_RECT_4(0, 0, 0, di), true);
+		} else {
+			return _UIPanelMeasure(panel);
+		}
+	} else if (message == UI_MSG_GET_HEIGHT) {
+		if (horizontal) {
+			return _UIPanelMeasure(panel);
+		} else {
+			int width = di && panel->scrollBar ? (di - UI_SIZE_SCROLL_BAR * element->window->scale) : di;
+			return _UIPanelLayout(panel, UI_RECT_4(0, width, 0, 0), true);
+		}
+	} else if (message == UI_MSG_PAINT) {
+		if (element->flags & UI_PANEL_GRAY) {
+			UIDrawBlock((UIPainter *) dp, element->bounds, ui.theme.panel1);
+		} else if (element->flags & UI_PANEL_WHITE) {
+			UIDrawBlock((UIPainter *) dp, element->bounds, ui.theme.panel2);
+		} 
+		
+		if (element->flags & UI_PANEL_BORDER) {
+			UIDrawBorder((UIPainter *) dp, element->bounds, ui.theme.border, UI_RECT_1((int) element->window->scale));
+		}
+	} else if (message == UI_MSG_MOUSE_WHEEL && panel->scrollBar) {
+		return UIElementMessage(&panel->scrollBar->e, message, di, dp);
+	} else if (message == UI_MSG_SCROLLED) {
+		UIElementRefresh(element);
+	}
+
+	return 0;
+}
+
+UIPanel *UIPanelCreate(UIElement *parent, uint32_t flags) {
+	UIPanel *panel = (UIPanel *) UIElementCreate(sizeof(UIPanel), parent, flags, _UIPanelMessage, "Panel");
+
+	if (flags & UI_PANEL_MEDIUM_SPACING) {
+		panel->border = UI_RECT_1(UI_SIZE_PANE_MEDIUM_BORDER);
+		panel->gap = UI_SIZE_PANE_MEDIUM_GAP;
+	} else if (flags & UI_PANEL_SMALL_SPACING) {
+		panel->border = UI_RECT_1(UI_SIZE_PANE_SMALL_BORDER);
+		panel->gap = UI_SIZE_PANE_SMALL_GAP;
+	}
+
+	if (flags & UI_PANEL_SCROLL) {
+		panel->scrollBar = UIScrollBarCreate(&panel->e, UI_ELEMENT_NON_CLIENT);
+	}
+
+	return panel;
+}
+
+void _UIWrapPanelLayoutRow(UIWrapPanel *panel, UIElement *child, UIElement *rowEnd, int rowY, int rowHeight) {
+	int rowPosition = 0;
+
+	while (child != rowEnd) {
+		int height = UIElementMessage(child, UI_MSG_GET_HEIGHT, 0, 0);
+		int width = UIElementMessage(child, UI_MSG_GET_WIDTH, 0, 0);
+		UIRectangle relative = UI_RECT_4(rowPosition, rowPosition + width, rowY + rowHeight / 2 - height / 2, rowY + rowHeight / 2 + height / 2);
+		UIElementMove(child, UIRectangleTranslate(relative, panel->e.bounds), false);
+		child = child->next;
+		rowPosition += width;
+	}
+}
+
+int _UIWrapPanelMessage(UIElement *element, UIMessage message, int di, void *dp) {
+	UIWrapPanel *panel = (UIWrapPanel *) element;
+	bool horizontal = element->flags & UI_PANEL_HORIZONTAL;
+
+	if (message == UI_MSG_LAYOUT || message == UI_MSG_GET_HEIGHT) {
+		int totalHeight = 0;
+		int rowPosition = 0;
+		int rowHeight = 0;
+		int rowLimit = message == UI_MSG_LAYOUT ? UI_RECT_WIDTH(element->bounds) : di;
+
+		UIElement *child = panel->e.children;
+		UIElement *rowStart = child;
+
+		while (child) {
+			if (~child->flags & UI_ELEMENT_HIDE) {
+				int height = UIElementMessage(child, UI_MSG_GET_HEIGHT, 0, 0);
+				int width = UIElementMessage(child, UI_MSG_GET_WIDTH, 0, 0);
+
+				if (rowLimit && rowPosition + width > rowLimit) {
+					_UIWrapPanelLayoutRow(panel, rowStart, child, totalHeight, rowHeight);
+					totalHeight += rowHeight;
+					rowPosition = rowHeight = 0;
+					rowStart = child;
+				}
+
+				if (height > rowHeight) {
+					rowHeight = height;
+				}
+
+				rowPosition += width;
+			}
+
+			child = child->next;
+		}
+
+		if (message == UI_MSG_GET_HEIGHT) {
+			return totalHeight + rowHeight;
+		} else {
+			_UIWrapPanelLayoutRow(panel, rowStart, child, totalHeight, rowHeight);
+		}
+	}
+
+	return 0;
+}
+
+UIWrapPanel *UIWrapPanelCreate(UIElement *parent, uint32_t flags) {
+	return (UIWrapPanel *) UIElementCreate(sizeof(UIWrapPanel), parent, flags, _UIWrapPanelMessage, "Wrap Panel");
+}
+
+void _UIButtonCalculateColors(UIElement *element, uint32_t *color, uint32_t *textColor) {
+	bool disabled = element->flags & UI_ELEMENT_DISABLED;
+	bool focused = element == element->window->focused;
+	bool pressed = element == element->window->pressed;
+	bool hovered = element == element->window->hovered;
+	*color = disabled ? ui.theme.buttonDisabled
+		: (pressed && hovered) ? ui.theme.buttonPressed 
+		: (pressed || hovered) ? ui.theme.buttonHovered 
+		: focused ? ui.theme.selected : ui.theme.buttonNormal;
+	*textColor = disabled ? ui.theme.textDisabled 
+		: *color == ui.theme.selected ? ui.theme.textSelected : ui.theme.text;
+}
+
+
+int _UIButtonMessage(UIElement *element, UIMessage message, int di, void *dp) {
+	UIButton *button = (UIButton *) element;
+	bool isMenuItem = element->flags & UI_BUTTON_MENU_ITEM;
+	bool isDropDown = element->flags & UI_BUTTON_DROP_DOWN;
+	
+	if (message == UI_MSG_GET_HEIGHT) {
+		if (isMenuItem) {
+			return UI_SIZE_MENU_ITEM_HEIGHT * element->window->scale;
+		} else {
+			return UI_SIZE_BUTTON_HEIGHT * element->window->scale;
+		}
+	} else if (message == UI_MSG_GET_WIDTH) {
+		int labelSize = UIMeasureStringWidth(button->label, button->labelBytes);
+		int paddedSize = labelSize + UI_SIZE_BUTTON_PADDING * element->window->scale;
+		if (isDropDown) paddedSize += ui.activeFont->glyphWidth * 2;
+		int minimumSize = ((element->flags & UI_BUTTON_SMALL) ? 0 
+				: isMenuItem ? UI_SIZE_MENU_ITEM_MINIMUM_WIDTH 
+				: UI_SIZE_BUTTON_MINIMUM_WIDTH) 
+			* element->window->scale;
+		return paddedSize > minimumSize ? paddedSize : minimumSize;
+	} else if (message == UI_MSG_PAINT) {
+		UIPainter *painter = (UIPainter *) dp;
+
+		uint32_t color, textColor;
+		_UIButtonCalculateColors(element, &color, &textColor);
+
+		UIDrawRectangle(painter, element->bounds, color, ui.theme.border, UI_RECT_1(isMenuItem ? 0 : 1));
+
+		if (element->flags & UI_BUTTON_CHECKED) {
+			UIDrawBlock(painter, UIRectangleAdd(element->bounds, 
+				UI_RECT_1I((int) (UI_SIZE_BUTTON_CHECKED_AREA * element->window->scale))), ui.theme.buttonPressed);
+		}
+
+		UIRectangle bounds = UIRectangleAdd(element->bounds, UI_RECT_2I((int) (UI_SIZE_MENU_ITEM_MARGIN * element->window->scale), 0));
+
+		if (isMenuItem) {
+			if (button->labelBytes == -1) {
+				button->labelBytes = _UIStringLength(button->label);
+			}
+
+			int tab = 0;
+			for (; tab < button->labelBytes && button->label[tab] != '\t'; tab++);
+
+			UIDrawString(painter, bounds, button->label, tab, textColor, UI_ALIGN_LEFT, NULL);
+
+			if (button->labelBytes > tab) {
+				UIDrawString(painter, bounds, button->label + tab + 1, button->labelBytes - tab - 1, textColor, UI_ALIGN_RIGHT, NULL);
+			}
+		} else if (isDropDown) {
+			UIDrawString(painter, bounds, button->label, button->labelBytes, textColor, UI_ALIGN_LEFT, NULL);
+			UIDrawString(painter, bounds, "\x19", 1, textColor, UI_ALIGN_RIGHT, NULL);
+		} else {
+			UIDrawString(painter, element->bounds, button->label, button->labelBytes, textColor, UI_ALIGN_CENTER, NULL);
+		}
+	} else if (message == UI_MSG_UPDATE) {
+		UIElementRepaint(element, NULL);
+	} else if (message == UI_MSG_DESTROY) {
+		UI_FREE(button->label);
+	} else if (message == UI_MSG_LEFT_DOWN) {
+		if (element->flags & UI_BUTTON_CAN_FOCUS) {
+			UIElementFocus(element);
+		}
+	} else if (message == UI_MSG_KEY_TYPED) {
+		UIKeyTyped *m = (UIKeyTyped *) dp;
+		
+		if (m->textBytes == 1 && m->text[0] == ' ') {
+			UIElementMessage(element, UI_MSG_CLICKED, 0, 0);
+			UIElementRepaint(element, NULL);
+		}
+	} else if (message == UI_MSG_CLICKED) {
+		if (button->invoke) {
+			button->invoke(element->cp);
+		}
+	}
+
+	return 0;
+}
+
+UIButton *UIButtonCreate(UIElement *parent, uint32_t flags, const char *label, ptrdiff_t labelBytes) {
+	UIButton *button = (UIButton *) UIElementCreate(sizeof(UIButton), parent, flags | UI_ELEMENT_TAB_STOP, _UIButtonMessage, "Button");
+	button->label = UIStringCopy(label, (button->labelBytes = labelBytes));
+	return button;
+}
+
+int _UICheckboxMessage(UIElement *element, UIMessage message, int di, void *dp) {
+	UICheckbox *box = (UICheckbox *) element;
+	
+	if (message == UI_MSG_GET_HEIGHT) {
+		return UI_SIZE_BUTTON_HEIGHT * element->window->scale;
+	} else if (message == UI_MSG_GET_WIDTH) {
+		int labelSize = UIMeasureStringWidth(box->label, box->labelBytes);
+		return (labelSize + UI_SIZE_CHECKBOX_BOX + UI_SIZE_CHECKBOX_GAP) * element->window->scale;
+	} else if (message == UI_MSG_PAINT) {
+		UIPainter *painter = (UIPainter *) dp;
+		uint32_t color, textColor;
+		_UIButtonCalculateColors(element, &color, &textColor);
+		int midY = (element->bounds.t + element->bounds.b) / 2;
+		UIRectangle boxBounds = UI_RECT_4(element->bounds.l, element->bounds.l + UI_SIZE_CHECKBOX_BOX, 
+				midY - UI_SIZE_CHECKBOX_BOX / 2, midY + UI_SIZE_CHECKBOX_BOX / 2);
+		UIDrawRectangle(painter, boxBounds, color, ui.theme.border, UI_RECT_1(1));
+		UIDrawString(painter, UIRectangleAdd(boxBounds, UI_RECT_4(1, 0, 0, 0)), 
+				box->check == UI_CHECK_CHECKED ? "*" : box->check == UI_CHECK_INDETERMINATE ? "-" : " ", -1, 
+				textColor, UI_ALIGN_CENTER, NULL);
+		UIDrawString(painter, UIRectangleAdd(element->bounds, UI_RECT_4(UI_SIZE_CHECKBOX_BOX + UI_SIZE_CHECKBOX_GAP, 0, 0, 0)), 
+				box->label, box->labelBytes, textColor, UI_ALIGN_LEFT, NULL);
+	} else if (message == UI_MSG_UPDATE) {
+		UIElementRepaint(element, NULL);
+	} else if (message == UI_MSG_DESTROY) {
+		UI_FREE(box->label);
+	} else if (message == UI_MSG_KEY_TYPED) {
+		UIKeyTyped *m = (UIKeyTyped *) dp;
+		
+		if (m->textBytes == 1 && m->text[0] == ' ') {
+			UIElementMessage(element, UI_MSG_CLICKED, 0, 0);
+			UIElementRepaint(element, NULL);
+		}
+	} else if (message == UI_MSG_CLICKED) {
+		box->check = (box->check + 1) % ((element->flags & UI_CHECKBOX_ALLOW_INDETERMINATE) ? 3 : 2);
+		UIElementRepaint(element, NULL);
+		if (box->invoke) box->invoke(element->cp);
+	}
+
+	return 0;
+}
+
+UICheckbox *UICheckboxCreate(UIElement *parent, uint32_t flags, const char *label, ptrdiff_t labelBytes) {
+	UICheckbox *box = (UICheckbox *) UIElementCreate(sizeof(UICheckbox), parent, flags | UI_ELEMENT_TAB_STOP, _UICheckboxMessage, "Checkbox");
+	box->label = UIStringCopy(label, (box->labelBytes = labelBytes));
+	return box;
+}
+
+int _UILabelMessage(UIElement *element, UIMessage message, int di, void *dp) {
+	UILabel *label = (UILabel *) element;
+	
+	if (message == UI_MSG_GET_HEIGHT) {
+		return UIMeasureStringHeight();
+	} else if (message == UI_MSG_GET_WIDTH) {
+		return UIMeasureStringWidth(label->label, label->labelBytes);
+	} else if (message == UI_MSG_PAINT) {
+		UIPainter *painter = (UIPainter *) dp;
+		UIDrawString(painter, element->bounds, label->label, label->labelBytes, ui.theme.text, UI_ALIGN_LEFT, NULL);
+	} else if (message == UI_MSG_DESTROY) {
+		UI_FREE(label->label);
+	}
+
+	return 0;
+}
+
+void UILabelSetContent(UILabel *label, const char *string, ptrdiff_t stringBytes) {
+	UI_FREE(label->label);
+	label->label = UIStringCopy(string, (label->labelBytes = stringBytes));
+}
+
+UILabel *UILabelCreate(UIElement *parent, uint32_t flags, const char *string, ptrdiff_t stringBytes) {
+	UILabel *label = (UILabel *) UIElementCreate(sizeof(UILabel), parent, flags, _UILabelMessage, "Label");
+	label->label = UIStringCopy(string, (label->labelBytes = stringBytes));
+	return label;
+}
+
+int _UISplitPaneMessage(UIElement *element, UIMessage message, int di, void *dp);
+
+int _UISplitterMessage(UIElement *element, UIMessage message, int di, void *dp) {
+	UISplitPane *splitPane = (UISplitPane *) element->parent;
+	bool vertical = splitPane->e.flags & UI_SPLIT_PANE_VERTICAL;
+
+	if (message == UI_MSG_PAINT) {
+		UIRectangle borders = vertical ? UI_RECT_2(0, 1) : UI_RECT_2(1, 0);
+		UIDrawRectangle((UIPainter *) dp, element->bounds, ui.theme.buttonNormal, ui.theme.border, borders);
+	} else if (message == UI_MSG_GET_CURSOR) {
+		return vertical ? UI_CURSOR_SPLIT_V : UI_CURSOR_SPLIT_H;
+	} else if (message == UI_MSG_MOUSE_DRAG) {
+		int cursor = vertical ? element->window->cursorY : element->window->cursorX;
+		int splitterSize = UI_SIZE_SPLITTER * element->window->scale;
+		int space = (vertical ? UI_RECT_HEIGHT(splitPane->e.bounds) : UI_RECT_WIDTH(splitPane->e.bounds)) - splitterSize;
+		float oldWeight = splitPane->weight;
+		splitPane->weight = (float) (cursor - splitterSize / 2 - (vertical ? splitPane->e.bounds.t : splitPane->e.bounds.l)) / space;
+		if (splitPane->weight < 0.05f) splitPane->weight = 0.05f;
+		if (splitPane->weight > 0.95f) splitPane->weight = 0.95f;
+
+		if (element->next->next->messageClass == _UISplitPaneMessage 
+				&& (element->next->next->flags & UI_SPLIT_PANE_VERTICAL) == (splitPane->e.flags & UI_SPLIT_PANE_VERTICAL)) {
+			UISplitPane *subSplitPane = (UISplitPane *) element->next->next;
+			subSplitPane->weight = (splitPane->weight - oldWeight - subSplitPane->weight + oldWeight * subSplitPane->weight) / (-1 + splitPane->weight);
+			if (subSplitPane->weight < 0.05f) subSplitPane->weight = 0.05f;
+			if (subSplitPane->weight > 0.95f) subSplitPane->weight = 0.95f;
+		}
+
+		UIElementRefresh(&splitPane->e);
+	}
+
+	return 0;
+}
+
+int _UISplitPaneMessage(UIElement *element, UIMessage message, int di, void *dp) {
+	UISplitPane *splitPane = (UISplitPane *) element;
+	bool vertical = splitPane->e.flags & UI_SPLIT_PANE_VERTICAL;
+
+	if (message == UI_MSG_LAYOUT) {
+		UIElement *splitter = element->children;
+		UI_ASSERT(splitter);
+		UIElement *left = splitter->next;
+		UI_ASSERT(left);
+		UIElement *right = left->next;
+		UI_ASSERT(right);
+		UI_ASSERT(!right->next);
+
+		int splitterSize = UI_SIZE_SPLITTER * element->window->scale;
+		int space = (vertical ? UI_RECT_HEIGHT(element->bounds) : UI_RECT_WIDTH(element->bounds)) - splitterSize;
+		int leftSize = space * splitPane->weight;
+		int rightSize = space - leftSize;
+
+		if (vertical) {
+			UIElementMove(left, UI_RECT_4(element->bounds.l, element->bounds.r, element->bounds.t, element->bounds.t + leftSize), false);
+			UIElementMove(splitter, UI_RECT_4(element->bounds.l, element->bounds.r, element->bounds.t + leftSize, element->bounds.t + leftSize + splitterSize), false);
+			UIElementMove(right, UI_RECT_4(element->bounds.l, element->bounds.r, element->bounds.b - rightSize, element->bounds.b), false);
+		} else {
+			UIElementMove(left, UI_RECT_4(element->bounds.l, element->bounds.l + leftSize, element->bounds.t, element->bounds.b), false);
+			UIElementMove(splitter, UI_RECT_4(element->bounds.l + leftSize, element->bounds.l + leftSize + splitterSize, element->bounds.t, element->bounds.b), false);
+			UIElementMove(right, UI_RECT_4(element->bounds.r - rightSize, element->bounds.r, element->bounds.t, element->bounds.b), false);
+		}
+	}
+
+	return 0;
+}
+
+UISplitPane *UISplitPaneCreate(UIElement *parent, uint32_t flags, float weight) {
+	UISplitPane *splitPane = (UISplitPane *) UIElementCreate(sizeof(UISplitPane), parent, flags, _UISplitPaneMessage, "Split Pane");
+	splitPane->weight = weight;
+	UIElementCreate(sizeof(UIElement), &splitPane->e, 0, _UISplitterMessage, "Splitter");
+	return splitPane;
+}
+
+int _UITabPaneMessage(UIElement *element, UIMessage message, int di, void *dp) {
+	UITabPane *tabPane = (UITabPane *) element;
+	
+	if (message == UI_MSG_PAINT) {
+		UIPainter *painter = (UIPainter *) dp;
+		UIRectangle top = element->bounds;
+		top.b = top.t + UI_SIZE_BUTTON_HEIGHT * element->window->scale;
+		UIDrawRectangle(painter, top, ui.theme.panel1, ui.theme.border, UI_RECT_4(0, 0, 0, 1));
+
+		UIRectangle tab = top;
+		tab.l += UI_SIZE_TAB_PANE_SPACE_LEFT * element->window->scale;
+		tab.t += UI_SIZE_TAB_PANE_SPACE_TOP * element->window->scale;
+
+		int position = 0;
+		int index = 0;
+
+		while (true) {
+			int end = position;
+			for (; tabPane->tabs[end] != '\t' && tabPane->tabs[end]; end++);
+
+			int width = UIMeasureStringWidth(tabPane->tabs, end - position);
+			tab.r = tab.l + width + UI_SIZE_BUTTON_PADDING;
+
+			uint32_t color = tabPane->active == index ? ui.theme.buttonPressed : ui.theme.buttonNormal;
+
+			UIRectangle t = tab;
+
+			if (tabPane->active == index) {
+				t.b++;
+				t.t--;
+			} else {
+				t.t++;
+			}
+
+			UIDrawRectangle(painter, t, color, ui.theme.border, UI_RECT_1(1));
+			UIDrawString(painter, tab, tabPane->tabs + position, end - position, ui.theme.text, UI_ALIGN_CENTER, NULL);
+			tab.l = tab.r - 1;
+
+			if (tabPane->tabs[end] == '\t') {
+				position = end + 1;
+				index++;
+			} else {
+				break;
+			}
+		}
+	} else if (message == UI_MSG_LEFT_DOWN) {
+		UIRectangle tab = element->bounds;
+		tab.b = tab.t + UI_SIZE_BUTTON_HEIGHT * element->window->scale;
+		tab.l += UI_SIZE_TAB_PANE_SPACE_LEFT * element->window->scale;
+		tab.t += UI_SIZE_TAB_PANE_SPACE_TOP * element->window->scale;
+
+		int position = 0;
+		int index = 0;
+
+		while (true) {
+			int end = position;
+			for (; tabPane->tabs[end] != '\t' && tabPane->tabs[end]; end++);
+
+			int width = UIMeasureStringWidth(tabPane->tabs, end - position);
+			tab.r = tab.l + width + UI_SIZE_BUTTON_PADDING;
+
+			if (UIRectangleContains(tab, element->window->cursorX, element->window->cursorY)) {
+				tabPane->active = index;
+				UIElementMessage(element, UI_MSG_LAYOUT, 0, 0);
+				UIElementRepaint(element, NULL);
+				break;
+			}
+
+			tab.l = tab.r - 1;
+
+			if (tabPane->tabs[end] == '\t') {
+				position = end + 1;
+				index++;
+			} else {
+				break;
+			}
+		}
+	} else if (message == UI_MSG_LAYOUT) {
+		UIElement *child = element->children;
+		int index = 0;
+
+		UIRectangle content = element->bounds;
+		content.t += UI_SIZE_BUTTON_HEIGHT * element->window->scale;
+
+		while (child) {
+			if (tabPane->active == index) {
+				child->flags &= ~UI_ELEMENT_HIDE;
+				UIElementMove(child, content, false);
+				UIElementMessage(child, UI_MSG_TAB_SELECTED, 0, 0);
+			} else {
+				child->flags |= UI_ELEMENT_HIDE;
+			}
+
+			child = child->next;
+			index++;
+		}
+	} else if (message == UI_MSG_GET_HEIGHT) {
+		UIElement *child = element->children;
+		int index = 0;
+		int baseHeight = UI_SIZE_BUTTON_HEIGHT * element->window->scale;
+
+		while (child) {
+			if (tabPane->active == index) {
+				return baseHeight + UIElementMessage(child, UI_MSG_GET_HEIGHT, di, dp);
+			}
+
+			child = child->next;
+			index++;
+		}
+	} else if (message == UI_MSG_DESTROY) {
+		UI_FREE(tabPane->tabs);
+	}
+
+	return 0;
+}
+
+UITabPane *UITabPaneCreate(UIElement *parent, uint32_t flags, const char *tabs) {
+	UITabPane *tabPane = (UITabPane *) UIElementCreate(sizeof(UITabPane), parent, flags, _UITabPaneMessage, "Tab Pane");
+	tabPane->tabs = UIStringCopy(tabs, -1);
+	return tabPane;
+}
+
+int _UISpacerMessage(UIElement *element, UIMessage message, int di, void *dp) {
+	UISpacer *spacer = (UISpacer *) element;
+	
+	if (message == UI_MSG_GET_HEIGHT) {
+		return spacer->height * element->window->scale;
+	} else if (message == UI_MSG_GET_WIDTH) {
+		return spacer->width * element->window->scale;
+	} else if (message == UI_MSG_PAINT && (element->flags & UI_SPACER_LINE)) {
+		UIDrawBlock((UIPainter *) dp, element->bounds, ui.theme.border);
+	}
+
+	return 0;
+}
+
+UISpacer *UISpacerCreate(UIElement *parent, uint32_t flags, int width, int height) {
+	UISpacer *spacer = (UISpacer *) UIElementCreate(sizeof(UISpacer), parent, flags, _UISpacerMessage, "Spacer");
+	spacer->width = width;
+	spacer->height = height;
+	return spacer;
+}
+
+int _UIScrollBarMessage(UIElement *element, UIMessage message, int di, void *dp) {
+	UIScrollBar *scrollBar = (UIScrollBar *) element;
+
+	if (message == UI_MSG_GET_WIDTH || message == UI_MSG_GET_HEIGHT) {
+		return UI_SIZE_SCROLL_BAR * element->window->scale;
+	} else if (message == UI_MSG_LAYOUT) {
+		UIElement *up = element->children;
+		UIElement *thumb = up->next;
+		UIElement *down = thumb->next;
+
+		if (scrollBar->page >= scrollBar->maximum || scrollBar->maximum <= 0 || scrollBar->page <= 0) {
+			up->flags |= UI_ELEMENT_HIDE;
+			thumb->flags |= UI_ELEMENT_HIDE;
+			down->flags |= UI_ELEMENT_HIDE;
+
+			scrollBar->position = 0;
+		} else {
+			up->flags &= ~UI_ELEMENT_HIDE;
+			thumb->flags &= ~UI_ELEMENT_HIDE;
+			down->flags &= ~UI_ELEMENT_HIDE;
+
+			int size = scrollBar->horizontal ? UI_RECT_WIDTH(element->bounds) : UI_RECT_HEIGHT(element->bounds);
+			int thumbSize = size * scrollBar->page / scrollBar->maximum;
+
+			if (thumbSize < UI_SIZE_SCROLL_MINIMUM_THUMB * element->window->scale) {
+				thumbSize = UI_SIZE_SCROLL_MINIMUM_THUMB * element->window->scale;
+			}
+
+			if (scrollBar->position < 0) {
+				scrollBar->position = 0;
+			} else if (scrollBar->position > scrollBar->maximum - scrollBar->page) {
+				scrollBar->position = scrollBar->maximum - scrollBar->page;
+			}
+
+			int thumbPosition = scrollBar->position / (scrollBar->maximum - scrollBar->page) * (size - thumbSize);
+
+			if (scrollBar->position == scrollBar->maximum - scrollBar->page) {
+				thumbPosition = size - thumbSize;
+			}
+
+			if (scrollBar->horizontal) {
+				UIRectangle r = element->bounds;
+				r.r = r.l + thumbPosition;
+				UIElementMove(up, r, false);
+				r.l = r.r, r.r = r.l + thumbSize;
+				UIElementMove(thumb, r, false);
+				r.l = r.r, r.r = element->bounds.r;
+				UIElementMove(down, r, false);
+			} else {
+				UIRectangle r = element->bounds;
+				r.b = r.t + thumbPosition;
+				UIElementMove(up, r, false);
+				r.t = r.b, r.b = r.t + thumbSize;
+				UIElementMove(thumb, r, false);
+				r.t = r.b, r.b = element->bounds.b;
+				UIElementMove(down, r, false);
+			}
+		}
+	} else if (message == UI_MSG_PAINT) {
+		if (scrollBar->page >= scrollBar->maximum || scrollBar->maximum <= 0 || scrollBar->page <= 0) {
+			UIDrawBlock((UIPainter *) dp, element->bounds, ui.theme.panel1);
+		}
+	} else if (message == UI_MSG_MOUSE_WHEEL) {
+		scrollBar->position += di;
+		UIElementRefresh(element);
+		UIElementMessage(element->parent, UI_MSG_SCROLLED, 0, 0);
+		return 1;
+	}
+
+	return 0;
+}
+
+int _UIScrollUpDownMessage(UIElement *element, UIMessage message, int di, void *dp) {
+	UIScrollBar *scrollBar = (UIScrollBar *) element->parent;
+	bool isDown = element->cp;
+
+	if (message == UI_MSG_PAINT) {
+		UIPainter *painter = (UIPainter *) dp;
+		uint32_t color = element == element->window->pressed ? ui.theme.buttonPressed 
+			: element == element->window->hovered ? ui.theme.buttonHovered : ui.theme.panel2;
+		UIDrawRectangle(painter, element->bounds, color, ui.theme.border, UI_RECT_1(0));
+		
+		if (scrollBar->horizontal) {
+			UIDrawGlyph(painter, isDown ? (element->bounds.r - ui.activeFont->glyphWidth - 2 * element->window->scale) 
+					: (element->bounds.l + 2 * element->window->scale), 
+				(element->bounds.t + element->bounds.b - ui.activeFont->glyphHeight) / 2,
+				isDown ? 26 : 27, ui.theme.text);
+		} else {
+			UIDrawGlyph(painter, (element->bounds.l + element->bounds.r - ui.activeFont->glyphWidth) / 2 + 1, 
+				isDown ? (element->bounds.b - ui.activeFont->glyphHeight - 2 * element->window->scale) 
+					: (element->bounds.t + 2 * element->window->scale), 
+				isDown ? 25 : 24, ui.theme.text);
+		}
+	} else if (message == UI_MSG_UPDATE) {
+		UIElementRepaint(element, NULL);
+	} else if (message == UI_MSG_LEFT_DOWN) {
+		UIElementAnimate(element, false);
+		scrollBar->lastAnimateTime = UI_CLOCK();
+	} else if (message == UI_MSG_LEFT_UP) {
+		UIElementAnimate(element, true);
+	} else if (message == UI_MSG_ANIMATE) {
+		UI_CLOCK_T previous = scrollBar->lastAnimateTime;
+		UI_CLOCK_T current = UI_CLOCK();
+		UI_CLOCK_T delta = current - previous;
+		double deltaSeconds = (double) delta / UI_CLOCKS_PER_SECOND;
+		if (deltaSeconds > 0.1) deltaSeconds = 0.1;
+		double deltaPixels = deltaSeconds * scrollBar->page * 3;
+		scrollBar->lastAnimateTime = current;
+		if (isDown) scrollBar->position += deltaPixels;
+		else scrollBar->position -= deltaPixels;
+		UIElementRefresh(&scrollBar->e);
+		UIElementMessage(scrollBar->e.parent, UI_MSG_SCROLLED, 0, 0);
+	}
+
+	return 0;
+}
+
+int _UIScrollThumbMessage(UIElement *element, UIMessage message, int di, void *dp) {
+	UIScrollBar *scrollBar = (UIScrollBar *) element->parent;
+
+	if (message == UI_MSG_PAINT) {
+		UIPainter *painter = (UIPainter *) dp;
+		uint32_t color = element == element->window->pressed ? ui.theme.buttonPressed 
+			: element == element->window->hovered ? ui.theme.buttonHovered : ui.theme.buttonNormal;
+		UIDrawRectangle(painter, element->bounds, color, ui.theme.border, UI_RECT_1(2));
+	} else if (message == UI_MSG_UPDATE) {
+		UIElementRepaint(element, NULL);
+	} else if (message == UI_MSG_MOUSE_DRAG && element->window->pressedButton == 1) {
+		if (!scrollBar->inDrag) {
+			scrollBar->inDrag = true;
+			
+			if (scrollBar->horizontal) {
+				scrollBar->dragOffset = element->bounds.l - scrollBar->e.bounds.l - element->window->cursorX;
+			} else {
+				scrollBar->dragOffset = element->bounds.t - scrollBar->e.bounds.t - element->window->cursorY;
+			}
+		}
+
+		int thumbPosition = (scrollBar->horizontal ? element->window->cursorX : element->window->cursorY) + scrollBar->dragOffset;
+		int size = scrollBar->horizontal ? (UI_RECT_WIDTH(scrollBar->e.bounds) - UI_RECT_WIDTH(element->bounds))
+				: (UI_RECT_HEIGHT(scrollBar->e.bounds) - UI_RECT_HEIGHT(element->bounds));
+		scrollBar->position = (double) thumbPosition / size * (scrollBar->maximum - scrollBar->page);
+		UIElementRefresh(&scrollBar->e);
+		UIElementMessage(scrollBar->e.parent, UI_MSG_SCROLLED, 0, 0);
+	} else if (message == UI_MSG_LEFT_UP) {
+		scrollBar->inDrag = false;
+	}
+
+	return 0;
+}
+
+UIScrollBar *UIScrollBarCreate(UIElement *parent, uint32_t flags) {
+	UIScrollBar *scrollBar = (UIScrollBar *) UIElementCreate(sizeof(UIScrollBar), parent, flags, _UIScrollBarMessage, "Scroll Bar");
+	bool horizontal = scrollBar->horizontal = flags & UI_SCROLL_BAR_HORIZONTAL;
+	UIElementCreate(sizeof(UIElement), &scrollBar->e, flags, _UIScrollUpDownMessage, !horizontal ? "Scroll Up" : "Scroll Left")->cp = (void *) (uintptr_t) 0;
+	UIElementCreate(sizeof(UIElement), &scrollBar->e, flags, _UIScrollThumbMessage, "Scroll Thumb");
+	UIElementCreate(sizeof(UIElement), &scrollBar->e, flags, _UIScrollUpDownMessage, !horizontal ? "Scroll Down" : "Scroll Right")->cp = (void *) (uintptr_t) 1;
+	return scrollBar;
+}
+
+bool _UICharIsAlpha(char c) {
+	return (c >= 'a' && c <= 'z') || (c >= 'A' && c <= 'Z');
+}
+
+bool _UICharIsDigit(char c) {
+	return c >= '0' && c <= '9';
+}
+
+bool _UICharIsAlphaOrDigitOrUnderscore(char c) {
+	return _UICharIsAlpha(c) || _UICharIsDigit(c) || c == '_';
+}
+
+int UICodeHitTest(UICode *code, int x, int y) {
+	x -= code->e.bounds.l;
+
+	if (x < 0 || x >= UI_RECT_WIDTH(code->e.bounds) - UI_SIZE_SCROLL_BAR * code->e.window->scale) {
+		return 0;
+	}
+
+	y -= code->e.bounds.t - code->vScroll->position;
+
+	UIFont *previousFont = UIFontActivate(code->font);
+	int lineHeight = UIMeasureStringHeight();
+	bool inMargin = x < UI_SIZE_CODE_MARGIN + UI_SIZE_CODE_MARGIN_GAP / 2 && (~code->e.flags & UI_CODE_NO_MARGIN);
+	UIFontActivate(previousFont);
+
+	if (y < 0 || y >= lineHeight * code->lineCount) {
+		return 0;
+	}
+
+	int line = y / lineHeight + 1;
+	return inMargin ? -line : line;
+}
+
+int UIDrawStringHighlighted(UIPainter *painter, UIRectangle lineBounds, const char *string, ptrdiff_t bytes, int tabSize) {
+	if (bytes == -1) bytes = _UIStringLength(string);
+	if (bytes > 10000) bytes = 10000;
+
+	uint32_t colors[] = {
+		ui.theme.codeDefault,
+		ui.theme.codeComment,
+		ui.theme.codeString,
+		ui.theme.codeNumber,
+		ui.theme.codeOperator,
+		ui.theme.codePreprocessor,
+	};
+
+	int x = lineBounds.l;
+	int y = (lineBounds.t + lineBounds.b - UIMeasureStringHeight()) / 2;
+	int ti = 0;
+	int lexState = 0;
+	bool inComment = false, inIdentifier = false, inChar = false, startedString = false;
+	uint32_t last = 0;
+
+	while (bytes--) {
+		char c = *string++;
+
+		last <<= 8;
+		last |= c;
+
+		if (lexState == 4) {
+			lexState = 0;
+		} else if (lexState == 1) {
+			if ((last & 0xFF0000) == ('*' << 16) && (last & 0xFF00) == ('/' << 8) && inComment) {
+				lexState = 0, inComment = false;
+			}
+		} else if (lexState == 3) {
+			if (!_UICharIsAlpha(c) && !_UICharIsDigit(c)) {
+				lexState = 0;
+			}
+		} else if (lexState == 2) {
+			if (!startedString) {
+				if (!inChar && ((last >> 8) & 0xFF) == '"' && ((last >> 16) & 0xFF) != '\\') {
+					lexState = 0;
+				} else if (inChar && ((last >> 8) & 0xFF) == '\'' && ((last >> 16) & 0xFF) != '\\') {
+					lexState = 0;
+				}
+			}
+
+			startedString = false;
+		}
+
+		if (lexState == 0) {
+			if (c == '#') {
+				lexState = 5;
+			} else if (c == '/' && *string == '/') {
+				lexState = 1;
+			} else if (c == '/' && *string == '*') {
+				lexState = 1, inComment = true;
+			} else if (c == '"') {
+				lexState = 2;
+				inChar = false;
+				startedString = true;
+			} else if (c == '\'') {
+				lexState = 2;
+				inChar = true;
+				startedString = true;
+			} else if (_UICharIsDigit(c) && !inIdentifier) {
+				lexState = 3;
+			} else if (!_UICharIsAlpha(c) && !_UICharIsDigit(c)) {
+				lexState = 4;
+				inIdentifier = false;
+			} else {
+				inIdentifier = true;
+			}
+		}
+
+		if (c == '\t') {
+			x += ui.activeFont->glyphWidth, ti++;
+			while (ti % tabSize) x += ui.activeFont->glyphWidth, ti++;
+		} else {
+			UIDrawGlyph(painter, x, y, c, colors[lexState]);
+			x += ui.activeFont->glyphWidth, ti++;
+		}
+	}
+
+	return x;
+}
+
+int _UICodeMessage(UIElement *element, UIMessage message, int di, void *dp) {
+	UICode *code = (UICode *) element;
+	
+	if (message == UI_MSG_LAYOUT) {
+		UIFont *previousFont = UIFontActivate(code->font);
+
+		if (code->moveScrollToFocusNextLayout) {
+			code->vScroll->position = (code->focused + 0.5) * UIMeasureStringHeight() - UI_RECT_HEIGHT(code->e.bounds) / 2;
+		}
+
+		UIRectangle scrollBarBounds = element->bounds;
+		scrollBarBounds.l = scrollBarBounds.r - UI_SIZE_SCROLL_BAR * code->e.window->scale;
+		code->vScroll->maximum = code->lineCount * UIMeasureStringHeight();
+		code->vScroll->page = UI_RECT_HEIGHT(element->bounds);
+		UIFontActivate(previousFont);
+		UIElementMove(&code->vScroll->e, scrollBarBounds, true);
+	} else if (message == UI_MSG_PAINT) {
+		UIFont *previousFont = UIFontActivate(code->font);
+
+		UIPainter *painter = (UIPainter *) dp;
+		UIRectangle lineBounds = element->bounds;
+		lineBounds.r -= UI_SIZE_SCROLL_BAR * code->e.window->scale;
+
+		if (~code->e.flags & UI_CODE_NO_MARGIN) {
+			lineBounds.l += UI_SIZE_CODE_MARGIN + UI_SIZE_CODE_MARGIN_GAP;
+		}
+
+		int lineHeight = UIMeasureStringHeight();
+		lineBounds.t -= (int64_t) code->vScroll->position % lineHeight;
+
+		UIDrawBlock(painter, element->bounds, ui.theme.codeBackground);
+
+		for (int i = code->vScroll->position / lineHeight; i < code->lineCount; i++) {
+			if (lineBounds.t > element->clip.b) {
+				break;
+			}
+
+			lineBounds.b = lineBounds.t + lineHeight;
+
+			if (~code->e.flags & UI_CODE_NO_MARGIN) {
+				char string[16];
+				int p = 16;
+				int lineNumber = i + 1;
+
+				while (lineNumber) {
+					string[--p] = (lineNumber % 10) + '0';
+					lineNumber /= 10;
+				}
+
+				UIRectangle marginBounds = lineBounds;
+				marginBounds.r = marginBounds.l - UI_SIZE_CODE_MARGIN_GAP;
+				marginBounds.l -= UI_SIZE_CODE_MARGIN + UI_SIZE_CODE_MARGIN_GAP;
+
+				uint32_t marginColor = UIElementMessage(element, UI_MSG_CODE_GET_MARGIN_COLOR, i + 1, 0);
+
+				if (marginColor) {
+					UIDrawBlock(painter, marginBounds, marginColor);
+				}
+
+				UIDrawString(painter, marginBounds, string + p, 16 - p, ui.theme.codeDefault, UI_ALIGN_RIGHT, NULL);
+			}
+
+			if (code->focused == i) {
+				UIDrawBlock(painter, lineBounds, ui.theme.codeFocused);
+			}
+
+			int x = UIDrawStringHighlighted(painter, lineBounds, code->content + code->lines[i].offset, code->lines[i].bytes, code->tabSize);
+			int y = (lineBounds.t + lineBounds.b - UIMeasureStringHeight()) / 2;
+
+			UICodeDecorateLine m = { 0 };
+			m.x = x, m.y = y, m.bounds = lineBounds, m.index = i + 1, m.painter = painter;
+			UIElementMessage(element, UI_MSG_CODE_DECORATE_LINE, 0, &m);
+
+			lineBounds.t += lineHeight;
+		}
+
+		UIFontActivate(previousFont);
+	} else if (message == UI_MSG_SCROLLED) {
+		code->moveScrollToFocusNextLayout = false;
+		UIElementRefresh(element);
+	} else if (message == UI_MSG_MOUSE_WHEEL) {
+		return UIElementMessage(&code->vScroll->e, message, di, dp);
+	} else if (message == UI_MSG_GET_CURSOR) {
+		if (UICodeHitTest(code, element->window->cursorX, element->window->cursorY) < 0) {
+			return UI_CURSOR_FLIPPED_ARROW;
+		}
+	} else if (message == UI_MSG_DESTROY) {
+		UI_FREE(code->content);
+		UI_FREE(code->lines);
+	}
+
+	return 0;
+}
+
+void UICodeFocusLine(UICode *code, int index) {
+	code->focused = index - 1;
+	code->moveScrollToFocusNextLayout = true;
+}
+
+void UICodeInsertContent(UICode *code, const char *content, ptrdiff_t byteCount, bool replace) {
+	UIFont *previousFont = UIFontActivate(code->font);
+
+	if (byteCount == -1) {
+		byteCount = _UIStringLength(content);
+	}
+
+	if (byteCount > 1000000000) {
+		byteCount = 1000000000;
+	}
+
+	if (replace) {
+		UI_FREE(code->content);
+		UI_FREE(code->lines);
+		code->content = NULL;
+		code->lines = NULL;
+		code->contentBytes = 0;
+		code->lineCount = 0;
+	}
+
+	code->content = (char *) UI_REALLOC(code->content, code->contentBytes + byteCount);
+
+	if (!byteCount) {
+		return;
+	}
+
+	int lineCount = content[byteCount - 1] != '\n';
+
+	for (int i = 0; i < byteCount; i++) {
+		code->content[i + code->contentBytes] = content[i];
+
+		if (content[i] == '\n') {
+			lineCount++;
+		}
+	}
+
+	code->lines = (UICodeLine *) UI_REALLOC(code->lines, sizeof(UICodeLine) * (code->lineCount + lineCount));
+	int offset = 0, lineIndex = 0;
+
+	for (intptr_t i = 0; i <= byteCount && lineIndex < lineCount; i++) {
+		if (content[i] == '\n' || i == byteCount) {
+			UICodeLine line = { 0 };
+			line.offset = offset + code->contentBytes;
+			line.bytes = i - offset;
+			code->lines[code->lineCount + lineIndex] = line;
+			lineIndex++;
+			offset = i + 1;
+		}
+	}
+
+	code->lineCount += lineCount;
+	code->contentBytes += byteCount;
+
+	if (!replace) {
+		code->vScroll->position = code->lineCount * UIMeasureStringHeight();
+	}
+
+	UIFontActivate(previousFont);
+}
+
+UICode *UICodeCreate(UIElement *parent, uint32_t flags) {
+	UICode *code = (UICode *) UIElementCreate(sizeof(UICode), parent, flags, _UICodeMessage, "Code");
+	code->font = ui.activeFont;
+	code->vScroll = UIScrollBarCreate(&code->e, 0);
+	code->focused = -1;
+	code->tabSize = 4;
+	return code;
+}
+
+int _UIGaugeMessage(UIElement *element, UIMessage message, int di, void *dp) {
+	UIGauge *gauge = (UIGauge *) element;
+
+	if (message == UI_MSG_GET_HEIGHT) {
+		return UI_SIZE_GAUGE_HEIGHT * element->window->scale;
+	} else if (message == UI_MSG_GET_WIDTH) {
+		return UI_SIZE_GAUGE_WIDTH * element->window->scale;
+	} else if (message == UI_MSG_PAINT) {
+		UIPainter *painter = (UIPainter *) dp;
+		UIDrawRectangle(painter, element->bounds, ui.theme.buttonNormal, ui.theme.border, UI_RECT_1(1));
+		UIRectangle filled = UIRectangleAdd(element->bounds, UI_RECT_1I(1));
+		filled.r = filled.l + UI_RECT_WIDTH(filled) * gauge->position;
+		UIDrawBlock(painter, filled, ui.theme.selected);
+	}
+
+	return 0;
+}
+
+UIGauge *UIGaugeCreate(UIElement *parent, uint32_t flags) {
+	return (UIGauge *) UIElementCreate(sizeof(UIGauge), parent, flags, _UIGaugeMessage, "Gauge");
+}
+
+int _UISliderMessage(UIElement *element, UIMessage message, int di, void *dp) {
+	UISlider *slider = (UISlider *) element;
+
+	if (message == UI_MSG_GET_HEIGHT) {
+		return UI_SIZE_SLIDER_HEIGHT * element->window->scale;
+	} else if (message == UI_MSG_GET_WIDTH) {
+		return UI_SIZE_SLIDER_WIDTH * element->window->scale;
+	} else if (message == UI_MSG_PAINT) {
+		UIPainter *painter = (UIPainter *) dp;
+		UIRectangle bounds = element->bounds;
+		int centerY = (bounds.t + bounds.b) / 2;
+		int trackSize = UI_SIZE_SLIDER_TRACK * element->window->scale;
+		int thumbSize = UI_SIZE_SLIDER_THUMB * element->window->scale;
+		int thumbPosition = (UI_RECT_WIDTH(bounds) - thumbSize) * slider->position;
+		UIRectangle track = UI_RECT_4(bounds.l, bounds.r, centerY - (trackSize + 1) / 2, centerY + trackSize / 2);
+		UIDrawRectangle(painter, track, ui.theme.buttonNormal, ui.theme.border, UI_RECT_1(1));
+		bool pressed = element == element->window->pressed;
+		bool hovered = element == element->window->hovered;
+		bool disabled = element->flags & UI_ELEMENT_DISABLED;
+		uint32_t color = disabled ? ui.theme.buttonDisabled : pressed ? ui.theme.buttonPressed : hovered ? ui.theme.buttonHovered : ui.theme.buttonNormal;
+		UIRectangle thumb = UI_RECT_4(bounds.l + thumbPosition, bounds.l + thumbPosition + thumbSize, centerY - (thumbSize + 1) / 2, centerY + thumbSize / 2);
+		UIDrawRectangle(painter, thumb, color, ui.theme.border, UI_RECT_1(1));
+	} else if (message == UI_MSG_LEFT_DOWN || (message == UI_MSG_MOUSE_DRAG && element->window->pressedButton == 1)) {
+		UIRectangle bounds = element->bounds;
+		int thumbSize = UI_SIZE_SLIDER_THUMB * element->window->scale;
+		slider->position = (float) (element->window->cursorX - thumbSize / 2 - bounds.l) / (UI_RECT_WIDTH(bounds) - thumbSize);
+		if (slider->steps > 1) slider->position = (int) (slider->position * (slider->steps - 1) + 0.5f) / (float) (slider->steps - 1);
+		if (slider->position < 0) slider->position = 0;
+		if (slider->position > 1) slider->position = 1;
+		UIElementMessage(element, UI_MSG_VALUE_CHANGED, 0, 0);
+		UIElementRepaint(element, NULL);
+	} else if (message == UI_MSG_UPDATE) {
+		UIElementRepaint(element, NULL);
+	}
+
+	return 0;
+}
+
+UISlider *UISliderCreate(UIElement *parent, uint32_t flags) {
+	return (UISlider *) UIElementCreate(sizeof(UISlider), parent, flags, _UISliderMessage, "Slider");
+}
+
+int UITableHitTest(UITable *table, int x, int y) {
+	x -= table->e.bounds.l;
+
+	if (x < 0 || x >= UI_RECT_WIDTH(table->e.bounds) - UI_SIZE_SCROLL_BAR * table->e.window->scale) {
+		return -1;
+	}
+
+	y -= (table->e.bounds.t + UI_SIZE_TABLE_HEADER * table->e.window->scale) - table->vScroll->position;
+
+	int rowHeight = UI_SIZE_TABLE_ROW * table->e.window->scale;
+
+	if (y < 0 || y >= rowHeight * table->itemCount) {
+		return -1;
+	}
+
+	return y / rowHeight;
+}
+
+int UITableHeaderHitTest(UITable *table, int x, int y) {
+	if (!table->columnCount) return -1;
+	UIRectangle header = table->e.bounds;
+	header.b = header.t + UI_SIZE_TABLE_HEADER * table->e.window->scale;
+	header.l += UI_SIZE_TABLE_COLUMN_GAP * table->e.window->scale;
+	int position = 0, index = 0;
+
+	while (true) {
+		int end = position;
+		for (; table->columns[end] != '\t' && table->columns[end]; end++);
+		header.r = header.l + table->columnWidths[index];
+		if (UIRectangleContains(header, x, y)) return index;
+		header.l += table->columnWidths[index] + UI_SIZE_TABLE_COLUMN_GAP * table->e.window->scale;
+		if (table->columns[end] != '\t') break;
+		position = end + 1, index++;
+	}
+
+	return -1;
+}
+
+bool UITableEnsureVisible(UITable *table, int index) {
+	int rowHeight = UI_SIZE_TABLE_ROW * table->e.window->scale;
+	int y = index * rowHeight;
+	y -= table->vScroll->position;
+	int height = UI_RECT_HEIGHT(table->e.bounds) - UI_SIZE_TABLE_HEADER * table->e.window->scale - rowHeight;
+
+	if (y < 0) {
+		table->vScroll->position += y;
+		UIElementRefresh(&table->e);
+		return true;
+	} else if (y > height) {
+		table->vScroll->position -= height - y;
+		UIElementRefresh(&table->e);
+		return true;
+	} else {
+		return false;
+	}
+}
+
+void UITableResizeColumns(UITable *table) {
+	int position = 0;
+	int count = 0;
+
+	while (true) {
+		int end = position;
+		for (; table->columns[end] != '\t' && table->columns[end]; end++);
+		count++;
+		if (table->columns[end] == '\t') position = end + 1;
+		else break;
+	}
+
+	UI_FREE(table->columnWidths);
+	table->columnWidths = (int *) UI_MALLOC(count * sizeof(int));
+	table->columnCount = count;
+
+	position = 0;
+
+	char buffer[256];
+	UITableGetItem m = { 0 };
+	m.buffer = buffer;
+	m.bufferBytes = sizeof(buffer);
+
+	while (true) {
+		int end = position;
+		for (; table->columns[end] != '\t' && table->columns[end]; end++);
+
+		int longest = UIMeasureStringWidth(table->columns + position, end - position);
+
+		for (int i = 0; i < table->itemCount; i++) {
+			m.index = i;
+			int bytes = UIElementMessage(&table->e, UI_MSG_TABLE_GET_ITEM, 0, &m);
+			int width = UIMeasureStringWidth(buffer, bytes);
+
+			if (width > longest) {
+				longest = width;
+			}
+		}
+
+		table->columnWidths[m.column] = longest;
+		m.column++;
+		if (table->columns[end] == '\t') position = end + 1;
+		else break;
+	}
+}
+
+int _UITableMessage(UIElement *element, UIMessage message, int di, void *dp) {
+	UITable *table = (UITable *) element;
+
+	if (message == UI_MSG_PAINT) {
+		UIPainter *painter = (UIPainter *) dp;
+		UIRectangle bounds = element->bounds;
+		bounds.r -= UI_SIZE_SCROLL_BAR * element->window->scale;
+		UIDrawBlock(painter, bounds, ui.theme.panel2);
+		char buffer[256];
+		UIRectangle row = bounds;
+		int rowHeight = UI_SIZE_TABLE_ROW * element->window->scale;
+		UITableGetItem m = { 0 };
+		m.buffer = buffer;
+		m.bufferBytes = sizeof(buffer);
+		row.t += UI_SIZE_TABLE_HEADER * table->e.window->scale;
+		row.t -= (int64_t) table->vScroll->position % rowHeight;
+		int hovered = UITableHitTest(table, element->window->cursorX, element->window->cursorY);
+
+		for (int i = table->vScroll->position / rowHeight; i < table->itemCount; i++) {
+			if (row.t > element->clip.b) {
+				break;
+			}
+			
+			row.b = row.t + rowHeight;
+			m.index = i;
+			m.isSelected = false;
+			m.column = 0;
+			int bytes = UIElementMessage(element, UI_MSG_TABLE_GET_ITEM, 0, &m);
+			uint32_t textColor = ui.theme.text;
+
+			if (m.isSelected) {
+				UIDrawBlock(painter, row, ui.theme.selected);
+				textColor = ui.theme.textSelected;
+			} else if (hovered == i) {
+				UIDrawBlock(painter, row, ui.theme.buttonHovered);
+			}
+
+			UIRectangle cell = row;
+			cell.l += UI_SIZE_TABLE_COLUMN_GAP * table->e.window->scale;
+
+			for (int j = 0; j < table->columnCount; j++) {
+				if (j) {
+					m.column = j;
+					bytes = UIElementMessage(element, UI_MSG_TABLE_GET_ITEM, 0, &m);
+				}
+
+				cell.r = cell.l + table->columnWidths[j];
+				if ((size_t) bytes > m.bufferBytes && bytes > 0) bytes = m.bufferBytes;
+				UIDrawString(painter, cell, buffer, bytes, textColor, UI_ALIGN_LEFT, NULL);
+				cell.l += table->columnWidths[j] + UI_SIZE_TABLE_COLUMN_GAP * table->e.window->scale;
+			}
+
+			row.t += rowHeight;
+		}
+
+		UIRectangle header = bounds;
+		header.b = header.t + UI_SIZE_TABLE_HEADER * table->e.window->scale;
+		UIDrawRectangle(painter, header, ui.theme.panel1, ui.theme.border, UI_RECT_4(0, 0, 0, 1));
+		header.l += UI_SIZE_TABLE_COLUMN_GAP * table->e.window->scale;
+
+		int position = 0;
+		int index = 0;
+
+		if (table->columnCount) {
+			while (true) {
+				int end = position;
+				for (; table->columns[end] != '\t' && table->columns[end]; end++);
+
+				header.r = header.l + table->columnWidths[index];
+				UIDrawString(painter, header, table->columns + position, end - position, ui.theme.text, UI_ALIGN_LEFT, NULL);
+				if (index == table->columnHighlight) UIDrawInvert(painter, header);
+				header.l += table->columnWidths[index] + UI_SIZE_TABLE_COLUMN_GAP * table->e.window->scale;
+
+				if (table->columns[end] == '\t') {
+					position = end + 1;
+					index++;
+				} else {
+					break;
+				}
+			}
+		}
+	} else if (message == UI_MSG_LAYOUT) {
+		UIRectangle scrollBarBounds = element->bounds;
+		scrollBarBounds.l = scrollBarBounds.r - UI_SIZE_SCROLL_BAR * element->window->scale;
+		table->vScroll->maximum = table->itemCount * UI_SIZE_TABLE_ROW * element->window->scale;
+		table->vScroll->page = UI_RECT_HEIGHT(element->bounds) - UI_SIZE_TABLE_HEADER * table->e.window->scale;
+		UIElementMove(&table->vScroll->e, scrollBarBounds, true);
+	} else if (message == UI_MSG_MOUSE_MOVE || message == UI_MSG_UPDATE) {
+		UIElementRepaint(element, NULL);
+	} else if (message == UI_MSG_SCROLLED) {
+		UIElementRefresh(element);
+	} else if (message == UI_MSG_MOUSE_WHEEL) {
+		return UIElementMessage(&table->vScroll->e, message, di, dp);
+	} else if (message == UI_MSG_DESTROY) {
+		UI_FREE(table->columns);
+		UI_FREE(table->columnWidths);
+	}
+
+	return 0;
+}
+
+UITable *UITableCreate(UIElement *parent, uint32_t flags, const char *columns) {
+	UITable *table = (UITable *) UIElementCreate(sizeof(UITable), parent, flags, _UITableMessage, "Table");
+	table->vScroll = UIScrollBarCreate(&table->e, 0);
+	table->columns = UIStringCopy(columns, -1);
+	table->columnHighlight = -1;
+	return table;
+}
+
+void UITextboxReplace(UITextbox *textbox, const char *text, ptrdiff_t bytes, bool sendChangedMessage) {
+	if (bytes == -1) {
+		bytes = _UIStringLength(text);
+	}
+
+	int deleteFrom = textbox->carets[0], deleteTo = textbox->carets[1];
+
+	if (deleteFrom > deleteTo) {
+		UI_SWAP(int, deleteFrom, deleteTo);
+	}
+
+	for (int i = deleteTo; i < textbox->bytes; i++) {
+		textbox->string[i - deleteTo + deleteFrom] = textbox->string[i];
+	}
+
+	textbox->bytes -= deleteTo - deleteFrom;
+	textbox->carets[0] = textbox->carets[1] = deleteFrom;
+
+	textbox->string = (char *) UI_REALLOC(textbox->string, textbox->bytes + bytes);
+
+	for (int i = textbox->bytes + bytes - 1; i >= textbox->carets[0] + bytes; i--) {
+		textbox->string[i] = textbox->string[i - bytes];
+	}
+
+	for (int i = textbox->carets[0]; i < textbox->carets[0] + bytes; i++) {
+		textbox->string[i] = text[i - textbox->carets[0]];
+	}
+
+	textbox->bytes += bytes;
+	textbox->carets[0] += bytes;
+	textbox->carets[1] = textbox->carets[0];
+
+	if (sendChangedMessage) {
+		UIElementMessage(&textbox->e, UI_MSG_VALUE_CHANGED, 0, 0);
+	}
+
+	textbox->e.window->textboxModifiedFlag = true;
+}
+
+void UITextboxClear(UITextbox *textbox, bool sendChangedMessage) {
+	textbox->carets[1] = 0;
+	textbox->carets[0] = textbox->bytes;
+	UITextboxReplace(textbox, "", 0, sendChangedMessage);
+}
+
+void UITextboxMoveCaret(UITextbox *textbox, bool backward, bool word) {
+	while (true) {
+		if (textbox->carets[0] > 0 && backward) {
+			textbox->carets[0]--;
+		} else if (textbox->carets[0] < textbox->bytes && !backward) {
+			textbox->carets[0]++;
+		} else {
+			return;
+		}
+
+		if (!word) {
+			return;
+		} else if (textbox->carets[0] != textbox->bytes && textbox->carets[0] != 0) {
+			char c1 = textbox->string[textbox->carets[0] - 1];
+			char c2 = textbox->string[textbox->carets[0]];
+
+			if (_UICharIsAlphaOrDigitOrUnderscore(c1) != _UICharIsAlphaOrDigitOrUnderscore(c2)) {
+				return;
+			}
+		}
+	}
+}
+
+int _UITextboxMessage(UIElement *element, UIMessage message, int di, void *dp) {
+	UITextbox *textbox = (UITextbox *) element;
+
+	if (message == UI_MSG_GET_HEIGHT) {
+		return UI_SIZE_TEXTBOX_HEIGHT * element->window->scale;
+	} else if (message == UI_MSG_GET_WIDTH) {
+		return UI_SIZE_TEXTBOX_WIDTH * element->window->scale;
+	} else if (message == UI_MSG_PAINT) {
+		int scaledMargin = UI_SIZE_TEXTBOX_MARGIN * element->window->scale;
+		int totalWidth = UIMeasureStringWidth(textbox->string, textbox->bytes) + scaledMargin * 2;
+		UIRectangle textBounds = UIRectangleAdd(element->bounds, UI_RECT_1I(scaledMargin));
+
+		if (textbox->scroll > totalWidth - UI_RECT_WIDTH(textBounds)) {
+			textbox->scroll = totalWidth - UI_RECT_WIDTH(textBounds);
+		}
+
+		if (textbox->scroll < 0) {
+			textbox->scroll = 0;
+		}
+
+		int caretX = UIMeasureStringWidth(textbox->string, textbox->carets[0]) - textbox->scroll;
+
+		if (caretX < 0) {
+			textbox->scroll = caretX + textbox->scroll;
+		} else if (caretX > UI_RECT_WIDTH(textBounds)) {
+			textbox->scroll = caretX - UI_RECT_WIDTH(textBounds) + textbox->scroll + 1;
+		}
+
+		UIPainter *painter = (UIPainter *) dp;
+		bool focused = element->window->focused == element;
+		bool disabled = element->flags & UI_ELEMENT_DISABLED;
+		UIDrawRectangle(painter, element->bounds, 
+			disabled ? ui.theme.buttonDisabled : focused ? ui.theme.textboxFocused : ui.theme.textboxNormal, 
+			ui.theme.border, UI_RECT_1(1));
+#ifdef __cplusplus
+		UIStringSelection selection = {};
+#else
+		UIStringSelection selection = { 0 };
+#endif
+		selection.carets[0] = textbox->carets[0];
+		selection.carets[1] = textbox->carets[1];
+		selection.colorBackground = ui.theme.selected;
+		selection.colorText = ui.theme.textSelected;
+		textBounds.l -= textbox->scroll;
+		UIDrawString(painter, textBounds, textbox->string, textbox->bytes, 
+			disabled ? ui.theme.textDisabled : ui.theme.text, UI_ALIGN_LEFT, focused ? &selection : NULL);
+	} else if (message == UI_MSG_GET_CURSOR) {
+		return UI_CURSOR_TEXT;
+	} else if (message == UI_MSG_LEFT_DOWN) {
+		UIElementFocus(element);
+	} else if (message == UI_MSG_UPDATE) {
+		UIElementRepaint(element, NULL);
+	} else if (message == UI_MSG_DESTROY) {
+		UI_FREE(textbox->string);
+	} else if (message == UI_MSG_KEY_TYPED) {
+		UIKeyTyped *m = (UIKeyTyped *) dp;
+		bool handled = true;
+
+		if (textbox->rejectNextKey) {
+			textbox->rejectNextKey = false;
+			handled = false;
+		} else if (m->code == UI_KEYCODE_BACKSPACE || m->code == UI_KEYCODE_DELETE) {
+			if (textbox->carets[0] == textbox->carets[1]) {
+				UITextboxMoveCaret(textbox, m->code == UI_KEYCODE_BACKSPACE, element->window->ctrl);
+			}
+
+			UITextboxReplace(textbox, NULL, 0, true);
+		} else if (m->code == UI_KEYCODE_LEFT || m->code == UI_KEYCODE_RIGHT) {
+			UITextboxMoveCaret(textbox, m->code == UI_KEYCODE_LEFT, element->window->ctrl);
+
+			if (!element->window->shift) {
+				textbox->carets[1] = textbox->carets[0];
+			}
+		} else if (m->code == UI_KEYCODE_HOME || m->code == UI_KEYCODE_END) {
+			if (m->code == UI_KEYCODE_HOME) {
+				textbox->carets[0] = 0;
+			} else {
+				textbox->carets[0] = textbox->bytes;
+			}
+
+			if (!element->window->shift) {
+				textbox->carets[1] = textbox->carets[0];
+			}
+		} else if (m->code == UI_KEYCODE_LETTER('A') && element->window->ctrl) {
+			textbox->carets[1] = 0;
+			textbox->carets[0] = textbox->bytes;
+		} else if (m->textBytes && !element->window->alt && !element->window->ctrl && m->text[0] >= 0x20) {
+			UITextboxReplace(textbox, m->text, m->textBytes, true);
+		} else if ((m->code == UI_KEYCODE_LETTER('C') || m->code == UI_KEYCODE_LETTER('X') || m->code == UI_KEYCODE_INSERT) 
+				&& element->window->ctrl && !element->window->alt && !element->window->shift) {
+			int   to = textbox->carets[0] > textbox->carets[1] ? textbox->carets[0] : textbox->carets[1];
+			int from = textbox->carets[0] < textbox->carets[1] ? textbox->carets[0] : textbox->carets[1];
+
+			if (from != to) {
+				char *pasteText = (char *) UI_CALLOC(to - from + 1);
+				for (int i = from; i < to; i++) pasteText[i - from] = textbox->string[i];
+				_UIClipboardWriteText(element->window, pasteText);
+			}
+			
+			if (m->code == UI_KEYCODE_LETTER('X')) {
+				UITextboxReplace(textbox, NULL, 0, true);
+			}
+		} else if ((m->code == UI_KEYCODE_LETTER('V') && element->window->ctrl && !element->window->alt && !element->window->shift)
+				|| (m->code == UI_KEYCODE_INSERT && !element->window->ctrl && !element->window->alt && element->window->shift)) {
+			size_t bytes;
+			char *text = _UIClipboardReadTextStart(element->window, &bytes);
+			if (text) UITextboxReplace(textbox, text, bytes, true);
+			_UIClipboardReadTextEnd(element->window, text);
+		} else {
+			handled = false;
+		}
+
+		if (handled) {
+			UIElementRepaint(element, NULL);
+			return 1;
+		}
+	}
+
+	return 0;
+}
+
+UITextbox *UITextboxCreate(UIElement *parent, uint32_t flags) {
+	return (UITextbox *) UIElementCreate(sizeof(UITextbox), parent, flags | UI_ELEMENT_TAB_STOP, _UITextboxMessage, "Textbox");
+}
+
+int _UIColorCircleMessage(UIElement *element, UIMessage message, int di, void *dp) {
+	UIColorPicker *colorPicker = (UIColorPicker *) element->parent;
+
+	if (message == UI_MSG_PAINT) {
+		UIPainter *painter = (UIPainter *) dp;
+
+		int startY = element->bounds.t, endY = element->bounds.b;
+		int startX = element->bounds.l, endX = element->bounds.r;
+		int size = endY - startY;
+
+		for (int i = startY; i < endY; i++) {
+			uint32_t *out = painter->bits + i * painter->width + startX;
+			int j = startX;
+			float y0 = i - startY - size / 2, x0 = -size / 2;
+			float angle = _UIArcTan2Float((i - startY) * 2.0f / size - 1, -1);
+
+			do {
+				float distanceFromCenterSquared = x0 * x0 + y0 * y0;
+				float hue = (angle + 3.14159f) * 0.954929658f;
+				float saturation = _UISquareRootFloat(distanceFromCenterSquared * 4.0f / size / size);
+
+				if (saturation <= 1 && UIRectangleContains(painter->clip, j, i)) {
+					UIColorToRGB(hue, saturation, colorPicker->value, out);
+					*out |= 0xFF000000;
+				}
+
+				out++, j++, x0++;
+
+				if (distanceFromCenterSquared) {
+					angle -= y0 / distanceFromCenterSquared;
+				} else {
+					angle = _UIArcTan2Float((i - startY) * 2.0f / size - 1, 0.01f);
+				}
+			} while (j < endX);
+		}
+
+		float angle = colorPicker->hue / 0.954929658f - 3.14159f;
+		float radius = colorPicker->saturation * size / 2;
+		int cx = (startX + endX) / 2 + radius * _UICosFloat(angle);
+		int cy = (startY + endY) / 2 + radius * _UISinFloat(angle);
+		UIDrawInvert(painter, UI_RECT_4(cx - 1, cx + 1, startY, endY));
+		UIDrawInvert(painter, UI_RECT_4(startX, endX, cy - 1, cy + 1));
+	} else if (message == UI_MSG_GET_CURSOR) {
+		return UI_CURSOR_CROSS_HAIR;
+	} else if (message == UI_MSG_LEFT_DOWN || message == UI_MSG_MOUSE_DRAG) {
+		int startY = element->bounds.t, endY = element->bounds.b, cursorY = element->window->cursorY;
+		int startX = element->bounds.l, endX = element->bounds.r, cursorX = element->window->cursorX;
+		int dx = (startX + endX) / 2, dy = (startY + endY) / 2;
+		int size = endY - startY;
+
+		float angle = _UIArcTan2Float((cursorY - startY) * 2.0f / size - 1, (cursorX - startX) * 2.0f / size - 1);
+		float distanceFromCenterSquared = (cursorX - dx) * (cursorX - dx) + (cursorY - dy) * (cursorY - dy);
+		colorPicker->hue = (angle + 3.14159f) * 0.954929658f;
+		colorPicker->saturation = _UISquareRootFloat(distanceFromCenterSquared * 4.0f / size / size);;
+		if (colorPicker->saturation > 1) colorPicker->saturation = 1;
+
+		UIElementMessage(&colorPicker->e, UI_MSG_VALUE_CHANGED, 0, 0);
+		UIElementRepaint(&colorPicker->e, NULL);
+	}
+
+	return 0;
+}
+
+int _UIColorSliderMessage(UIElement *element, UIMessage message, int di, void *dp) {
+	UIColorPicker *colorPicker = (UIColorPicker *) element->parent;
+	float opacitySlider = element->flags & 1;
+
+	if (message == UI_MSG_PAINT) {
+		UIPainter *painter = (UIPainter *) dp;
+
+		int startY = element->bounds.t, endY = element->bounds.b;
+		int startX = element->bounds.l, endX = element->bounds.r;
+		int size = endY - startY;
+
+		for (int i = startY; i < endY; i++) {
+			if (i < painter->clip.t || i >= painter->clip.b) continue;
+			uint32_t *out = painter->bits + i * painter->width + startX;
+			int j = element->clip.l;
+			uint32_t color;
+			float p = 1.0f - (float) (i - startY) / size;
+
+			if (opacitySlider) {
+				UIColorToRGB(colorPicker->hue, colorPicker->saturation, colorPicker->value, &color);
+				color = UI_COLOR_FROM_FLOAT(p * (UI_COLOR_RED_F(color) - 0.5f) + 0.5f, 
+					p * (UI_COLOR_GREEN_F(color) - 0.5f) + 0.5f, 
+					p * (UI_COLOR_BLUE_F(color) - 0.5f) + 0.5f);
+			} else {
+				UIColorToRGB(colorPicker->hue, colorPicker->saturation, p, &color);
+			}
+
+			color |= 0xFF000000;
+
+			do {
+				*out = color;
+				out++, j++;
+			} while (j < element->clip.r);
+		}
+
+		int cy = (size - 1) * (1 - (opacitySlider ? colorPicker->opacity : colorPicker->value)) + startY;
+		UIDrawInvert(painter, UI_RECT_4(startX, endX, cy - 1, cy + 1));
+	} else if (message == UI_MSG_GET_CURSOR) {
+		return UI_CURSOR_CROSS_HAIR;
+	} else if (message == UI_MSG_LEFT_DOWN || message == UI_MSG_MOUSE_DRAG) {
+		int startY = element->bounds.t, endY = element->bounds.b, cursorY = element->window->cursorY;
+		float *value = opacitySlider ? &colorPicker->opacity : &colorPicker->value;
+		*value = 1 - (float) (cursorY - startY) / (endY - startY);
+		if (*value < 0) *value = 0;
+		if (*value > 1) *value = 1;
+		UIElementMessage(&colorPicker->e, UI_MSG_VALUE_CHANGED, 0, 0);
+		UIElementRepaint(&colorPicker->e, NULL);
+	}
+
+	return 0;
+}
+
+int _UIColorPickerMessage(UIElement *element, UIMessage message, int di, void *dp) {
+	bool hasOpacity = element->flags & UI_COLOR_PICKER_HAS_OPACITY;
+
+	if (message == UI_MSG_GET_WIDTH) {
+		return (hasOpacity ? 280 : 240) * element->window->scale;
+	} else if (message == UI_MSG_GET_HEIGHT) {
+		return 200 * element->window->scale;
+	} else if (message == UI_MSG_LAYOUT) {
+		UIRectangle bounds = element->bounds;
+
+		int sliderSize = 35 * element->window->scale;
+		int gap = 5 * element->window->scale;
+
+		if (hasOpacity) {
+			UIElementMove(element->children, UI_RECT_4(bounds.l, bounds.r - (sliderSize + gap) * 2, bounds.t, bounds.b), false);
+			UIElementMove(element->children->next, UI_RECT_4(bounds.r - sliderSize * 2 - gap, bounds.r - sliderSize - gap, bounds.t, bounds.b), false);
+			UIElementMove(element->children->next->next, UI_RECT_4(bounds.r - sliderSize, bounds.r, bounds.t, bounds.b), false);
+		} else {
+			UIElementMove(element->children, UI_RECT_4(bounds.l, bounds.r - sliderSize - gap, bounds.t, bounds.b), false);
+			UIElementMove(element->children->next, UI_RECT_4(bounds.r - sliderSize, bounds.r, bounds.t, bounds.b), false);
+		}
+	}
+
+	return 0;
+}
+
+UIColorPicker *UIColorPickerCreate(UIElement *parent, uint32_t flags) {
+	UIColorPicker *colorPicker = (UIColorPicker *) UIElementCreate(sizeof(UIColorPicker), parent, flags, _UIColorPickerMessage, "ColorPicker");
+	UIElementCreate(sizeof(UIElement), &colorPicker->e, 0, _UIColorCircleMessage, "ColorCircle");
+	UIElementCreate(sizeof(UIElement), &colorPicker->e, 0, _UIColorSliderMessage, "ColorSlider");
+
+	if (flags & UI_COLOR_PICKER_HAS_OPACITY) {
+		UIElementCreate(sizeof(UIElement), &colorPicker->e, 1, _UIColorSliderMessage, "ColorSlider");
+	}
+
+	return colorPicker;
+}
+
+#define UI_MDI_CHILD_CALCULATE_LAYOUT() \
+	int titleSize = UI_SIZE_MDI_CHILD_TITLE * element->window->scale; \
+	int borderSize = UI_SIZE_MDI_CHILD_BORDER * element->window->scale; \
+	UIRectangle title = UIRectangleAdd(element->bounds, UI_RECT_4(borderSize, -borderSize, 0, 0)); \
+	title.b = title.t + titleSize; \
+	UIRectangle content = UIRectangleAdd(element->bounds, UI_RECT_4(borderSize, -borderSize, titleSize, -borderSize));
+
+int _UIMDIChildHitTest(UIMDIChild *mdiChild, int x, int y) {
+	UIElement *element = &mdiChild->e;
+	UI_MDI_CHILD_CALCULATE_LAYOUT();
+	int cornerSize = UI_SIZE_MDI_CHILD_CORNER * element->window->scale;
+	if (!UIRectangleContains(element->bounds, x, y) || UIRectangleContains(content, x, y)) return -1;
+	else if (x < element->bounds.l + cornerSize && y < element->bounds.t + cornerSize) return 0b1010;
+	else if (x > element->bounds.r - cornerSize && y < element->bounds.t + cornerSize) return 0b0110;
+	else if (x < element->bounds.l + cornerSize && y > element->bounds.b - cornerSize) return 0b1001;
+	else if (x > element->bounds.r - cornerSize && y > element->bounds.b - cornerSize) return 0b0101;
+	else if (x < element->bounds.l + borderSize) return 0b1000;
+	else if (x > element->bounds.r - borderSize) return 0b0100;
+	else if (y < element->bounds.t + borderSize) return 0b0010;
+	else if (y > element->bounds.b - borderSize) return 0b0001;
+	else if (UIRectangleContains(title, x, y)) return 0b1111;
+	else return -1;
+}
+
+void _UIMDIChildCloseButton(void *_child) {
+	UIElement *child = (UIElement *) _child;
+	
+	if (!UIElementMessage(child, UI_MSG_WINDOW_CLOSE, 0, 0)) {
+		UIElementDestroy(child);
+		UIElementRefresh(child->parent);
+	}
+}
+
+int _UIMDIChildMessage(UIElement *element, UIMessage message, int di, void *dp) {
+	UIMDIChild *mdiChild = (UIMDIChild *) element;
+
+	if (message == UI_MSG_PAINT) {
+		UI_MDI_CHILD_CALCULATE_LAYOUT();
+		UIPainter *painter = (UIPainter *) dp;
+		UIRectangle borders = UI_RECT_4(borderSize, borderSize, titleSize, borderSize);
+		UIDrawBorder(painter, element->bounds, ui.theme.buttonNormal, borders);
+		UIDrawBorder(painter, element->bounds, ui.theme.border, UI_RECT_1((int) element->window->scale));
+		UIDrawBorder(painter, UIRectangleAdd(content, UI_RECT_1I(-1)), ui.theme.border, UI_RECT_1((int) element->window->scale));
+		UIDrawString(painter, title, mdiChild->title, mdiChild->titleBytes, ui.theme.text, UI_ALIGN_LEFT, NULL);
+	} else if (message == UI_MSG_GET_WIDTH) {
+		UIElement *child = element->children;
+		while (child && child->next) child = child->next;
+		int width = 2 * UI_SIZE_MDI_CHILD_BORDER;
+		width += (child ? UIElementMessage(child, message, di ? (di - UI_SIZE_MDI_CHILD_TITLE + UI_SIZE_MDI_CHILD_BORDER) : 0, dp) : 0);
+		if (width < UI_SIZE_MDI_CHILD_MINIMUM_WIDTH) width = UI_SIZE_MDI_CHILD_MINIMUM_WIDTH;
+		return width;
+	} else if (message == UI_MSG_GET_HEIGHT) {
+		UIElement *child = element->children;
+		while (child && child->next) child = child->next;
+		int height = UI_SIZE_MDI_CHILD_TITLE + UI_SIZE_MDI_CHILD_BORDER;
+		height += (child ? UIElementMessage(child, message, di ? (di - 2 * UI_SIZE_MDI_CHILD_BORDER) : 0, dp) : 0);
+		if (height < UI_SIZE_MDI_CHILD_MINIMUM_HEIGHT) height = UI_SIZE_MDI_CHILD_MINIMUM_HEIGHT;
+		return height;
+	} else if (message == UI_MSG_LAYOUT) {
+		UI_MDI_CHILD_CALCULATE_LAYOUT();
+
+		UIElement *child = element->children;
+		int position = title.r;
+
+		while (child && child->next) {
+			int width = UIElementMessage(child, UI_MSG_GET_WIDTH, 0, 0);
+			UIElementMove(child, UI_RECT_4(position - width, position, title.t, title.b), false);
+			position -= width, child = child->next;
+		}
+
+		if (child) {
+			UIElementMove(child, content, false);
+		}
+	} else if (message == UI_MSG_GET_CURSOR) {
+		int hitTest = _UIMDIChildHitTest(mdiChild, element->window->cursorX, element->window->cursorY);
+		if (hitTest == 0b1000) return UI_CURSOR_RESIZE_LEFT;
+		if (hitTest == 0b0010) return UI_CURSOR_RESIZE_UP;
+		if (hitTest == 0b0110) return UI_CURSOR_RESIZE_UP_RIGHT;
+		if (hitTest == 0b1010) return UI_CURSOR_RESIZE_UP_LEFT;
+		if (hitTest == 0b0100) return UI_CURSOR_RESIZE_RIGHT;
+		if (hitTest == 0b0001) return UI_CURSOR_RESIZE_DOWN;
+		if (hitTest == 0b1001) return UI_CURSOR_RESIZE_DOWN_LEFT;
+		if (hitTest == 0b0101) return UI_CURSOR_RESIZE_DOWN_RIGHT;
+		return UI_CURSOR_ARROW;
+	} else if (message == UI_MSG_LEFT_DOWN) {
+		mdiChild->dragHitTest = _UIMDIChildHitTest(mdiChild, element->window->cursorX, element->window->cursorY);
+		mdiChild->dragOffset = UIRectangleAdd(element->bounds, UI_RECT_2(-element->window->cursorX, -element->window->cursorY));
+	} else if (message == UI_MSG_LEFT_UP) {
+		if (mdiChild->bounds.l < 0) mdiChild->bounds.r -= mdiChild->bounds.l, mdiChild->bounds.l = 0;
+		if (mdiChild->bounds.t < 0) mdiChild->bounds.b -= mdiChild->bounds.t, mdiChild->bounds.t = 0;
+		UIElementRefresh(element->parent);
+	} else if (message == UI_MSG_MOUSE_DRAG) {
+		if (mdiChild->dragHitTest > 0) {
+#define _UI_MDI_CHILD_MOVE_EDGE(bit, edge, cursor, size, opposite, negate, minimum, offset) \
+	if (mdiChild->dragHitTest & bit) mdiChild->bounds.edge = mdiChild->dragOffset.edge + element->window->cursor - element->parent->bounds.offset; \
+	if ((mdiChild->dragHitTest & bit) && size(mdiChild->bounds) < minimum) mdiChild->bounds.edge = mdiChild->bounds.opposite negate minimum;
+			_UI_MDI_CHILD_MOVE_EDGE(0b1000, l, cursorX, UI_RECT_WIDTH, r, -, UI_SIZE_MDI_CHILD_MINIMUM_WIDTH, l);
+			_UI_MDI_CHILD_MOVE_EDGE(0b0100, r, cursorX, UI_RECT_WIDTH, l, +, UI_SIZE_MDI_CHILD_MINIMUM_WIDTH, l);
+			_UI_MDI_CHILD_MOVE_EDGE(0b0010, t, cursorY, UI_RECT_HEIGHT, b, -, UI_SIZE_MDI_CHILD_MINIMUM_HEIGHT, t);
+			_UI_MDI_CHILD_MOVE_EDGE(0b0001, b, cursorY, UI_RECT_HEIGHT, t, +, UI_SIZE_MDI_CHILD_MINIMUM_HEIGHT, t);
+			UIElementRefresh(element->parent);
+		}
+	} else if (message == UI_MSG_DESTROY) {
+		UI_FREE(mdiChild->title);
+		UIMDIClient *client = (UIMDIClient *) element->parent;
+		if (client->e.children == element) client->e.children = element->next;
+		if (mdiChild->previous) mdiChild->previous->e.next = element->next;
+		if (element->next) ((UIMDIChild *) element->next)->previous = mdiChild->previous;
+		if (client->active == mdiChild) client->active = mdiChild->previous;
+	}
+
+	return 0;
+}
+
+int _UIMDIClientMessage(UIElement *element, UIMessage message, int di, void *dp) {
+	UIMDIClient *client = (UIMDIClient *) element;
+
+	if (message == UI_MSG_PAINT) {
+		UIDrawBlock((UIPainter *) dp, element->bounds, (element->flags & UI_MDI_CLIENT_TRANSPARENT) ? 0 : ui.theme.panel2);
+	} else if (message == UI_MSG_LAYOUT) {
+		UIElement *child = element->children;
+
+		while (child) {
+			UI_ASSERT(child->messageClass == _UIMDIChildMessage);
+
+			UIMDIChild *mdiChild = (UIMDIChild *) child;
+
+			if (UIRectangleEquals(mdiChild->bounds, UI_RECT_1(0))) {
+				int width = UIElementMessage(&mdiChild->e, UI_MSG_GET_WIDTH, 0, 0);
+				int height = UIElementMessage(&mdiChild->e, UI_MSG_GET_HEIGHT, width, 0);
+				if (client->cascade + width > element->bounds.r || client->cascade + height > element->bounds.b) client->cascade = 0;
+				mdiChild->bounds = UI_RECT_4(client->cascade, client->cascade + width, client->cascade, client->cascade + height);
+				client->cascade += UI_SIZE_MDI_CASCADE * element->window->scale;
+			}
+
+			UIRectangle bounds = UIRectangleAdd(mdiChild->bounds, UI_RECT_2(element->bounds.l, element->bounds.t));
+			UIElementMove(child, bounds, false);
+			child = child->next;
+		}
+	} else if (message == UI_MSG_FIND_BY_POINT) {
+		UIFindByPoint *m = (UIFindByPoint *) dp;
+		UIMDIChild *child = client->active;
+
+		while (child) {
+			if (UIRectangleContains(child->e.bounds, m->x, m->y)) {
+				m->result = UIElementFindByPoint(&child->e, m->x, m->y);
+				return 1;
+			}
+
+			child = child->previous;
+		}
+
+		return 1;
+	} else if (message == UI_MSG_PRESSED_DESCENDENT) {
+		UIMDIChild *child = (UIMDIChild *) dp;
+
+		if (child && child != client->active) {
+			if (client->e.children == &child->e) client->e.children = child->e.next;
+			if (child->previous) child->previous->e.next = child->e.next;
+			if (child->e.next) ((UIMDIChild *) child->e.next)->previous = child->previous;
+			if (client->active) client->active->e.next = &child->e;
+			child->previous = client->active;
+			child->e.next = NULL;
+			client->active = child;
+			((UIMDIChild *) client->e.children)->previous = NULL;
+			UIElementRefresh(element);
+		}
+	}
+
+	return 0;
+}
+
+UIMDIChild *UIMDIChildCreate(UIElement *parent, uint32_t flags, UIRectangle initialBounds, const char *title, ptrdiff_t titleBytes) {
+	UI_ASSERT(parent->messageClass == _UIMDIClientMessage);
+
+	UIMDIChild *mdiChild = (UIMDIChild *) UIElementCreate(sizeof(UIMDIChild), parent, flags, _UIMDIChildMessage, "MDIChild");
+	UIMDIClient *mdiClient = (UIMDIClient *) parent;
+
+	mdiChild->bounds = initialBounds;
+	mdiChild->title = UIStringCopy(title, (mdiChild->titleBytes = titleBytes));
+	mdiChild->previous = mdiClient->active;
+	mdiClient->active = mdiChild;
+
+	if (flags & UI_MDI_CHILD_CLOSE_BUTTON) {
+		UIButton *closeButton = UIButtonCreate(&mdiChild->e, UI_BUTTON_SMALL | UI_ELEMENT_NON_CLIENT, "X", 1);
+		closeButton->invoke = _UIMDIChildCloseButton;
+		closeButton->e.cp = mdiChild;
+	}
+
+	return mdiChild;
+}
+
+UIMDIClient *UIMDIClientCreate(UIElement *parent, uint32_t flags) {
+	return (UIMDIClient *) UIElementCreate(sizeof(UIMDIClient), parent, flags, _UIMDIClientMessage, "MDIClient");
+}
+
+int _UIExpandPaneMessage(UIElement *element, UIMessage message, int di, void *dp) {
+	UIExpandPane *pane = (UIExpandPane *) element;
+	
+	if (message == UI_MSG_GET_HEIGHT) {
+		int height = UIElementMessage(&pane->button->e, message, di, dp);
+
+		if (pane->expanded) {
+			height += UIElementMessage(&pane->panel->e, message, di, dp);
+		}
+
+		return height;
+	} else if (message == UI_MSG_LAYOUT) {
+		UIRectangle bounds = pane->e.bounds;
+		int buttonHeight = UIElementMessage(&pane->button->e, UI_MSG_GET_HEIGHT, UI_RECT_WIDTH(bounds), NULL);
+		UIElementMove(&pane->button->e, UI_RECT_4(bounds.l, bounds.r, bounds.t, bounds.t + buttonHeight), false);
+
+		if (pane->expanded) {
+			pane->panel->e.flags &= ~UI_ELEMENT_HIDE;
+			UIElementMove(&pane->panel->e, UI_RECT_4(bounds.l, bounds.r, bounds.t + buttonHeight, bounds.b), false);
+		} else {
+			pane->panel->e.flags |= UI_ELEMENT_HIDE;
+		}
+	} else if (message == UI_MSG_CLIENT_PARENT) {
+		*(UIElement **) dp = &pane->panel->e;
+	}
+
+	return 0;
+}
+
+void _UIExpandPaneButtonInvoke(void *cp) {
+	UIExpandPane *pane = (UIExpandPane *) cp;
+	pane->expanded = !pane->expanded;
+	if (pane->expanded) pane->button->e.flags |= UI_BUTTON_CHECKED;
+	else pane->button->e.flags &= ~UI_BUTTON_CHECKED;
+
+	UIElement *ancestor = &pane->e;
+
+	while (ancestor) {
+		UIElementRefresh(ancestor);
+
+		if ((ancestor->messageClass == _UIPanelMessage && (ancestor->flags & UI_PANEL_SCROLL)) 
+				|| (ancestor->messageClass == _UIMDIChildMessage)
+				|| (ancestor->flags & UI_ELEMENT_V_FILL)) {
+			break;
+		}
+
+		ancestor = ancestor->parent;
+	}
+}
+
+UIExpandPane *UIExpandPaneCreate(UIElement *parent, uint32_t flags, const char *label, ptrdiff_t labelBytes, uint32_t panelFlags) {
+	UIExpandPane *pane = (UIExpandPane *) UIElementCreate(sizeof(UIExpandPane), parent, flags, _UIExpandPaneMessage, "ExpandPane");
+	pane->button = UIButtonCreate(parent, UI_ELEMENT_NON_CLIENT, label, labelBytes);
+	pane->button->e.cp = pane;
+	pane->button->invoke = _UIExpandPaneButtonInvoke;
+	pane->panel = UIPanelCreate(parent, UI_ELEMENT_NON_CLIENT | panelFlags);
+	return pane;
+}
+
+void _UIImageDisplayUpdateViewport(UIImageDisplay *display) {
+	UIRectangle bounds = display->e.bounds;
+	bounds.r -= bounds.l, bounds.b -= bounds.t;
+	
+	float minimumZoomX = 1, minimumZoomY = 1;
+	if (display->width  > bounds.r) minimumZoomX = (float) bounds.r / display->width;
+	if (display->height > bounds.b) minimumZoomY = (float) bounds.b / display->height;
+	float minimumZoom = minimumZoomX < minimumZoomY ? minimumZoomX : minimumZoomY;
+	
+	if (display->zoom < minimumZoom || (display->e.flags & _UI_IMAGE_DISPLAY_ZOOM_FIT)) {
+		display->zoom = minimumZoom;
+		display->e.flags |= _UI_IMAGE_DISPLAY_ZOOM_FIT;
+	}
+	
+	if (display->panX < 0) display->panX = 0;
+	if (display->panY < 0) display->panY = 0;
+	if (display->panX > display->width  - bounds.r / display->zoom) display->panX = display->width  - bounds.r / display->zoom;
+	if (display->panY > display->height - bounds.b / display->zoom) display->panY = display->height - bounds.b / display->zoom;
+	
+	if (bounds.r && display->width  * display->zoom <= bounds.r) display->panX = display->width  / 2 - bounds.r / display->zoom / 2;
+	if (bounds.b && display->height * display->zoom <= bounds.b) display->panY = display->height / 2 - bounds.b / display->zoom / 2;
+}
+
+int _UIImageDisplayMessage(UIElement *element, UIMessage message, int di, void *dp) {
+	UIImageDisplay *display = (UIImageDisplay *) element;
+	
+	if (message == UI_MSG_GET_HEIGHT) {
+		return display->height;
+	} else if (message == UI_MSG_GET_WIDTH) {
+		return display->width;
+	} else if (message == UI_MSG_DESTROY) {
+		UI_FREE(display->bits);
+	} else if (message == UI_MSG_PAINT) {
+		UIPainter *painter = (UIPainter *) dp;
+		
+		int w = UI_RECT_WIDTH(element->bounds), h = UI_RECT_HEIGHT(element->bounds);
+		int x = _UILinearMap(0, display->panX, display->panX + w / display->zoom, 0, w) + element->bounds.l;
+		int y = _UILinearMap(0, display->panY, display->panY + h / display->zoom, 0, h) + element->bounds.t;
+		
+		UIRectangle image = UI_RECT_4(x, x + (int) (display->width * display->zoom), y, (int) (y + display->height * display->zoom));
+		UIRectangle bounds = UIRectangleIntersection(painter->clip, UIRectangleIntersection(display->e.bounds, image));
+		if (!UI_RECT_VALID(bounds)) return 0;
+		
+		if (display->zoom == 1) {
+			uint32_t *lineStart = (uint32_t *) painter->bits + bounds.t * painter->width + bounds.l;
+			uint32_t *sourceLineStart = display->bits + (bounds.l - image.l) + display->width * (bounds.t - image.t);
+	
+			for (int i = 0; i < bounds.b - bounds.t; i++, lineStart += painter->width, sourceLineStart += display->width) {
+				uint32_t *destination = lineStart;
+				uint32_t *source = sourceLineStart;
+				int j = bounds.r - bounds.l;
+	
+				do {
+					*destination = *source;
+					destination++;
+					source++;
+				} while (--j);
+			}
+		} else {
+			float zr = 1.0f / display->zoom;
+			uint32_t *destination = (uint32_t *) painter->bits;
+			
+			for (int i = bounds.t; i < bounds.b; i++) {
+				int ty = (i - image.t) * zr;
+				
+				for (int j = bounds.l; j < bounds.r; j++) {
+					int tx = (j - image.l) * zr;
+					destination[i * painter->width + j] = display->bits[ty * display->width + tx];
+				}
+			}
+		}
+	} else if (message == UI_MSG_MOUSE_WHEEL && (element->flags & UI_IMAGE_DISPLAY_INTERACTIVE)) {
+		display->e.flags &= ~_UI_IMAGE_DISPLAY_ZOOM_FIT;
+		int divisions = -di / 72;
+		float factor = 1;
+		float perDivision = element->window->ctrl ? 2.0f : element->window->alt ? 1.01f : 1.2f;
+		while (divisions > 0) factor *= perDivision, divisions--;
+		while (divisions < 0) factor /= perDivision, divisions++;
+		if (display->zoom * factor > 64) factor = 64 / display->zoom;
+		int mx = element->window->cursorX - element->bounds.l;
+		int my = element->window->cursorY - element->bounds.t;
+		display->zoom *= factor;
+		display->panX -= mx / display->zoom * (1 - factor);
+		display->panY -= my / display->zoom * (1 - factor);
+		_UIImageDisplayUpdateViewport(display);
+		UIElementRepaint(&display->e, NULL);
+	} else if (message == UI_MSG_LAYOUT && (element->flags & UI_IMAGE_DISPLAY_INTERACTIVE)) {
+		UIRectangle bounds = display->e.bounds;
+		bounds.r -= bounds.l, bounds.b -= bounds.t;
+		display->panX -= (bounds.r - display->previousWidth ) / 2 / display->zoom;
+		display->panY -= (bounds.b - display->previousHeight) / 2 / display->zoom;
+		display->previousWidth = bounds.r, display->previousHeight = bounds.b;
+		_UIImageDisplayUpdateViewport(display);
+	} else if (message == UI_MSG_GET_CURSOR && (element->flags & UI_IMAGE_DISPLAY_INTERACTIVE)
+			&& (UI_RECT_WIDTH(element->bounds) < display->width * display->zoom 
+				|| UI_RECT_HEIGHT(element->bounds) < display->height * display->zoom)) {
+		return UI_CURSOR_HAND;
+	} else if (message == UI_MSG_MOUSE_DRAG) {
+		display->panX -= (element->window->cursorX - display->previousPanPointX) / display->zoom;
+		display->panY -= (element->window->cursorY - display->previousPanPointY) / display->zoom;
+		_UIImageDisplayUpdateViewport(display);
+		display->previousPanPointX = element->window->cursorX;
+		display->previousPanPointY = element->window->cursorY;
+		UIElementRepaint(element, NULL);
+	} else if (message == UI_MSG_LEFT_DOWN) {           
+		display->e.flags &= ~_UI_IMAGE_DISPLAY_ZOOM_FIT;  
+		display->previousPanPointX = element->window->cursorX;
+		display->previousPanPointY = element->window->cursorY;
+	}
+
+	return 0;
+}
+
+void UIImageDisplaySetContent(UIImageDisplay *display, uint32_t *bits, size_t width, size_t height, size_t stride) {
+	UI_FREE(display->bits);
+
+	display->bits = (uint32_t *) UI_MALLOC(width * height * 4);
+	display->width = width;
+	display->height = height;
+
+	uint32_t *destination = display->bits;
+	uint32_t *source = bits;
+
+	for (uintptr_t row = 0; row < height; row++, source += stride / 4) {
+		for (uintptr_t i = 0; i < width; i++) {
+			*destination++ = source[i];
+		}
+	}
+}
+
+UIImageDisplay *UIImageDisplayCreate(UIElement *parent, uint32_t flags, uint32_t *bits, size_t width, size_t height, size_t stride) {
+	UIImageDisplay *display = (UIImageDisplay *) UIElementCreate(sizeof(UIImageDisplay), parent, flags, _UIImageDisplayMessage, "ImageDisplay");
+	display->zoom = 1.0f;
+	UIImageDisplaySetContent(display, bits, width, height, stride);
+	return display;
+}
+
+int _UIDialogWrapperMessage(UIElement *element, UIMessage message, int di, void *dp) {
+	if (message == UI_MSG_LAYOUT) {
+		int width = UIElementMessage(element->children, UI_MSG_GET_WIDTH, 0, 0);
+		int height = UIElementMessage(element->children, UI_MSG_GET_HEIGHT, width, 0);
+		int cx = (element->bounds.l + element->bounds.r) / 2;
+		int cy = (element->bounds.t + element->bounds.b) / 2;
+		UIRectangle bounds = UI_RECT_4(cx - (width + 1) / 2, cx + width / 2, cy - (height + 1) / 2, cy + height / 2);
+		UIElementMove(element->children, bounds, false);
+		UIElementRepaint(element, NULL);
+	} else if (message == UI_MSG_PAINT) {
+		UIRectangle bounds = UIRectangleAdd(element->children->bounds, UI_RECT_1I(-1));
+		UIDrawBorder((UIPainter *) dp, bounds, ui.theme.border, UI_RECT_1(1));
+		UIDrawBorder((UIPainter *) dp, UIRectangleAdd(bounds, UI_RECT_1(1)), ui.theme.border, UI_RECT_1(1));
+	} else if (message == UI_MSG_KEY_TYPED) {
+		UIKeyTyped *typed = (UIKeyTyped *) dp;
+
+		if (element->window->ctrl) return 0;
+		if (element->window->shift) return 0;
+
+		char c0 = 0, c1 = 0;
+
+		if (typed->textBytes == 1 && typed->text[0] >= 'a' && typed->text[0] <= 'z') {
+			c0 = typed->text[0], c1 = typed->text[0] - 'a' + 'A';
+		} else {
+			return 0;
+		}
+
+		UIElement *row = element->children->children;
+		UIElement *target = NULL;
+		bool duplicate = false;
+
+		while (row) {
+			UIElement *item = row->children;
+
+			while (item) {
+				if (item->messageClass == _UIButtonMessage) {
+					UIButton *button = (UIButton *) item;
+
+					if (button->label && button->labelBytes && (button->label[0] == c0 || button->label[0] == c1)) {
+						if (!target) {
+							target = &button->e;
+						} else {
+							duplicate = true;
+						}
+					}
+				}
+
+				item = item->next;
+			}
+
+			row = row->next;
+		}
+
+		if (target) {
+			if (duplicate) {
+				UIElementFocus(target);
+			} else {
+				UIElementMessage(target, UI_MSG_CLICKED, 0, 0);
+			}
+
+			return 1;
+		}
+	}
+
+	return 0;
+}
+
+void _UIDialogButtonInvoke(void *cp) {
+	ui.dialogResult = (const char *) cp;
+}
+
+int _UIDialogTextboxMessage(UIElement *element, UIMessage message, int di, void *dp) {
+	if (message == UI_MSG_VALUE_CHANGED) {
+		UITextbox *textbox = (UITextbox *) element;
+		char **buffer = (char **) element->cp;
+		*buffer = (char *) UI_REALLOC(*buffer, textbox->bytes + 1);
+		(*buffer)[textbox->bytes] = 0;
+
+		for (ptrdiff_t i = 0; i < textbox->bytes; i++) {
+			(*buffer)[i] = textbox->string[i];
+		}
+	}
+
+	return 0;
+}
+
+const char *UIDialogShow(UIWindow *window, uint32_t flags, const char *format, ...) {
+	// TODO Enter and escape.
+
+	// Create the dialog wrapper and panel.
+
+	UI_ASSERT(!window->dialog);
+	window->dialog = UIElementCreate(sizeof(UIElement), &window->e, 0, _UIDialogWrapperMessage, "DialogWrapper");
+	UIPanel *panel = UIPanelCreate(window->dialog, UI_PANEL_MEDIUM_SPACING | UI_PANEL_GRAY | UI_PANEL_EXPAND);
+	panel->border = UI_RECT_1(UI_SIZE_PANE_MEDIUM_BORDER * 2);
+	window->e.children->flags |= UI_ELEMENT_DISABLED;
+
+	// Create the dialog contents.
+
+	va_list arguments;
+	va_start(arguments, format);
+	UIPanel *row = NULL;
+	UIElement *focus = NULL;
+
+	for (int i = 0; format[i]; i++) {
+		if (i == 0 || format[i - 1] == '\n') {
+			row = UIPanelCreate(&panel->e, UI_PANEL_HORIZONTAL);
+			row->gap = UI_SIZE_PANE_SMALL_GAP;
+		}
+
+		if (format[i] == ' ' || format[i] == '\n') {
+		} else if (format[i] == '%') {
+			i++;
+
+			if (format[i] == 'b' /* button */) {
+				const char *label = va_arg(arguments, const char *);
+				UIButton *button = UIButtonCreate(&row->e, 0, label, -1);
+				if (!focus) focus = &button->e;
+				button->invoke = _UIDialogButtonInvoke;
+				button->e.cp = (void *) label;
+			} else if (format[i] == 's' /* label from string */) {
+				const char *label = va_arg(arguments, const char *);
+				UILabelCreate(&row->e, 0, label, -1);
+			} else if (format[i] == 't' /* textbox */) {
+				char **buffer = va_arg(arguments, char **);
+				UITextbox *textbox = UITextboxCreate(&row->e, UI_ELEMENT_H_FILL);
+				if (!focus) focus = &textbox->e;
+				if (*buffer) UITextboxReplace(textbox, *buffer, _UIStringLength(*buffer), false);
+				textbox->e.cp = buffer;
+				textbox->e.messageUser = _UIDialogTextboxMessage;
+			} else if (format[i] == 'f' /* horizontal fill */) {
+				UISpacerCreate(&row->e, UI_ELEMENT_H_FILL, 0, 0);
+			} else if (format[i] == 'l' /* horizontal line */) {
+				UISpacerCreate(&row->e, UI_SPACER_LINE | UI_ELEMENT_H_FILL, 0, 1);
+			} else if (format[i] == 'u' /* user */) {
+				UIDialogUserCallback callback = va_arg(arguments, UIDialogUserCallback);
+				callback(&row->e);
+			}
+		} else {
+			int j = i;
+			while (format[j] && format[j] != '%' && format[j] != '\n') j++;
+			UILabelCreate(&row->e, 0, format + i, j - i);
+			i = j - 1;
+		}
+	}
+
+	va_end(arguments);
+
+	window->dialogOldFocus = window->focused;
+	UIElementFocus(focus ? focus : window->dialog);
+
+	// Run the modal message loop.
+
+	int result;
+	ui.dialogResult = NULL;
+	for (int i = 1; i <= 3; i++) _UIWindowSetPressed(window, NULL, i);
+	UIElementRefresh(&window->e);
+	_UIUpdate();
+	while (!ui.dialogResult && _UIMessageLoopSingle(&result));
+	ui.quit = !ui.dialogResult;
+
+	// Destroy the dialog.
+
+	window->e.children->flags &= ~UI_ELEMENT_DISABLED;
+	UIElementDestroy(window->dialog);
+	window->dialog = NULL;
+	UIElementRefresh(&window->e);
+	if (window->dialogOldFocus) UIElementFocus(window->dialogOldFocus);
+	return ui.dialogResult ? ui.dialogResult : "";
+}
+
+bool _UIMenusClose() {
+	UIWindow *window = ui.windows;
+	bool anyClosed = false;
+
+	while (window) {
+		if (window->e.flags & UI_WINDOW_MENU) {
+			UIElementDestroy(&window->e);
+			anyClosed = true;
+		}
+
+		window = window->next;
+	}
+
+	return anyClosed;
+}
+
+#ifndef UI_ESSENCE
+int _UIMenuItemMessage(UIElement *element, UIMessage message, int di, void *dp) {
+	if (message == UI_MSG_CLICKED) {
+		_UIMenusClose();
+	}
+
+	return 0;
+}
+
+int _UIMenuMessage(UIElement *element, UIMessage message, int di, void *dp) {
+	UIMenu *menu = (UIMenu *) element;
+
+	if (message == UI_MSG_GET_WIDTH) {
+		UIElement *child = element->children;
+		int width = 0;
+
+		while (child) {
+			if (~child->flags & UI_ELEMENT_NON_CLIENT) {
+				int w = UIElementMessage(child, UI_MSG_GET_WIDTH, 0, 0);
+				if (w > width) width = w;
+			}
+
+			child = child->next;
+		}
+
+		return width + 4 + UI_SIZE_SCROLL_BAR;
+	} else if (message == UI_MSG_GET_HEIGHT) {
+		UIElement *child = element->children;
+		int height = 0;
+
+		while (child) {
+			if (~child->flags & UI_ELEMENT_NON_CLIENT) {
+				height += UIElementMessage(child, UI_MSG_GET_HEIGHT, 0, 0);
+			}
+
+			child = child->next;
+		}
+
+		return height + 4;
+	} else if (message == UI_MSG_PAINT) {
+		UIDrawBlock((UIPainter *) dp, element->bounds, ui.theme.border);
+	} else if (message == UI_MSG_LAYOUT) {
+		UIElement *child = element->children;
+		int position = element->bounds.t + 2 - menu->vScroll->position;
+		int totalHeight = 0;
+		int scrollBarSize = (menu->e.flags & UI_MENU_NO_SCROLL) ? 0 : UI_SIZE_SCROLL_BAR;
+
+		while (child) {
+			if (~child->flags & UI_ELEMENT_NON_CLIENT) {
+				int height = UIElementMessage(child, UI_MSG_GET_HEIGHT, 0, 0);
+				UIElementMove(child, UI_RECT_4(element->bounds.l + 2, element->bounds.r - scrollBarSize - 2, 
+							position, position + height), false);
+				position += height;
+				totalHeight += height;
+			}
+
+			child = child->next;
+		}
+
+		UIRectangle scrollBarBounds = element->bounds;
+		scrollBarBounds.l = scrollBarBounds.r - scrollBarSize * element->window->scale;
+		menu->vScroll->maximum = totalHeight;
+		menu->vScroll->page = UI_RECT_HEIGHT(element->bounds);
+		UIElementMove(&menu->vScroll->e, scrollBarBounds, true);
+	} else if (message == UI_MSG_KEY_TYPED) {
+		UIKeyTyped *m = (UIKeyTyped *) dp;
+
+		if (m->code == UI_KEYCODE_ESCAPE) {
+			_UIMenusClose();
+			return 1;
+		}
+	} else if (message == UI_MSG_MOUSE_WHEEL) {
+		return UIElementMessage(&menu->vScroll->e, message, di, dp);
+	} else if (message == UI_MSG_SCROLLED) {
+		UIElementRefresh(element);
+	}
+
+	return 0;
+}
+
+void UIMenuAddItem(UIMenu *menu, uint32_t flags, const char *label, ptrdiff_t labelBytes, void (*invoke)(void *cp), void *cp) {
+	UIButton *button = UIButtonCreate(&menu->e, flags | UI_BUTTON_MENU_ITEM, label, labelBytes);
+	button->invoke = invoke;
+	button->e.messageUser = _UIMenuItemMessage;
+	button->e.cp = cp;
+}
+
+void _UIMenuPrepare(UIMenu *menu, int *width, int *height) {
+	*width = UIElementMessage(&menu->e, UI_MSG_GET_WIDTH, 0, 0);
+	*height = UIElementMessage(&menu->e, UI_MSG_GET_HEIGHT, 0, 0);
+
+	if (menu->e.flags & UI_MENU_PLACE_ABOVE) {
+		menu->pointY -= *height;
+	}
+}
+
+UIMenu *UIMenuCreate(UIElement *parent, uint32_t flags) {
+	UIWindow *window = UIWindowCreate(parent->window, UI_WINDOW_MENU, 0, 0, 0);
+	UIMenu *menu = (UIMenu *) UIElementCreate(sizeof(UIMenu), &window->e, flags, _UIMenuMessage, "Menu");
+	menu->vScroll = UIScrollBarCreate(&menu->e, UI_ELEMENT_NON_CLIENT);
+
+	if (parent->parent) {
+		UIRectangle screenBounds = UIElementScreenBounds(parent);
+		menu->pointX = screenBounds.l;
+		menu->pointY = (flags & UI_MENU_PLACE_ABOVE) ? (screenBounds.t + 1) : (screenBounds.b - 1);
+	} else {
+		int x = 0, y = 0;
+		_UIWindowGetScreenPosition(parent->window, &x, &y);
+
+		menu->pointX = parent->window->cursorX + x;
+		menu->pointY = parent->window->cursorY + y;
+	}
+
+	return menu;
+}
+#endif
+
+UIRectangle UIElementScreenBounds(UIElement *element) {
+	int x = 0, y = 0;
+	_UIWindowGetScreenPosition(element->window, &x, &y);
+	return UIRectangleAdd(element->bounds, UI_RECT_2(x, y));
+}
+
+void UIWindowRegisterShortcut(UIWindow *window, UIShortcut shortcut) {
+	if (window->shortcutCount + 1 > window->shortcutAllocated) {
+		window->shortcutAllocated = (window->shortcutCount + 1) * 2;
+		window->shortcuts = (UIShortcut *) UI_REALLOC(window->shortcuts, window->shortcutAllocated * sizeof(UIShortcut));
+	}
+
+	window->shortcuts[window->shortcutCount++] = shortcut;
+}
+
+void _UIElementPaint(UIElement *element, UIPainter *painter) {
+	if (element->flags & UI_ELEMENT_HIDE) {
+		return;
+	}
+
+	// Clip painting to the element's clip.
+
+	painter->clip = UIRectangleIntersection(element->clip, painter->clip);
+
+	if (!UI_RECT_VALID(painter->clip)) {
+		return;
+	}
+
+	// Paint the element.
+
+	UIElementMessage(element, UI_MSG_PAINT, 0, painter);
+
+	// Paint its children.
+
+	UIElement *child = element->children;
+	UIRectangle previousClip = painter->clip;
+
+	while (child) {
+		painter->clip = previousClip;
+		_UIElementPaint(child, painter);
+		child = child->next;
+	}
+}
+
+void UIElementFocus(UIElement *element) {
+	UIElement *previous = element->window->focused;
+	if (previous == element) return;
+	element->window->focused = element;
+	if (previous) UIElementMessage(previous, UI_MSG_UPDATE, UI_UPDATE_FOCUSED, 0);
+	if (element) UIElementMessage(element, UI_MSG_UPDATE, UI_UPDATE_FOCUSED, 0);
+
+#ifdef UI_DEBUG
+	_UIInspectorRefresh();
+#endif
+}
+
+void _UIWindowSetPressed(UIWindow *window, UIElement *element, int button) {
+	UIElement *previous = window->pressed;
+	window->pressed = element;
+	window->pressedButton = button;
+	if (previous) UIElementMessage(previous, UI_MSG_UPDATE, UI_UPDATE_PRESSED, 0);
+	if (element) UIElementMessage(element, UI_MSG_UPDATE, UI_UPDATE_PRESSED, 0);
+
+	UIElement *ancestor = element;
+	UIElement *child = NULL;
+
+	while (ancestor) {
+		UIElementMessage(ancestor, UI_MSG_PRESSED_DESCENDENT, 0, child);
+		child = ancestor;
+		ancestor = ancestor->parent;
+	}
+}
+
+bool _UIDestroy(UIElement *element) {
+	if (element->flags & UI_ELEMENT_DESTROY_DESCENDENT) {
+		element->flags &= ~UI_ELEMENT_DESTROY_DESCENDENT;
+
+		UIElement *child = element->children;
+		UIElement **link = &element->children;
+
+		while (child) {
+			UIElement *next = child->next;
+
+			if (_UIDestroy(child)) {
+				*link = next;
+			} else {
+				link = &child->next;
+			}
+
+			child = next;
+		}
+	}
+
+	if (element->flags & UI_ELEMENT_DESTROY) {
+		UIElementMessage(element, UI_MSG_DESTROY, 0, 0);
+
+		if (element->window->pressed == element) {
+			_UIWindowSetPressed(element->window, NULL, 0);
+		}
+
+		if (element->window->hovered == element) {
+			element->window->hovered = &element->window->e;
+		}
+
+		if (element->window->focused == element) {
+			element->window->focused = NULL;
+		}
+
+		if (element->window->dialogOldFocus == element) {
+			element->window->dialogOldFocus = NULL;
+		}
+
+		if (ui.animating == element) {
+			ui.animating = NULL;
+		}
+
+		UI_FREE(element);
+		return true;
+	} else {
+		return false;
+	}
+}
+
+void _UIUpdate() {
+	UIWindow *window = ui.windows;
+	UIWindow **link = &ui.windows;
+
+	while (window) {
+		UIWindow *next = window->next;
+
+		if (_UIDestroy(&window->e)) {
+			*link = next;
+		} else {
+			link = &window->next;
+
+			if (UI_RECT_VALID(window->updateRegion)) {
+#ifdef __cplusplus
+				UIPainter painter = {};
+#else
+				UIPainter painter = { 0 };
+#endif
+				painter.bits = window->bits;
+				painter.width = window->width;
+				painter.height = window->height;
+				painter.clip = UIRectangleIntersection(UI_RECT_2S(window->width, window->height), window->updateRegion);
+				_UIElementPaint(&window->e, &painter);
+				_UIWindowEndPaint(window, &painter);
+				window->updateRegion = UI_RECT_1(0);
+
+#ifdef UI_DEBUG
+				window->lastFullFillCount = (float) painter.fillCount / (UI_RECT_WIDTH(window->updateRegion) * UI_RECT_HEIGHT(window->updateRegion));
+#endif
+			}
+		}
+
+		window = next;
+	}
+}
+
+UIElement *UIElementFindByPoint(UIElement *element, int x, int y) {
+	UIFindByPoint m = { 0 };
+	m.x = x, m.y = y;
+
+	if (UIElementMessage(element, UI_MSG_FIND_BY_POINT, 0, &m)) {
+		return m.result ? m.result : element;
+	}
+
+	UIElement *child = element->children;
+
+	while (child) {
+		if ((~child->flags & UI_ELEMENT_HIDE) && UIRectangleContains(child->clip, x, y)) {
+			return UIElementFindByPoint(child, x, y);
+		}
+
+		child = child->next;
+	}
+
+	return element;
+}
+
+void _UIProcessAnimations() {
+	if (ui.animating) {
+		UIElementMessage(ui.animating, UI_MSG_ANIMATE, 0, 0);
+		_UIUpdate();
+	}
+}
+
+bool _UIMenusOpen() {
+	UIWindow *window = ui.windows;
+
+	while (window) {
+		if (window->e.flags & UI_WINDOW_MENU) {
+			return true;
+		}
+
+		window = window->next;
+	}
+
+	return false;
+}
+
+void _UIWindowDestroyCommon(UIWindow *window) {
+	UI_FREE(window->bits);
+	UI_FREE(window->shortcuts);
+}
+
+UIElement *_UIElementLastChild(UIElement *element) {
+	if (!element->children) {
+		return NULL;
+	}
+
+	UIElement *child = element->children;
+
+	while (child->next) {
+		child = child->next;
+	}
+
+	return child;
+}
+
+UIElement *_UIElementPreviousSibling(UIElement *element) {
+	if (!element->parent) {
+		return NULL;
+	}
+
+	UIElement *sibling = element->parent->children;
+
+	if (sibling == element) {
+		return NULL;
+	}
+
+	while (sibling->next != element) {
+		sibling = sibling->next;
+		UI_ASSERT(sibling);
+	}
+	
+	return sibling;
+}
+
+bool _UIWindowInputEvent(UIWindow *window, UIMessage message, int di, void *dp) {
+	bool handled = true;
+
+	if (window->pressed) {
+		if (message == UI_MSG_MOUSE_MOVE) {
+			UIElementMessage(window->pressed, UI_MSG_MOUSE_DRAG, di, dp);
+		} else if (message == UI_MSG_LEFT_UP && window->pressedButton == 1) {
+			if (window->hovered == window->pressed) {
+				UIElementMessage(window->pressed, UI_MSG_CLICKED, di, dp);
+				if (ui.quit || ui.dialogResult) goto end;
+			}
+
+			if (window->pressed) {
+				UIElementMessage(window->pressed, UI_MSG_LEFT_UP, di, dp);
+				if (ui.quit || ui.dialogResult) goto end;
+				_UIWindowSetPressed(window, NULL, 1);
+			}
+		} else if (message == UI_MSG_MIDDLE_UP && window->pressedButton == 2) {
+			UIElementMessage(window->pressed, UI_MSG_MIDDLE_UP, di, dp);
+			if (ui.quit || ui.dialogResult) goto end;
+			_UIWindowSetPressed(window, NULL, 2);
+		} else if (message == UI_MSG_RIGHT_UP && window->pressedButton == 3) {
+			UIElementMessage(window->pressed, UI_MSG_RIGHT_UP, di, dp);
+			if (ui.quit || ui.dialogResult) goto end;
+			_UIWindowSetPressed(window, NULL, 3);
+		}
+	}
+
+	if (window->pressed) {
+		bool inside = UIRectangleContains(window->pressed->clip, window->cursorX, window->cursorY);
+
+		if (inside && window->hovered == &window->e) {
+			window->hovered = window->pressed;
+			UIElementMessage(window->pressed, UI_MSG_UPDATE, UI_UPDATE_HOVERED, 0);
+		} else if (!inside && window->hovered == window->pressed) {
+			window->hovered = &window->e;
+			UIElementMessage(window->pressed, UI_MSG_UPDATE, UI_UPDATE_HOVERED, 0);
+		}
+
+		if (ui.quit || ui.dialogResult) goto end;
+	}
+
+	if (!window->pressed) {
+		UIElement *hovered = UIElementFindByPoint(&window->e, window->cursorX, window->cursorY);
+
+		if (message == UI_MSG_MOUSE_MOVE) {
+			UIElementMessage(hovered, UI_MSG_MOUSE_MOVE, di, dp);
+
+			int cursor = UIElementMessage(window->hovered, UI_MSG_GET_CURSOR, di, dp);
+
+			if (cursor != window->cursorStyle) {
+				window->cursorStyle = cursor;
+				_UIWindowSetCursor(window, cursor);
+			}
+		} else if (message == UI_MSG_LEFT_DOWN) {
+			if ((window->e.flags & UI_WINDOW_MENU) || !_UIMenusClose()) {
+				_UIWindowSetPressed(window, hovered, 1);
+				UIElementMessage(hovered, UI_MSG_LEFT_DOWN, di, dp);
+			}
+		} else if (message == UI_MSG_MIDDLE_DOWN) {
+			if ((window->e.flags & UI_WINDOW_MENU) || !_UIMenusClose()) {
+				_UIWindowSetPressed(window, hovered, 2);
+				UIElementMessage(hovered, UI_MSG_MIDDLE_DOWN, di, dp);
+			}
+		} else if (message == UI_MSG_RIGHT_DOWN) {
+			if ((window->e.flags & UI_WINDOW_MENU) || !_UIMenusClose()) {
+				_UIWindowSetPressed(window, hovered, 3);
+				UIElementMessage(hovered, UI_MSG_RIGHT_DOWN, di, dp);
+			}
+		} else if (message == UI_MSG_MOUSE_WHEEL) {
+			UIElement *element = hovered;
+
+			while (element) {
+				if (UIElementMessage(element, UI_MSG_MOUSE_WHEEL, di, dp)) {
+					break;
+				}
+
+				element = element->parent;
+			}
+		} else if (message == UI_MSG_KEY_TYPED) {
+			handled = false;
+
+			if (window->focused) {
+				UIElement *element = window->focused;
+
+				while (element) {
+					if (UIElementMessage(element, UI_MSG_KEY_TYPED, di, dp)) {
+						handled = true;
+						break;
+					}
+
+					element = element->parent;
+				}
+			} else {
+				if (UIElementMessage(&window->e, UI_MSG_KEY_TYPED, di, dp)) {
+					handled = true;
+				}
+			}
+
+			if (!handled && !_UIMenusOpen()) {
+				UIKeyTyped *m = (UIKeyTyped *) dp;
+
+				if (m->code == UI_KEYCODE_TAB && !window->ctrl && !window->alt) {
+					UIElement *start = window->focused ? window->focused : &window->e;
+					UIElement *element = start;
+
+					do {
+						if (element->children && !(element->flags & (UI_ELEMENT_HIDE | UI_ELEMENT_DISABLED))) {
+							element = window->shift ? _UIElementLastChild(element) : element->children;
+							continue;
+						} 
+
+						while (element) {
+							if (window->shift ? (element->parent && element->parent->children != element) : !!element->next) {
+								element = window->shift ? _UIElementPreviousSibling(element) : element->next;
+								break;
+							} else {
+								element = element->parent;
+							}
+						}
+
+						if (!element) {
+							element = &window->e;
+						}
+					} while (element != start && ((~element->flags & UI_ELEMENT_TAB_STOP) 
+						|| (element->flags & (UI_ELEMENT_HIDE | UI_ELEMENT_DISABLED))));
+
+					if (~element->flags & UI_ELEMENT_WINDOW) {
+						UIElementFocus(element);
+					}
+
+					handled = true;
+				} else if (!window->dialog) {
+					for (intptr_t i = window->shortcutCount - 1; i >= 0; i--) {
+						UIShortcut *shortcut = window->shortcuts + i;
+
+						if (shortcut->code == m->code && shortcut->ctrl == window->ctrl 
+								&& shortcut->shift == window->shift && shortcut->alt == window->alt) {
+							shortcut->invoke(shortcut->cp);
+							handled = true;
+							break;
+						}
+					}
+				}
+			}
+		}
+
+		if (ui.quit || ui.dialogResult) goto end;
+
+		if (hovered != window->hovered) {
+			UIElement *previous = window->hovered;
+			window->hovered = hovered;
+			UIElementMessage(previous, UI_MSG_UPDATE, UI_UPDATE_HOVERED, 0);
+			UIElementMessage(window->hovered, UI_MSG_UPDATE, UI_UPDATE_HOVERED, 0);
+		}
+	}
+
+	end: _UIUpdate();
+	return handled;
+}
+
+UIFont *UIFontCreate(const char *cPath, uint32_t size) {
+	UIFont *font = (UIFont *) UI_CALLOC(sizeof(UIFont));
+
+#ifdef UI_FREETYPE
+	if (cPath) {
+		if (!FT_New_Face(ui.ft, cPath, 0, &font->font)) {
+			FT_Set_Char_Size(font->font, 0, size * 64, 100, 100);
+			FT_Load_Char(font->font, 'a', FT_LOAD_DEFAULT);
+			font->glyphWidth = font->font->glyph->advance.x / 64;
+			font->glyphHeight = (font->font->size->metrics.ascender - font->font->size->metrics.descender) / 64;
+			font->isFreeType = true;
+			return font;
+		}
+	}
+#endif
+	
+	font->glyphWidth = 9;
+	font->glyphHeight = 16;
+	return font;
+}
+
+UIFont *UIFontActivate(UIFont *font) {
+	UIFont *previous = ui.activeFont;
+	ui.activeFont = font;
+	return previous;
+}
+
+void _UIInitialiseCommon() {
+	ui.theme = _uiThemeDark;
+
+#ifdef UI_FREETYPE
+	FT_Init_FreeType(&ui.ft);
+	UIFontActivate(UIFontCreate(_UI_TO_STRING_2(UI_FONT_PATH), 11));
+#else
+	UIFontActivate(UIFontCreate(0, 0));
+#endif
+}
+
+void _UIWindowAdd(UIWindow *window) {
+	window->scale = 1.0f;
+	window->e.window = window;
+	window->hovered = &window->e;
+	window->next = ui.windows;
+	ui.windows = window;
+}
+
+int _UIWindowMessageCommon(UIElement *element, UIMessage message, int di, void *dp) {
+	if (message == UI_MSG_LAYOUT && element->children) {
+		UIElementMove(element->children, element->bounds, false);
+		if (element->window->dialog) UIElementMove(element->window->dialog, element->bounds, false);
+		UIElementRepaint(element, NULL);
+	} else if (message == UI_MSG_FIND_BY_POINT) {
+		UIFindByPoint *m = (UIFindByPoint *) dp;
+		if (element->window->dialog) m->result = UIElementFindByPoint(element->window->dialog, m->x, m->y);
+		else if (!element->children) m->result = NULL;
+		else m->result = UIElementFindByPoint(element->children, m->x, m->y);
+		return 1;
+	}
+
+	return 0;
+}
+
+#ifdef UI_DEBUG
+
+void UIInspectorLog(const char *cFormat, ...) {
+	va_list arguments;
+	va_start(arguments, cFormat);
+	char buffer[4096];
+	vsnprintf(buffer, sizeof(buffer), cFormat, arguments);
+	UICodeInsertContent(ui.inspectorLog, buffer, -1, false);
+	va_end(arguments);
+	UIElementRefresh(&ui.inspectorLog->e);
+}
+
+UIElement *_UIInspectorFindNthElement(UIElement *element, int *index, int *depth) {
+	if (*index == 0) {
+		return element;
+	}
+
+	*index = *index - 1;
+	
+	UIElement *child = element->children;
+
+	while (child) {
+		if (!(child->flags & (UI_ELEMENT_DESTROY | UI_ELEMENT_HIDE))) {
+			UIElement *result = _UIInspectorFindNthElement(child, index, depth);
+
+			if (result) {
+				if (depth) {
+					*depth = *depth + 1;
+				}
+
+				return result;
+			}
+		}
+
+		child = child->next;
+	}
+
+	return NULL;
+}
+
+int _UIInspectorTableMessage(UIElement *element, UIMessage message, int di, void *dp) {
+	if (!ui.inspectorTarget) {
+		return 0;
+	}
+
+	if (message == UI_MSG_TABLE_GET_ITEM) {
+		UITableGetItem *m = (UITableGetItem *) dp;
+		int index = m->index;
+		int depth = 0;
+		UIElement *element = _UIInspectorFindNthElement(&ui.inspectorTarget->e, &index, &depth);
+		if (!element) return 0;
+
+		if (m->column == 0) {
+			return snprintf(m->buffer, m->bufferBytes, "%.*s%s", depth * 2, "                ", element->cClassName);
+		} else if (m->column == 1) {
+			return snprintf(m->buffer, m->bufferBytes, "%d:%d, %d:%d", UI_RECT_ALL(element->bounds));
+		} else if (m->column == 2) {
+			return snprintf(m->buffer, m->bufferBytes, "%d%c", element->id, element->window->focused == element ? '*' : ' ');
+		}
+	} else if (message == UI_MSG_MOUSE_MOVE) {
+		int index = UITableHitTest(ui.inspectorTable, element->window->cursorX, element->window->cursorY);
+		UIElement *element = NULL;
+		if (index >= 0) element = _UIInspectorFindNthElement(&ui.inspectorTarget->e, &index, NULL);
+		UIWindow *window = ui.inspectorTarget;
+		UIPainter painter = { 0 };
+		window->updateRegion = window->e.bounds;
+		painter.bits = window->bits;
+		painter.width = window->width;
+		painter.height = window->height;
+		painter.clip = UI_RECT_2S(window->width, window->height);
+
+		for (int i = 0; i < window->width * window->height; i++) {
+			window->bits[i] = 0xFF00FF;
+		}
+
+		_UIElementPaint(&window->e, &painter);
+		painter.clip = UI_RECT_2S(window->width, window->height);
+
+		if (element) {
+			UIDrawInvert(&painter, element->bounds);
+			UIDrawInvert(&painter, UIRectangleAdd(element->bounds, UI_RECT_1I(4)));
+		}
+
+		_UIWindowEndPaint(window, &painter);
+	}
+
+	return 0;
+}
+
+void _UIInspectorCreate() {
+	ui.inspector = UIWindowCreate(0, UI_WINDOW_INSPECTOR, "Inspector", 0, 0);
+	UISplitPane *splitPane = UISplitPaneCreate(&ui.inspector->e, 0, 0.5f);
+	ui.inspectorTable = UITableCreate(&splitPane->e, 0, "Class\tBounds\tID");
+	ui.inspectorTable->e.messageUser = _UIInspectorTableMessage;
+	ui.inspectorLog = UICodeCreate(&splitPane->e, 0);
+}
+
+int _UIInspectorCountElements(UIElement *element) {
+	UIElement *child = element->children;
+	int count = 1;
+
+	while (child) {
+		if (!(child->flags & (UI_ELEMENT_DESTROY | UI_ELEMENT_HIDE))) {
+			count += _UIInspectorCountElements(child);
+		}
+
+		child = child->next;
+	}
+
+	return count;
+}
+
+void _UIInspectorRefresh() {
+	if (!ui.inspectorTarget || !ui.inspector || !ui.inspectorTable) return;
+	ui.inspectorTable->itemCount = _UIInspectorCountElements(&ui.inspectorTarget->e);
+	UITableResizeColumns(ui.inspectorTable);
+	UIElementRefresh(&ui.inspectorTable->e);
+}
+
+void _UIInspectorSetFocusedWindow(UIWindow *window) {
+	if (!ui.inspector || !ui.inspectorTable) return;
+
+	if (window->e.flags & UI_WINDOW_INSPECTOR) {
+		return;
+	}
+
+	if (ui.inspectorTarget != window) {
+		ui.inspectorTarget = window;
+		_UIInspectorRefresh();
+	}
+}
+
+#else
+
+void _UIInspectorCreate() {}
+void _UIInspectorSetFocusedWindow(UIWindow *window) {}
+void _UIInspectorRefresh() {}
+
+#endif
+
+#ifdef UI_AUTOMATION_TESTS
+
+int UIAutomationRunTests();
+
+void UIAutomationProcessMessage() {
+	int result;
+	_UIMessageLoopSingle(&result);
+}
+
+void UIAutomationKeyboardTypeSingle(intptr_t code, bool ctrl, bool shift, bool alt) {
+	UIWindow *window = ui.windows; // TODO Get the focused window.
+	UIKeyTyped m = { 0 };
+	m.code = code;
+	window->ctrl = ctrl;
+	window->alt = alt;
+	window->shift = shift;
+	_UIWindowInputEvent(window, UI_MSG_KEY_TYPED, 0, &m);
+	window->ctrl = false;
+	window->alt = false;
+	window->shift = false;
+}
+
+void UIAutomationKeyboardType(const char *string) {
+	UIWindow *window = ui.windows; // TODO Get the focused window.
+
+	UIKeyTyped m = { 0 };
+	char c[2];
+	m.text = c;
+	m.textBytes = 1;
+	c[1] = 0;
+
+	for (int i = 0; string[i]; i++) {
+		window->ctrl = false;
+		window->alt = false;
+		window->shift = (c[0] >= 'A' && c[0] <= 'Z');
+		c[0] = string[i];
+		m.code = (c[0] >= 'A' && c[0] <= 'Z') ? UI_KEYCODE_LETTER(c[0]) 
+			: c[0] == '\n' ? UI_KEYCODE_ENTER 
+			: c[0] == '\t' ? UI_KEYCODE_TAB 
+			: c[0] == ' ' ? UI_KEYCODE_SPACE 
+			: (c[0] >= '0' && c[0] <= '9') ? UI_KEYCODE_DIGIT(c[0]) : 0;
+		_UIWindowInputEvent(window, UI_MSG_KEY_TYPED, 0, &m);
+	}
+
+	window->ctrl = false;
+	window->alt = false;
+	window->shift = false;
+}
+
+bool UIAutomationCheckCodeLineMatches(UICode *code, int lineIndex, const char *input) {
+	if (lineIndex < 1 || lineIndex > code->lineCount) return false;
+	int bytes = 0;
+	for (int i = 0; input[i]; i++) bytes++;
+	if (bytes != code->lines[lineIndex - 1].bytes) return false;
+	for (int i = 0; input[i]; i++) if (code->content[code->lines[lineIndex - 1].offset + i] != input[i]) return false;
+	return true;
+}
+
+bool UIAutomationCheckTableItemMatches(UITable *table, int row, int column, const char *input) {
+	int bytes = 0;
+	for (int i = 0; input[i]; i++) bytes++;
+	if (row < 0 || row >= table->itemCount) return false;
+	if (column < 0 || column >= table->columnCount) return false;
+	char *buffer = (char *) UI_MALLOC(bytes + 1);
+	UITableGetItem m = { 0 };
+	m.buffer = buffer;
+	m.bufferBytes = bytes + 1;
+	m.column = column;
+	m.index = row;
+	int length = UIElementMessage(&table->e, UI_MSG_TABLE_GET_ITEM, 0, &m);
+	if (length != bytes) return false;
+	for (int i = 0; input[i]; i++) if (buffer[i] != input[i]) return false;
+	return true;
+}
+
+#endif
+
+int UIMessageLoop() {
+	_UIInspectorCreate();
+	_UIUpdate();
+#ifdef UI_AUTOMATION_TESTS
+	return UIAutomationRunTests();
+#else
+	int result = 0;
+	while (!ui.quit && _UIMessageLoopSingle(&result)) ui.dialogResult = NULL;
+	return result;
+#endif
+}
+
+#ifdef UI_LINUX
+
+const int UI_KEYCODE_A = XK_a;
+const int UI_KEYCODE_BACKSPACE = XK_BackSpace;
+const int UI_KEYCODE_DELETE = XK_Delete;
+const int UI_KEYCODE_DOWN = XK_Down;
+const int UI_KEYCODE_END = XK_End;
+const int UI_KEYCODE_ENTER = XK_Return;
+const int UI_KEYCODE_ESCAPE = XK_Escape;
+const int UI_KEYCODE_F1 = XK_F1;
+const int UI_KEYCODE_HOME = XK_Home;
+const int UI_KEYCODE_LEFT = XK_Left;
+const int UI_KEYCODE_RIGHT = XK_Right;
+const int UI_KEYCODE_SPACE = XK_space;
+const int UI_KEYCODE_TAB = XK_Tab;
+const int UI_KEYCODE_UP = XK_Up;
+const int UI_KEYCODE_INSERT = XK_Insert;
+const int UI_KEYCODE_0 = XK_0;
+
+int _UIWindowMessage(UIElement *element, UIMessage message, int di, void *dp) {
+	if (message == UI_MSG_DESTROY) {
+		UIWindow *window = (UIWindow *) element;
+		_UIWindowDestroyCommon(window);
+		window->image->data = NULL;
+		XDestroyImage(window->image);
+		XDestroyIC(window->xic);
+		XDestroyWindow(ui.display, ((UIWindow *) element)->window);
+	}
+
+	return _UIWindowMessageCommon(element, message, di, dp);
+}
+
+UIWindow *UIWindowCreate(UIWindow *owner, uint32_t flags, const char *cTitle, int _width, int _height) {
+	_UIMenusClose();
+
+	UIWindow *window = (UIWindow *) UIElementCreate(sizeof(UIWindow), NULL, flags | UI_ELEMENT_WINDOW, _UIWindowMessage, "Window");
+	_UIWindowAdd(window);
+	if (owner) window->scale = owner->scale;
+
+	int width = (flags & UI_WINDOW_MENU) ? 1 : _width ? _width : 800;
+	int height = (flags & UI_WINDOW_MENU) ? 1 : _height ? _height : 600;
+
+	XSetWindowAttributes attributes = {};
+	attributes.override_redirect = flags & UI_WINDOW_MENU;
+
+	window->window = XCreateWindow(ui.display, DefaultRootWindow(ui.display), 0, 0, width, height, 0, 0, 
+		InputOutput, CopyFromParent, CWOverrideRedirect, &attributes);
+	if (cTitle) XStoreName(ui.display, window->window, cTitle);
+	XSelectInput(ui.display, window->window, SubstructureNotifyMask | ExposureMask | PointerMotionMask 
+		| ButtonPressMask | ButtonReleaseMask | KeyPressMask | KeyReleaseMask | StructureNotifyMask
+		| EnterWindowMask | LeaveWindowMask | ButtonMotionMask | KeymapStateMask | FocusChangeMask | PropertyChangeMask);
+
+	if (flags & UI_WINDOW_MAXIMIZE) {
+		Atom atoms[2] = { XInternAtom(ui.display, "_NET_WM_STATE_MAXIMIZED_HORZ", 0), XInternAtom(ui.display, "_NET_WM_STATE_MAXIMIZED_VERT", 0) };
+		XChangeProperty(ui.display, window->window, XInternAtom(ui.display, "_NET_WM_STATE", 0), XA_ATOM, 32, PropModeReplace, (unsigned char *) atoms, 2);
+	}
+
+	if (~flags & UI_WINDOW_MENU) {
+		XMapRaised(ui.display, window->window);
+	}
+
+	if (flags & UI_WINDOW_CENTER_IN_OWNER) {
+		int x = 0, y = 0;
+		_UIWindowGetScreenPosition(owner, &x, &y);
+		XMoveResizeWindow(ui.display, window->window, x + owner->width / 2 - width / 2, y + owner->height / 2 - height / 2, width, height);
+	}
+
+	XSetWMProtocols(ui.display, window->window, &ui.windowClosedID, 1);
+	window->image = XCreateImage(ui.display, ui.visual, 24, ZPixmap, 0, NULL, 10, 10, 32, 0);
+
+	window->xic = XCreateIC(ui.xim, XNInputStyle, XIMPreeditNothing | XIMStatusNothing, XNClientWindow, window->window, XNFocusWindow, window->window, NULL);
+
+	int dndVersion = 4;
+	XChangeProperty(ui.display, window->window, ui.dndAwareID, XA_ATOM, 32 /* bits */, PropModeReplace, (uint8_t *) &dndVersion, 1);
+
+	return window;
+}
+
+Display *_UIX11GetDisplay() {
+	return ui.display;
+}
+
+UIWindow *_UIFindWindow(Window window) {
+	UIWindow *w = ui.windows;
+
+	while (w) {
+		if (w->window == window) {
+			return w;
+		}
+
+		w = w->next;
+	}
+
+	return NULL;
+}
+
+void _UIClipboardWriteText(UIWindow *window, char *text) {
+	UI_FREE(ui.pasteText);
+	ui.pasteText = text;
+	XSetSelectionOwner(ui.display, ui.clipboardID, window->window, 0);
+}
+
+char *_UIClipboardReadTextStart(UIWindow *window, size_t *bytes) {
+	Window clipboardOwner = XGetSelectionOwner(ui.display, ui.clipboardID);
+
+	if (clipboardOwner == None) {
+		return NULL;
+	}
+
+	if (_UIFindWindow(clipboardOwner)) {
+		*bytes = strlen(ui.pasteText);
+		char *copy = (char *) UI_MALLOC(*bytes);
+		memcpy(copy, ui.pasteText, *bytes);
+		return copy;
+	}
+
+	XConvertSelection(ui.display, ui.clipboardID, XA_STRING, ui.xSelectionDataID, window->window, CurrentTime);
+	XSync(ui.display, 0);
+	XNextEvent(ui.display, &ui.copyEvent);
+
+	// Hack to get around the fact that PropertyNotify arrives before SelectionNotify.
+	// We need PropertyNotify for incremental transfers.
+	while (ui.copyEvent.type == PropertyNotify) {
+		XNextEvent(ui.display, &ui.copyEvent);
+	}
+
+	if (ui.copyEvent.type == SelectionNotify && ui.copyEvent.xselection.selection == ui.clipboardID && ui.copyEvent.xselection.property) {
+		Atom target;
+		// This `itemAmount` is actually `bytes_after_return`
+		unsigned long size, itemAmount;
+		char *data;
+		int format;
+		XGetWindowProperty(ui.copyEvent.xselection.display, ui.copyEvent.xselection.requestor, ui.copyEvent.xselection.property, 0L, ~0L, 0, 
+				AnyPropertyType, &target, &format, &size, &itemAmount, (unsigned char **) &data);
+
+		// We have to allocate for incremental transfers but we don't have to allocate for non-incremental transfers.
+		// I'm allocating for both here to make _UIClipboardReadTextEnd work the same for both
+		if (target != ui.incrID) {
+			*bytes = size;
+			char *copy = (char *) UI_MALLOC(*bytes);
+			memcpy(copy, data, *bytes);
+			XFree(data);
+			XDeleteProperty(ui.copyEvent.xselection.display, ui.copyEvent.xselection.requestor, ui.copyEvent.xselection.property);
+			return copy;
+		}
+
+		XFree(data);
+		XDeleteProperty(ui.display, ui.copyEvent.xselection.requestor, ui.copyEvent.xselection.property);
+		XSync(ui.display, 0);
+
+		*bytes = 0;
+		char *fullData = NULL;
+
+		while (true) {
+			// TODO Timeout.
+			XNextEvent(ui.display, &ui.copyEvent);
+
+			if (ui.copyEvent.type == PropertyNotify) {
+				// The other case - PropertyDelete would be caused by us and can be ignored
+				if (ui.copyEvent.xproperty.state == PropertyNewValue) {
+					unsigned long chunkSize;
+
+					// Note that this call deletes the property.
+					XGetWindowProperty(ui.display, ui.copyEvent.xproperty.window, ui.copyEvent.xproperty.atom, 0L, ~0L, 
+						True, AnyPropertyType, &target, &format, &chunkSize, &itemAmount, (unsigned char **) &data);
+					
+					if (chunkSize == 0) {
+						return fullData;
+					} else {
+						ptrdiff_t currentOffset = *bytes;
+						*bytes += chunkSize;
+						fullData = (char *) UI_REALLOC(fullData, *bytes);
+						memcpy(fullData + currentOffset, data, chunkSize);
+					}
+
+					XFree(data);
+				}
+			}
+		}
+	} else {
+		// TODO What should happen in this case? Is the next event always going to be the selection event?
+		return NULL;
+	}
+}
+
+void _UIClipboardReadTextEnd(UIWindow *window, char *text) {
+	if (text) {
+		//XFree(text);
+		//XDeleteProperty(ui.copyEvent.xselection.display, ui.copyEvent.xselection.requestor, ui.copyEvent.xselection.property);
+		UI_FREE(text);
+	}
+}
+
+void UIInitialise() {
+	_UIInitialiseCommon();
+
+	XInitThreads();
+
+	ui.display = XOpenDisplay(NULL);
+	ui.visual = XDefaultVisual(ui.display, 0);
+
+	ui.windowClosedID = XInternAtom(ui.display, "WM_DELETE_WINDOW", 0);
+	ui.primaryID = XInternAtom(ui.display, "PRIMARY", 0);
+	ui.dndEnterID = XInternAtom(ui.display, "XdndEnter", 0);
+	ui.dndPositionID = XInternAtom(ui.display, "XdndPosition", 0);
+	ui.dndStatusID = XInternAtom(ui.display, "XdndStatus", 0);
+	ui.dndActionCopyID = XInternAtom(ui.display, "XdndActionCopy", 0);
+	ui.dndDropID = XInternAtom(ui.display, "XdndDrop", 0);
+	ui.dndSelectionID = XInternAtom(ui.display, "XdndSelection", 0);
+	ui.dndFinishedID = XInternAtom(ui.display, "XdndFinished", 0);
+	ui.dndAwareID = XInternAtom(ui.display, "XdndAware", 0);
+	ui.uriListID = XInternAtom(ui.display, "text/uri-list", 0);
+	ui.plainTextID = XInternAtom(ui.display, "text/plain", 0);
+	ui.clipboardID = XInternAtom(ui.display, "CLIPBOARD", 0);
+	ui.xSelectionDataID = XInternAtom(ui.display, "XSEL_DATA", 0);
+	ui.textID = XInternAtom(ui.display, "TEXT", 0);
+	ui.targetID = XInternAtom(ui.display, "TARGETS", 0);
+	ui.incrID = XInternAtom(ui.display, "INCR", 0);
+
+	ui.cursors[UI_CURSOR_ARROW] = XCreateFontCursor(ui.display, XC_left_ptr);
+	ui.cursors[UI_CURSOR_TEXT] = XCreateFontCursor(ui.display, XC_xterm);
+	ui.cursors[UI_CURSOR_SPLIT_V] = XCreateFontCursor(ui.display, XC_sb_v_double_arrow);
+	ui.cursors[UI_CURSOR_SPLIT_H] = XCreateFontCursor(ui.display, XC_sb_h_double_arrow);
+	ui.cursors[UI_CURSOR_FLIPPED_ARROW] = XCreateFontCursor(ui.display, XC_right_ptr);
+	ui.cursors[UI_CURSOR_CROSS_HAIR] = XCreateFontCursor(ui.display, XC_crosshair);
+	ui.cursors[UI_CURSOR_HAND] = XCreateFontCursor(ui.display, XC_hand1);
+	ui.cursors[UI_CURSOR_RESIZE_UP] = XCreateFontCursor(ui.display, XC_top_side);
+	ui.cursors[UI_CURSOR_RESIZE_LEFT] = XCreateFontCursor(ui.display, XC_left_side);
+	ui.cursors[UI_CURSOR_RESIZE_UP_RIGHT] = XCreateFontCursor(ui.display, XC_top_right_corner);
+	ui.cursors[UI_CURSOR_RESIZE_UP_LEFT] = XCreateFontCursor(ui.display, XC_top_left_corner);
+	ui.cursors[UI_CURSOR_RESIZE_DOWN] = XCreateFontCursor(ui.display, XC_bottom_side);
+	ui.cursors[UI_CURSOR_RESIZE_RIGHT] = XCreateFontCursor(ui.display, XC_right_side);
+	ui.cursors[UI_CURSOR_RESIZE_DOWN_LEFT] = XCreateFontCursor(ui.display, XC_bottom_left_corner);
+	ui.cursors[UI_CURSOR_RESIZE_DOWN_RIGHT] = XCreateFontCursor(ui.display, XC_bottom_right_corner);
+
+	XSetLocaleModifiers("");
+
+	ui.xim = XOpenIM(ui.display, 0, 0, 0);
+
+	if(!ui.xim){
+		XSetLocaleModifiers("@im=none");
+		ui.xim = XOpenIM(ui.display, 0, 0, 0);
+	}
+}
+
+void _UIWindowSetCursor(UIWindow *window, int cursor) {
+	XDefineCursor(ui.display, window->window, ui.cursors[cursor]);
+}
+
+void _UIX11ResetCursor(UIWindow *window) {
+	XDefineCursor(ui.display, window->window, ui.cursors[UI_CURSOR_ARROW]);
+}
+
+void _UIWindowEndPaint(UIWindow *window, UIPainter *painter) {
+	(void) painter;
+
+	XPutImage(ui.display, window->window, DefaultGC(ui.display, 0), window->image, 
+		UI_RECT_TOP_LEFT(window->updateRegion), UI_RECT_TOP_LEFT(window->updateRegion),
+		UI_RECT_SIZE(window->updateRegion));
+}
+
+void _UIWindowGetScreenPosition(UIWindow *window, int *_x, int *_y) {
+	Window child;
+	XTranslateCoordinates(ui.display, window->window, DefaultRootWindow(ui.display), 0, 0, _x, _y, &child);
+}
+
+void UIMenuShow(UIMenu *menu) {
+	int width, height;
+	_UIMenuPrepare(menu, &width, &height);
+
+	for (int i = 0; i < ScreenCount(ui.display); i++) {
+		Screen *screen = ScreenOfDisplay(ui.display, i);
+
+		int x, y;
+		Window child;
+		XTranslateCoordinates(ui.display, screen->root, DefaultRootWindow(ui.display), 0, 0, &x, &y, &child);
+
+		if (menu->pointX >= x && menu->pointX < x + screen->width 
+				&& menu->pointY >= y && menu->pointY < y + screen->height) {
+			if (menu->pointX + width > x + screen->width) menu->pointX = x + screen->width - width;
+			if (menu->pointY + height > y + screen->height) menu->pointY = y + screen->height - height;
+			if (menu->pointX < x) menu->pointX = x;
+			if (menu->pointY < y) menu->pointY = y;
+			if (menu->pointX + width > x + screen->width) width = x + screen->width - menu->pointX;
+			if (menu->pointY + height > y + screen->height) height = y + screen->height - menu->pointY;
+			break;
+		}
+	}
+
+	Atom properties[] = {
+		XInternAtom(ui.display, "_NET_WM_WINDOW_TYPE", true),
+		XInternAtom(ui.display, "_NET_WM_WINDOW_TYPE_DROPDOWN_MENU", true),
+		XInternAtom(ui.display, "_MOTIF_WM_HINTS", true),
+	};
+
+	XChangeProperty(ui.display, menu->e.window->window, properties[0], XA_ATOM, 32, PropModeReplace, (uint8_t *) properties, 2);
+	XSetTransientForHint(ui.display, menu->e.window->window, DefaultRootWindow(ui.display));
+
+	struct Hints {
+		int flags;
+		int functions;
+		int decorations;
+		int inputMode;
+		int status;
+	};
+
+	struct Hints hints = { 0 };
+	hints.flags = 2;
+	XChangeProperty(ui.display, menu->e.window->window, properties[2], properties[2], 32, PropModeReplace, (uint8_t *) &hints, 5);
+
+	XMapWindow(ui.display, menu->e.window->window);
+	XMoveResizeWindow(ui.display, menu->e.window->window, menu->pointX, menu->pointY, width, height);
+}
+
+void UIWindowPack(UIWindow *window, int _width) {
+	int width = _width ? _width : UIElementMessage(window->e.children, UI_MSG_GET_WIDTH, 0, 0);
+	int height = UIElementMessage(window->e.children, UI_MSG_GET_HEIGHT, width, 0);
+	XResizeWindow(ui.display, window->window, width, height);
+}
+
+bool _UIProcessEvent(XEvent *event) {
+	if (event->type == ClientMessage && (Atom) event->xclient.data.l[0] == ui.windowClosedID) {
+		UIWindow *window = _UIFindWindow(event->xclient.window);
+		if (!window) return false;
+		bool exit = !UIElementMessage(&window->e, UI_MSG_WINDOW_CLOSE, 0, 0);
+		if (exit) return true;
+		_UIUpdate();
+		return false;
+	} else if (event->type == Expose) {
+		UIWindow *window = _UIFindWindow(event->xexpose.window);
+		if (!window) return false;
+		XPutImage(ui.display, window->window, DefaultGC(ui.display, 0), window->image, 0, 0, 0, 0, window->width, window->height);
+	} else if (event->type == ConfigureNotify) {
+		UIWindow *window = _UIFindWindow(event->xconfigure.window);
+		if (!window) return false;
+
+		if (window->width != event->xconfigure.width || window->height != event->xconfigure.height) {
+			window->width = event->xconfigure.width;
+			window->height = event->xconfigure.height;
+			window->bits = (uint32_t *) UI_REALLOC(window->bits, window->width * window->height * 4);
+			window->image->width = window->width;
+			window->image->height = window->height;
+			window->image->bytes_per_line = window->width * 4;
+			window->image->data = (char *) window->bits;
+			window->e.bounds = UI_RECT_2S(window->width, window->height);
+			window->e.clip = UI_RECT_2S(window->width, window->height);
+#ifdef UI_DEBUG
+			for (int i = 0; i < window->width * window->height; i++) window->bits[i] = 0xFF00FF;
+#endif
+			UIElementMessage(&window->e, UI_MSG_LAYOUT, 0, 0);
+			_UIUpdate();
+		}
+	} else if (event->type == MotionNotify) {
+		UIWindow *window = _UIFindWindow(event->xmotion.window);
+		if (!window) return false;
+		window->cursorX = event->xmotion.x;
+		window->cursorY = event->xmotion.y;
+		_UIWindowInputEvent(window, UI_MSG_MOUSE_MOVE, 0, 0);
+	} else if (event->type == LeaveNotify) {
+		UIWindow *window = _UIFindWindow(event->xcrossing.window);
+		if (!window) return false;
+
+		if (!window->pressed) {
+			window->cursorX = -1;
+			window->cursorY = -1;
+		}
+
+		_UIWindowInputEvent(window, UI_MSG_MOUSE_MOVE, 0, 0);
+	} else if (event->type == ButtonPress || event->type == ButtonRelease) {
+		UIWindow *window = _UIFindWindow(event->xbutton.window);
+		if (!window) return false;
+		window->cursorX = event->xbutton.x;
+		window->cursorY = event->xbutton.y;
+
+		if (event->xbutton.button >= 1 && event->xbutton.button <= 3) {
+			_UIWindowInputEvent(window, (UIMessage) ((event->type == ButtonPress ? UI_MSG_LEFT_DOWN : UI_MSG_LEFT_UP) 
+				+ event->xbutton.button * 2 - 2), 0, 0);
+		} else if (event->xbutton.button == 4) {
+			_UIWindowInputEvent(window, UI_MSG_MOUSE_WHEEL, -72, 0);
+		} else if (event->xbutton.button == 5) {
+			_UIWindowInputEvent(window, UI_MSG_MOUSE_WHEEL, 72, 0);
+		}
+
+		_UIInspectorSetFocusedWindow(window);
+	} else if (event->type == KeyPress) {
+		UIWindow *window = _UIFindWindow(event->xkey.window);
+		if (!window) return false;
+
+		if (event->xkey.x == 0x7123 && event->xkey.y == 0x7456) {
+			// HACK! See UIWindowPostMessage.
+			UIElementMessage(&window->e, (UIMessage) event->xkey.state, 0, 
+				(void *) (((uintptr_t) (event->xkey.time & 0xFFFFFFFF) << 32) 
+					| ((uintptr_t) (event->xkey.x_root & 0xFFFF) << 0) 
+					| ((uintptr_t) (event->xkey.y_root & 0xFFFF) << 16)));
+			_UIUpdate();
+		} else {
+			char text[32];
+			KeySym symbol = NoSymbol;
+			Status status;
+			// printf("%ld, %s\n", symbol, text);
+			UIKeyTyped m = { 0 };
+			m.textBytes = Xutf8LookupString(window->xic, &event->xkey, text, sizeof(text) - 1, &symbol, &status); 
+			m.text = text;
+			m.code = XLookupKeysym(&event->xkey, 0);
+
+			if (symbol == XK_Control_L || symbol == XK_Control_R) {
+				window->ctrl = true;
+				window->ctrlCode = event->xkey.keycode;
+				_UIWindowInputEvent(window, UI_MSG_MOUSE_MOVE, 0, 0);
+			} else if (symbol == XK_Shift_L || symbol == XK_Shift_R) {
+				window->shift = true;
+				window->shiftCode = event->xkey.keycode;
+				_UIWindowInputEvent(window, UI_MSG_MOUSE_MOVE, 0, 0);
+			} else if (symbol == XK_Alt_L || symbol == XK_Alt_R) {
+				window->alt = true;
+				window->altCode = event->xkey.keycode;
+				_UIWindowInputEvent(window, UI_MSG_MOUSE_MOVE, 0, 0);
+			} else if (symbol == XK_KP_Left) {
+				m.code = UI_KEYCODE_LEFT;
+			} else if (symbol == XK_KP_Right) {
+				m.code = UI_KEYCODE_RIGHT;
+			} else if (symbol == XK_KP_Up) {
+				m.code = UI_KEYCODE_UP;
+			} else if (symbol == XK_KP_Down) {
+				m.code = UI_KEYCODE_DOWN;
+			} else if (symbol == XK_KP_Home) {
+				m.code = UI_KEYCODE_HOME;
+			} else if (symbol == XK_KP_End) {
+				m.code = UI_KEYCODE_END;
+			} else if (symbol == XK_KP_Enter) {
+				m.code = UI_KEYCODE_ENTER;
+			} else if (symbol == XK_KP_Delete) {
+				m.code = UI_KEYCODE_DELETE;
+			}
+
+			_UIWindowInputEvent(window, UI_MSG_KEY_TYPED, 0, &m);
+		}
+	} else if (event->type == KeyRelease) {
+		UIWindow *window = _UIFindWindow(event->xkey.window);
+		if (!window) return false;
+
+		if (event->xkey.keycode == window->ctrlCode) {
+			window->ctrl = false;
+			_UIWindowInputEvent(window, UI_MSG_MOUSE_MOVE, 0, 0);
+		} else if (event->xkey.keycode == window->shiftCode) {
+			window->shift = false;
+			_UIWindowInputEvent(window, UI_MSG_MOUSE_MOVE, 0, 0);
+		} else if (event->xkey.keycode == window->altCode) {
+			window->alt = false;
+			_UIWindowInputEvent(window, UI_MSG_MOUSE_MOVE, 0, 0);
+		}
+	} else if (event->type == FocusIn) {
+		UIWindow *window = _UIFindWindow(event->xfocus.window);
+		if (!window) return false;
+		window->ctrl = window->shift = window->alt = false;
+		UIElementMessage(&window->e, UI_MSG_WINDOW_ACTIVATE, 0, 0);
+	} else if (event->type == ClientMessage && event->xclient.message_type == ui.dndEnterID) {
+		UIWindow *window = _UIFindWindow(event->xclient.window);
+		if (!window) return false;
+		window->dragSource = (Window) event->xclient.data.l[0];
+	} else if (event->type == ClientMessage && event->xclient.message_type == ui.dndPositionID) {
+		UIWindow *window = _UIFindWindow(event->xclient.window);
+		if (!window) return false;
+		XClientMessageEvent m = { 0 };
+		m.type = ClientMessage;
+		m.display = event->xclient.display;
+		m.window = (Window) event->xclient.data.l[0];
+		m.message_type = ui.dndStatusID;
+		m.format = 32;
+		m.data.l[0] = window->window;
+		m.data.l[1] = true;
+		m.data.l[4] = ui.dndActionCopyID;
+		XSendEvent(ui.display, m.window, False, NoEventMask, (XEvent *) &m);
+		XFlush(ui.display);
+	} else if (event->type == ClientMessage && event->xclient.message_type == ui.dndDropID) {
+		UIWindow *window = _UIFindWindow(event->xclient.window);
+		if (!window) return false;
+
+		// TODO Dropping text.
+
+		if (!XConvertSelection(ui.display, ui.dndSelectionID, ui.uriListID, ui.primaryID, window->window, event->xclient.data.l[2])) {
+			XClientMessageEvent m = { 0 };
+			m.type = ClientMessage;
+			m.display = ui.display;
+			m.window = window->dragSource;
+			m.message_type = ui.dndFinishedID;
+			m.format = 32;
+			m.data.l[0] = window->window;
+			m.data.l[1] = 0;
+			m.data.l[2] = ui.dndActionCopyID;
+			XSendEvent(ui.display, m.window, False, NoEventMask, (XEvent *) &m);
+			XFlush(ui.display);
+		}
+	} else if (event->type == SelectionNotify) {
+		UIWindow *window = _UIFindWindow(event->xselection.requestor);
+		if (!window) return false;
+		if (!window->dragSource) return false;
+
+		Atom type = None;
+		int format = 0;
+		uint64_t count = 0, bytesLeft = 0;
+		uint8_t *data = NULL;
+		XGetWindowProperty(ui.display, window->window, ui.primaryID, 0, 65536, False, AnyPropertyType, &type, &format, &count, &bytesLeft, &data);
+
+		if (format == 8 /* bits per character */) {
+			if (event->xselection.target == ui.uriListID) {
+				char *copy = (char *) UI_MALLOC(count);
+				int fileCount = 0;
+
+				for (int i = 0; i < (int) count; i++) {
+					copy[i] = data[i];
+
+					if (i && data[i - 1] == '\r' && data[i] == '\n') {
+						fileCount++;
+					}
+				}
+
+				char **files = (char **) UI_MALLOC(sizeof(char *) * fileCount);
+				fileCount = 0;
+
+				for (int i = 0; i < (int) count; i++) {
+					char *s = copy + i;
+					while (!(i && data[i - 1] == '\r' && data[i] == '\n' && i < (int) count)) i++;
+					copy[i - 1] = 0;
+
+					for (int j = 0; s[j]; j++) {
+						if (s[j] == '%' && s[j + 1] && s[j + 2]) {
+							char n[3];
+							n[0] = s[j + 1], n[1] = s[j + 2], n[2] = 0;
+							s[j] = strtol(n, NULL, 16);
+							if (!s[j]) break;
+							memmove(s + j + 1, s + j + 3, strlen(s) - j - 2);
+						}
+					}
+
+					if (s[0] == 'f' && s[1] == 'i' && s[2] == 'l' && s[3] == 'e' && s[4] == ':' && s[5] == '/' && s[6] == '/') {
+						files[fileCount++] = s + 7;
+					}
+				}
+
+				UIElementMessage(&window->e, UI_MSG_WINDOW_DROP_FILES, fileCount, files);
+
+				UI_FREE(files);
+				UI_FREE(copy);
+			} else if (event->xselection.target == ui.plainTextID) {
+				// TODO.
+			}
+		}
+
+		XFree(data);
+
+		XClientMessageEvent m = { 0 };
+		m.type = ClientMessage;
+		m.display = ui.display;
+		m.window = window->dragSource;
+		m.message_type = ui.dndFinishedID;
+		m.format = 32;
+		m.data.l[0] = window->window;
+		m.data.l[1] = true;
+		m.data.l[2] = ui.dndActionCopyID;
+		XSendEvent(ui.display, m.window, False, NoEventMask, (XEvent *) &m);
+		XFlush(ui.display);
+
+		window->dragSource = 0; // Drag complete.
+		_UIUpdate();
+	} else if (event->type == SelectionRequest) {
+		UIWindow *window = _UIFindWindow(event->xclient.window);
+		if (!window) return false;
+
+		if ((XGetSelectionOwner(ui.display, ui.clipboardID) == window->window) 
+				&& (event->xselectionrequest.selection == ui.clipboardID)) {
+			XSelectionRequestEvent requestEvent = event->xselectionrequest;
+			Atom utf8ID = XInternAtom(ui.display, "UTF8_STRING", 1);
+			if (utf8ID == None) utf8ID = XA_STRING;
+
+			Atom type = requestEvent.target;
+			type = (type == ui.textID) ? XA_STRING : type;
+			int changePropertyResult = 0;
+
+			if(requestEvent.target == XA_STRING || requestEvent.target == ui.textID || requestEvent.target == utf8ID) {
+				changePropertyResult = XChangeProperty(requestEvent.display, requestEvent.requestor, requestEvent.property, 
+						type, 8, PropModeReplace, (const unsigned char *) ui.pasteText, strlen(ui.pasteText));
+			} else if (requestEvent.target == ui.targetID) {
+				changePropertyResult = XChangeProperty(requestEvent.display, requestEvent.requestor, requestEvent.property, 
+						XA_ATOM, 32, PropModeReplace, (unsigned char *) &utf8ID, 1);
+			}
+
+			if(changePropertyResult == 0 || changePropertyResult == 1) {
+				XSelectionEvent sendEvent = {
+					.type = SelectionNotify,
+					.serial = requestEvent.serial,
+					.send_event = requestEvent.send_event,
+					.display = requestEvent.display,
+					.requestor = requestEvent.requestor,
+					.selection = requestEvent.selection,
+					.target = requestEvent.target,
+					.property = requestEvent.property,
+					.time = requestEvent.time
+				};
+
+				XSendEvent(ui.display, requestEvent.requestor, 0, 0, (XEvent *) &sendEvent);
+			}
+		}
+	}
+
+	return false;
+}
+
+bool _UIMessageLoopSingle(int *result) {
+	XEvent events[64];
+
+	if (ui.animating) {
+		if (XPending(ui.display)) {
+			XNextEvent(ui.display, events + 0);
+		} else {
+			_UIProcessAnimations();
+			return true;
+		}
+	} else {
+		XNextEvent(ui.display, events + 0);
+	}
+
+	int p = 1;
+
+	int configureIndex = -1, motionIndex = -1, exposeIndex = -1;
+
+	while (p < 64 && XPending(ui.display)) {
+		XNextEvent(ui.display, events + p);
+
+#define _UI_MERGE_EVENTS(a, b) \
+	if (events[p].type == a) { \
+		if (b != -1) events[b].type = 0; \
+		b = p; \
+	}
+
+		_UI_MERGE_EVENTS(ConfigureNotify, configureIndex);
+		_UI_MERGE_EVENTS(MotionNotify, motionIndex);
+		_UI_MERGE_EVENTS(Expose, exposeIndex);
+
+		p++;
+	}
+
+	for (int i = 0; i < p; i++) {
+		if (!events[i].type) {
+			continue;
+		}
+
+		if (_UIProcessEvent(events + i)) {
+			return false;
+		}
+	}
+
+	return true;
+}
+
+void UIWindowPostMessage(UIWindow *window, UIMessage message, void *_dp) {
+	// HACK! Xlib doesn't seem to have a nice way to do this,
+	// so send a specially crafted key press event instead.
+	// TODO Maybe ClientMessage is what this should use?
+	uintptr_t dp = (uintptr_t) _dp;
+	XKeyEvent event = { 0 };
+	event.display = ui.display;
+	event.window = window->window;
+	event.root = DefaultRootWindow(ui.display);
+	event.subwindow = None;
+	event.time = dp >> 32;
+	event.x = 0x7123;
+	event.y = 0x7456;
+	event.x_root = (dp >> 0) & 0xFFFF;
+	event.y_root = (dp >> 16) & 0xFFFF;
+	event.same_screen = True;
+	event.keycode = 1;
+	event.state = message;
+	event.type = KeyPress;
+	XSendEvent(ui.display, window->window, True, KeyPressMask, (XEvent *) &event);
+	XFlush(ui.display);
+}
+
+#endif
+
+#ifdef UI_WINDOWS
+
+const int UI_KEYCODE_A = 'A';
+const int UI_KEYCODE_0 = '0';
+const int UI_KEYCODE_BACKSPACE = VK_BACK;
+const int UI_KEYCODE_DELETE = VK_DELETE;
+const int UI_KEYCODE_DOWN = VK_DOWN;
+const int UI_KEYCODE_END = VK_END;
+const int UI_KEYCODE_ENTER = VK_RETURN;
+const int UI_KEYCODE_ESCAPE = VK_ESCAPE;
+const int UI_KEYCODE_F1 = VK_F1;
+const int UI_KEYCODE_HOME = VK_HOME;
+const int UI_KEYCODE_LEFT = VK_LEFT;
+const int UI_KEYCODE_RIGHT = VK_RIGHT;
+const int UI_KEYCODE_SPACE = VK_SPACE;
+const int UI_KEYCODE_TAB = VK_TAB;
+const int UI_KEYCODE_UP = VK_UP;
+const int UI_KEYCODE_INSERT = VK_INSERT;
+
+int _UIWindowMessage(UIElement *element, UIMessage message, int di, void *dp) {
+	if (message == UI_MSG_DESTROY) {
+		UIWindow *window = (UIWindow *) element;
+		_UIWindowDestroyCommon(window);
+		SetWindowLongPtr(window->hwnd, GWLP_USERDATA, 0);
+		DestroyWindow(window->hwnd);
+	}
+
+	return _UIWindowMessageCommon(element, message, di, dp);
+}
+
+LRESULT CALLBACK _UIWindowProcedure(HWND hwnd, UINT message, WPARAM wParam, LPARAM lParam) {
+	UIWindow *window = (UIWindow *) GetWindowLongPtr(hwnd, GWLP_USERDATA);
+
+	if (!window || ui.assertionFailure) {
+		return DefWindowProc(hwnd, message, wParam, lParam);
+	}
+
+	if (message == WM_CLOSE) {
+		if (UIElementMessage(&window->e, UI_MSG_WINDOW_CLOSE, 0, 0)) {
+			_UIUpdate();
+			return 0;
+		} else {
+			PostQuitMessage(0);
+		}
+	} else if (message == WM_SIZE) {
+		RECT client;
+		GetClientRect(hwnd, &client);
+		window->width = client.right;
+		window->height = client.bottom;
+		window->bits = (uint32_t *) UI_REALLOC(window->bits, window->width * window->height * 4);
+		window->e.bounds = UI_RECT_2S(window->width, window->height);
+		window->e.clip = UI_RECT_2S(window->width, window->height);
+		UIElementMessage(&window->e, UI_MSG_LAYOUT, 0, 0);
+		_UIUpdate();
+	} else if (message == WM_MOUSEMOVE) {
+		if (!window->trackingLeave) {
+			window->trackingLeave = true;
+			TRACKMOUSEEVENT leave = { 0 };
+			leave.cbSize = sizeof(TRACKMOUSEEVENT);
+			leave.dwFlags = TME_LEAVE;
+			leave.hwndTrack = hwnd;
+			TrackMouseEvent(&leave);
+		}
+
+		POINT cursor;
+		GetCursorPos(&cursor);
+		ScreenToClient(hwnd, &cursor);
+		window->cursorX = cursor.x;
+		window->cursorY = cursor.y;
+		_UIWindowInputEvent(window, UI_MSG_MOUSE_MOVE, 0, 0);
+	} else if (message == WM_MOUSELEAVE) {
+		window->trackingLeave = false;
+
+		if (!window->pressed) {
+			window->cursorX = -1;
+			window->cursorY = -1;
+		}
+
+		_UIWindowInputEvent(window, UI_MSG_MOUSE_MOVE, 0, 0);
+	} else if (message == WM_LBUTTONDOWN) {
+		SetCapture(hwnd);
+		_UIWindowInputEvent(window, UI_MSG_LEFT_DOWN, 0, 0);
+	} else if (message == WM_LBUTTONUP) {
+		if (window->pressedButton == 1) ReleaseCapture();
+		_UIWindowInputEvent(window, UI_MSG_LEFT_UP, 0, 0);
+	} else if (message == WM_MBUTTONDOWN) {
+		SetCapture(hwnd);
+		_UIWindowInputEvent(window, UI_MSG_MIDDLE_DOWN, 0, 0);
+	} else if (message == WM_MBUTTONUP) {
+		if (window->pressedButton == 2) ReleaseCapture();
+		_UIWindowInputEvent(window, UI_MSG_MIDDLE_UP, 0, 0);
+	} else if (message == WM_RBUTTONDOWN) {
+		SetCapture(hwnd);
+		_UIWindowInputEvent(window, UI_MSG_RIGHT_DOWN, 0, 0);
+	} else if (message == WM_RBUTTONUP) {
+		if (window->pressedButton == 3) ReleaseCapture();
+		_UIWindowInputEvent(window, UI_MSG_RIGHT_UP, 0, 0);
+	} else if (message == WM_MOUSEWHEEL) {
+		int delta = (int) wParam >> 16;
+		_UIWindowInputEvent(window, UI_MSG_MOUSE_WHEEL, -delta, 0);
+	} else if (message == WM_KEYDOWN) {
+		window->ctrl = GetKeyState(VK_CONTROL) & 0x8000;
+		window->shift = GetKeyState(VK_SHIFT) & 0x8000;
+		window->alt = GetKeyState(VK_MENU) & 0x8000;
+
+		UIKeyTyped m = { 0 };
+		m.code = wParam;
+		_UIWindowInputEvent(window, UI_MSG_KEY_TYPED, 0, &m);
+	} else if (message == WM_CHAR) {
+		UIKeyTyped m = { 0 };
+		char c = wParam;
+		m.text = &c;
+		m.textBytes = 1;
+		_UIWindowInputEvent(window, UI_MSG_KEY_TYPED, 0, &m);
+	} else if (message == WM_PAINT) {
+		PAINTSTRUCT paint;
+		HDC dc = BeginPaint(hwnd, &paint);
+		BITMAPINFOHEADER info = { 0 };
+		info.biSize = sizeof(info);
+		info.biWidth = window->width, info.biHeight = -window->height;
+		info.biPlanes = 1, info.biBitCount = 32;
+		StretchDIBits(dc, 0, 0, UI_RECT_SIZE(window->e.bounds), 0, 0, UI_RECT_SIZE(window->e.bounds),
+			window->bits, (BITMAPINFO *) &info, DIB_RGB_COLORS, SRCCOPY);
+		EndPaint(hwnd, &paint);
+	} else if (message == WM_SETCURSOR && LOWORD(lParam) == HTCLIENT) {
+		SetCursor(ui.cursors[window->cursorStyle]);
+		return 1;
+	} else if (message == WM_SETFOCUS || message == WM_KILLFOCUS) {
+		_UIMenusClose();
+
+		if (message == WM_SETFOCUS) {
+			_UIInspectorSetFocusedWindow(window);
+			UIElementMessage(&window->e, UI_MSG_WINDOW_ACTIVATE, 0, 0);
+		}
+	} else if (message == WM_MOUSEACTIVATE && (window->e.flags & UI_WINDOW_MENU)) {
+		return MA_NOACTIVATE;
+	} else if (message == WM_DROPFILES) {
+		HDROP drop = (HDROP) wParam;
+		int count = DragQueryFile(drop, 0xFFFFFFFF, NULL, 0);
+		char **files = (char **) UI_MALLOC(sizeof(char *) * count);
+		
+		for (int i = 0; i < count; i++) {
+			int length = DragQueryFile(drop, i, NULL, 0);
+			files[i] = (char *) UI_MALLOC(length + 1);
+			files[i][length] = 0;
+			DragQueryFile(drop, i, files[i], length + 1);
+		}
+		
+		UIElementMessage(&window->e, UI_MSG_WINDOW_DROP_FILES, count, files);
+		for (int i = 0; i < count; i++) UI_FREE(files[i]);		
+		UI_FREE(files);
+		DragFinish(drop);
+		_UIUpdate();
+	} else if (message == WM_APP + 1) {
+		UIElementMessage(&window->e, (UIMessage) wParam, 0, (void *) lParam);
+		_UIUpdate();
+	} else {
+		if (message == WM_NCLBUTTONDOWN || message == WM_NCMBUTTONDOWN || message == WM_NCRBUTTONDOWN) {
+			if (~window->e.flags & UI_WINDOW_MENU) {
+				_UIMenusClose();
+				_UIUpdate();
+			}
+		}
+
+		return DefWindowProc(hwnd, message, wParam, lParam);
+	}
+
+	return 0;
+}
+
+void UIInitialise() {
+	ui.heap = GetProcessHeap();
+	
+	_UIInitialiseCommon();
+
+	ui.cursors[UI_CURSOR_ARROW] = LoadCursor(NULL, IDC_ARROW);
+	ui.cursors[UI_CURSOR_TEXT] = LoadCursor(NULL, IDC_IBEAM);
+	ui.cursors[UI_CURSOR_SPLIT_V] = LoadCursor(NULL, IDC_SIZENS);
+	ui.cursors[UI_CURSOR_SPLIT_H] = LoadCursor(NULL, IDC_SIZEWE);
+	ui.cursors[UI_CURSOR_FLIPPED_ARROW] = LoadCursor(NULL, IDC_ARROW);
+	ui.cursors[UI_CURSOR_CROSS_HAIR] = LoadCursor(NULL, IDC_CROSS);
+	ui.cursors[UI_CURSOR_HAND] = LoadCursor(NULL, IDC_HAND);
+	ui.cursors[UI_CURSOR_RESIZE_UP] = LoadCursor(NULL, IDC_SIZENS);
+	ui.cursors[UI_CURSOR_RESIZE_LEFT] = LoadCursor(NULL, IDC_SIZEWE);
+	ui.cursors[UI_CURSOR_RESIZE_UP_RIGHT] = LoadCursor(NULL, IDC_SIZENESW);
+	ui.cursors[UI_CURSOR_RESIZE_UP_LEFT] = LoadCursor(NULL, IDC_SIZENWSE);
+	ui.cursors[UI_CURSOR_RESIZE_DOWN] = LoadCursor(NULL, IDC_SIZENS);
+	ui.cursors[UI_CURSOR_RESIZE_RIGHT] = LoadCursor(NULL, IDC_SIZEWE);
+	ui.cursors[UI_CURSOR_RESIZE_DOWN_LEFT] = LoadCursor(NULL, IDC_SIZENESW);
+	ui.cursors[UI_CURSOR_RESIZE_DOWN_RIGHT] = LoadCursor(NULL, IDC_SIZENWSE);
+
+	WNDCLASS windowClass = { 0 };
+	windowClass.lpfnWndProc = _UIWindowProcedure;
+	windowClass.lpszClassName = "normal";
+	RegisterClass(&windowClass);
+	windowClass.style |= CS_DROPSHADOW;
+	windowClass.lpszClassName = "shadow";
+	RegisterClass(&windowClass);
+}
+
+bool _UIMessageLoopSingle(int *result) {
+	MSG message = { 0 };
+
+	if (ui.animating) {
+		if (PeekMessage(&message, NULL, 0, 0, PM_REMOVE)) {
+			if (message.message == WM_QUIT) {
+				*result = message.wParam;
+				return false;
+			}
+
+			TranslateMessage(&message);
+			DispatchMessage(&message);
+		} else {
+			_UIProcessAnimations();
+		}
+	} else {
+		if (!GetMessage(&message, NULL, 0, 0)) {
+			*result = message.wParam;
+			return false;
+		}
+
+		TranslateMessage(&message);
+		DispatchMessage(&message);
+	}
+
+	return true;
+}
+
+void UIMenuShow(UIMenu *menu) {
+	int width, height;
+	_UIMenuPrepare(menu, &width, &height);
+	MoveWindow(menu->e.window->hwnd, menu->pointX, menu->pointY, width, height, FALSE);
+	ShowWindow(menu->e.window->hwnd, SW_SHOWNOACTIVATE);
+}
+
+UIWindow *UIWindowCreate(UIWindow *owner, uint32_t flags, const char *cTitle, int width, int height) {
+	_UIMenusClose();
+
+	UIWindow *window = (UIWindow *) UIElementCreate(sizeof(UIWindow), NULL, flags | UI_ELEMENT_WINDOW, _UIWindowMessage, "Window");
+	_UIWindowAdd(window);
+	if (owner) window->scale = owner->scale;
+
+	if (flags & UI_WINDOW_MENU) {
+		UI_ASSERT(owner);
+
+		window->hwnd = CreateWindowEx(WS_EX_TOPMOST | WS_EX_NOACTIVATE, "shadow", 0, WS_POPUP, 
+			0, 0, 0, 0, owner->hwnd, NULL, NULL, NULL);
+	} else {
+		window->hwnd = CreateWindowEx(WS_EX_ACCEPTFILES, "normal", cTitle, WS_OVERLAPPEDWINDOW, 
+			CW_USEDEFAULT, CW_USEDEFAULT, width ? width : CW_USEDEFAULT, height ? height : CW_USEDEFAULT,
+			owner ? owner->hwnd : NULL, NULL, NULL, NULL);
+	}
+
+	SetWindowLongPtr(window->hwnd, GWLP_USERDATA, (LONG_PTR) window);
+
+	if (~flags & UI_WINDOW_MENU) {
+		ShowWindow(window->hwnd, SW_SHOW);
+		PostMessage(window->hwnd, WM_SIZE, 0, 0);
+	}
+
+	return window;
+}
+
+void _UIWindowEndPaint(UIWindow *window, UIPainter *painter) {
+	HDC dc = GetDC(window->hwnd);
+	BITMAPINFOHEADER info = { 0 };
+	info.biSize = sizeof(info);
+	info.biWidth = window->width, info.biHeight = window->height;
+	info.biPlanes = 1, info.biBitCount = 32;
+	StretchDIBits(dc, 
+		UI_RECT_TOP_LEFT(window->updateRegion), UI_RECT_SIZE(window->updateRegion), 
+		window->updateRegion.l, window->updateRegion.b + 1, 
+		UI_RECT_WIDTH(window->updateRegion), -UI_RECT_HEIGHT(window->updateRegion),
+		window->bits, (BITMAPINFO *) &info, DIB_RGB_COLORS, SRCCOPY);
+	ReleaseDC(window->hwnd, dc);
+}
+
+void _UIWindowSetCursor(UIWindow *window, int cursor) {
+	SetCursor(ui.cursors[cursor]);
+}
+
+void _UIWindowGetScreenPosition(UIWindow *window, int *_x, int *_y) {
+	POINT p;
+	p.x = 0;
+	p.y = 0;
+	ClientToScreen(window->hwnd, &p);
+	*_x = p.x;
+	*_y = p.y;
+}
+
+void UIWindowPostMessage(UIWindow *window, UIMessage message, void *_dp) {
+	PostMessage(window->hwnd, WM_APP + 1, (WPARAM) message, (LPARAM) _dp);
+}
+
+void *_UIHeapReAlloc(void *pointer, size_t size) {
+	if (pointer) {
+		if (size) {
+			return HeapReAlloc(ui.heap, 0, pointer, size);
+		} else {
+			UI_FREE(pointer);
+			return NULL;
+		}
+	} else {
+		if (size) {
+			return UI_MALLOC(size);
+		} else {
+			return NULL;
+		}
+	}
+}
+
+void _UIClipboardWriteText(UIWindow *window, char *text) {
+	if (OpenClipboard(window->hwnd)) {
+		EmptyClipboard();
+		HGLOBAL memory = GlobalAlloc(GMEM_MOVEABLE | GMEM_ZEROINIT, _UIStringLength(text) + 1);
+		char *copy = (char *) GlobalLock(memory);
+		for (uintptr_t i = 0; text[i]; i++) copy[i] = text[i];
+		GlobalUnlock(copy);
+		SetClipboardData(CF_TEXT, memory);
+		CloseClipboard();
+	}
+}
+
+char *_UIClipboardReadTextStart(UIWindow *window, size_t *bytes) {
+	if (!OpenClipboard(window->hwnd)) {
+		return NULL;
+	}
+	
+	HANDLE memory = GetClipboardData(CF_TEXT);
+	
+	if (!memory) {
+		CloseClipboard();
+		return NULL;
+	}
+	
+	char *buffer = (char *) GlobalLock(memory);
+	
+	if (!buffer) {
+		CloseClipboard();
+		return NULL;
+	}
+	
+	size_t byteCount = GlobalSize(memory);
+	
+	if (byteCount < 1) {
+		GlobalUnlock(memory);
+		CloseClipboard();
+		return NULL;
+	}
+
+	char *copy = (char *) UI_MALLOC(byteCount + 1);
+	for (uintptr_t i = 0; i < byteCount; i++) copy[i] = buffer[i];
+	copy[byteCount] = 0; // Just in case.
+	
+	GlobalUnlock(memory);
+	CloseClipboard();
+	
+	if (bytes) *bytes = _UIStringLength(copy);
+	return copy;
+}
+
+void _UIClipboardReadTextEnd(UIWindow *window, char *text) {
+	UI_FREE(text);
+}
+
+#endif
+
+#ifdef UI_ESSENCE
+
+const int UI_KEYCODE_A = ES_SCANCODE_A;
+const int UI_KEYCODE_0 = ES_SCANCODE_0;
+const int UI_KEYCODE_BACKSPACE = ES_SCANCODE_BACKSPACE;
+const int UI_KEYCODE_DELETE = ES_SCANCODE_DELETE;
+const int UI_KEYCODE_DOWN = ES_SCANCODE_DOWN_ARROW;
+const int UI_KEYCODE_END = ES_SCANCODE_END;
+const int UI_KEYCODE_ENTER = ES_SCANCODE_ENTER;
+const int UI_KEYCODE_ESCAPE = ES_SCANCODE_ESCAPE;
+const int UI_KEYCODE_F1 = ES_SCANCODE_F1;
+const int UI_KEYCODE_HOME = ES_SCANCODE_HOME;
+const int UI_KEYCODE_LEFT = ES_SCANCODE_LEFT_ARROW;
+const int UI_KEYCODE_RIGHT = ES_SCANCODE_RIGHT_ARROW;
+const int UI_KEYCODE_SPACE = ES_SCANCODE_SPACE;
+const int UI_KEYCODE_TAB = ES_SCANCODE_TAB;
+const int UI_KEYCODE_UP = ES_SCANCODE_UP_ARROW;
+const int UI_KEYCODE_INSERT = ES_SCANCODE_INSERT;
+
+int _UIWindowMessage(UIElement *element, UIMessage message, int di, void *dp) {
+	if (message == UI_MSG_DESTROY) {
+		// TODO Non-main windows.
+		element->window = NULL;
+		EsInstanceDestroy(ui.instance);
+	}
+
+	return _UIWindowMessageCommon(element, message, di, dp);
+}
+
+void UIInitialise() {
+	_UIInitialiseCommon();
+
+	while (true) {
+		EsMessage *message = EsMessageReceive();
+
+		if (message->type == ES_MSG_INSTANCE_CREATE) {
+			ui.instance = EsInstanceCreate(message, NULL, 0);
+			break;
+		}
+	}
+}
+
+bool _UIMessageLoopSingle(int *result) {
+	if (ui.animating) {
+		// TODO.
+	} else {
+		_UIMessageProcess(EsMessageReceive());
+	}
+
+	return true;
+}
+
+UIMenu *UIMenuCreate(UIElement *parent, uint32_t flags) {
+	ui.menuIndex = 0;
+	return EsMenuCreate(parent->window->window, ES_MENU_AT_CURSOR);
+}
+
+void _UIMenuItemCallback(EsMenu *menu, EsGeneric context) {
+	((void (*)(void *)) ui.menuData[context.u * 2 + 0])(ui.menuData[context.u * 2 + 1]);
+}
+
+void UIMenuAddItem(UIMenu *menu, uint32_t flags, const char *label, ptrdiff_t labelBytes, void (*invoke)(void *cp), void *cp) {
+	EsAssert(ui.menuIndex < 128);
+	ui.menuData[ui.menuIndex * 2 + 0] = (void *) invoke;
+	ui.menuData[ui.menuIndex * 2 + 1] = cp;
+	EsMenuAddItem(menu, (flags & UI_BUTTON_CHECKED) ? ES_MENU_ITEM_CHECKED : ES_FLAGS_DEFAULT, 
+			label, labelBytes, _UIMenuItemCallback, ui.menuIndex);
+	ui.menuIndex++;
+}
+
+void UIMenuShow(UIMenu *menu) {
+	EsMenuShow(menu);
+}
+
+int _UIWindowCanvasMessage(EsElement *element, EsMessage *message) {
+	UIWindow *window = (UIWindow *) element->window->userData.p;
+
+	if (!window) {
+		return 0;
+	} else if (message->type == ES_MSG_PAINT) {
+		EsRectangle bounds = ES_RECT_4PD(message->painter->offsetX, message->painter->offsetY, window->width, window->height);
+		EsDrawBitmap(message->painter, bounds, window->bits, window->width * 4, 0xFFFF);
+	} else if (message->type == ES_MSG_LAYOUT) {
+		EsElementGetSize(element, &window->width, &window->height);
+		window->bits = (uint32_t *) UI_REALLOC(window->bits, window->width * window->height * 4);
+		window->e.bounds = UI_RECT_2S(window->width, window->height);
+		window->e.clip = UI_RECT_2S(window->width, window->height);
+		UIElementMessage(&window->e, UI_MSG_LAYOUT, 0, 0);
+		_UIUpdate();
+	} else if (message->type == ES_MSG_SCROLL_WHEEL) {
+		_UIWindowInputEvent(window, UI_MSG_MOUSE_WHEEL, -message->scrollWheel.dy, 0);
+	} else if (message->type == ES_MSG_MOUSE_MOVED || message->type == ES_MSG_HOVERED_END
+			|| message->type == ES_MSG_MOUSE_LEFT_DRAG || message->type == ES_MSG_MOUSE_RIGHT_DRAG || message->type == ES_MSG_MOUSE_MIDDLE_DRAG) {
+		EsPoint point = EsMouseGetPosition(element); 
+		window->cursorX = point.x, window->cursorY = point.y;
+		_UIWindowInputEvent(window, UI_MSG_MOUSE_MOVE, 0, 0);
+	} else if (message->type == ES_MSG_KEY_UP) {
+		window->ctrl = EsKeyboardIsCtrlHeld();
+		window->shift = EsKeyboardIsShiftHeld();
+		window->alt = EsKeyboardIsAltHeld();
+	} else if (message->type == ES_MSG_KEY_DOWN) {
+		window->ctrl = EsKeyboardIsCtrlHeld();
+		window->shift = EsKeyboardIsShiftHeld();
+		window->alt = EsKeyboardIsAltHeld();
+		UIKeyTyped m = { 0 };
+		char c[64];
+		m.text = c;
+		m.textBytes = EsMessageGetInputText(message, c);
+		m.code = message->keyboard.scancode;
+		return _UIWindowInputEvent(window, UI_MSG_KEY_TYPED, 0, &m) ? ES_HANDLED : 0;
+	} else if (message->type == ES_MSG_MOUSE_LEFT_CLICK) {
+		_UIInspectorSetFocusedWindow(window);
+	} else if (message->type == ES_MSG_USER_START) {
+		UIElementMessage(&window->e, (UIMessage) message->user.context1.u, 0, (void *) message->user.context2.p);
+		_UIUpdate();
+	} else if (message->type == ES_MSG_GET_CURSOR) {
+		message->cursorStyle = ES_CURSOR_NORMAL;
+		if (window->cursor == UI_CURSOR_TEXT)              message->cursorStyle = ES_CURSOR_TEXT;
+		if (window->cursor == UI_CURSOR_SPLIT_V)           message->cursorStyle = ES_CURSOR_SPLIT_VERTICAL;
+		if (window->cursor == UI_CURSOR_SPLIT_H)           message->cursorStyle = ES_CURSOR_SPLIT_HORIZONTAL;
+		if (window->cursor == UI_CURSOR_FLIPPED_ARROW)     message->cursorStyle = ES_CURSOR_SELECT_LINES;
+		if (window->cursor == UI_CURSOR_CROSS_HAIR)        message->cursorStyle = ES_CURSOR_CROSS_HAIR_PICK;
+		if (window->cursor == UI_CURSOR_HAND)              message->cursorStyle = ES_CURSOR_HAND_HOVER;
+		if (window->cursor == UI_CURSOR_RESIZE_UP)         message->cursorStyle = ES_CURSOR_RESIZE_VERTICAL;
+		if (window->cursor == UI_CURSOR_RESIZE_LEFT)       message->cursorStyle = ES_CURSOR_RESIZE_HORIZONTAL;
+		if (window->cursor == UI_CURSOR_RESIZE_UP_RIGHT)   message->cursorStyle = ES_CURSOR_RESIZE_DIAGONAL_1;
+		if (window->cursor == UI_CURSOR_RESIZE_UP_LEFT)    message->cursorStyle = ES_CURSOR_RESIZE_DIAGONAL_2;
+		if (window->cursor == UI_CURSOR_RESIZE_DOWN)       message->cursorStyle = ES_CURSOR_RESIZE_VERTICAL;
+		if (window->cursor == UI_CURSOR_RESIZE_RIGHT)      message->cursorStyle = ES_CURSOR_RESIZE_HORIZONTAL;
+		if (window->cursor == UI_CURSOR_RESIZE_DOWN_RIGHT) message->cursorStyle = ES_CURSOR_RESIZE_DIAGONAL_1;
+		if (window->cursor == UI_CURSOR_RESIZE_DOWN_LEFT)  message->cursorStyle = ES_CURSOR_RESIZE_DIAGONAL_2;
+	}
+	
+	else if (message->type == ES_MSG_MOUSE_LEFT_DOWN)   _UIWindowInputEvent(window, UI_MSG_LEFT_DOWN, 0, 0);
+	else if (message->type == ES_MSG_MOUSE_LEFT_UP)     _UIWindowInputEvent(window, UI_MSG_LEFT_UP, 0, 0);
+	else if (message->type == ES_MSG_MOUSE_MIDDLE_DOWN) _UIWindowInputEvent(window, UI_MSG_MIDDLE_DOWN, 0, 0);
+	else if (message->type == ES_MSG_MOUSE_MIDDLE_UP)   _UIWindowInputEvent(window, UI_MSG_MIDDLE_UP, 0, 0);
+	else if (message->type == ES_MSG_MOUSE_RIGHT_DOWN)  _UIWindowInputEvent(window, UI_MSG_RIGHT_DOWN, 0, 0);
+	else if (message->type == ES_MSG_MOUSE_RIGHT_UP)    _UIWindowInputEvent(window, UI_MSG_RIGHT_UP, 0, 0);
+
+	else return 0;
+
+	return ES_HANDLED;
+}
+
+UIWindow *UIWindowCreate(UIWindow *owner, uint32_t flags, const char *cTitle, int width, int height) {
+	_UIMenusClose();
+
+	UIWindow *window = (UIWindow *) UIElementCreate(sizeof(UIWindow), NULL, flags | UI_ELEMENT_WINDOW, _UIWindowMessage, "Window");
+	_UIWindowAdd(window);
+	if (owner) window->scale = owner->scale;
+
+	if (flags & UI_WINDOW_MENU) {
+		// TODO.
+	} else {
+		// TODO Non-main windows.
+		window->window = ui.instance->window;
+		window->window->userData = window;
+		window->canvas = EsCustomElementCreate(window->window, ES_CELL_FILL | ES_ELEMENT_FOCUSABLE);
+		window->canvas->messageUser = _UIWindowCanvasMessage;
+		EsWindowSetTitle(window->window, cTitle, -1);
+		EsElementFocus(window->canvas);
+	}
+
+	return window;
+}
+
+void _UIWindowEndPaint(UIWindow *window, UIPainter *painter) {
+	EsElementRepaint(window->canvas, &window->updateRegion);
+}
+
+void _UIWindowSetCursor(UIWindow *window, int cursor) {
+	window->cursor = cursor;
+}
+
+void _UIWindowGetScreenPosition(UIWindow *window, int *_x, int *_y) {
+	EsRectangle r = EsElementGetScreenBounds(window->window);
+	*_x = r.l, *_y = r.t;
+}
+
+void UIWindowPostMessage(UIWindow *window, UIMessage message, void *_dp) {
+	EsMessage m = {};
+	m.type = ES_MSG_USER_START;
+	m.user.context1.u = message;
+	m.user.context2.p = _dp;
+	EsMessagePost(window->canvas, &m);
+}
+
+void _UIClipboardWriteText(UIWindow *window, char *text) {
+	EsClipboardAddText(ES_CLIPBOARD_PRIMARY, text, -1);
+	UI_FREE(text);
+}
+
+char *_UIClipboardReadTextStart(UIWindow *window, size_t *bytes) {
+	return EsClipboardReadText(ES_CLIPBOARD_PRIMARY, bytes, NULL);
+}
+
+void _UIClipboardReadTextEnd(UIWindow *window, char *text) {
+	EsHeapFree(text);
+}
+
+#endif
+
+#endif
diff --git a/main.c b/main.c
index 91075f1..c350331 100644
--- a/main.c
+++ b/main.c
@@ -1,67 +1,134 @@
-#include "memory.h"
-#include "plat.h"
-#include "rcache.h"
-#include "ui.h"
+#include <stdio.h>
+#include "config.h"
 #include "library.h"
+#include "luigi.h"
+#include "memory.h"
 
-UI* ui;
-Library* lib;
+Library lib;
 
-Font* ffont;
-void prog_init(void* memory) {
-	Heap* fh;
-	unsigned char* bin;
-	int bin_size;
-	int fhs = 1024 * 1024 * 4;
-	Font* font;
-	memory_init(memory);
-	rc_init();
-	fh = galloc(sizeof *fh);
-	init_heap(fh, galloc(fhs), fhs);
-	bin = load_binary(default_font_location, &bin_size);
-	font = new_font(fh, bin, default_font_size);
-	ui = stack_alloc(sizeof *ui);
-	init_ui(ui, font);
-	lib = stack_alloc(sizeof *lib);
-	init_library(lib);
-	build_library(lib, library_folder);
+int libtab_msg(
+	UIElement *el,
+	UIMessage msg,
+	int di,
+	void *dp
+) {
+	if (msg == UI_MSG_TABLE_GET_ITEM) {
+		UITableGetItem *m = (UITableGetItem*)dp;
+		Song* song;
+		song = &lib.songs[lib.indices[m->index]];
+		m->isSelected = 0;
+		switch (m->column) {
+			case 0:
+				return snprintf(
+					m->buffer,
+					m->bufferBytes,
+					"%s",
+					song->name
+				);
+			case 1:
+				return snprintf(
+					m->buffer,
+					m->bufferBytes,
+					"%s",
+					song->artist
+				);
+			case 2:
+				return snprintf(
+					m->buffer,
+					m->bufferBytes,
+					"%s",
+					song->album
+				);
+			case 3:
+				return snprintf(
+					m->buffer,
+					m->bufferBytes,
+					"%s",
+					song->path
+				);
+			default: return 0;
+		}
+	}
+	return 0;
 }
 
-void prog_update(void) {
-	Rectangle l, r, lc, list;
-	int h, i, c;
-	rc_begin();
-	ui_begin(ui, &l);
-	rc_add_cmd_rect(&l, theme_background_colour);
-
-	r = rectcut_right(&l, l.w / 2);
-	h = font_height(ui->font) + theme_padding * 2;
-	c = lib->song_count;
-	lc.x = 0;
-	lc.y = 0;
-	lc.w = r.w - theme_padding * 2;
-	lc.h = h * c + theme_padding * (c - 1);
-	ui_container(
-		ui,
-		&r,
-		&lc,
-		&lc
+int prog_main(void* mem) {
+	Arena liba;
+	UIWindow* wi;
+	UISplitPane* split1, * split2, * split3;
+	UIPanel* plib, * pctrl, * plist, * pqueue;
+	UITable* libtab;
+	memory_init(mem);
+	init_arena(
+		&liba,
+		galloc(library_memory_size),
+		library_memory_size
+	);
+	build_library(&liba, &lib, library_path);
+	UIInitialise();
+	wi = UIWindowCreate(
+		0,
+		0,
+		app_name,
+		default_window_w,
+		default_window_h
+	);
+	split1 = UISplitPaneCreate(
+		&wi->e,
+		0,
+		0.3f
+	);
+	split3 = UISplitPaneCreate(
+		&split1->e,
+		UI_SPLIT_PANE_VERTICAL,
+		0.5f
 	);
-	list = lc;
-	list.h = h;
-	for (i = 0; i < max_songs; i++) {
-		Song* song = &lib->songs[i];
-		if (song->file[0]) {
-			ui_button(ui, &list, song->name);
-			list.y += list.h + theme_padding;
-		}
-	}
-	rc_add_cmd_reset_clip();
 
-	ui_end(ui);
-	rc_flush();
-}
+	split2 = UISplitPaneCreate(
+		&split1->e,
+		0,
+		0.5f
+	);
+	pctrl  = UIPanelCreate(&split3->e, UI_PANEL_GRAY);
+	pctrl->gap = 5;
+	pctrl->border = UI_RECT_1(5);
+	
+	pqueue = UIPanelCreate(&split3->e, UI_PANEL_GRAY);
+	pqueue->gap = 5;
+	pqueue->border = UI_RECT_1(5);
+	UILabelCreate(
+		&pqueue->e,
+		0,
+		"Queue",
+		5
+	);
 
-void prog_deinit(void) {
+	plist  = UIPanelCreate(&split2->e, UI_PANEL_GRAY);
+	plist->gap = 5;
+	plist->border = UI_RECT_1(5);
+	UILabelCreate(
+		&plist->e,
+		0,
+		"Playlist",
+		8
+	);
+	plib   = UIPanelCreate(&split2->e, UI_PANEL_GRAY);
+	plib->gap = 5;
+	plib->border = UI_RECT_1(5);
+	UILabelCreate(
+		&plib->e,
+		0,
+		"Library",
+		7
+	);
+	libtab = UITableCreate(
+		&plib->e,
+		UI_ELEMENT_H_FILL | UI_ELEMENT_V_FILL,
+		"Track\tArtist\tAlbum\tFilename"
+	);
+	libtab->itemCount = lib.cnt;
+	libtab->e.messageUser = libtab_msg;
+	UITableResizeColumns(libtab);
 
+	return UIMessageLoop();
 }
diff --git a/maths.h b/maths.h
deleted file mode 100644
index b446c9a..0000000
--- a/maths.h
+++ /dev/null
@@ -1,9 +0,0 @@
-#ifndef maths_h
-#define maths_h
-
-#define mini(a_, b_) ((a_) < (b_) ? (a_) : (b_))
-#define maxi(a_, b_) ((a_) > (b_) ? (a_) : (b_))
-#define sign(n_) ((n_ < 0) ? -1 : 1)
-#define clamp(v_, min_, max_) (maxi(min_, mini(max_, v_)))
-
-#endif
diff --git a/memory.c b/memory.c
index 9dd99c2..f35ed51 100644
--- a/memory.c
+++ b/memory.c
@@ -75,7 +75,6 @@ static void* heap_alloc_impla(Heap* h, int size) {
 	as = size + sizeof *header;
 	p = 0;
 	m = sizeof *header;
-	h->usage += as;
 	for (i = 0; i < h->blocks; i++) {
 		header = (int*)(h->buffer + p);
 		s = header[0];
@@ -87,9 +86,11 @@ static void* heap_alloc_impla(Heap* h, int size) {
 				header[0] = (s << 1);
 				nh = (int*)(h->buffer + p + s);
 				nh[0] = (as << 1) | 1;
+				h->usage += as;
 				h->blocks++;
 				return nh + 1;
 			} else if (s == as) {
+				h->usage += as;
 				header[0] |= 1;
 				return header + 1;
 			}
@@ -203,6 +204,11 @@ void heap_free(Heap* h, void* ptr) {
 	heap_free_aligned(h, ptr);
 }
 
+int heap_block_size(void* ptr) {
+	int h = ((int*)ptr)[-1];
+	return h >> 1;
+}
+
 void* galloc(int size) {
 #if use_system_malloc
 	return malloc(size);
diff --git a/memory.h b/memory.h
index 7ee6f53..90b89d2 100644
--- a/memory.h
+++ b/memory.h
@@ -23,6 +23,7 @@ void* heap_alloc_aligned(Heap* h, int size, int align);
 void heap_free_aligned(Heap* h, void* ptr);
 void* heap_alloc(Heap* h, int size);
 void heap_free(Heap* h, void* ptr);
+int heap_block_size(void* ptr);
 
 void* galloc(int size);
 void gfree(void* ptr);
diff --git a/plat.c b/plat.c
index e9c71ba..c947b71 100644
--- a/plat.c
+++ b/plat.c
@@ -1,82 +1,66 @@
 #include "plat.h"
 #include "config.h"
 
-extern void prog_init(void*);
-extern void prog_update(void);
-extern void prog_deinit(void);
-
 #ifdef plat_posix
 #define _POSIX_SOURCE
 #define _GNU_SOURCE
-#include "std_printers.c"
-#include <dirent.h>
-#include <fcntl.h>
-#include <signal.h>
+#include <stdarg.h>
 #include <stdio.h>
+#include <unistd.h>
 #include <stdlib.h>
+#include <dirent.h>
 #include <string.h>
-#include <sys/stat.h>
-#include <sys/time.h>
-#include <sys/types.h>
-#include <time.h>
-#include <unistd.h>
 
-static clockid_t global_clock;
-static unsigned long global_freq;
+extern int fileno(FILE*);
 
-void init_timer(void) {
-	struct timespec ts;
+void print(const char* fmt, ...) {
+	va_list args;
+	va_start(args, fmt);
+	vfprintf(stdout, fmt, args);
+	va_end(args);
+}
 
-	global_clock = CLOCK_REALTIME;
-	global_freq = 1000000000;
+void print_err(const char* fmt, ...) {
+	va_list args;
+	va_start(args, fmt);
 
-#if defined(_POSIX_MONOTONIC_CLOCK)
-	if (clock_gettime(CLOCK_MONOTONIC, &ts) == 0) {
-		global_clock = CLOCK_MONOTONIC;
+	if (isatty(fileno(stderr))) {
+		fprintf(stderr, "\033[31;31m");
 	}
-#endif
-}
 
-static unsigned long get_timer() {
-	struct timespec ts;
+	vfprintf(stderr, fmt, args);
 
-	clock_gettime(global_clock, &ts);
-	return
-		(unsigned long)ts.tv_sec * global_freq +
-		(unsigned long)ts.tv_nsec;
-}
+	if (isatty(fileno(stderr))) {
+		fprintf(stderr, "\033[0m");
+	}
 
-unsigned long get_current_time() {
-	return get_timer();
+	va_end(args);
 }
 
-void sleep_ns(unsigned long ns) {
-	struct timespec t = { 0 };
-	t.tv_nsec = ns;
-	nanosleep(&t, &t);
-}
+void print_war(const char* fmt, ...) {
+	va_list args;
+	va_start(args, fmt);
+
+	if (isatty(fileno(stderr))) {
+		fprintf(stderr, "\033[31;35m");
+	}
 
-unsigned char* load_binary(const char* n, int* size) {
-	FILE* f;
-	int s;
-	void* b;
-	f = fopen(n, "r");
-	if (!f) {
-		print_err("Failed to open %s.\n", n);
-		pbreak(error_file_not_found);
-		return 0;
+	vfprintf(stderr, fmt, args);
+
+	if (isatty(fileno(stderr))) {
+		fprintf(stderr, "\033[0m");
 	}
-	fseek(f, 0, SEEK_END);
-	s = ftell(f);
-	rewind(f);
-	b = malloc(s);
-	s = fread(b, 1, s, f);
-	if (size) *size = s;
-	return b;
+
+	va_end(args);
 }
 
-void free_file(void* p) {
-	free(p);
+void pbreak(Error code) {
+#if defined(DEBUG) && defined(plat_x86)
+	__asm__("int3;");
+	(void)code;
+#else
+	exit(code);
+#endif
 }
 
 void iter_dir(const char* path, Dir_Iter fn, void* u) {
@@ -104,359 +88,18 @@ void iter_dir(const char* path, Dir_Iter fn, void* u) {
 		else
 			fn(u, buf);
 	}
+	closedir(di);
 }
 
-#endif
-
-#ifdef plat_x11
-
-#define Font RX11Font
-#include <stdlib.h>
-#include <X11/Xlib.h>
-#include <X11/Xutil.h>
-#undef RX11Font
-
-#include "rcache.h"
-
-#define txt_buf_max 32
-
-struct {
-	int w, h, run;
-	Display* d;
-	Window wi;
-	GC gc;
-	XImage* bb;
-	Colour* bp, * fb;
-	Atom wm_p, wm_d;
-	char txt_buf[txt_buf_max];
-	char raw_buf[txt_buf_max];
-	int txt_len;
-	int mx, my, scrx, scry;
-} app;
-static MBtn mheld_btns[mbtn_count];
-static MBtn mpressed_btns[mbtn_count];
-static MBtn mreleased_btns[mbtn_count];
-
-static void reset_keys(void) {
-	int i;
-	for (i = 0; i < mbtn_count; i++) {
-		mpressed_btns[i] = 0;
-		mreleased_btns[i] = 0;
-	}
-}
-
-void init_app_render(const XWindowAttributes* wa) {
-	Colour* p, * fb;
-	Display* d;
-	XImage* bb;
-	int w, h;
-	d = app.d;
-	w = app.w;
-	h = app.h;
-	p  = (Colour*)malloc(sizeof *p  * w * h);
-	fb = (Colour*)malloc(sizeof *fb * w * h);
-	bb = app.bb;
-	if (bb) {
-		XDestroyImage(bb);
-	}
-	if (!p) {
-		print_err("Out of memory.\n");
-		pbreak(error_out_of_memory);
-	}
-	bb = XCreateImage(
-		d,
-		wa->visual,
-		wa->depth,
-		ZPixmap,
-		0,
-		(char*)p,
-		w,
-		h,
-		32,
-		w * sizeof *p
-	);
-	if (!bb) {
-		print_err("Failed to create X11 backbuffer.\n");
-		pbreak(error_platform_error);
-	}
-	app.bp = p;
-	app.fb = fb;
-	app.bb = bb;
-}
-
-void init_app(void) {
-	Window w, r;
-	Display* d;
-	GC gc;
-	unsigned rm, bm;
-	XWindowAttributes wa;
-	app.run = 0;
-	d = XOpenDisplay(0);
-	if (!d) {
-		print_err("Failed to open X11 display.\n");
-		pbreak(error_platform_error);
-		return;
-	}
-	r = DefaultRootWindow(d);
-	app.wm_p = XInternAtom(
-		d,
-		"WM_PROTOCOLS",
-		0
-	);
-	app.wm_d = XInternAtom(
-		d,
-		"WM_DELETE_WINDOW",
-		0
-	);
-	w = XCreateSimpleWindow(
-		d,
-		r,
-		0,
-		0,
-		default_window_w,
-		default_window_h,
-		0,
-		WhitePixel(d, 0),
-		BlackPixel(d, 0)
-	);
-	XSetWMProtocols(d, w, &app.wm_d, 1);
-	XStoreName(d, w, app_name);
-	XSelectInput(
-		d,
-		w,
-		ExposureMask      |
-		KeyPressMask      |
-		KeyReleaseMask    |
-		PointerMotionMask |
-		ButtonPressMask   |
-		ButtonReleaseMask
-	);
-	XClearWindow(d, w);
-	XMapRaised(d, w);
-	gc = XCreateGC(d, w, 0, 0);
-	if (!gc) {
-		print_err("Failed to create graphics context.\n");
-		pbreak(error_platform_error);
-		return;
-	}
-	XGetWindowAttributes(d, w, &wa);
-	if (wa.depth != 24 && wa.depth != 32) {
-		print_err("Only true colour displays are supported.\n");
-		pbreak(error_platform_error);
-	}
-	rm = wa.visual->red_mask  & 0x1000000;
-	bm = wa.visual->blue_mask & 0xffffff;
-	if ((rm == 0xff && bm == 0xff0000)) {
-	    print_war("Detected BGR. Colours will look fucked.\n");
-	}
-	app.d = d;
-	app.gc = gc;
-	app.w = wa.width;
-	app.h = wa.height;
-	app.run = 1;
-	app.wi = w;
-	app.bb = 0;
-	init_app_render(&wa);
-}
-
-void deinit_app(void) {
-	XDestroyWindow(app.d, app.wi);
-	XDestroyImage(app.bb);
-	XCloseDisplay(app.d);
-}
+extern int prog_main(void*);
 
-void update_events(void) {
-	XWindowAttributes wa;
-	Display* d;
-	Window w;
-	XEvent e;
-	Rectangle r;
-	int mb;
-	d = app.d;
-	w = app.wi;
-	reset_keys();
-	app.scrx = app.scry = 0;
-	while (XPending(d)) {
-		XNextEvent(d, &e);
-		switch (e.type) {
-			case ClientMessage:
-				if (
-					(Atom)e.xclient.data.l[0] ==
-					app.wm_d
-				) {
-					app.run = 0;
-				}
-				break;
-			case Expose:
-				XGetWindowAttributes(d, w, &wa);
-				if (
-					wa.width  != app.w ||
-					wa.height != app.h
-				) {
-					app.w = wa.width;
-					app.h = wa.height;
-					init_app_render(&wa);
-				}
-				r.x = 0;
-				r.y = 0;
-				r.w = app.w;
-				r.h = app.h;
-				rc_invalidate(&r);
-				break;
-			case MotionNotify:
-				app.mx = e.xmotion.x;
-				app.my = e.xmotion.y;
-				break;
-			case ButtonPress:
-				switch (e.xbutton.button) {
-					case 1:
-					case 2:
-					case 3:
-						mb = e.xbutton.button - 1;
-						mheld_btns[mb] = 1;
-						mpressed_btns[mb] = 1;
-						break;
-				}
-				break;
-			case ButtonRelease:
-				switch (e.xbutton.button) {
-					case 1:
-					case 2:
-					case 3:
-						mb = e.xbutton.button - 1;
-						mheld_btns[mb] = 0;
-						mreleased_btns[mb] = 1;
-						break;
-					case 4:
-						app.scry--;
-						break;
-					case 5:
-						app.scry++;
-						break;
-					case 6:
-						app.scrx--;
-						break;
-					case 7:
-						app.scrx++;
-						break;
-				}
-				break;
-			default:
-				break;
-		}
-	}
-}
-
-int main(int argc, const char** argv) {
+int main() {
+	int r;
 	void* mem;
-	unsigned long now, next;
-	long ts;
-	(void)argc;
-	(void)argv;
-	init_timer();
 	mem = malloc(memory_size);
-	if (!mem) {
-		print_err("Out of memory.\n");
-		return error_out_of_memory;
-	}
-	init_app();
-	next = get_current_time();
-	prog_init(mem);
-	while (app.run) {
-		now = get_current_time();
-		while (now >= next) {
-			update_events();
-			prog_update();
-			next += ms_per_frame * 1000000;
-		}
-		ts = next - now;
-		if (ts > 0)
-			sleep_ns(ts);
-	}
-	prog_deinit();
-	deinit_app();
+	r = prog_main(mem);
 	free(mem);
-	return 0;
-}
-
-void plat_present(
-	int x,
-	int y,
-	int w,
-	int h
-) {
-	const Colour* src;
-	Colour* dst;
-	int i, j;
-	int ex, ey;
-	int s;
-	ex = x + w;
-	ey = y + h;
-	dst = app.bp + x + y * app.w;
-	src = app.fb + x + y * app.w;
-	s = app.w - w;
-	for (j = y; j < ey; j++) {
-		for (i = x; i < ex; i++) {
-			dst->r = src->b;
-			dst->g = src->g;
-			dst->b = src->r;
-			dst->a = src->a;
-			dst++;
-			src++;
-		}
-		dst += s;
-		src += s;
-	}
-	XPutImage(
-		app.d,
-		app.wi,
-		app.gc,
-		app.bb,
-		x, y,
-		x, y,
-		w, h
-	);
-}
-
-Colour* get_fb(void) {
-	return app.fb;
-}
-
-int get_render_w(void) {
-	return app.w;
-}
-
-int get_render_h(void) {
-	return app.h;
-}
-
-int get_mouse_x(void) {
-	return app.mx;
-}
-
-int get_mouse_y(void) {
-	return app.my;
-}
-
-int get_mscroll_x(void) {
-	return app.scrx;
-}
-
-int get_mscroll_y(void) {
-	return app.scry;
-}
-
-int mbtn_pressed(MBtn btn) {
-	return mheld_btns[btn];
-}
-
-int mbtn_just_pressed(MBtn btn) {
-	return mpressed_btns[btn];
-}
-
-int mbtn_just_released(MBtn btn) {
-	return mreleased_btns[btn];
+	return r;
 }
 
 #endif
-
diff --git a/plat.h b/plat.h
index 8b23eea..f001d70 100644
--- a/plat.h
+++ b/plat.h
@@ -1,44 +1,13 @@
 #ifndef plat_h
+#define plat_h
 
-#include "render.h"
 #include "error.h"
 
-int get_mouse_x();
-int get_mouse_y();
-
-typedef enum {
-	mbtn_left,
-	mbtn_middle,
-	mbtn_right,
-	mbtn_count
-} MBtn;
-int mbtn_pressed(MBtn btn);
-int mbtn_just_pressed(MBtn btn);
-int mbtn_just_released(MBtn btn);
-int get_mscroll_x(void);
-int get_mscroll_y(void);
-int get_mouse_x(void);
-int get_mouse_y(void);
-
-const char* get_text_input(int* len);
-
 void print(const char* fmt, ...);
 void print_err(const char* fmt, ...);
 void print_war(const char* fmt, ...);
 void pbreak(Error code);
 
-Colour* get_fb(void);
-
-unsigned long get_current_time();
-void sleep_ns(unsigned long ns);
-
-void plat_present(int x, int y, int w, int h);
-int get_render_w(void);
-int get_render_h(void);
-
-unsigned char* load_binary(const char* n, int* size);
-void free_file(void* p);
-
 typedef void (*Dir_Iter)(void* uptr, const char* path);
 void iter_dir(const char* path, Dir_Iter fn, void* u);
 
diff --git a/rcache.c b/rcache.c
deleted file mode 100644
index 9170f71..0000000
--- a/rcache.c
+++ /dev/null
@@ -1,467 +0,0 @@
-#include "error.h"
-#include "memory.h"
-#include "maths.h"
-#include "plat.h"
-#include "rcache.h"
-#include "render.h"
-
-#if rcache_enable
-unsigned* hgrid, * ohgrid, * hgrid1, * hgrid2;
-char* cmd_buf;
-int cmd_count;
-int rcache_cell_w, rcache_cell_h;
-
-void rc_init(void) {
-	int i, e;
-	rcache_cell_w = get_render_w() / rcache_gw;
-	rcache_cell_h = get_render_h() / rcache_gh;
-	cmd_buf = stack_alloc(rcache_cmd_buf_size);
-	e = rcache_gw * rcache_gh * sizeof *hgrid1;
-	hgrid1 = stack_alloc(e);
-	hgrid2 = stack_alloc(e);
-	hgrid = hgrid1;
-	ohgrid = hgrid2;
-	cmd_count = 0;
-	e /= sizeof *hgrid1;
-	for (i = 0; i < e; i++)
-		hgrid1[i] = hgrid2[i] = 0;
-}
-
-RC_Cmd* rc_add_cmd(RC_Cmd* cmd) {
-	int i;
-	RC_Cmd* r;
-	cmd->size = (cmd->size + 7) & -8;
-#ifdef DEBUG
-	if (cmd_count + cmd->size >= rcache_cmd_buf_size) {
-		print_err("Out of render command space.");
-		pbreak(error_out_of_memory);
-		return 0;
-	}
-#endif
-	r = (RC_Cmd*)&cmd_buf[cmd_count];
-	for (i = 0; i < cmd->size; i++) {
-		cmd_buf[cmd_count++] = ((char*)cmd)[i];
-	}
-	return r;
-}
-
-void rc_add_cmd_bmp(
-	const Bitmap* bmp,
-	int x,
-	int y,
-	const Rectangle* src
-) {
-	RC_Cmd_Bmp cmd;
-	cmd.cmd.type = rc_cmd_bmp;
-	cmd.cmd.size = sizeof cmd;
-	cmd.cmd.x = x;
-	cmd.cmd.y = y;
-	cmd.cmd.w = src->w;
-	cmd.cmd.h = src->h;
-	cmd.sx = src->x;
-	cmd.sy = src->y;
-	cmd.bmp = bmp;
-	rc_add_cmd((RC_Cmd*)&cmd);
-}
-
-void rc_add_cmd_bmp_col(
-	const Bitmap* bmp,
-	int x,
-	int y,
-	const Rectangle* src,
-	Colour col
-) {
-	RC_Cmd_Bmp_Mod cmd;
-	cmd.cmd.type = rc_cmd_bmp_mod;
-	cmd.cmd.size = sizeof cmd;
-	cmd.cmd.x = x;
-	cmd.cmd.y = y;
-	cmd.cmd.w = src->w;
-	cmd.cmd.h = src->h;
-	cmd.sx = src->x;
-	cmd.sy = src->y;
-	cmd.bmp = bmp;
-	cmd.col = col;
-	rc_add_cmd((RC_Cmd*)&cmd);
-}
-
-void rc_add_cmd_bmp_cp(
-	const Bitmap* bmp,
-	int x,
-	int y,
-	const Rectangle* src
-) {
-	RC_Cmd_Bmp_Cp cmd;
-	cmd.cmd.type = rc_cmd_bmp_cp;
-	cmd.cmd.size = sizeof cmd;
-	cmd.cmd.x = x;
-	cmd.cmd.y = y;
-	cmd.cmd.w = src->w;
-	cmd.cmd.h = src->h;
-	cmd.sx = src->x;
-	cmd.sy = src->y;
-	cmd.bmp = *bmp;
-	rc_add_cmd((RC_Cmd*)&cmd);
-}
-
-void rc_add_cmd_bmp_cp_col(
-	const Bitmap* bmp,
-	int x,
-	int y,
-	const Rectangle* src,
-	Colour col
-) {
-	RC_Cmd_Bmp_Cp_Mod cmd;
-	cmd.cmd.type = rc_cmd_bmp_cp_mod;
-	cmd.cmd.size = sizeof cmd;
-	cmd.cmd.x = x;
-	cmd.cmd.y = y;
-	cmd.cmd.w = src->w;
-	cmd.cmd.h = src->h;
-	cmd.sx = src->x;
-	cmd.sy = src->y;
-	cmd.bmp = *bmp;
-	cmd.col = col;
-	rc_add_cmd((RC_Cmd*)&cmd);
-}
-
-void rc_add_cmd_rfont_text(
-	Font* font,
-	int x,
-	int y,
-	const char* text
-) {
-	RC_Cmd_RFont_Text cmd;
-	RC_Cmd_RFont_Text* dc;
-	Rectangle r;
-	const char* c;
-	char* d;
-	r = text_rect(font, text);
-	cmd.cmd.type = rc_cmd_rfont_text;
-	cmd.cmd.x = x;
-	cmd.cmd.y = y;
-	cmd.cmd.w = r.w;
-	cmd.cmd.h = r.h;
-	cmd.font = font;
-	for (c = text; *c; c++);
-	cmd.cmd.size = sizeof cmd + (c - text) + 1;
-	dc = (void*)rc_add_cmd((RC_Cmd*)&cmd);
-	for (
-		c = text, d = (char*)(dc + 1);
-		*c;
-		c++, d++
-	) *d = *c;
-	*d = 0;
-}
-
-void rc_add_cmd_rfont_text_col(
-	Font* font,
-	int x,
-	int y,
-	const char* text,
-	Colour col
-) {
-	RC_Cmd_RFont_Text_Col cmd;
-	RC_Cmd_RFont_Text_Col* dc;
-	Rectangle r;
-	const char* c;
-	char* d;
-	r = text_rect(font, text);
-	cmd.cmd.type = rc_cmd_rfont_text_col;
-	cmd.cmd.x = x;
-	cmd.cmd.y = y;
-	cmd.cmd.w = r.w;
-	cmd.cmd.h = r.h;
-	cmd.font = font;
-	cmd.col = col;
-	for (c = text; *c; c++);
-	cmd.cmd.size = sizeof cmd + (c - text) + 1;
-	dc = (void*)rc_add_cmd((RC_Cmd*)&cmd);
-	for (
-		c = text, d = (char*)(dc + 1);
-		*c;
-		c++, d++
-	) *d = *c;
-	*d = 0;
-}
-
-void rc_add_cmd_clip(
-	const Rectangle* rect
-) {
-	RC_Cmd cmd;
-	cmd.type = rc_cmd_set_clip;
-	cmd.size = sizeof cmd;
-	cmd.x = rect->x;
-	cmd.y = rect->y;
-	cmd.w = rect->w - rect->x;
-	cmd.h = rect->h - rect->y;
-	rc_add_cmd(&cmd);
-}
-
-void rc_add_cmd_reset_clip(void) {
-	RC_Cmd cmd;
-	cmd.type = rc_cmd_set_clip;
-	cmd.size = sizeof cmd;
-	cmd.x = 0;
-	cmd.y = 0;
-	cmd.w = get_render_w();
-	cmd.h = get_render_h();
-	rc_add_cmd(&cmd);
-}
-
-void rc_add_cmd_rect(
-	const Rectangle* r,
-	Colour col
-) {
-	RC_Cmd_Rect cmd;
-	cmd.cmd.type = rc_cmd_rect;
-	cmd.cmd.size = sizeof cmd;
-	cmd.cmd.x = r->x;
-	cmd.cmd.y = r->y;
-	cmd.cmd.w = r->w;
-	cmd.cmd.h = r->h;
-	cmd.col = col;
-	rc_add_cmd((RC_Cmd*)&cmd);
-}
-
-static void hash(unsigned* hash, const void* data, int size) {
-	const unsigned char *p = data;
-	while (size--)
-		*hash = (*hash ^ *p++) * 16777619;
-}
-
-void update_hgrid(unsigned* g) {
-	int x, y, x1, y1, x2, y2, w;
-	unsigned* d;
-	RC_Cmd* cmd, * end;
-	w = rcache_gw;
-	x = rcache_gw * rcache_gh;
-	for (y = 0; y < x; y++)
-		g[y] = rcache_hash_seed;
-	cmd = (RC_Cmd*)cmd_buf;
-	end = (RC_Cmd*)(cmd_buf + cmd_count); 
-	for (
-		;
-		cmd != end;
-		cmd = (RC_Cmd*)((char*)cmd + cmd->size)
-	) {
-		if (
-			cmd->type == rc_cmd_set_clip ||
-			cmd->type == rc_cmd_reset_clip
-		) continue;
-		x1 = maxi(0, cmd->x / rcache_cell_w);
-		y1 = maxi(0, cmd->y / rcache_cell_h);
-		if (x1 >= rcache_gw) continue;
-		if (y1 >= rcache_gh) continue;
-		x2 = mini(
-			(cmd->x + cmd->w) / rcache_cell_w,
-			rcache_gw - 1
-		);
-		y2 = mini(
-			(cmd->y + cmd->h) / rcache_cell_h,
-			rcache_gh - 1
-		);
-		if (x2 < 0) continue;
-		if (y2 < 0) continue;
-		for (y = y1; y <= y2; y++)
-			for (x = x1; x <= x2; x++) {
-				d = &g[x + y * w];
-				hash(d, cmd, cmd->size);
-			}
-	}
-}
-
-void rc_flush_rect(int x, int y, int w, int h) {
-	RC_Cmd* cmd, * end;
-	RC_Cmd_Bmp* bmpcmd;
-	RC_Cmd_Bmp_Mod* bmpmcmd;
-	RC_Cmd_Bmp_Cp* bmpcpcmd;
-	RC_Cmd_Bmp_Cp_Mod* bmpmcpcmd;
-	RC_Cmd_RFont_Text* txtrfcmd;
-	RC_Cmd_RFont_Text_Col* txtrfccmd;
-	Rectangle c, r;
-	cmd = (RC_Cmd*)cmd_buf;
-	end = (RC_Cmd*)(cmd_buf + cmd_count); 
-	int n, rw, rh;
-	rw = get_render_w();
-	rh = get_render_h();
-	if (x > rw) return;
-	if (y > rh) return;
-	x = maxi(x, 0);
-	y = maxi(y, 0);
-	if ((n = x + w - rw) > 0)
-		w -= n;
-	if ((n = y + h - rh) > 0)
-		h -= n;
-	if (w <= 0) return;
-	if (h <= 0) return;
-	c.x = x;
-	c.y = y;
-	c.w = x + w;
-	c.h = y + h;
-	render_clip(&c);
-	render_clear();
-	while (cmd != end) {
-		switch (cmd->type) {
-			case rc_cmd_bmp:
-				bmpcmd = (RC_Cmd_Bmp*)cmd;
-				r.x = bmpcmd->sx;
-				r.y = bmpcmd->sy;
-				r.w = cmd->w;
-				r.h = cmd->h;
-				render_bitmap(
-					bmpcmd->bmp,
-					cmd->x,
-					cmd->y,
-					&r
-				);
-				break;
-			case rc_cmd_bmp_mod:
-				bmpmcmd = (RC_Cmd_Bmp_Mod*)cmd;
-				r.x = bmpmcmd->sx;
-				r.y = bmpmcmd->sy;
-				r.w = cmd->w;
-				r.h = cmd->h;
-				render_bitmap_col(
-					bmpmcmd->bmp,
-					cmd->x,
-					cmd->y,
-					&r,
-					bmpmcmd->col
-				);
-				break;
-			case rc_cmd_bmp_cp:
-				bmpcpcmd = (RC_Cmd_Bmp_Cp*)cmd;
-				r.x = bmpcpcmd->sx;
-				r.y = bmpcpcmd->sy;
-				r.w = cmd->w;
-				r.h = cmd->h;
-				render_bitmap(
-					&bmpcpcmd->bmp,
-					cmd->x,
-					cmd->y,
-					&r
-				);
-				break;
-			case rc_cmd_bmp_cp_mod:
-				bmpmcpcmd = (RC_Cmd_Bmp_Cp_Mod*)cmd;
-				r.x = bmpmcpcmd->sx;
-				r.y = bmpmcpcmd->sy;
-				r.w = cmd->w;
-				r.h = cmd->h;
-				render_bitmap_col(
-					&bmpmcpcmd->bmp,
-					cmd->x,
-					cmd->y,
-					&r,
-					bmpmcpcmd->col
-				);
-				break;
-			case rc_cmd_rfont_text:
-				txtrfcmd = (RC_Cmd_RFont_Text*)cmd;
-				rfont_text(
-					txtrfcmd->font,
-					cmd->x,
-					cmd->y,
-					(char*)(txtrfcmd + 1)
-				);
-				break;
-			case rc_cmd_rfont_text_col:
-				txtrfccmd = (RC_Cmd_RFont_Text_Col*)cmd;
-				rfont_text_col(
-					txtrfccmd->font,
-					cmd->x,
-					cmd->y,
-					(char*)(txtrfccmd + 1),
-					txtrfccmd->col
-				);
-				break;
-			case rc_cmd_set_clip:
-				r.x = cmd->x;
-				r.y = cmd->y;
-				r.w = cmd->w;
-				r.h = cmd->h;
-				r = rect_intersect(&c, &r);
-				r.w += r.x;
-				r.h += r.y;
-				render_clip(&r);
-				break;
-			case rc_cmd_reset_clip:
-				render_clip(&c);
-				break;
-			case rc_cmd_rect:
-				r.x = cmd->x;
-				r.y = cmd->y;
-				r.w = cmd->w;
-				r.h = cmd->h;
-				render_rect(
-					&r,
-					((RC_Cmd_Rect*)cmd)->col
-				);
-				break;
-		}
-		cmd = (RC_Cmd*)((char*)cmd + cmd->size);
-	}
-	plat_present(x, y, w, h);
-	render_reset_clip();
-}
-
-void rc_flush(void) {
-	int x, y, w, h, i;
-	w = rcache_gw;
-	h = rcache_gh;
-	update_hgrid(hgrid);
-	for (y = 0; y < h; y++)
-		for (x = 0; x < w; x++) {
-			i = x + y * w;
-			if (hgrid[i] != ohgrid[i])
-				rc_flush_rect(
-					x * rcache_cell_w,
-					y * rcache_cell_h,
-					rcache_cell_w,
-					rcache_cell_h
-				);
-		}
-	cmd_count = 0;
-	x = rcache_gw * rcache_gh;
-	for (y = 0; y < x; y++)
-		ohgrid[y] = hgrid[y];
-	ohgrid = hgrid;
-	hgrid = hgrid == hgrid1? hgrid2: hgrid1;
-}
-
-void rc_begin(void) {
-	rcache_cell_w = (get_render_w() / rcache_gw) + 1;
-	rcache_cell_h = (get_render_h() / rcache_gh) + 1;
-}
-
-void rc_invalidate(const Rectangle* r) {
-	int x, y, x1, y1, x2, y2, w;
-	x1 = maxi(0, r->x / rcache_cell_w);
-	y1 = maxi(0, r->y / rcache_cell_h);
-	x2 = mini(
-		(r->x + r->w) / rcache_cell_w,
-		rcache_gw - 1
-	);
-	y2 = mini(
-		(r->y + r->h) / rcache_cell_h,
-		rcache_gh - 1
-	);
-	w = rcache_gw;
-	for (y = y1; y <= y2; y++)
-		for (x = x1; x <= x2; x++) {
-			ohgrid[x + y * w] = -1;
-		}
-}
-#else
-
-void rc_init(void) {
-	render_init();
-}
-
-void rc_flush(void) {
-	plat_present(0, 0, get_render_w(), get_render_h());
-	render_begin();
-}
-
-#endif
diff --git a/rcache.h b/rcache.h
deleted file mode 100644
index ff24f0c..0000000
--- a/rcache.h
+++ /dev/null
@@ -1,148 +0,0 @@
-#ifndef rcache_h
-#define rcache_h
-
-#include "render.h"
-
-#if rcache_enable
-typedef enum {
-	rc_cmd_bmp,
-	rc_cmd_bmp_mod,
-	rc_cmd_bmp_cp,
-	rc_cmd_bmp_cp_mod,
-	rc_cmd_rfont_text,
-	rc_cmd_rfont_text_col,
-	rc_cmd_set_clip,
-	rc_cmd_reset_clip,
-	rc_cmd_rect
-} RC_Cmd_Type;
-
-typedef struct {
-	RC_Cmd_Type type;
-	int size;
-	int x, y, w, h;
-} RC_Cmd;
-
-typedef struct {
-	RC_Cmd cmd;
-	const Bitmap* bmp;
-	int sx, sy;
-} RC_Cmd_Bmp;
-
-typedef struct {
-	RC_Cmd cmd;
-	const Bitmap* bmp;
-	int sx, sy;
-	Colour col;
-} RC_Cmd_Bmp_Mod;
-
-typedef struct {
-	RC_Cmd cmd;
-	Bitmap bmp;
-	int sx, sy;
-} RC_Cmd_Bmp_Cp;
-
-typedef struct {
-	RC_Cmd cmd;
-	Bitmap bmp;
-	int sx, sy;
-	Colour col;
-} RC_Cmd_Bmp_Cp_Mod;
-
-typedef struct {
-	RC_Cmd cmd;
-	Font* font;
-} RC_Cmd_RFont_Text;
-
-typedef struct {
-	RC_Cmd cmd;
-	Colour col;
-	Font* font;
-} RC_Cmd_RFont_Text_Col;
-
-typedef struct {
-	RC_Cmd cmd;
-	Colour col;
-} RC_Cmd_Rect;
-
-typedef RC_Cmd RC_Cmd_Set_Clip;
-typedef RC_Cmd RC_Cmd_Reset_Clip;
-
-void rc_init(void);
-RC_Cmd* rc_add_cmd(RC_Cmd* cmd);
-void rc_add_cmd_bmp(
-	const Bitmap* bmp,
-	int x,
-	int y,
-	const Rectangle* src
-);
-void rc_add_cmd_bmp_col(
-	const Bitmap* bmp,
-	int x,
-	int y,
-	const Rectangle* src,
-	Colour col
-);
-void rc_add_cmd_bmp_cp(
-	const Bitmap* bmp,
-	int x,
-	int y,
-	const Rectangle* src
-);
-void rc_add_cmd_bmp_cp_col(
-	const Bitmap* bmp,
-	int x,
-	int y,
-	const Rectangle* src,
-	Colour col
-);
-void rc_add_cmd_rfont_text(
-	Font* font,
-	int x,
-	int y,
-	const char* text
-);
-void rc_add_cmd_rfont_text_col(
-	Font* font,
-	int x,
-	int y,
-	const char* text,
-	Colour col
-);
-void rc_add_cmd_clip(
-	const Rectangle* rect
-);
-void rc_add_cmd_reset_clip(void);
-void rc_add_cmd_rect(
-	const Rectangle* r,
-	Colour col
-);
-void rc_begin(void);
-void rc_flush(void);
-void rc_invalidate(const Rectangle* r);
-
-#else
-void rc_init(void);
-#define rc_add_cmd_bmp(bmp, x, y, src) \
-	render_bitmap(bmp, x, y, src)
-#define rc_add_cmd_bmp_col(bmp, x, y, src, col) \
-	render_bitmap_col(bmp, x, y, src, col)
-#define rc_add_cmd_bmp_cp(bmp, x, y, src) \
-	render_bitmap(bmp, x, y, src)
-#define rc_add_cmd_bmp_cp_col(bmp, x, y, src, col) \
-	render_bitmap_col(bmp, x, y, src, col)
-#define rc_add_cmd_rfont_text(font, x, y, text) \
-	rfont_text(font, x, y, text)
-#define rc_add_cmd_rfont_text_col(font, x, y, text, col) \
-	rfont_text_col(font, x, y, text, col)
-#define rc_add_cmd_clip(rect) \
-	render_clip(rect)
-#define rc_add_cmd_reset_clip() \
-	render_reset_clip()
-#define rc_add_cmd_rect(r, col) \
-	render_rect(r, col)
-#define rc_invalidate(r)
-#define rc_begin()
-void rc_flush(void);
-#endif
-
-#endif
diff --git a/rect.c b/rect.c
deleted file mode 100644
index bcc95cd..0000000
--- a/rect.c
+++ /dev/null
@@ -1,96 +0,0 @@
-#include "maths.h"
-#include "rect.h"
-
-Rectangle make_rect(int x, int y, int w, int h) {
-	Rectangle r;
-	r.x = x;
-	r.y = y;
-	r.w = w;
-	r.h = h;
-	return r;
-}
-
-int rects_overlap(const Rectangle* a, const Rectangle* b) {
-	return
-		a->x + a->w > b->x &&
-		a->y + a->h > b->y &&
-		a->x < b->x + b->w &&
-		a->y < b->y + b->h;
-}
-
-int rects_overlap2(
-	int x0,
-	int y0,
-	int w0,
-	int h0,
-	int x1,
-	int y1,
-	int w1,
-	int h1
-) {
-	return
-		x0 + w0 > x1 &&
-		y0 + h0 > y1 &&
-		x0 < x1 + w1 &&
-		y0 < y1 + h1;
-}
-
-int point_rect_overlap(
-	const Rectangle* r,
-	int px,
-	int py
-) {
-	return
-		px >= r->x &&
-		py >= r->y &&
-		px <= r->x + r->w &&
-		py <= r->y + r->h;
-}
-
-int point_rect_overlap2(
-	int x,
-	int y,
-	int w,
-	int h,
-	int px,
-	int py
-) {
-	return
-		px > x &&
-		py > y &&
-		px < x + w &&
-		py < y + h;
-}
-
-Rectangle rect_intersect(
-	const Rectangle* a,
-	const Rectangle* b
-) {
-	int x1, y1, x2, y2;
-	Rectangle r;
-  x1 = maxi(a->x, b->x);
-  y1 = maxi(a->y, b->y);
-  x2 = mini(a->x + a->w, b->x + b->w);
-  y2 = mini(a->y + a->h, b->y + b->h);
-  r.x = x1;
-  r.y = y1;
-  r.w = maxi(0, x2 - x1);
-  r.h = maxi(0, y2 - y1);
-  return r;
-}
-
-Rectangle* rect_merge(
-	Rectangle* d,
-	const Rectangle* r
-) {
-	int x1, y1, x2, y2;
-	x1 = mini(d->x, r->x);
-	y1 = mini(d->y, r->y);
-	x2 = maxi(d->x + d->w, r->x + r->w);
-	y2 = maxi(d->y + d->h, r->y + r->h);
-	d->x = x1;
-	d->y = y1;
-	d->w = x2 - x1;
-	d->h = y2 - y1;
-	return d;
-}
diff --git a/rect.h b/rect.h
deleted file mode 100644
index 4d6a88f..0000000
--- a/rect.h
+++ /dev/null
@@ -1,43 +0,0 @@
-#ifndef rect_h
-#define rect_h
-
-typedef struct {
-	int x, y, w, h;
-} Rectangle;
-
-Rectangle make_rect(int x, int y, int w, int h);
-
-int rects_overlap(const Rectangle* a, const Rectangle* b);
-int rects_overlap2(
-	int x0,
-	int y0,
-	int w0,
-	int h0,
-	int x1,
-	int y1,
-	int w1,
-	int h1
-);
-int point_rect_overlap(
-	const Rectangle* r,
-	int px,
-	int py
-);
-int point_rect_overlap2(
-	int x,
-	int y,
-	int w,
-	int h,
-	int px,
-	int py
-);
-Rectangle rect_intersect(
-	const Rectangle* a,
-	const Rectangle* b
-);
-Rectangle* rect_merge(
-	Rectangle* d,
-	const Rectangle* r
-);
-
-#endif
diff --git a/render.c b/render.c
deleted file mode 100644
index e291e7d..0000000
--- a/render.c
+++ /dev/null
@@ -1,380 +0,0 @@
-#include "plat.h"
-#include "maths.h"
-#include "render.h"
-
-static struct {
-	Rectangle clip;
-} renderer;
-
-Colour make_colour(unsigned rgb, unsigned char a) {
-	Colour r;
-	r.r = (unsigned char)(rgb >> 16);
-	r.g = (unsigned char)(rgb >> 8);
-	r.b = (unsigned char)rgb;
-	r.a = a;
-	return r;
-}
-
-Colour make_black() {
-	Colour r = { 0, 0, 0, 255 };
-	return r;
-}
-
-Colour make_white() {
-	return make_colour(0xffffff, 255);
-}
-
-void init_bitmap(
-	Bitmap* bitmap,
-	Colour* pixels,
-	int w,
-	int h
-) {
-	bitmap->pixels = pixels;
-	bitmap->w = w;
-	bitmap->h = h;
-}
-
-void render_init(void) {
-	int i, e;
-	Colour b, * p;
-	b = make_black();
-	p = get_fb();
-	e = get_render_w() * get_render_h();
-	for (i = 0; i < e; i++)
-		p[i] = b;
-}
-
-void render_begin(void) {
-	int i, e;
-	Colour b, * p;
-	b = make_black();
-	p = get_fb();
-	e = get_render_w() * get_render_h();
-	for (i = 0; i < e; i++)
-		p[i] = b;
-	render_reset_clip();
-}
-
-void render_clear(void) {
-	render_clear_col(make_black());
-}
-
-void render_clear_col(Colour col) {
-	Colour* d;
-	int ox, x, y, ex, ey, w;
-	d = get_fb();
-	x  = renderer.clip.x;
-	y  = renderer.clip.y;
-	ex = renderer.clip.w;
-	ey = renderer.clip.h;
-	ox = x;
-	w = get_render_w();
-	for (; y < ey; y++)
-		for (x = ox; x < ex; x++)
-			d[x + y * w] = col;
-}
-
-void render_clip(const Rectangle* rect) {
-	renderer.clip = *rect;
-	renderer.clip.x = maxi(renderer.clip.x, 0);
-	renderer.clip.y = maxi(renderer.clip.y, 0);
-	renderer.clip.w = mini(
-		renderer.clip.w,
-		get_render_w()
-	);
-	renderer.clip.h = mini(
-		renderer.clip.h,
-		get_render_h()
-	);
-}
-
-void render_reset_clip(void) {
-	renderer.clip.x = 0;
-	renderer.clip.y = 0;
-	renderer.clip.w = get_render_w();
-	renderer.clip.h = get_render_h();
-}
-
-Colour* get_render_pixels(void) {
-	return get_fb();
-}
-
-Colour blend(Colour dst, Colour src) {
-	int ima;
-
-	ima = 0xff - src.a;
-	dst.r = (unsigned char)(((src.r * src.a) + (dst.r * ima)) >> 8);
-	dst.g = (unsigned char)(((src.g * src.a) + (dst.g * ima)) >> 8);
-	dst.b = (unsigned char)(((src.b * src.a) + (dst.b * ima)) >> 8);
-
-	return dst;
-}
-
-Colour blend_mod(Colour dst, Colour src, Colour mod) {
-	int ima;
-
-	src.a = (src.a * mod.a) >> 8;
-	ima = 0xff - src.a;
-	dst.r = (unsigned char)(((src.r * mod.r * src.a) >> 16) + ((dst.r * ima) >> 8));
-	dst.g = (unsigned char)(((src.g * mod.g * src.a) >> 16) + ((dst.g * ima) >> 8));
-	dst.b = (unsigned char)(((src.b * mod.b * src.a) >> 16) + ((dst.b * ima) >> 8));
-
-	return dst;
-}
-
-#define blit_vars(t_) \
-	int i, j, stride, sstride, n; \
-	Colour* dst; \
-	const t_* src; \
-	Rectangle sub;
-#define blit( \
-	src_, \
-	dst_, \
-	pitch_, \
-	clip_, \
-	src_w_, \
-	src_h_, \
-	expr_ \
-) \
-	sub = *rect; \
-	if (sub.w <= 0) { goto end; } \
-	if (sub.h <= 0) { goto end; } \
-	if (sub.w > (src_w_)) { goto end; } \
-	if (sub.h > (src_h_)) { goto end; } \
-	if ((n = (clip_).x - x) > 0) { \
-		sub.w -= n; \
-		sub.x += n; \
-		x += n; \
-	} \
-	if ((n = (clip_).y - y) > 0) { \
-		sub.h -= n; \
-		sub.y += n; \
-		y += n; \
-	} \
-	if ((n = x + sub.w - (clip_).w) > 0) { \
-		sub.w -= n; \
-	} \
-	if ((n = y + sub.h - (clip_).h) > 0) { \
-		sub.h -= n; \
-	} \
-	if (sub.w <= 0) { goto end; } \
-	if (sub.h <= 0) { goto end; } \
-	dst = (dst_) + (x + y * (pitch_)); \
-	src = (src_) + (sub.x + sub.y * (src_w_)); \
-	stride = (pitch_) - sub.w; \
-	sstride = (src_w_) - sub.w; \
-	for (i = 0; i < sub.h; i++) { \
-		for (j = 0; j < sub.w; j++) { \
-			expr_; \
-			dst++; \
-			src++; \
-		} \
-		dst += stride; \
-		src += sstride; \
-	} \
-	end:
-
-#define blit_rect_vars \
-	int x1, y1, x2, y2, dr, j, i; \
-	Colour* dst;
-#define blit_rect(dst_, pitch_, clip_) \
-	x1 = rect->x < (clip_).x ? (clip_).x : rect->x; \
-	y1 = rect->y < (clip_).y ? (clip_).y : rect->y; \
-	x2 = rect->x + rect->w; \
-	y2 = rect->y + rect->h; \
-	x2 = x2 > (clip_).w? (clip_).w: x2; \
-	y2 = y2 > (clip_).h? (clip_).h: y2; \
-	dst = dst_; \
-	dst += x1 + y1 * pitch_; \
-	dr = pitch_ - (x2 - x1); \
-	for (j = y1; j < y2; j++) { \
-		for (i = x1; i < x2; i++) { \
-			*dst = blend(*dst, colour); \
-			dst++; \
-		} \
-		dst += dr; \
-	}
-
-void render_bitmap(
-	const Bitmap* bitmap,
-	int x,
-	int y,
-	const Rectangle* rect
-) {
-	blit_vars(Colour)
-	int renderer_w;
-	Colour* renderer_pixels;
-	renderer_w = get_render_w();
-	renderer_pixels = get_fb();
-	blit(
-		bitmap->pixels,
-		renderer_pixels,
-		renderer_w,
-		renderer.clip,
-		bitmap->w,
-		bitmap->h,
-		*dst = blend(*dst, *src)
-	);
-}
-
-void render_bitmap_col(
-	const Bitmap* bitmap,
-	int x,
-	int y,
-	const Rectangle* rect,
-	Colour colour
-) {
-	blit_vars(Colour)
-	int renderer_w;
-	Colour* renderer_pixels;
-	renderer_w = get_render_w();
-	renderer_pixels = get_fb();
-	blit(
-		bitmap->pixels,
-		renderer_pixels,
-		renderer_w,
-		renderer.clip,
-		bitmap->w,
-		bitmap->h,
-		*dst = blend_mod(*dst, *src, colour)
-	);
-}
-
-void render_mask(
-	const unsigned char* pixels,
-	int x,
-	int y,
-	int w,
-	int h,
-	const Rectangle* rect,
-	Colour colour
-) {
-	blit_vars(unsigned char)
-	int renderer_w;
-	Colour c;
-	Colour* renderer_pixels;
-	renderer_w = get_render_w();
-	renderer_pixels = get_fb();
-	c = colour;
-	blit(
-		pixels,
-		renderer_pixels,
-		renderer_w,
-		renderer.clip,
-		w,
-		h,
-		{
-			c.a = colour.a * *src;
-			*dst = blend(*dst, c);
-		}
-	);
-}
-
-void render_rect(
-	const Rectangle* rect,
-	Colour colour
-) {
-	blit_rect_vars
-	int renderer_w;
-	Colour* renderer_pixels;
-	renderer_w = get_render_w();
-	renderer_pixels = get_fb();
-	blit_rect(
-		renderer_pixels,
-		renderer_w,
-		renderer.clip
-	);
-}
-
-void rcopy(
-	Bitmap* to,
-	const Bitmap* from,
-	int x,
-	int y,
-	const Rectangle* rect
-) {
-	Rectangle clip;
-	blit_vars(Colour);
-	clip.x = 0;
-	clip.y = 0;
-	clip.w = to->w;
-	clip.h = to->h;
-	blit(
-		from->pixels,
-		to->pixels,
-		to->w,
-		clip,
-		from->w,
-		from->h,
-		*dst = blend(*dst, *src)
-	);
-}
-
-void rcopy_col(
-	Bitmap* to,
-	const Bitmap* from,
-	int x,
-	int y,
-	const Rectangle* rect,
-	Colour colour
-) {
-	Rectangle clip;
-	blit_vars(Colour)
-	clip.x = 0;
-	clip.y = 0;
-	clip.w = to->w;
-	clip.h = to->h;
-	blit(
-		from->pixels,
-		to->pixels,
-		to->w,
-		clip,
-		from->w,
-		from->h,
-		*dst = blend_mod(*dst, *src, colour);
-	);
-}
-
-void rcopy_ac(
-	Bitmap* to,
-	const Bitmap* from,
-	int x,
-	int y,
-	const Rectangle* rect,
-	unsigned char t
-) {
-	Rectangle clip;
-	blit_vars(Colour)
-	clip.x = 0;
-	clip.y = 0;
-	clip.w = to->w;
-	clip.h = to->h;
-	blit(
-		from->pixels,
-		to->pixels,
-		to->w,
-		clip,
-		from->w,
-		from->h,
-		*dst = src->a > t ? *src : *dst;
-	);
-}
-
-void rcopy_rect(
-	Bitmap* to,
-	const Rectangle* rect,
-	Colour colour
-) {
-	Rectangle clip;
-	blit_rect_vars
-	clip.x = 0;
-	clip.y = 0;
-	clip.w = to->w;
-	clip.h = to->h;
-	blit_rect(
-		to->pixels,
-		to->w,
-		clip
-	);
-}
diff --git a/render.h b/render.h
deleted file mode 100644
index 62044a0..0000000
--- a/render.h
+++ /dev/null
@@ -1,140 +0,0 @@
-#ifndef render_h
-#define render_h
-
-#include "config.h"
-#include "rect.h"
-#include "memory.h"
-
-typedef struct {
-	unsigned char r, g, b, a;
-} Colour;
-
-Colour make_colour(unsigned rgb, unsigned char a);
-Colour make_black();
-Colour make_white();
-
-typedef struct Bitmap {
-	Colour* pixels;
-	int w, h;
-} Bitmap;
-
-void init_bitmap(
-	Bitmap* bitmap,
-	Colour* pixels,
-	int w,
-	int h
-);
-
-#define max_glyphset 256
-
-struct Font;
-typedef struct Font Font;
-
-int font_height(Font* f);
-
-void get_font_rect(
-	const Font* font,
-	Rectangle* dst,
-	char c
-);
-
-Font* new_font(
-	Heap* h,
-	const unsigned char* raw,
-	int size
-);
-Rectangle text_rect(Font* font, const char* text);
-
-void rfont_text(
-	Font* font,
-	int x,
-	int y,
-	const char* text
-);
-void rfont_text_col(
-	Font* font,
-	int x,
-	int y,
-	const char* text,
-	Colour colour
-);
-
-void render_init(void);
-
-void render_begin(void);
-void render_clear(void);
-void render_clear_col(Colour col);
-void render_clip(const Rectangle* rect);
-void render_reset_clip(void);
-
-Colour blend(Colour dst, Colour src);
-Colour blend_mod(Colour dst, Colour src, Colour mod);
-
-Colour* get_render_pixels(void);
-
-extern int sprite_camera_x;
-extern int sprite_camera_y;
-
-void render_bitmap(
-	const Bitmap* bitmap,
-	int x,
-	int y,
-	const Rectangle* rect
-);
-
-void render_bitmap_col(
-	const Bitmap* bitmap,
-	int x,
-	int y,
-	const Rectangle* rect,
-	Colour colour
-);
-
-void render_rect(
-	const Rectangle* rect,
-	Colour colour
-);
-
-void render_mask(
-	const unsigned char* pixels,
-	int x,
-	int y,
-	int w,
-	int h,
-	const Rectangle* rect,
-	Colour colour
-);
-
-void rcopy(
-	Bitmap* dst,
-	const Bitmap* src,
-	int x,
-	int y,
-	const Rectangle* rect
-);
-
-void rcopy_col(
-	Bitmap* dst,
-	const Bitmap* src,
-	int x,
-	int y,
-	const Rectangle* rect,
-	Colour colour
-);
-
-void rcopy_ac(
-	Bitmap* dst,
-	const Bitmap* src,
-	int x,
-	int y,
-	const Rectangle* rect,
-	unsigned char t
-);
-
-void rcopy_rect(
-	Bitmap* dst,
-	const Rectangle* rect,
-	Colour colour
-);
-
-#endif
diff --git a/stb_image.h b/stb_image.h
index a632d54..9eedabe 100644
--- a/stb_image.h
+++ b/stb_image.h
@@ -1,4 +1,4 @@
-/* stb_image - v2.29 - public domain image loader - http://nothings.org/stb
+/* stb_image - v2.30 - public domain image loader - http://nothings.org/stb
                                   no warranty implied; use at your own risk
 
    Do this:
@@ -48,6 +48,7 @@ LICENSE
 
 RECENT REVISION HISTORY:
 
+      2.30  (2024-05-31) avoid erroneous gcc warning
       2.29  (2023-05-xx) optimizations
       2.28  (2023-01-29) many error fixes, security errors, just tons of stuff
       2.27  (2021-07-11) document stbi_info better, 16-bit PNM support, bug fixes
@@ -5159,9 +5160,11 @@ static int stbi__parse_png_file(stbi__png *z, int scan, int req_comp)
                // non-paletted with tRNS = constant alpha. if header-scanning, we can stop now.
                if (scan == STBI__SCAN_header) { ++s->img_n; return 1; }
                if (z->depth == 16) {
-                  for (k = 0; k < s->img_n; ++k) tc16[k] = (stbi__uint16)stbi__get16be(s); // copy the values as-is
+                  for (k = 0; k < s->img_n && k < 3; ++k) // extra loop test to suppress false GCC warning
+                     tc16[k] = (stbi__uint16)stbi__get16be(s); // copy the values as-is
                } else {
-                  for (k = 0; k < s->img_n; ++k) tc[k] = (stbi_uc)(stbi__get16be(s) & 255) * stbi__depth_scale_table[z->depth]; // non 8-bit images will be larger
+                  for (k = 0; k < s->img_n && k < 3; ++k)
+                     tc[k] = (stbi_uc)(stbi__get16be(s) & 255) * stbi__depth_scale_table[z->depth]; // non 8-bit images will be larger
                }
             }
             break;
diff --git a/stb_rect_pack.h b/stb_rect_pack.h
deleted file mode 100644
index 6a633ce..0000000
--- a/stb_rect_pack.h
+++ /dev/null
@@ -1,623 +0,0 @@
-// stb_rect_pack.h - v1.01 - public domain - rectangle packing
-// Sean Barrett 2014
-//
-// Useful for e.g. packing rectangular textures into an atlas.
-// Does not do rotation.
-//
-// Before #including,
-//
-//    #define STB_RECT_PACK_IMPLEMENTATION
-//
-// in the file that you want to have the implementation.
-//
-// Not necessarily the awesomest packing method, but better than
-// the totally naive one in stb_truetype (which is primarily what
-// this is meant to replace).
-//
-// Has only had a few tests run, may have issues.
-//
-// More docs to come.
-//
-// No memory allocations; uses qsort() and assert() from stdlib.
-// Can override those by defining STBRP_SORT and STBRP_ASSERT.
-//
-// This library currently uses the Skyline Bottom-Left algorithm.
-//
-// Please note: better rectangle packers are welcome! Please
-// implement them to the same API, but with a different init
-// function.
-//
-// Credits
-//
-//  Library
-//    Sean Barrett
-//  Minor features
-//    Martins Mozeiko
-//    github:IntellectualKitty
-//
-//  Bugfixes / warning fixes
-//    Jeremy Jaussaud
-//    Fabian Giesen
-//
-// Version history:
-//
-//     1.01  (2021-07-11)  always use large rect mode, expose STBRP__MAXVAL in public section
-//     1.00  (2019-02-25)  avoid small space waste; gracefully fail too-wide rectangles
-//     0.99  (2019-02-07)  warning fixes
-//     0.11  (2017-03-03)  return packing success/fail result
-//     0.10  (2016-10-25)  remove cast-away-const to avoid warnings
-//     0.09  (2016-08-27)  fix compiler warnings
-//     0.08  (2015-09-13)  really fix bug with empty rects (w=0 or h=0)
-//     0.07  (2015-09-13)  fix bug with empty rects (w=0 or h=0)
-//     0.06  (2015-04-15)  added STBRP_SORT to allow replacing qsort
-//     0.05:  added STBRP_ASSERT to allow replacing assert
-//     0.04:  fixed minor bug in STBRP_LARGE_RECTS support
-//     0.01:  initial release
-//
-// LICENSE
-//
-//   See end of file for license information.
-
-//////////////////////////////////////////////////////////////////////////////
-//
-//       INCLUDE SECTION
-//
-
-#ifndef STB_INCLUDE_STB_RECT_PACK_H
-#define STB_INCLUDE_STB_RECT_PACK_H
-
-#define STB_RECT_PACK_VERSION  1
-
-#ifdef STBRP_STATIC
-#define STBRP_DEF static
-#else
-#define STBRP_DEF extern
-#endif
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-
-typedef struct stbrp_context stbrp_context;
-typedef struct stbrp_node    stbrp_node;
-typedef struct stbrp_rect    stbrp_rect;
-
-typedef int            stbrp_coord;
-
-#define STBRP__MAXVAL  0x7fffffff
-// Mostly for internal use, but this is the maximum supported coordinate value.
-
-STBRP_DEF int stbrp_pack_rects (stbrp_context *context, stbrp_rect *rects, int num_rects);
-// Assign packed locations to rectangles. The rectangles are of type
-// 'stbrp_rect' defined below, stored in the array 'rects', and there
-// are 'num_rects' many of them.
-//
-// Rectangles which are successfully packed have the 'was_packed' flag
-// set to a non-zero value and 'x' and 'y' store the minimum location
-// on each axis (i.e. bottom-left in cartesian coordinates, top-left
-// if you imagine y increasing downwards). Rectangles which do not fit
-// have the 'was_packed' flag set to 0.
-//
-// You should not try to access the 'rects' array from another thread
-// while this function is running, as the function temporarily reorders
-// the array while it executes.
-//
-// To pack into another rectangle, you need to call stbrp_init_target
-// again. To continue packing into the same rectangle, you can call
-// this function again. Calling this multiple times with multiple rect
-// arrays will probably produce worse packing results than calling it
-// a single time with the full rectangle array, but the option is
-// available.
-//
-// The function returns 1 if all of the rectangles were successfully
-// packed and 0 otherwise.
-
-struct stbrp_rect
-{
-   // reserved for your use:
-   int            id;
-
-   // input:
-   stbrp_coord    w, h;
-
-   // output:
-   stbrp_coord    x, y;
-   int            was_packed;  // non-zero if valid packing
-
-}; // 16 bytes, nominally
-
-
-STBRP_DEF void stbrp_init_target (stbrp_context *context, int width, int height, stbrp_node *nodes, int num_nodes);
-// Initialize a rectangle packer to:
-//    pack a rectangle that is 'width' by 'height' in dimensions
-//    using temporary storage provided by the array 'nodes', which is 'num_nodes' long
-//
-// You must call this function every time you start packing into a new target.
-//
-// There is no "shutdown" function. The 'nodes' memory must stay valid for
-// the following stbrp_pack_rects() call (or calls), but can be freed after
-// the call (or calls) finish.
-//
-// Note: to guarantee best results, either:
-//       1. make sure 'num_nodes' >= 'width'
-//   or  2. call stbrp_allow_out_of_mem() defined below with 'allow_out_of_mem = 1'
-//
-// If you don't do either of the above things, widths will be quantized to multiples
-// of small integers to guarantee the algorithm doesn't run out of temporary storage.
-//
-// If you do #2, then the non-quantized algorithm will be used, but the algorithm
-// may run out of temporary storage and be unable to pack some rectangles.
-
-STBRP_DEF void stbrp_setup_allow_out_of_mem (stbrp_context *context, int allow_out_of_mem);
-// Optionally call this function after init but before doing any packing to
-// change the handling of the out-of-temp-memory scenario, described above.
-// If you call init again, this will be reset to the default (false).
-
-
-STBRP_DEF void stbrp_setup_heuristic (stbrp_context *context, int heuristic);
-// Optionally select which packing heuristic the library should use. Different
-// heuristics will produce better/worse results for different data sets.
-// If you call init again, this will be reset to the default.
-
-enum
-{
-   STBRP_HEURISTIC_Skyline_default=0,
-   STBRP_HEURISTIC_Skyline_BL_sortHeight = STBRP_HEURISTIC_Skyline_default,
-   STBRP_HEURISTIC_Skyline_BF_sortHeight
-};
-
-
-//////////////////////////////////////////////////////////////////////////////
-//
-// the details of the following structures don't matter to you, but they must
-// be visible so you can handle the memory allocations for them
-
-struct stbrp_node
-{
-   stbrp_coord  x,y;
-   stbrp_node  *next;
-};
-
-struct stbrp_context
-{
-   int width;
-   int height;
-   int align;
-   int init_mode;
-   int heuristic;
-   int num_nodes;
-   stbrp_node *active_head;
-   stbrp_node *free_head;
-   stbrp_node extra[2]; // we allocate two extra nodes so optimal user-node-count is 'width' not 'width+2'
-};
-
-#ifdef __cplusplus
-}
-#endif
-
-#endif
-
-//////////////////////////////////////////////////////////////////////////////
-//
-//     IMPLEMENTATION SECTION
-//
-
-#ifdef STB_RECT_PACK_IMPLEMENTATION
-#ifndef STBRP_SORT
-#include <stdlib.h>
-#define STBRP_SORT qsort
-#endif
-
-#ifndef STBRP_ASSERT
-#include <assert.h>
-#define STBRP_ASSERT assert
-#endif
-
-#ifdef _MSC_VER
-#define STBRP__NOTUSED(v)  (void)(v)
-#define STBRP__CDECL       __cdecl
-#else
-#define STBRP__NOTUSED(v)  (void)sizeof(v)
-#define STBRP__CDECL
-#endif
-
-enum
-{
-   STBRP__INIT_skyline = 1
-};
-
-STBRP_DEF void stbrp_setup_heuristic(stbrp_context *context, int heuristic)
-{
-   switch (context->init_mode) {
-      case STBRP__INIT_skyline:
-         STBRP_ASSERT(heuristic == STBRP_HEURISTIC_Skyline_BL_sortHeight || heuristic == STBRP_HEURISTIC_Skyline_BF_sortHeight);
-         context->heuristic = heuristic;
-         break;
-      default:
-         STBRP_ASSERT(0);
-   }
-}
-
-STBRP_DEF void stbrp_setup_allow_out_of_mem(stbrp_context *context, int allow_out_of_mem)
-{
-   if (allow_out_of_mem)
-      // if it's ok to run out of memory, then don't bother aligning them;
-      // this gives better packing, but may fail due to OOM (even though
-      // the rectangles easily fit). @TODO a smarter approach would be to only
-      // quantize once we've hit OOM, then we could get rid of this parameter.
-      context->align = 1;
-   else {
-      // if it's not ok to run out of memory, then quantize the widths
-      // so that num_nodes is always enough nodes.
-      //
-      // I.e. num_nodes * align >= width
-      //                  align >= width / num_nodes
-      //                  align = ceil(width/num_nodes)
-
-      context->align = (context->width + context->num_nodes-1) / context->num_nodes;
-   }
-}
-
-STBRP_DEF void stbrp_init_target(stbrp_context *context, int width, int height, stbrp_node *nodes, int num_nodes)
-{
-   int i;
-
-   for (i=0; i < num_nodes-1; ++i)
-      nodes[i].next = &nodes[i+1];
-   nodes[i].next = NULL;
-   context->init_mode = STBRP__INIT_skyline;
-   context->heuristic = STBRP_HEURISTIC_Skyline_default;
-   context->free_head = &nodes[0];
-   context->active_head = &context->extra[0];
-   context->width = width;
-   context->height = height;
-   context->num_nodes = num_nodes;
-   stbrp_setup_allow_out_of_mem(context, 0);
-
-   // node 0 is the full width, node 1 is the sentinel (lets us not store width explicitly)
-   context->extra[0].x = 0;
-   context->extra[0].y = 0;
-   context->extra[0].next = &context->extra[1];
-   context->extra[1].x = (stbrp_coord) width;
-   context->extra[1].y = (1<<30);
-   context->extra[1].next = NULL;
-}
-
-// find minimum y position if it starts at x1
-static int stbrp__skyline_find_min_y(stbrp_context *c, stbrp_node *first, int x0, int width, int *pwaste)
-{
-   stbrp_node *node = first;
-   int x1 = x0 + width;
-   int min_y, visited_width, waste_area;
-
-   STBRP__NOTUSED(c);
-
-   STBRP_ASSERT(first->x <= x0);
-
-   #if 0
-   // skip in case we're past the node
-   while (node->next->x <= x0)
-      ++node;
-   #else
-   STBRP_ASSERT(node->next->x > x0); // we ended up handling this in the caller for efficiency
-   #endif
-
-   STBRP_ASSERT(node->x <= x0);
-
-   min_y = 0;
-   waste_area = 0;
-   visited_width = 0;
-   while (node->x < x1) {
-      if (node->y > min_y) {
-         // raise min_y higher.
-         // we've accounted for all waste up to min_y,
-         // but we'll now add more waste for everything we've visted
-         waste_area += visited_width * (node->y - min_y);
-         min_y = node->y;
-         // the first time through, visited_width might be reduced
-         if (node->x < x0)
-            visited_width += node->next->x - x0;
-         else
-            visited_width += node->next->x - node->x;
-      } else {
-         // add waste area
-         int under_width = node->next->x - node->x;
-         if (under_width + visited_width > width)
-            under_width = width - visited_width;
-         waste_area += under_width * (min_y - node->y);
-         visited_width += under_width;
-      }
-      node = node->next;
-   }
-
-   *pwaste = waste_area;
-   return min_y;
-}
-
-typedef struct
-{
-   int x,y;
-   stbrp_node **prev_link;
-} stbrp__findresult;
-
-static stbrp__findresult stbrp__skyline_find_best_pos(stbrp_context *c, int width, int height)
-{
-   int best_waste = (1<<30), best_x, best_y = (1 << 30);
-   stbrp__findresult fr;
-   stbrp_node **prev, *node, *tail, **best = NULL;
-
-   // align to multiple of c->align
-   width = (width + c->align - 1);
-   width -= width % c->align;
-   STBRP_ASSERT(width % c->align == 0);
-
-   // if it can't possibly fit, bail immediately
-   if (width > c->width || height > c->height) {
-      fr.prev_link = NULL;
-      fr.x = fr.y = 0;
-      return fr;
-   }
-
-   node = c->active_head;
-   prev = &c->active_head;
-   while (node->x + width <= c->width) {
-      int y,waste;
-      y = stbrp__skyline_find_min_y(c, node, node->x, width, &waste);
-      if (c->heuristic == STBRP_HEURISTIC_Skyline_BL_sortHeight) { // actually just want to test BL
-         // bottom left
-         if (y < best_y) {
-            best_y = y;
-            best = prev;
-         }
-      } else {
-         // best-fit
-         if (y + height <= c->height) {
-            // can only use it if it first vertically
-            if (y < best_y || (y == best_y && waste < best_waste)) {
-               best_y = y;
-               best_waste = waste;
-               best = prev;
-            }
-         }
-      }
-      prev = &node->next;
-      node = node->next;
-   }
-
-   best_x = (best == NULL) ? 0 : (*best)->x;
-
-   // if doing best-fit (BF), we also have to try aligning right edge to each node position
-   //
-   // e.g, if fitting
-   //
-   //     ____________________
-   //    |____________________|
-   //
-   //            into
-   //
-   //   |                         |
-   //   |             ____________|
-   //   |____________|
-   //
-   // then right-aligned reduces waste, but bottom-left BL is always chooses left-aligned
-   //
-   // This makes BF take about 2x the time
-
-   if (c->heuristic == STBRP_HEURISTIC_Skyline_BF_sortHeight) {
-      tail = c->active_head;
-      node = c->active_head;
-      prev = &c->active_head;
-      // find first node that's admissible
-      while (tail->x < width)
-         tail = tail->next;
-      while (tail) {
-         int xpos = tail->x - width;
-         int y,waste;
-         STBRP_ASSERT(xpos >= 0);
-         // find the left position that matches this
-         while (node->next->x <= xpos) {
-            prev = &node->next;
-            node = node->next;
-         }
-         STBRP_ASSERT(node->next->x > xpos && node->x <= xpos);
-         y = stbrp__skyline_find_min_y(c, node, xpos, width, &waste);
-         if (y + height <= c->height) {
-            if (y <= best_y) {
-               if (y < best_y || waste < best_waste || (waste==best_waste && xpos < best_x)) {
-                  best_x = xpos;
-                  STBRP_ASSERT(y <= best_y);
-                  best_y = y;
-                  best_waste = waste;
-                  best = prev;
-               }
-            }
-         }
-         tail = tail->next;
-      }
-   }
-
-   fr.prev_link = best;
-   fr.x = best_x;
-   fr.y = best_y;
-   return fr;
-}
-
-static stbrp__findresult stbrp__skyline_pack_rectangle(stbrp_context *context, int width, int height)
-{
-   // find best position according to heuristic
-   stbrp__findresult res = stbrp__skyline_find_best_pos(context, width, height);
-   stbrp_node *node, *cur;
-
-   // bail if:
-   //    1. it failed
-   //    2. the best node doesn't fit (we don't always check this)
-   //    3. we're out of memory
-   if (res.prev_link == NULL || res.y + height > context->height || context->free_head == NULL) {
-      res.prev_link = NULL;
-      return res;
-   }
-
-   // on success, create new node
-   node = context->free_head;
-   node->x = (stbrp_coord) res.x;
-   node->y = (stbrp_coord) (res.y + height);
-
-   context->free_head = node->next;
-
-   // insert the new node into the right starting point, and
-   // let 'cur' point to the remaining nodes needing to be
-   // stiched back in
-
-   cur = *res.prev_link;
-   if (cur->x < res.x) {
-      // preserve the existing one, so start testing with the next one
-      stbrp_node *next = cur->next;
-      cur->next = node;
-      cur = next;
-   } else {
-      *res.prev_link = node;
-   }
-
-   // from here, traverse cur and free the nodes, until we get to one
-   // that shouldn't be freed
-   while (cur->next && cur->next->x <= res.x + width) {
-      stbrp_node *next = cur->next;
-      // move the current node to the free list
-      cur->next = context->free_head;
-      context->free_head = cur;
-      cur = next;
-   }
-
-   // stitch the list back in
-   node->next = cur;
-
-   if (cur->x < res.x + width)
-      cur->x = (stbrp_coord) (res.x + width);
-
-#ifdef _DEBUG
-   cur = context->active_head;
-   while (cur->x < context->width) {
-      STBRP_ASSERT(cur->x < cur->next->x);
-      cur = cur->next;
-   }
-   STBRP_ASSERT(cur->next == NULL);
-
-   {
-      int count=0;
-      cur = context->active_head;
-      while (cur) {
-         cur = cur->next;
-         ++count;
-      }
-      cur = context->free_head;
-      while (cur) {
-         cur = cur->next;
-         ++count;
-      }
-      STBRP_ASSERT(count == context->num_nodes+2);
-   }
-#endif
-
-   return res;
-}
-
-static int STBRP__CDECL rect_height_compare(const void *a, const void *b)
-{
-   const stbrp_rect *p = (const stbrp_rect *) a;
-   const stbrp_rect *q = (const stbrp_rect *) b;
-   if (p->h > q->h)
-      return -1;
-   if (p->h < q->h)
-      return  1;
-   return (p->w > q->w) ? -1 : (p->w < q->w);
-}
-
-static int STBRP__CDECL rect_original_order(const void *a, const void *b)
-{
-   const stbrp_rect *p = (const stbrp_rect *) a;
-   const stbrp_rect *q = (const stbrp_rect *) b;
-   return (p->was_packed < q->was_packed) ? -1 : (p->was_packed > q->was_packed);
-}
-
-STBRP_DEF int stbrp_pack_rects(stbrp_context *context, stbrp_rect *rects, int num_rects)
-{
-   int i, all_rects_packed = 1;
-
-   // we use the 'was_packed' field internally to allow sorting/unsorting
-   for (i=0; i < num_rects; ++i) {
-      rects[i].was_packed = i;
-   }
-
-   // sort according to heuristic
-   STBRP_SORT(rects, num_rects, sizeof(rects[0]), rect_height_compare);
-
-   for (i=0; i < num_rects; ++i) {
-      if (rects[i].w == 0 || rects[i].h == 0) {
-         rects[i].x = rects[i].y = 0;  // empty rect needs no space
-      } else {
-         stbrp__findresult fr = stbrp__skyline_pack_rectangle(context, rects[i].w, rects[i].h);
-         if (fr.prev_link) {
-            rects[i].x = (stbrp_coord) fr.x;
-            rects[i].y = (stbrp_coord) fr.y;
-         } else {
-            rects[i].x = rects[i].y = STBRP__MAXVAL;
-         }
-      }
-   }
-
-   // unsort
-   STBRP_SORT(rects, num_rects, sizeof(rects[0]), rect_original_order);
-
-   // set was_packed flags and all_rects_packed status
-   for (i=0; i < num_rects; ++i) {
-      rects[i].was_packed = !(rects[i].x == STBRP__MAXVAL && rects[i].y == STBRP__MAXVAL);
-      if (!rects[i].was_packed)
-         all_rects_packed = 0;
-   }
-
-   // return the all_rects_packed status
-   return all_rects_packed;
-}
-#endif
-
-/*
-------------------------------------------------------------------------------
-This software is available under 2 licenses -- choose whichever you prefer.
-------------------------------------------------------------------------------
-ALTERNATIVE A - MIT License
-Copyright (c) 2017 Sean Barrett
-Permission is hereby granted, free of charge, to any person obtaining a copy of
-this software and associated documentation files (the "Software"), to deal in
-the Software without restriction, including without limitation the rights to
-use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies
-of the Software, and to permit persons to whom the Software is furnished to do
-so, subject to the following conditions:
-The above copyright notice and this permission notice shall be included in all
-copies or substantial portions of the Software.
-THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
-IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
-FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
-AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
-LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
-OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
-SOFTWARE.
-------------------------------------------------------------------------------
-ALTERNATIVE B - Public Domain (www.unlicense.org)
-This is free and unencumbered software released into the public domain.
-Anyone is free to copy, modify, publish, use, compile, sell, or distribute this
-software, either in source code form or as a compiled binary, for any purpose,
-commercial or non-commercial, and by any means.
-In jurisdictions that recognize copyright laws, the author or authors of this
-software dedicate any and all copyright interest in the software to the public
-domain. We make this dedication for the benefit of the public at large and to
-the detriment of our heirs and successors. We intend this dedication to be an
-overt act of relinquishment in perpetuity of all present and future rights to
-this software under copyright law.
-THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
-IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
-FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
-AUTHORS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
-ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION
-WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
-------------------------------------------------------------------------------
-*/
diff --git a/stb_truetype.h b/stb_truetype.h
deleted file mode 100644
index bbf2284..0000000
--- a/stb_truetype.h
+++ /dev/null
@@ -1,5077 +0,0 @@
-// stb_truetype.h - v1.26 - public domain
-// authored from 2009-2021 by Sean Barrett / RAD Game Tools
-//
-// =======================================================================
-//
-//    NO SECURITY GUARANTEE -- DO NOT USE THIS ON UNTRUSTED FONT FILES
-//
-// This library does no range checking of the offsets found in the file,
-// meaning an attacker can use it to read arbitrary memory.
-//
-// =======================================================================
-//
-//   This library processes TrueType files:
-//        parse files
-//        extract glyph metrics
-//        extract glyph shapes
-//        render glyphs to one-channel bitmaps with antialiasing (box filter)
-//        render glyphs to one-channel SDF bitmaps (signed-distance field/function)
-//
-//   Todo:
-//        non-MS cmaps
-//        crashproof on bad data
-//        hinting? (no longer patented)
-//        cleartype-style AA?
-//        optimize: use simple memory allocator for intermediates
-//        optimize: build edge-list directly from curves
-//        optimize: rasterize directly from curves?
-//
-// ADDITIONAL CONTRIBUTORS
-//
-//   Mikko Mononen: compound shape support, more cmap formats
-//   Tor Andersson: kerning, subpixel rendering
-//   Dougall Johnson: OpenType / Type 2 font handling
-//   Daniel Ribeiro Maciel: basic GPOS-based kerning
-//
-//   Misc other:
-//       Ryan Gordon
-//       Simon Glass
-//       github:IntellectualKitty
-//       Imanol Celaya
-//       Daniel Ribeiro Maciel
-//
-//   Bug/warning reports/fixes:
-//       "Zer" on mollyrocket       Fabian "ryg" Giesen   github:NiLuJe
-//       Cass Everitt               Martins Mozeiko       github:aloucks
-//       stoiko (Haemimont Games)   Cap Petschulat        github:oyvindjam
-//       Brian Hook                 Omar Cornut           github:vassvik
-//       Walter van Niftrik         Ryan Griege
-//       David Gow                  Peter LaValle
-//       David Given                Sergey Popov
-//       Ivan-Assen Ivanov          Giumo X. Clanjor
-//       Anthony Pesch              Higor Euripedes
-//       Johan Duparc               Thomas Fields
-//       Hou Qiming                 Derek Vinyard
-//       Rob Loach                  Cort Stratton
-//       Kenney Phillis Jr.         Brian Costabile
-//       Ken Voskuil (kaesve)
-//
-// VERSION HISTORY
-//
-//   1.26 (2021-08-28) fix broken rasterizer
-//   1.25 (2021-07-11) many fixes
-//   1.24 (2020-02-05) fix warning
-//   1.23 (2020-02-02) query SVG data for glyphs; query whole kerning table (but only kern not GPOS)
-//   1.22 (2019-08-11) minimize missing-glyph duplication; fix kerning if both 'GPOS' and 'kern' are defined
-//   1.21 (2019-02-25) fix warning
-//   1.20 (2019-02-07) PackFontRange skips missing codepoints; GetScaleFontVMetrics()
-//   1.19 (2018-02-11) GPOS kerning, STBTT_fmod
-//   1.18 (2018-01-29) add missing function
-//   1.17 (2017-07-23) make more arguments const; doc fix
-//   1.16 (2017-07-12) SDF support
-//   1.15 (2017-03-03) make more arguments const
-//   1.14 (2017-01-16) num-fonts-in-TTC function
-//   1.13 (2017-01-02) support OpenType fonts, certain Apple fonts
-//   1.12 (2016-10-25) suppress warnings about casting away const with -Wcast-qual
-//   1.11 (2016-04-02) fix unused-variable warning
-//   1.10 (2016-04-02) user-defined fabs(); rare memory leak; remove duplicate typedef
-//   1.09 (2016-01-16) warning fix; avoid crash on outofmem; use allocation userdata properly
-//   1.08 (2015-09-13) document stbtt_Rasterize(); fixes for vertical & horizontal edges
-//   1.07 (2015-08-01) allow PackFontRanges to accept arrays of sparse codepoints;
-//                     variant PackFontRanges to pack and render in separate phases;
-//                     fix stbtt_GetFontOFfsetForIndex (never worked for non-0 input?);
-//                     fixed an assert() bug in the new rasterizer
-//                     replace assert() with STBTT_assert() in new rasterizer
-//
-//   Full history can be found at the end of this file.
-//
-// LICENSE
-//
-//   See end of file for license information.
-//
-// USAGE
-//
-//   Include this file in whatever places need to refer to it. In ONE C/C++
-//   file, write:
-//      #define STB_TRUETYPE_IMPLEMENTATION
-//   before the #include of this file. This expands out the actual
-//   implementation into that C/C++ file.
-//
-//   To make the implementation private to the file that generates the implementation,
-//      #define STBTT_STATIC
-//
-//   Simple 3D API (don't ship this, but it's fine for tools and quick start)
-//           stbtt_BakeFontBitmap()               -- bake a font to a bitmap for use as texture
-//           stbtt_GetBakedQuad()                 -- compute quad to draw for a given char
-//
-//   Improved 3D API (more shippable):
-//           #include "stb_rect_pack.h"           -- optional, but you really want it
-//           stbtt_PackBegin()
-//           stbtt_PackSetOversampling()          -- for improved quality on small fonts
-//           stbtt_PackFontRanges()               -- pack and renders
-//           stbtt_PackEnd()
-//           stbtt_GetPackedQuad()
-//
-//   "Load" a font file from a memory buffer (you have to keep the buffer loaded)
-//           stbtt_InitFont()
-//           stbtt_GetFontOffsetForIndex()        -- indexing for TTC font collections
-//           stbtt_GetNumberOfFonts()             -- number of fonts for TTC font collections
-//
-//   Render a unicode codepoint to a bitmap
-//           stbtt_GetCodepointBitmap()           -- allocates and returns a bitmap
-//           stbtt_MakeCodepointBitmap()          -- renders into bitmap you provide
-//           stbtt_GetCodepointBitmapBox()        -- how big the bitmap must be
-//
-//   Character advance/positioning
-//           stbtt_GetCodepointHMetrics()
-//           stbtt_GetFontVMetrics()
-//           stbtt_GetFontVMetricsOS2()
-//           stbtt_GetCodepointKernAdvance()
-//
-//   Starting with version 1.06, the rasterizer was replaced with a new,
-//   faster and generally-more-precise rasterizer. The new rasterizer more
-//   accurately measures pixel coverage for anti-aliasing, except in the case
-//   where multiple shapes overlap, in which case it overestimates the AA pixel
-//   coverage. Thus, anti-aliasing of intersecting shapes may look wrong. If
-//   this turns out to be a problem, you can re-enable the old rasterizer with
-//        #define STBTT_RASTERIZER_VERSION 1
-//   which will incur about a 15% speed hit.
-//
-// ADDITIONAL DOCUMENTATION
-//
-//   Immediately after this block comment are a series of sample programs.
-//
-//   After the sample programs is the "header file" section. This section
-//   includes documentation for each API function.
-//
-//   Some important concepts to understand to use this library:
-//
-//      Codepoint
-//         Characters are defined by unicode codepoints, e.g. 65 is
-//         uppercase A, 231 is lowercase c with a cedilla, 0x7e30 is
-//         the hiragana for "ma".
-//
-//      Glyph
-//         A visual character shape (every codepoint is rendered as
-//         some glyph)
-//
-//      Glyph index
-//         A font-specific integer ID representing a glyph
-//
-//      Baseline
-//         Glyph shapes are defined relative to a baseline, which is the
-//         bottom of uppercase characters. Characters extend both above
-//         and below the baseline.
-//
-//      Current Point
-//         As you draw text to the screen, you keep track of a "current point"
-//         which is the origin of each character. The current point's vertical
-//         position is the baseline. Even "baked fonts" use this model.
-//
-//      Vertical Font Metrics
-//         The vertical qualities of the font, used to vertically position
-//         and space the characters. See docs for stbtt_GetFontVMetrics.
-//
-//      Font Size in Pixels or Points
-//         The preferred interface for specifying font sizes in stb_truetype
-//         is to specify how tall the font's vertical extent should be in pixels.
-//         If that sounds good enough, skip the next paragraph.
-//
-//         Most font APIs instead use "points", which are a common typographic
-//         measurement for describing font size, defined as 72 points per inch.
-//         stb_truetype provides a point API for compatibility. However, true
-//         "per inch" conventions don't make much sense on computer displays
-//         since different monitors have different number of pixels per
-//         inch. For example, Windows traditionally uses a convention that
-//         there are 96 pixels per inch, thus making 'inch' measurements have
-//         nothing to do with inches, and thus effectively defining a point to
-//         be 1.333 pixels. Additionally, the TrueType font data provides
-//         an explicit scale factor to scale a given font's glyphs to points,
-//         but the author has observed that this scale factor is often wrong
-//         for non-commercial fonts, thus making fonts scaled in points
-//         according to the TrueType spec incoherently sized in practice.
-//
-// DETAILED USAGE:
-//
-//  Scale:
-//    Select how high you want the font to be, in points or pixels.
-//    Call ScaleForPixelHeight or ScaleForMappingEmToPixels to compute
-//    a scale factor SF that will be used by all other functions.
-//
-//  Baseline:
-//    You need to select a y-coordinate that is the baseline of where
-//    your text will appear. Call GetFontBoundingBox to get the baseline-relative
-//    bounding box for all characters. SF*-y0 will be the distance in pixels
-//    that the worst-case character could extend above the baseline, so if
-//    you want the top edge of characters to appear at the top of the
-//    screen where y=0, then you would set the baseline to SF*-y0.
-//
-//  Current point:
-//    Set the current point where the first character will appear. The
-//    first character could extend left of the current point; this is font
-//    dependent. You can either choose a current point that is the leftmost
-//    point and hope, or add some padding, or check the bounding box or
-//    left-side-bearing of the first character to be displayed and set
-//    the current point based on that.
-//
-//  Displaying a character:
-//    Compute the bounding box of the character. It will contain signed values
-//    relative to <current_point, baseline>. I.e. if it returns x0,y0,x1,y1,
-//    then the character should be displayed in the rectangle from
-//    <current_point+SF*x0, baseline+SF*y0> to <current_point+SF*x1,baseline+SF*y1).
-//
-//  Advancing for the next character:
-//    Call GlyphHMetrics, and compute 'current_point += SF * advance'.
-//
-//
-// ADVANCED USAGE
-//
-//   Quality:
-//
-//    - Use the functions with Subpixel at the end to allow your characters
-//      to have subpixel positioning. Since the font is anti-aliased, not
-//      hinted, this is very import for quality. (This is not possible with
-//      baked fonts.)
-//
-//    - Kerning is now supported, and if you're supporting subpixel rendering
-//      then kerning is worth using to give your text a polished look.
-//
-//   Performance:
-//
-//    - Convert Unicode codepoints to glyph indexes and operate on the glyphs;
-//      if you don't do this, stb_truetype is forced to do the conversion on
-//      every call.
-//
-//    - There are a lot of memory allocations. We should modify it to take
-//      a temp buffer and allocate from the temp buffer (without freeing),
-//      should help performance a lot.
-//
-// NOTES
-//
-//   The system uses the raw data found in the .ttf file without changing it
-//   and without building auxiliary data structures. This is a bit inefficient
-//   on little-endian systems (the data is big-endian), but assuming you're
-//   caching the bitmaps or glyph shapes this shouldn't be a big deal.
-//
-//   It appears to be very hard to programmatically determine what font a
-//   given file is in a general way. I provide an API for this, but I don't
-//   recommend it.
-//
-//
-// PERFORMANCE MEASUREMENTS FOR 1.06:
-//
-//                      32-bit     64-bit
-//   Previous release:  8.83 s     7.68 s
-//   Pool allocations:  7.72 s     6.34 s
-//   Inline sort     :  6.54 s     5.65 s
-//   New rasterizer  :  5.63 s     5.00 s
-
-//////////////////////////////////////////////////////////////////////////////
-//////////////////////////////////////////////////////////////////////////////
-////
-////  SAMPLE PROGRAMS
-////
-//
-//  Incomplete text-in-3d-api example, which draws quads properly aligned to be lossless.
-//  See "tests/truetype_demo_win32.c" for a complete version.
-#if 0
-#define STB_TRUETYPE_IMPLEMENTATION  // force following include to generate implementation
-#include "stb_truetype.h"
-
-unsigned char ttf_buffer[1<<20];
-unsigned char temp_bitmap[512*512];
-
-stbtt_bakedchar cdata[96]; // ASCII 32..126 is 95 glyphs
-GLuint ftex;
-
-void my_stbtt_initfont(void)
-{
-   fread(ttf_buffer, 1, 1<<20, fopen("c:/windows/fonts/times.ttf", "rb"));
-   stbtt_BakeFontBitmap(ttf_buffer,0, 32.0, temp_bitmap,512,512, 32,96, cdata); // no guarantee this fits!
-   // can free ttf_buffer at this point
-   glGenTextures(1, &ftex);
-   glBindTexture(GL_TEXTURE_2D, ftex);
-   glTexImage2D(GL_TEXTURE_2D, 0, GL_ALPHA, 512,512, 0, GL_ALPHA, GL_UNSIGNED_BYTE, temp_bitmap);
-   // can free temp_bitmap at this point
-   glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
-}
-
-void my_stbtt_print(float x, float y, char *text)
-{
-   // assume orthographic projection with units = screen pixels, origin at top left
-   glEnable(GL_BLEND);
-   glBlendFunc(GL_SRC_ALPHA, GL_ONE_MINUS_SRC_ALPHA);
-   glEnable(GL_TEXTURE_2D);
-   glBindTexture(GL_TEXTURE_2D, ftex);
-   glBegin(GL_QUADS);
-   while (*text) {
-      if (*text >= 32 && *text < 128) {
-         stbtt_aligned_quad q;
-         stbtt_GetBakedQuad(cdata, 512,512, *text-32, &x,&y,&q,1);//1=opengl & d3d10+,0=d3d9
-         glTexCoord2f(q.s0,q.t0); glVertex2f(q.x0,q.y0);
-         glTexCoord2f(q.s1,q.t0); glVertex2f(q.x1,q.y0);
-         glTexCoord2f(q.s1,q.t1); glVertex2f(q.x1,q.y1);
-         glTexCoord2f(q.s0,q.t1); glVertex2f(q.x0,q.y1);
-      }
-      ++text;
-   }
-   glEnd();
-}
-#endif
-//
-//
-//////////////////////////////////////////////////////////////////////////////
-//
-// Complete program (this compiles): get a single bitmap, print as ASCII art
-//
-#if 0
-#include <stdio.h>
-#define STB_TRUETYPE_IMPLEMENTATION  // force following include to generate implementation
-#include "stb_truetype.h"
-
-char ttf_buffer[1<<25];
-
-int main(int argc, char **argv)
-{
-   stbtt_fontinfo font;
-   unsigned char *bitmap;
-   int w,h,i,j,c = (argc > 1 ? atoi(argv[1]) : 'a'), s = (argc > 2 ? atoi(argv[2]) : 20);
-
-   fread(ttf_buffer, 1, 1<<25, fopen(argc > 3 ? argv[3] : "c:/windows/fonts/arialbd.ttf", "rb"));
-
-   stbtt_InitFont(&font, ttf_buffer, stbtt_GetFontOffsetForIndex(ttf_buffer,0));
-   bitmap = stbtt_GetCodepointBitmap(&font, 0,stbtt_ScaleForPixelHeight(&font, s), c, &w, &h, 0,0);
-
-   for (j=0; j < h; ++j) {
-      for (i=0; i < w; ++i)
-         putchar(" .:ioVM@"[bitmap[j*w+i]>>5]);
-      putchar('\n');
-   }
-   return 0;
-}
-#endif
-//
-// Output:
-//
-//     .ii.
-//    @@@@@@.
-//   V@Mio@@o
-//   :i.  V@V
-//     :oM@@M
-//   :@@@MM@M
-//   @@o  o@M
-//  :@@.  M@M
-//   @@@o@@@@
-//   :M@@V:@@.
-//
-//////////////////////////////////////////////////////////////////////////////
-//
-// Complete program: print "Hello World!" banner, with bugs
-//
-#if 0
-char buffer[24<<20];
-unsigned char screen[20][79];
-
-int main(int arg, char **argv)
-{
-   stbtt_fontinfo font;
-   int i,j,ascent,baseline,ch=0;
-   float scale, xpos=2; // leave a little padding in case the character extends left
-   char *text = "Heljo World!"; // intentionally misspelled to show 'lj' brokenness
-
-   fread(buffer, 1, 1000000, fopen("c:/windows/fonts/arialbd.ttf", "rb"));
-   stbtt_InitFont(&font, buffer, 0);
-
-   scale = stbtt_ScaleForPixelHeight(&font, 15);
-   stbtt_GetFontVMetrics(&font, &ascent,0,0);
-   baseline = (int) (ascent*scale);
-
-   while (text[ch]) {
-      int advance,lsb,x0,y0,x1,y1;
-      float x_shift = xpos - (float) floor(xpos);
-      stbtt_GetCodepointHMetrics(&font, text[ch], &advance, &lsb);
-      stbtt_GetCodepointBitmapBoxSubpixel(&font, text[ch], scale,scale,x_shift,0, &x0,&y0,&x1,&y1);
-      stbtt_MakeCodepointBitmapSubpixel(&font, &screen[baseline + y0][(int) xpos + x0], x1-x0,y1-y0, 79, scale,scale,x_shift,0, text[ch]);
-      // note that this stomps the old data, so where character boxes overlap (e.g. 'lj') it's wrong
-      // because this API is really for baking character bitmaps into textures. if you want to render
-      // a sequence of characters, you really need to render each bitmap to a temp buffer, then
-      // "alpha blend" that into the working buffer
-      xpos += (advance * scale);
-      if (text[ch+1])
-         xpos += scale*stbtt_GetCodepointKernAdvance(&font, text[ch],text[ch+1]);
-      ++ch;
-   }
-
-   for (j=0; j < 20; ++j) {
-      for (i=0; i < 78; ++i)
-         putchar(" .:ioVM@"[screen[j][i]>>5]);
-      putchar('\n');
-   }
-
-   return 0;
-}
-#endif
-
-
-//////////////////////////////////////////////////////////////////////////////
-//////////////////////////////////////////////////////////////////////////////
-////
-////   INTEGRATION WITH YOUR CODEBASE
-////
-////   The following sections allow you to supply alternate definitions
-////   of C library functions used by stb_truetype, e.g. if you don't
-////   link with the C runtime library.
-
-#ifdef STB_TRUETYPE_IMPLEMENTATION
-   // #define your own (u)stbtt_int8/16/32 before including to override this
-   #ifndef stbtt_uint8
-   typedef unsigned char   stbtt_uint8;
-   typedef signed   char   stbtt_int8;
-   typedef unsigned short  stbtt_uint16;
-   typedef signed   short  stbtt_int16;
-   typedef unsigned int    stbtt_uint32;
-   typedef signed   int    stbtt_int32;
-   #endif
-
-   typedef char stbtt__check_size32[sizeof(stbtt_int32)==4 ? 1 : -1];
-   typedef char stbtt__check_size16[sizeof(stbtt_int16)==2 ? 1 : -1];
-
-   // e.g. #define your own STBTT_ifloor/STBTT_iceil() to avoid math.h
-   #ifndef STBTT_ifloor
-   #include <math.h>
-   #define STBTT_ifloor(x)   ((int) floor(x))
-   #define STBTT_iceil(x)    ((int) ceil(x))
-   #endif
-
-   #ifndef STBTT_sqrt
-   #include <math.h>
-   #define STBTT_sqrt(x)      sqrt(x)
-   #define STBTT_pow(x,y)     pow(x,y)
-   #endif
-
-   #ifndef STBTT_fmod
-   #include <math.h>
-   #define STBTT_fmod(x,y)    fmod(x,y)
-   #endif
-
-   #ifndef STBTT_cos
-   #include <math.h>
-   #define STBTT_cos(x)       cos(x)
-   #define STBTT_acos(x)      acos(x)
-   #endif
-
-   #ifndef STBTT_fabs
-   #include <math.h>
-   #define STBTT_fabs(x)      fabs(x)
-   #endif
-
-   // #define your own functions "STBTT_malloc" / "STBTT_free" to avoid malloc.h
-   #ifndef STBTT_malloc
-   #include <stdlib.h>
-   #define STBTT_malloc(x,u)  ((void)(u),malloc(x))
-   #define STBTT_free(x,u)    ((void)(u),free(x))
-   #endif
-
-   #ifndef STBTT_assert
-   #include <assert.h>
-   #define STBTT_assert(x)    assert(x)
-   #endif
-
-   #ifndef STBTT_strlen
-   #include <string.h>
-   #define STBTT_strlen(x)    strlen(x)
-   #endif
-
-   #ifndef STBTT_memcpy
-   #include <string.h>
-   #define STBTT_memcpy       memcpy
-   #define STBTT_memset       memset
-   #endif
-#endif
-
-///////////////////////////////////////////////////////////////////////////////
-///////////////////////////////////////////////////////////////////////////////
-////
-////   INTERFACE
-////
-////
-
-#ifndef __STB_INCLUDE_STB_TRUETYPE_H__
-#define __STB_INCLUDE_STB_TRUETYPE_H__
-
-#ifdef STBTT_STATIC
-#define STBTT_DEF static
-#else
-#define STBTT_DEF extern
-#endif
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-
-// private structure
-typedef struct
-{
-   unsigned char *data;
-   int cursor;
-   int size;
-} stbtt__buf;
-
-//////////////////////////////////////////////////////////////////////////////
-//
-// TEXTURE BAKING API
-//
-// If you use this API, you only have to call two functions ever.
-//
-
-typedef struct
-{
-   unsigned short x0,y0,x1,y1; // coordinates of bbox in bitmap
-   float xoff,yoff,xadvance;
-} stbtt_bakedchar;
-
-STBTT_DEF int stbtt_BakeFontBitmap(const unsigned char *data, int offset,  // font location (use offset=0 for plain .ttf)
-                                float pixel_height,                     // height of font in pixels
-                                unsigned char *pixels, int pw, int ph,  // bitmap to be filled in
-                                int first_char, int num_chars,          // characters to bake
-                                stbtt_bakedchar *chardata);             // you allocate this, it's num_chars long
-// if return is positive, the first unused row of the bitmap
-// if return is negative, returns the negative of the number of characters that fit
-// if return is 0, no characters fit and no rows were used
-// This uses a very crappy packing.
-
-typedef struct
-{
-   float x0,y0,s0,t0; // top-left
-   float x1,y1,s1,t1; // bottom-right
-} stbtt_aligned_quad;
-
-STBTT_DEF void stbtt_GetBakedQuad(const stbtt_bakedchar *chardata, int pw, int ph,  // same data as above
-                               int char_index,             // character to display
-                               float *xpos, float *ypos,   // pointers to current position in screen pixel space
-                               stbtt_aligned_quad *q,      // output: quad to draw
-                               int opengl_fillrule);       // true if opengl fill rule; false if DX9 or earlier
-// Call GetBakedQuad with char_index = 'character - first_char', and it
-// creates the quad you need to draw and advances the current position.
-//
-// The coordinate system used assumes y increases downwards.
-//
-// Characters will extend both above and below the current position;
-// see discussion of "BASELINE" above.
-//
-// It's inefficient; you might want to c&p it and optimize it.
-
-STBTT_DEF void stbtt_GetScaledFontVMetrics(const unsigned char *fontdata, int index, float size, float *ascent, float *descent, float *lineGap);
-// Query the font vertical metrics without having to create a font first.
-
-
-//////////////////////////////////////////////////////////////////////////////
-//
-// NEW TEXTURE BAKING API
-//
-// This provides options for packing multiple fonts into one atlas, not
-// perfectly but better than nothing.
-
-typedef struct
-{
-   unsigned short x0,y0,x1,y1; // coordinates of bbox in bitmap
-   float xoff,yoff,xadvance;
-   float xoff2,yoff2;
-} stbtt_packedchar;
-
-typedef struct stbtt_pack_context stbtt_pack_context;
-typedef struct stbtt_fontinfo stbtt_fontinfo;
-#ifndef STB_RECT_PACK_VERSION
-typedef struct stbrp_rect stbrp_rect;
-#endif
-
-STBTT_DEF int  stbtt_PackBegin(stbtt_pack_context *spc, unsigned char *pixels, int width, int height, int stride_in_bytes, int padding, void *alloc_context);
-// Initializes a packing context stored in the passed-in stbtt_pack_context.
-// Future calls using this context will pack characters into the bitmap passed
-// in here: a 1-channel bitmap that is width * height. stride_in_bytes is
-// the distance from one row to the next (or 0 to mean they are packed tightly
-// together). "padding" is the amount of padding to leave between each
-// character (normally you want '1' for bitmaps you'll use as textures with
-// bilinear filtering).
-//
-// Returns 0 on failure, 1 on success.
-
-STBTT_DEF void stbtt_PackEnd  (stbtt_pack_context *spc);
-// Cleans up the packing context and frees all memory.
-
-#define STBTT_POINT_SIZE(x)   (-(x))
-
-STBTT_DEF int  stbtt_PackFontRange(stbtt_pack_context *spc, const unsigned char *fontdata, int font_index, float font_size,
-                                int first_unicode_char_in_range, int num_chars_in_range, stbtt_packedchar *chardata_for_range);
-// Creates character bitmaps from the font_index'th font found in fontdata (use
-// font_index=0 if you don't know what that is). It creates num_chars_in_range
-// bitmaps for characters with unicode values starting at first_unicode_char_in_range
-// and increasing. Data for how to render them is stored in chardata_for_range;
-// pass these to stbtt_GetPackedQuad to get back renderable quads.
-//
-// font_size is the full height of the character from ascender to descender,
-// as computed by stbtt_ScaleForPixelHeight. To use a point size as computed
-// by stbtt_ScaleForMappingEmToPixels, wrap the point size in STBTT_POINT_SIZE()
-// and pass that result as 'font_size':
-//       ...,                  20 , ... // font max minus min y is 20 pixels tall
-//       ..., STBTT_POINT_SIZE(20), ... // 'M' is 20 pixels tall
-
-typedef struct
-{
-   float font_size;
-   int first_unicode_codepoint_in_range;  // if non-zero, then the chars are continuous, and this is the first codepoint
-   int *array_of_unicode_codepoints;       // if non-zero, then this is an array of unicode codepoints
-   int num_chars;
-   stbtt_packedchar *chardata_for_range; // output
-   unsigned char h_oversample, v_oversample; // don't set these, they're used internally
-} stbtt_pack_range;
-
-STBTT_DEF int  stbtt_PackFontRanges(stbtt_pack_context *spc, const unsigned char *fontdata, int font_index, stbtt_pack_range *ranges, int num_ranges);
-// Creates character bitmaps from multiple ranges of characters stored in
-// ranges. This will usually create a better-packed bitmap than multiple
-// calls to stbtt_PackFontRange. Note that you can call this multiple
-// times within a single PackBegin/PackEnd.
-
-STBTT_DEF void stbtt_PackSetOversampling(stbtt_pack_context *spc, unsigned int h_oversample, unsigned int v_oversample);
-// Oversampling a font increases the quality by allowing higher-quality subpixel
-// positioning, and is especially valuable at smaller text sizes.
-//
-// This function sets the amount of oversampling for all following calls to
-// stbtt_PackFontRange(s) or stbtt_PackFontRangesGatherRects for a given
-// pack context. The default (no oversampling) is achieved by h_oversample=1
-// and v_oversample=1. The total number of pixels required is
-// h_oversample*v_oversample larger than the default; for example, 2x2
-// oversampling requires 4x the storage of 1x1. For best results, render
-// oversampled textures with bilinear filtering. Look at the readme in
-// stb/tests/oversample for information about oversampled fonts
-//
-// To use with PackFontRangesGather etc., you must set it before calls
-// call to PackFontRangesGatherRects.
-
-STBTT_DEF void stbtt_PackSetSkipMissingCodepoints(stbtt_pack_context *spc, int skip);
-// If skip != 0, this tells stb_truetype to skip any codepoints for which
-// there is no corresponding glyph. If skip=0, which is the default, then
-// codepoints without a glyph recived the font's "missing character" glyph,
-// typically an empty box by convention.
-
-STBTT_DEF void stbtt_GetPackedQuad(const stbtt_packedchar *chardata, int pw, int ph,  // same data as above
-                               int char_index,             // character to display
-                               float *xpos, float *ypos,   // pointers to current position in screen pixel space
-                               stbtt_aligned_quad *q,      // output: quad to draw
-                               int align_to_integer);
-
-STBTT_DEF int  stbtt_PackFontRangesGatherRects(stbtt_pack_context *spc, const stbtt_fontinfo *info, stbtt_pack_range *ranges, int num_ranges, stbrp_rect *rects);
-STBTT_DEF void stbtt_PackFontRangesPackRects(stbtt_pack_context *spc, stbrp_rect *rects, int num_rects);
-STBTT_DEF int  stbtt_PackFontRangesRenderIntoRects(stbtt_pack_context *spc, const stbtt_fontinfo *info, stbtt_pack_range *ranges, int num_ranges, stbrp_rect *rects);
-// Calling these functions in sequence is roughly equivalent to calling
-// stbtt_PackFontRanges(). If you more control over the packing of multiple
-// fonts, or if you want to pack custom data into a font texture, take a look
-// at the source to of stbtt_PackFontRanges() and create a custom version
-// using these functions, e.g. call GatherRects multiple times,
-// building up a single array of rects, then call PackRects once,
-// then call RenderIntoRects repeatedly. This may result in a
-// better packing than calling PackFontRanges multiple times
-// (or it may not).
-
-// this is an opaque structure that you shouldn't mess with which holds
-// all the context needed from PackBegin to PackEnd.
-struct stbtt_pack_context {
-   void *user_allocator_context;
-   void *pack_info;
-   int   width;
-   int   height;
-   int   stride_in_bytes;
-   int   padding;
-   int   skip_missing;
-   unsigned int   h_oversample, v_oversample;
-   unsigned char *pixels;
-   void  *nodes;
-};
-
-//////////////////////////////////////////////////////////////////////////////
-//
-// FONT LOADING
-//
-//
-
-STBTT_DEF int stbtt_GetNumberOfFonts(const unsigned char *data);
-// This function will determine the number of fonts in a font file.  TrueType
-// collection (.ttc) files may contain multiple fonts, while TrueType font
-// (.ttf) files only contain one font. The number of fonts can be used for
-// indexing with the previous function where the index is between zero and one
-// less than the total fonts. If an error occurs, -1 is returned.
-
-STBTT_DEF int stbtt_GetFontOffsetForIndex(const unsigned char *data, int index);
-// Each .ttf/.ttc file may have more than one font. Each font has a sequential
-// index number starting from 0. Call this function to get the font offset for
-// a given index; it returns -1 if the index is out of range. A regular .ttf
-// file will only define one font and it always be at offset 0, so it will
-// return '0' for index 0, and -1 for all other indices.
-
-// The following structure is defined publicly so you can declare one on
-// the stack or as a global or etc, but you should treat it as opaque.
-struct stbtt_fontinfo
-{
-   void           * userdata;
-   unsigned char  * data;              // pointer to .ttf file
-   int              fontstart;         // offset of start of font
-
-   int numGlyphs;                     // number of glyphs, needed for range checking
-
-   int loca,head,glyf,hhea,hmtx,kern,gpos,svg; // table locations as offset from start of .ttf
-   int index_map;                     // a cmap mapping for our chosen character encoding
-   int indexToLocFormat;              // format needed to map from glyph index to glyph
-
-   stbtt__buf cff;                    // cff font data
-   stbtt__buf charstrings;            // the charstring index
-   stbtt__buf gsubrs;                 // global charstring subroutines index
-   stbtt__buf subrs;                  // private charstring subroutines index
-   stbtt__buf fontdicts;              // array of font dicts
-   stbtt__buf fdselect;               // map from glyph to fontdict
-};
-
-STBTT_DEF int stbtt_InitFont(stbtt_fontinfo *info, const unsigned char *data, int offset);
-// Given an offset into the file that defines a font, this function builds
-// the necessary cached info for the rest of the system. You must allocate
-// the stbtt_fontinfo yourself, and stbtt_InitFont will fill it out. You don't
-// need to do anything special to free it, because the contents are pure
-// value data with no additional data structures. Returns 0 on failure.
-
-
-//////////////////////////////////////////////////////////////////////////////
-//
-// CHARACTER TO GLYPH-INDEX CONVERSIOn
-
-STBTT_DEF int stbtt_FindGlyphIndex(const stbtt_fontinfo *info, int unicode_codepoint);
-// If you're going to perform multiple operations on the same character
-// and you want a speed-up, call this function with the character you're
-// going to process, then use glyph-based functions instead of the
-// codepoint-based functions.
-// Returns 0 if the character codepoint is not defined in the font.
-
-
-//////////////////////////////////////////////////////////////////////////////
-//
-// CHARACTER PROPERTIES
-//
-
-STBTT_DEF float stbtt_ScaleForPixelHeight(const stbtt_fontinfo *info, float pixels);
-// computes a scale factor to produce a font whose "height" is 'pixels' tall.
-// Height is measured as the distance from the highest ascender to the lowest
-// descender; in other words, it's equivalent to calling stbtt_GetFontVMetrics
-// and computing:
-//       scale = pixels / (ascent - descent)
-// so if you prefer to measure height by the ascent only, use a similar calculation.
-
-STBTT_DEF float stbtt_ScaleForMappingEmToPixels(const stbtt_fontinfo *info, float pixels);
-// computes a scale factor to produce a font whose EM size is mapped to
-// 'pixels' tall. This is probably what traditional APIs compute, but
-// I'm not positive.
-
-STBTT_DEF void stbtt_GetFontVMetrics(const stbtt_fontinfo *info, int *ascent, int *descent, int *lineGap);
-// ascent is the coordinate above the baseline the font extends; descent
-// is the coordinate below the baseline the font extends (i.e. it is typically negative)
-// lineGap is the spacing between one row's descent and the next row's ascent...
-// so you should advance the vertical position by "*ascent - *descent + *lineGap"
-//   these are expressed in unscaled coordinates, so you must multiply by
-//   the scale factor for a given size
-
-STBTT_DEF int  stbtt_GetFontVMetricsOS2(const stbtt_fontinfo *info, int *typoAscent, int *typoDescent, int *typoLineGap);
-// analogous to GetFontVMetrics, but returns the "typographic" values from the OS/2
-// table (specific to MS/Windows TTF files).
-//
-// Returns 1 on success (table present), 0 on failure.
-
-STBTT_DEF void stbtt_GetFontBoundingBox(const stbtt_fontinfo *info, int *x0, int *y0, int *x1, int *y1);
-// the bounding box around all possible characters
-
-STBTT_DEF void stbtt_GetCodepointHMetrics(const stbtt_fontinfo *info, int codepoint, int *advanceWidth, int *leftSideBearing);
-// leftSideBearing is the offset from the current horizontal position to the left edge of the character
-// advanceWidth is the offset from the current horizontal position to the next horizontal position
-//   these are expressed in unscaled coordinates
-
-STBTT_DEF int  stbtt_GetCodepointKernAdvance(const stbtt_fontinfo *info, int ch1, int ch2);
-// an additional amount to add to the 'advance' value between ch1 and ch2
-
-STBTT_DEF int stbtt_GetCodepointBox(const stbtt_fontinfo *info, int codepoint, int *x0, int *y0, int *x1, int *y1);
-// Gets the bounding box of the visible part of the glyph, in unscaled coordinates
-
-STBTT_DEF void stbtt_GetGlyphHMetrics(const stbtt_fontinfo *info, int glyph_index, int *advanceWidth, int *leftSideBearing);
-STBTT_DEF int  stbtt_GetGlyphKernAdvance(const stbtt_fontinfo *info, int glyph1, int glyph2);
-STBTT_DEF int  stbtt_GetGlyphBox(const stbtt_fontinfo *info, int glyph_index, int *x0, int *y0, int *x1, int *y1);
-// as above, but takes one or more glyph indices for greater efficiency
-
-typedef struct stbtt_kerningentry
-{
-   int glyph1; // use stbtt_FindGlyphIndex
-   int glyph2;
-   int advance;
-} stbtt_kerningentry;
-
-STBTT_DEF int  stbtt_GetKerningTableLength(const stbtt_fontinfo *info);
-STBTT_DEF int  stbtt_GetKerningTable(const stbtt_fontinfo *info, stbtt_kerningentry* table, int table_length);
-// Retrieves a complete list of all of the kerning pairs provided by the font
-// stbtt_GetKerningTable never writes more than table_length entries and returns how many entries it did write.
-// The table will be sorted by (a.glyph1 == b.glyph1)?(a.glyph2 < b.glyph2):(a.glyph1 < b.glyph1)
-
-//////////////////////////////////////////////////////////////////////////////
-//
-// GLYPH SHAPES (you probably don't need these, but they have to go before
-// the bitmaps for C declaration-order reasons)
-//
-
-#ifndef STBTT_vmove // you can predefine these to use different values (but why?)
-   enum {
-      STBTT_vmove=1,
-      STBTT_vline,
-      STBTT_vcurve,
-      STBTT_vcubic
-   };
-#endif
-
-#ifndef stbtt_vertex // you can predefine this to use different values
-                   // (we share this with other code at RAD)
-   #define stbtt_vertex_type short // can't use stbtt_int16 because that's not visible in the header file
-   typedef struct
-   {
-      stbtt_vertex_type x,y,cx,cy,cx1,cy1;
-      unsigned char type,padding;
-   } stbtt_vertex;
-#endif
-
-STBTT_DEF int stbtt_IsGlyphEmpty(const stbtt_fontinfo *info, int glyph_index);
-// returns non-zero if nothing is drawn for this glyph
-
-STBTT_DEF int stbtt_GetCodepointShape(const stbtt_fontinfo *info, int unicode_codepoint, stbtt_vertex **vertices);
-STBTT_DEF int stbtt_GetGlyphShape(const stbtt_fontinfo *info, int glyph_index, stbtt_vertex **vertices);
-// returns # of vertices and fills *vertices with the pointer to them
-//   these are expressed in "unscaled" coordinates
-//
-// The shape is a series of contours. Each one starts with
-// a STBTT_moveto, then consists of a series of mixed
-// STBTT_lineto and STBTT_curveto segments. A lineto
-// draws a line from previous endpoint to its x,y; a curveto
-// draws a quadratic bezier from previous endpoint to
-// its x,y, using cx,cy as the bezier control point.
-
-STBTT_DEF void stbtt_FreeShape(const stbtt_fontinfo *info, stbtt_vertex *vertices);
-// frees the data allocated above
-
-STBTT_DEF unsigned char *stbtt_FindSVGDoc(const stbtt_fontinfo *info, int gl);
-STBTT_DEF int stbtt_GetCodepointSVG(const stbtt_fontinfo *info, int unicode_codepoint, const char **svg);
-STBTT_DEF int stbtt_GetGlyphSVG(const stbtt_fontinfo *info, int gl, const char **svg);
-// fills svg with the character's SVG data.
-// returns data size or 0 if SVG not found.
-
-//////////////////////////////////////////////////////////////////////////////
-//
-// BITMAP RENDERING
-//
-
-STBTT_DEF void stbtt_FreeBitmap(unsigned char *bitmap, void *userdata);
-// frees the bitmap allocated below
-
-STBTT_DEF unsigned char *stbtt_GetCodepointBitmap(const stbtt_fontinfo *info, float scale_x, float scale_y, int codepoint, int *width, int *height, int *xoff, int *yoff);
-// allocates a large-enough single-channel 8bpp bitmap and renders the
-// specified character/glyph at the specified scale into it, with
-// antialiasing. 0 is no coverage (transparent), 255 is fully covered (opaque).
-// *width & *height are filled out with the width & height of the bitmap,
-// which is stored left-to-right, top-to-bottom.
-//
-// xoff/yoff are the offset it pixel space from the glyph origin to the top-left of the bitmap
-
-STBTT_DEF unsigned char *stbtt_GetCodepointBitmapSubpixel(const stbtt_fontinfo *info, float scale_x, float scale_y, float shift_x, float shift_y, int codepoint, int *width, int *height, int *xoff, int *yoff);
-// the same as stbtt_GetCodepoitnBitmap, but you can specify a subpixel
-// shift for the character
-
-STBTT_DEF void stbtt_MakeCodepointBitmap(const stbtt_fontinfo *info, unsigned char *output, int out_w, int out_h, int out_stride, float scale_x, float scale_y, int codepoint);
-// the same as stbtt_GetCodepointBitmap, but you pass in storage for the bitmap
-// in the form of 'output', with row spacing of 'out_stride' bytes. the bitmap
-// is clipped to out_w/out_h bytes. Call stbtt_GetCodepointBitmapBox to get the
-// width and height and positioning info for it first.
-
-STBTT_DEF void stbtt_MakeCodepointBitmapSubpixel(const stbtt_fontinfo *info, unsigned char *output, int out_w, int out_h, int out_stride, float scale_x, float scale_y, float shift_x, float shift_y, int codepoint);
-// same as stbtt_MakeCodepointBitmap, but you can specify a subpixel
-// shift for the character
-
-STBTT_DEF void stbtt_MakeCodepointBitmapSubpixelPrefilter(const stbtt_fontinfo *info, unsigned char *output, int out_w, int out_h, int out_stride, float scale_x, float scale_y, float shift_x, float shift_y, int oversample_x, int oversample_y, float *sub_x, float *sub_y, int codepoint);
-// same as stbtt_MakeCodepointBitmapSubpixel, but prefiltering
-// is performed (see stbtt_PackSetOversampling)
-
-STBTT_DEF void stbtt_GetCodepointBitmapBox(const stbtt_fontinfo *font, int codepoint, float scale_x, float scale_y, int *ix0, int *iy0, int *ix1, int *iy1);
-// get the bbox of the bitmap centered around the glyph origin; so the
-// bitmap width is ix1-ix0, height is iy1-iy0, and location to place
-// the bitmap top left is (leftSideBearing*scale,iy0).
-// (Note that the bitmap uses y-increases-down, but the shape uses
-// y-increases-up, so CodepointBitmapBox and CodepointBox are inverted.)
-
-STBTT_DEF void stbtt_GetCodepointBitmapBoxSubpixel(const stbtt_fontinfo *font, int codepoint, float scale_x, float scale_y, float shift_x, float shift_y, int *ix0, int *iy0, int *ix1, int *iy1);
-// same as stbtt_GetCodepointBitmapBox, but you can specify a subpixel
-// shift for the character
-
-// the following functions are equivalent to the above functions, but operate
-// on glyph indices instead of Unicode codepoints (for efficiency)
-STBTT_DEF unsigned char *stbtt_GetGlyphBitmap(const stbtt_fontinfo *info, float scale_x, float scale_y, int glyph, int *width, int *height, int *xoff, int *yoff);
-STBTT_DEF unsigned char *stbtt_GetGlyphBitmapSubpixel(const stbtt_fontinfo *info, float scale_x, float scale_y, float shift_x, float shift_y, int glyph, int *width, int *height, int *xoff, int *yoff);
-STBTT_DEF void stbtt_MakeGlyphBitmap(const stbtt_fontinfo *info, unsigned char *output, int out_w, int out_h, int out_stride, float scale_x, float scale_y, int glyph);
-STBTT_DEF void stbtt_MakeGlyphBitmapSubpixel(const stbtt_fontinfo *info, unsigned char *output, int out_w, int out_h, int out_stride, float scale_x, float scale_y, float shift_x, float shift_y, int glyph);
-STBTT_DEF void stbtt_MakeGlyphBitmapSubpixelPrefilter(const stbtt_fontinfo *info, unsigned char *output, int out_w, int out_h, int out_stride, float scale_x, float scale_y, float shift_x, float shift_y, int oversample_x, int oversample_y, float *sub_x, float *sub_y, int glyph);
-STBTT_DEF void stbtt_GetGlyphBitmapBox(const stbtt_fontinfo *font, int glyph, float scale_x, float scale_y, int *ix0, int *iy0, int *ix1, int *iy1);
-STBTT_DEF void stbtt_GetGlyphBitmapBoxSubpixel(const stbtt_fontinfo *font, int glyph, float scale_x, float scale_y,float shift_x, float shift_y, int *ix0, int *iy0, int *ix1, int *iy1);
-
-
-// @TODO: don't expose this structure
-typedef struct
-{
-   int w,h,stride;
-   unsigned char *pixels;
-} stbtt__bitmap;
-
-// rasterize a shape with quadratic beziers into a bitmap
-STBTT_DEF void stbtt_Rasterize(stbtt__bitmap *result,        // 1-channel bitmap to draw into
-                               float flatness_in_pixels,     // allowable error of curve in pixels
-                               stbtt_vertex *vertices,       // array of vertices defining shape
-                               int num_verts,                // number of vertices in above array
-                               float scale_x, float scale_y, // scale applied to input vertices
-                               float shift_x, float shift_y, // translation applied to input vertices
-                               int x_off, int y_off,         // another translation applied to input
-                               int invert,                   // if non-zero, vertically flip shape
-                               void *userdata);              // context for to STBTT_MALLOC
-
-//////////////////////////////////////////////////////////////////////////////
-//
-// Signed Distance Function (or Field) rendering
-
-STBTT_DEF void stbtt_FreeSDF(unsigned char *bitmap, void *userdata);
-// frees the SDF bitmap allocated below
-
-STBTT_DEF unsigned char * stbtt_GetGlyphSDF(const stbtt_fontinfo *info, float scale, int glyph, int padding, unsigned char onedge_value, float pixel_dist_scale, int *width, int *height, int *xoff, int *yoff);
-STBTT_DEF unsigned char * stbtt_GetCodepointSDF(const stbtt_fontinfo *info, float scale, int codepoint, int padding, unsigned char onedge_value, float pixel_dist_scale, int *width, int *height, int *xoff, int *yoff);
-// These functions compute a discretized SDF field for a single character, suitable for storing
-// in a single-channel texture, sampling with bilinear filtering, and testing against
-// larger than some threshold to produce scalable fonts.
-//        info              --  the font
-//        scale             --  controls the size of the resulting SDF bitmap, same as it would be creating a regular bitmap
-//        glyph/codepoint   --  the character to generate the SDF for
-//        padding           --  extra "pixels" around the character which are filled with the distance to the character (not 0),
-//                                 which allows effects like bit outlines
-//        onedge_value      --  value 0-255 to test the SDF against to reconstruct the character (i.e. the isocontour of the character)
-//        pixel_dist_scale  --  what value the SDF should increase by when moving one SDF "pixel" away from the edge (on the 0..255 scale)
-//                                 if positive, > onedge_value is inside; if negative, < onedge_value is inside
-//        width,height      --  output height & width of the SDF bitmap (including padding)
-//        xoff,yoff         --  output origin of the character
-//        return value      --  a 2D array of bytes 0..255, width*height in size
-//
-// pixel_dist_scale & onedge_value are a scale & bias that allows you to make
-// optimal use of the limited 0..255 for your application, trading off precision
-// and special effects. SDF values outside the range 0..255 are clamped to 0..255.
-//
-// Example:
-//      scale = stbtt_ScaleForPixelHeight(22)
-//      padding = 5
-//      onedge_value = 180
-//      pixel_dist_scale = 180/5.0 = 36.0
-//
-//      This will create an SDF bitmap in which the character is about 22 pixels
-//      high but the whole bitmap is about 22+5+5=32 pixels high. To produce a filled
-//      shape, sample the SDF at each pixel and fill the pixel if the SDF value
-//      is greater than or equal to 180/255. (You'll actually want to antialias,
-//      which is beyond the scope of this example.) Additionally, you can compute
-//      offset outlines (e.g. to stroke the character border inside & outside,
-//      or only outside). For example, to fill outside the character up to 3 SDF
-//      pixels, you would compare against (180-36.0*3)/255 = 72/255. The above
-//      choice of variables maps a range from 5 pixels outside the shape to
-//      2 pixels inside the shape to 0..255; this is intended primarily for apply
-//      outside effects only (the interior range is needed to allow proper
-//      antialiasing of the font at *smaller* sizes)
-//
-// The function computes the SDF analytically at each SDF pixel, not by e.g.
-// building a higher-res bitmap and approximating it. In theory the quality
-// should be as high as possible for an SDF of this size & representation, but
-// unclear if this is true in practice (perhaps building a higher-res bitmap
-// and computing from that can allow drop-out prevention).
-//
-// The algorithm has not been optimized at all, so expect it to be slow
-// if computing lots of characters or very large sizes.
-
-
-
-//////////////////////////////////////////////////////////////////////////////
-//
-// Finding the right font...
-//
-// You should really just solve this offline, keep your own tables
-// of what font is what, and don't try to get it out of the .ttf file.
-// That's because getting it out of the .ttf file is really hard, because
-// the names in the file can appear in many possible encodings, in many
-// possible languages, and e.g. if you need a case-insensitive comparison,
-// the details of that depend on the encoding & language in a complex way
-// (actually underspecified in truetype, but also gigantic).
-//
-// But you can use the provided functions in two possible ways:
-//     stbtt_FindMatchingFont() will use *case-sensitive* comparisons on
-//             unicode-encoded names to try to find the font you want;
-//             you can run this before calling stbtt_InitFont()
-//
-//     stbtt_GetFontNameString() lets you get any of the various strings
-//             from the file yourself and do your own comparisons on them.
-//             You have to have called stbtt_InitFont() first.
-
-
-STBTT_DEF int stbtt_FindMatchingFont(const unsigned char *fontdata, const char *name, int flags);
-// returns the offset (not index) of the font that matches, or -1 if none
-//   if you use STBTT_MACSTYLE_DONTCARE, use a font name like "Arial Bold".
-//   if you use any other flag, use a font name like "Arial"; this checks
-//     the 'macStyle' header field; i don't know if fonts set this consistently
-#define STBTT_MACSTYLE_DONTCARE     0
-#define STBTT_MACSTYLE_BOLD         1
-#define STBTT_MACSTYLE_ITALIC       2
-#define STBTT_MACSTYLE_UNDERSCORE   4
-#define STBTT_MACSTYLE_NONE         8   // <= not same as 0, this makes us check the bitfield is 0
-
-STBTT_DEF int stbtt_CompareUTF8toUTF16_bigendian(const char *s1, int len1, const char *s2, int len2);
-// returns 1/0 whether the first string interpreted as utf8 is identical to
-// the second string interpreted as big-endian utf16... useful for strings from next func
-
-STBTT_DEF const char *stbtt_GetFontNameString(const stbtt_fontinfo *font, int *length, int platformID, int encodingID, int languageID, int nameID);
-// returns the string (which may be big-endian double byte, e.g. for unicode)
-// and puts the length in bytes in *length.
-//
-// some of the values for the IDs are below; for more see the truetype spec:
-//     http://developer.apple.com/textfonts/TTRefMan/RM06/Chap6name.html
-//     http://www.microsoft.com/typography/otspec/name.htm
-
-enum { // platformID
-   STBTT_PLATFORM_ID_UNICODE   =0,
-   STBTT_PLATFORM_ID_MAC       =1,
-   STBTT_PLATFORM_ID_ISO       =2,
-   STBTT_PLATFORM_ID_MICROSOFT =3
-};
-
-enum { // encodingID for STBTT_PLATFORM_ID_UNICODE
-   STBTT_UNICODE_EID_UNICODE_1_0    =0,
-   STBTT_UNICODE_EID_UNICODE_1_1    =1,
-   STBTT_UNICODE_EID_ISO_10646      =2,
-   STBTT_UNICODE_EID_UNICODE_2_0_BMP=3,
-   STBTT_UNICODE_EID_UNICODE_2_0_FULL=4
-};
-
-enum { // encodingID for STBTT_PLATFORM_ID_MICROSOFT
-   STBTT_MS_EID_SYMBOL        =0,
-   STBTT_MS_EID_UNICODE_BMP   =1,
-   STBTT_MS_EID_SHIFTJIS      =2,
-   STBTT_MS_EID_UNICODE_FULL  =10
-};
-
-enum { // encodingID for STBTT_PLATFORM_ID_MAC; same as Script Manager codes
-   STBTT_MAC_EID_ROMAN        =0,   STBTT_MAC_EID_ARABIC       =4,
-   STBTT_MAC_EID_JAPANESE     =1,   STBTT_MAC_EID_HEBREW       =5,
-   STBTT_MAC_EID_CHINESE_TRAD =2,   STBTT_MAC_EID_GREEK        =6,
-   STBTT_MAC_EID_KOREAN       =3,   STBTT_MAC_EID_RUSSIAN      =7
-};
-
-enum { // languageID for STBTT_PLATFORM_ID_MICROSOFT; same as LCID...
-       // problematic because there are e.g. 16 english LCIDs and 16 arabic LCIDs
-   STBTT_MS_LANG_ENGLISH     =0x0409,   STBTT_MS_LANG_ITALIAN     =0x0410,
-   STBTT_MS_LANG_CHINESE     =0x0804,   STBTT_MS_LANG_JAPANESE    =0x0411,
-   STBTT_MS_LANG_DUTCH       =0x0413,   STBTT_MS_LANG_KOREAN      =0x0412,
-   STBTT_MS_LANG_FRENCH      =0x040c,   STBTT_MS_LANG_RUSSIAN     =0x0419,
-   STBTT_MS_LANG_GERMAN      =0x0407,   STBTT_MS_LANG_SPANISH     =0x0409,
-   STBTT_MS_LANG_HEBREW      =0x040d,   STBTT_MS_LANG_SWEDISH     =0x041D
-};
-
-enum { // languageID for STBTT_PLATFORM_ID_MAC
-   STBTT_MAC_LANG_ENGLISH      =0 ,   STBTT_MAC_LANG_JAPANESE     =11,
-   STBTT_MAC_LANG_ARABIC       =12,   STBTT_MAC_LANG_KOREAN       =23,
-   STBTT_MAC_LANG_DUTCH        =4 ,   STBTT_MAC_LANG_RUSSIAN      =32,
-   STBTT_MAC_LANG_FRENCH       =1 ,   STBTT_MAC_LANG_SPANISH      =6 ,
-   STBTT_MAC_LANG_GERMAN       =2 ,   STBTT_MAC_LANG_SWEDISH      =5 ,
-   STBTT_MAC_LANG_HEBREW       =10,   STBTT_MAC_LANG_CHINESE_SIMPLIFIED =33,
-   STBTT_MAC_LANG_ITALIAN      =3 ,   STBTT_MAC_LANG_CHINESE_TRAD =19
-};
-
-#ifdef __cplusplus
-}
-#endif
-
-#endif // __STB_INCLUDE_STB_TRUETYPE_H__
-
-///////////////////////////////////////////////////////////////////////////////
-///////////////////////////////////////////////////////////////////////////////
-////
-////   IMPLEMENTATION
-////
-////
-
-#ifdef STB_TRUETYPE_IMPLEMENTATION
-
-#ifndef STBTT_MAX_OVERSAMPLE
-#define STBTT_MAX_OVERSAMPLE   8
-#endif
-
-#if STBTT_MAX_OVERSAMPLE > 255
-#error "STBTT_MAX_OVERSAMPLE cannot be > 255"
-#endif
-
-typedef int stbtt__test_oversample_pow2[(STBTT_MAX_OVERSAMPLE & (STBTT_MAX_OVERSAMPLE-1)) == 0 ? 1 : -1];
-
-#ifndef STBTT_RASTERIZER_VERSION
-#define STBTT_RASTERIZER_VERSION 2
-#endif
-
-#ifdef _MSC_VER
-#define STBTT__NOTUSED(v)  (void)(v)
-#else
-#define STBTT__NOTUSED(v)  (void)sizeof(v)
-#endif
-
-//////////////////////////////////////////////////////////////////////////
-//
-// stbtt__buf helpers to parse data from file
-//
-
-static stbtt_uint8 stbtt__buf_get8(stbtt__buf *b)
-{
-   if (b->cursor >= b->size)
-      return 0;
-   return b->data[b->cursor++];
-}
-
-static stbtt_uint8 stbtt__buf_peek8(stbtt__buf *b)
-{
-   if (b->cursor >= b->size)
-      return 0;
-   return b->data[b->cursor];
-}
-
-static void stbtt__buf_seek(stbtt__buf *b, int o)
-{
-   STBTT_assert(!(o > b->size || o < 0));
-   b->cursor = (o > b->size || o < 0) ? b->size : o;
-}
-
-static void stbtt__buf_skip(stbtt__buf *b, int o)
-{
-   stbtt__buf_seek(b, b->cursor + o);
-}
-
-static stbtt_uint32 stbtt__buf_get(stbtt__buf *b, int n)
-{
-   stbtt_uint32 v = 0;
-   int i;
-   STBTT_assert(n >= 1 && n <= 4);
-   for (i = 0; i < n; i++)
-      v = (v << 8) | stbtt__buf_get8(b);
-   return v;
-}
-
-static stbtt__buf stbtt__new_buf(const void *p, size_t size)
-{
-   stbtt__buf r;
-   STBTT_assert(size < 0x40000000);
-   r.data = (stbtt_uint8*) p;
-   r.size = (int) size;
-   r.cursor = 0;
-   return r;
-}
-
-#define stbtt__buf_get16(b)  stbtt__buf_get((b), 2)
-#define stbtt__buf_get32(b)  stbtt__buf_get((b), 4)
-
-static stbtt__buf stbtt__buf_range(const stbtt__buf *b, int o, int s)
-{
-   stbtt__buf r = stbtt__new_buf(NULL, 0);
-   if (o < 0 || s < 0 || o > b->size || s > b->size - o) return r;
-   r.data = b->data + o;
-   r.size = s;
-   return r;
-}
-
-static stbtt__buf stbtt__cff_get_index(stbtt__buf *b)
-{
-   int count, start, offsize;
-   start = b->cursor;
-   count = stbtt__buf_get16(b);
-   if (count) {
-      offsize = stbtt__buf_get8(b);
-      STBTT_assert(offsize >= 1 && offsize <= 4);
-      stbtt__buf_skip(b, offsize * count);
-      stbtt__buf_skip(b, stbtt__buf_get(b, offsize) - 1);
-   }
-   return stbtt__buf_range(b, start, b->cursor - start);
-}
-
-static stbtt_uint32 stbtt__cff_int(stbtt__buf *b)
-{
-   int b0 = stbtt__buf_get8(b);
-   if (b0 >= 32 && b0 <= 246)       return b0 - 139;
-   else if (b0 >= 247 && b0 <= 250) return (b0 - 247)*256 + stbtt__buf_get8(b) + 108;
-   else if (b0 >= 251 && b0 <= 254) return -(b0 - 251)*256 - stbtt__buf_get8(b) - 108;
-   else if (b0 == 28)               return stbtt__buf_get16(b);
-   else if (b0 == 29)               return stbtt__buf_get32(b);
-   STBTT_assert(0);
-   return 0;
-}
-
-static void stbtt__cff_skip_operand(stbtt__buf *b) {
-   int v, b0 = stbtt__buf_peek8(b);
-   STBTT_assert(b0 >= 28);
-   if (b0 == 30) {
-      stbtt__buf_skip(b, 1);
-      while (b->cursor < b->size) {
-         v = stbtt__buf_get8(b);
-         if ((v & 0xF) == 0xF || (v >> 4) == 0xF)
-            break;
-      }
-   } else {
-      stbtt__cff_int(b);
-   }
-}
-
-static stbtt__buf stbtt__dict_get(stbtt__buf *b, int key)
-{
-   stbtt__buf_seek(b, 0);
-   while (b->cursor < b->size) {
-      int start = b->cursor, end, op;
-      while (stbtt__buf_peek8(b) >= 28)
-         stbtt__cff_skip_operand(b);
-      end = b->cursor;
-      op = stbtt__buf_get8(b);
-      if (op == 12)  op = stbtt__buf_get8(b) | 0x100;
-      if (op == key) return stbtt__buf_range(b, start, end-start);
-   }
-   return stbtt__buf_range(b, 0, 0);
-}
-
-static void stbtt__dict_get_ints(stbtt__buf *b, int key, int outcount, stbtt_uint32 *out)
-{
-   int i;
-   stbtt__buf operands = stbtt__dict_get(b, key);
-   for (i = 0; i < outcount && operands.cursor < operands.size; i++)
-      out[i] = stbtt__cff_int(&operands);
-}
-
-static int stbtt__cff_index_count(stbtt__buf *b)
-{
-   stbtt__buf_seek(b, 0);
-   return stbtt__buf_get16(b);
-}
-
-static stbtt__buf stbtt__cff_index_get(stbtt__buf b, int i)
-{
-   int count, offsize, start, end;
-   stbtt__buf_seek(&b, 0);
-   count = stbtt__buf_get16(&b);
-   offsize = stbtt__buf_get8(&b);
-   STBTT_assert(i >= 0 && i < count);
-   STBTT_assert(offsize >= 1 && offsize <= 4);
-   stbtt__buf_skip(&b, i*offsize);
-   start = stbtt__buf_get(&b, offsize);
-   end = stbtt__buf_get(&b, offsize);
-   return stbtt__buf_range(&b, 2+(count+1)*offsize+start, end - start);
-}
-
-//////////////////////////////////////////////////////////////////////////
-//
-// accessors to parse data from file
-//
-
-// on platforms that don't allow misaligned reads, if we want to allow
-// truetype fonts that aren't padded to alignment, define ALLOW_UNALIGNED_TRUETYPE
-
-#define ttBYTE(p)     (* (stbtt_uint8 *) (p))
-#define ttCHAR(p)     (* (stbtt_int8 *) (p))
-#define ttFixed(p)    ttLONG(p)
-
-static stbtt_uint16 ttUSHORT(stbtt_uint8 *p) { return p[0]*256 + p[1]; }
-static stbtt_int16 ttSHORT(stbtt_uint8 *p)   { return p[0]*256 + p[1]; }
-static stbtt_uint32 ttULONG(stbtt_uint8 *p)  { return (p[0]<<24) + (p[1]<<16) + (p[2]<<8) + p[3]; }
-static stbtt_int32 ttLONG(stbtt_uint8 *p)    { return (p[0]<<24) + (p[1]<<16) + (p[2]<<8) + p[3]; }
-
-#define stbtt_tag4(p,c0,c1,c2,c3) ((p)[0] == (c0) && (p)[1] == (c1) && (p)[2] == (c2) && (p)[3] == (c3))
-#define stbtt_tag(p,str)           stbtt_tag4(p,str[0],str[1],str[2],str[3])
-
-static int stbtt__isfont(stbtt_uint8 *font)
-{
-   // check the version number
-   if (stbtt_tag4(font, '1',0,0,0))  return 1; // TrueType 1
-   if (stbtt_tag(font, "typ1"))   return 1; // TrueType with type 1 font -- we don't support this!
-   if (stbtt_tag(font, "OTTO"))   return 1; // OpenType with CFF
-   if (stbtt_tag4(font, 0,1,0,0)) return 1; // OpenType 1.0
-   if (stbtt_tag(font, "true"))   return 1; // Apple specification for TrueType fonts
-   return 0;
-}
-
-// @OPTIMIZE: binary search
-static stbtt_uint32 stbtt__find_table(stbtt_uint8 *data, stbtt_uint32 fontstart, const char *tag)
-{
-   stbtt_int32 num_tables = ttUSHORT(data+fontstart+4);
-   stbtt_uint32 tabledir = fontstart + 12;
-   stbtt_int32 i;
-   for (i=0; i < num_tables; ++i) {
-      stbtt_uint32 loc = tabledir + 16*i;
-      if (stbtt_tag(data+loc+0, tag))
-         return ttULONG(data+loc+8);
-   }
-   return 0;
-}
-
-static int stbtt_GetFontOffsetForIndex_internal(unsigned char *font_collection, int index)
-{
-   // if it's just a font, there's only one valid index
-   if (stbtt__isfont(font_collection))
-      return index == 0 ? 0 : -1;
-
-   // check if it's a TTC
-   if (stbtt_tag(font_collection, "ttcf")) {
-      // version 1?
-      if (ttULONG(font_collection+4) == 0x00010000 || ttULONG(font_collection+4) == 0x00020000) {
-         stbtt_int32 n = ttLONG(font_collection+8);
-         if (index >= n)
-            return -1;
-         return ttULONG(font_collection+12+index*4);
-      }
-   }
-   return -1;
-}
-
-static int stbtt_GetNumberOfFonts_internal(unsigned char *font_collection)
-{
-   // if it's just a font, there's only one valid font
-   if (stbtt__isfont(font_collection))
-      return 1;
-
-   // check if it's a TTC
-   if (stbtt_tag(font_collection, "ttcf")) {
-      // version 1?
-      if (ttULONG(font_collection+4) == 0x00010000 || ttULONG(font_collection+4) == 0x00020000) {
-         return ttLONG(font_collection+8);
-      }
-   }
-   return 0;
-}
-
-static stbtt__buf stbtt__get_subrs(stbtt__buf cff, stbtt__buf fontdict)
-{
-   stbtt_uint32 subrsoff = 0, private_loc[2] = { 0, 0 };
-   stbtt__buf pdict;
-   stbtt__dict_get_ints(&fontdict, 18, 2, private_loc);
-   if (!private_loc[1] || !private_loc[0]) return stbtt__new_buf(NULL, 0);
-   pdict = stbtt__buf_range(&cff, private_loc[1], private_loc[0]);
-   stbtt__dict_get_ints(&pdict, 19, 1, &subrsoff);
-   if (!subrsoff) return stbtt__new_buf(NULL, 0);
-   stbtt__buf_seek(&cff, private_loc[1]+subrsoff);
-   return stbtt__cff_get_index(&cff);
-}
-
-// since most people won't use this, find this table the first time it's needed
-static int stbtt__get_svg(stbtt_fontinfo *info)
-{
-   stbtt_uint32 t;
-   if (info->svg < 0) {
-      t = stbtt__find_table(info->data, info->fontstart, "SVG ");
-      if (t) {
-         stbtt_uint32 offset = ttULONG(info->data + t + 2);
-         info->svg = t + offset;
-      } else {
-         info->svg = 0;
-      }
-   }
-   return info->svg;
-}
-
-static int stbtt_InitFont_internal(stbtt_fontinfo *info, unsigned char *data, int fontstart)
-{
-   stbtt_uint32 cmap, t;
-   stbtt_int32 i,numTables;
-
-   info->data = data;
-   info->fontstart = fontstart;
-   info->cff = stbtt__new_buf(NULL, 0);
-
-   cmap = stbtt__find_table(data, fontstart, "cmap");       // required
-   info->loca = stbtt__find_table(data, fontstart, "loca"); // required
-   info->head = stbtt__find_table(data, fontstart, "head"); // required
-   info->glyf = stbtt__find_table(data, fontstart, "glyf"); // required
-   info->hhea = stbtt__find_table(data, fontstart, "hhea"); // required
-   info->hmtx = stbtt__find_table(data, fontstart, "hmtx"); // required
-   info->kern = stbtt__find_table(data, fontstart, "kern"); // not required
-   info->gpos = stbtt__find_table(data, fontstart, "GPOS"); // not required
-
-   if (!cmap || !info->head || !info->hhea || !info->hmtx)
-      return 0;
-   if (info->glyf) {
-      // required for truetype
-      if (!info->loca) return 0;
-   } else {
-      // initialization for CFF / Type2 fonts (OTF)
-      stbtt__buf b, topdict, topdictidx;
-      stbtt_uint32 cstype = 2, charstrings = 0, fdarrayoff = 0, fdselectoff = 0;
-      stbtt_uint32 cff;
-
-      cff = stbtt__find_table(data, fontstart, "CFF ");
-      if (!cff) return 0;
-
-      info->fontdicts = stbtt__new_buf(NULL, 0);
-      info->fdselect = stbtt__new_buf(NULL, 0);
-
-      // @TODO this should use size from table (not 512MB)
-      info->cff = stbtt__new_buf(data+cff, 512*1024*1024);
-      b = info->cff;
-
-      // read the header
-      stbtt__buf_skip(&b, 2);
-      stbtt__buf_seek(&b, stbtt__buf_get8(&b)); // hdrsize
-
-      // @TODO the name INDEX could list multiple fonts,
-      // but we just use the first one.
-      stbtt__cff_get_index(&b);  // name INDEX
-      topdictidx = stbtt__cff_get_index(&b);
-      topdict = stbtt__cff_index_get(topdictidx, 0);
-      stbtt__cff_get_index(&b);  // string INDEX
-      info->gsubrs = stbtt__cff_get_index(&b);
-
-      stbtt__dict_get_ints(&topdict, 17, 1, &charstrings);
-      stbtt__dict_get_ints(&topdict, 0x100 | 6, 1, &cstype);
-      stbtt__dict_get_ints(&topdict, 0x100 | 36, 1, &fdarrayoff);
-      stbtt__dict_get_ints(&topdict, 0x100 | 37, 1, &fdselectoff);
-      info->subrs = stbtt__get_subrs(b, topdict);
-
-      // we only support Type 2 charstrings
-      if (cstype != 2) return 0;
-      if (charstrings == 0) return 0;
-
-      if (fdarrayoff) {
-         // looks like a CID font
-         if (!fdselectoff) return 0;
-         stbtt__buf_seek(&b, fdarrayoff);
-         info->fontdicts = stbtt__cff_get_index(&b);
-         info->fdselect = stbtt__buf_range(&b, fdselectoff, b.size-fdselectoff);
-      }
-
-      stbtt__buf_seek(&b, charstrings);
-      info->charstrings = stbtt__cff_get_index(&b);
-   }
-
-   t = stbtt__find_table(data, fontstart, "maxp");
-   if (t)
-      info->numGlyphs = ttUSHORT(data+t+4);
-   else
-      info->numGlyphs = 0xffff;
-
-   info->svg = -1;
-
-   // find a cmap encoding table we understand *now* to avoid searching
-   // later. (todo: could make this installable)
-   // the same regardless of glyph.
-   numTables = ttUSHORT(data + cmap + 2);
-   info->index_map = 0;
-   for (i=0; i < numTables; ++i) {
-      stbtt_uint32 encoding_record = cmap + 4 + 8 * i;
-      // find an encoding we understand:
-      switch(ttUSHORT(data+encoding_record)) {
-         case STBTT_PLATFORM_ID_MICROSOFT:
-            switch (ttUSHORT(data+encoding_record+2)) {
-               case STBTT_MS_EID_UNICODE_BMP:
-               case STBTT_MS_EID_UNICODE_FULL:
-                  // MS/Unicode
-                  info->index_map = cmap + ttULONG(data+encoding_record+4);
-                  break;
-            }
-            break;
-        case STBTT_PLATFORM_ID_UNICODE:
-            // Mac/iOS has these
-            // all the encodingIDs are unicode, so we don't bother to check it
-            info->index_map = cmap + ttULONG(data+encoding_record+4);
-            break;
-      }
-   }
-   if (info->index_map == 0)
-      return 0;
-
-   info->indexToLocFormat = ttUSHORT(data+info->head + 50);
-   return 1;
-}
-
-STBTT_DEF int stbtt_FindGlyphIndex(const stbtt_fontinfo *info, int unicode_codepoint)
-{
-   stbtt_uint8 *data = info->data;
-   stbtt_uint32 index_map = info->index_map;
-
-   stbtt_uint16 format = ttUSHORT(data + index_map + 0);
-   if (format == 0) { // apple byte encoding
-      stbtt_int32 bytes = ttUSHORT(data + index_map + 2);
-      if (unicode_codepoint < bytes-6)
-         return ttBYTE(data + index_map + 6 + unicode_codepoint);
-      return 0;
-   } else if (format == 6) {
-      stbtt_uint32 first = ttUSHORT(data + index_map + 6);
-      stbtt_uint32 count = ttUSHORT(data + index_map + 8);
-      if ((stbtt_uint32) unicode_codepoint >= first && (stbtt_uint32) unicode_codepoint < first+count)
-         return ttUSHORT(data + index_map + 10 + (unicode_codepoint - first)*2);
-      return 0;
-   } else if (format == 2) {
-      STBTT_assert(0); // @TODO: high-byte mapping for japanese/chinese/korean
-      return 0;
-   } else if (format == 4) { // standard mapping for windows fonts: binary search collection of ranges
-      stbtt_uint16 segcount = ttUSHORT(data+index_map+6) >> 1;
-      stbtt_uint16 searchRange = ttUSHORT(data+index_map+8) >> 1;
-      stbtt_uint16 entrySelector = ttUSHORT(data+index_map+10);
-      stbtt_uint16 rangeShift = ttUSHORT(data+index_map+12) >> 1;
-
-      // do a binary search of the segments
-      stbtt_uint32 endCount = index_map + 14;
-      stbtt_uint32 search = endCount;
-
-      if (unicode_codepoint > 0xffff)
-         return 0;
-
-      // they lie from endCount .. endCount + segCount
-      // but searchRange is the nearest power of two, so...
-      if (unicode_codepoint >= ttUSHORT(data + search + rangeShift*2))
-         search += rangeShift*2;
-
-      // now decrement to bias correctly to find smallest
-      search -= 2;
-      while (entrySelector) {
-         stbtt_uint16 end;
-         searchRange >>= 1;
-         end = ttUSHORT(data + search + searchRange*2);
-         if (unicode_codepoint > end)
-            search += searchRange*2;
-         --entrySelector;
-      }
-      search += 2;
-
-      {
-         stbtt_uint16 offset, start, last;
-         stbtt_uint16 item = (stbtt_uint16) ((search - endCount) >> 1);
-
-         start = ttUSHORT(data + index_map + 14 + segcount*2 + 2 + 2*item);
-         last = ttUSHORT(data + endCount + 2*item);
-         if (unicode_codepoint < start || unicode_codepoint > last)
-            return 0;
-
-         offset = ttUSHORT(data + index_map + 14 + segcount*6 + 2 + 2*item);
-         if (offset == 0)
-            return (stbtt_uint16) (unicode_codepoint + ttSHORT(data + index_map + 14 + segcount*4 + 2 + 2*item));
-
-         return ttUSHORT(data + offset + (unicode_codepoint-start)*2 + index_map + 14 + segcount*6 + 2 + 2*item);
-      }
-   } else if (format == 12 || format == 13) {
-      stbtt_uint32 ngroups = ttULONG(data+index_map+12);
-      stbtt_int32 low,high;
-      low = 0; high = (stbtt_int32)ngroups;
-      // Binary search the right group.
-      while (low < high) {
-         stbtt_int32 mid = low + ((high-low) >> 1); // rounds down, so low <= mid < high
-         stbtt_uint32 start_char = ttULONG(data+index_map+16+mid*12);
-         stbtt_uint32 end_char = ttULONG(data+index_map+16+mid*12+4);
-         if ((stbtt_uint32) unicode_codepoint < start_char)
-            high = mid;
-         else if ((stbtt_uint32) unicode_codepoint > end_char)
-            low = mid+1;
-         else {
-            stbtt_uint32 start_glyph = ttULONG(data+index_map+16+mid*12+8);
-            if (format == 12)
-               return start_glyph + unicode_codepoint-start_char;
-            else // format == 13
-               return start_glyph;
-         }
-      }
-      return 0; // not found
-   }
-   // @TODO
-   STBTT_assert(0);
-   return 0;
-}
-
-STBTT_DEF int stbtt_GetCodepointShape(const stbtt_fontinfo *info, int unicode_codepoint, stbtt_vertex **vertices)
-{
-   return stbtt_GetGlyphShape(info, stbtt_FindGlyphIndex(info, unicode_codepoint), vertices);
-}
-
-static void stbtt_setvertex(stbtt_vertex *v, stbtt_uint8 type, stbtt_int32 x, stbtt_int32 y, stbtt_int32 cx, stbtt_int32 cy)
-{
-   v->type = type;
-   v->x = (stbtt_int16) x;
-   v->y = (stbtt_int16) y;
-   v->cx = (stbtt_int16) cx;
-   v->cy = (stbtt_int16) cy;
-}
-
-static int stbtt__GetGlyfOffset(const stbtt_fontinfo *info, int glyph_index)
-{
-   int g1,g2;
-
-   STBTT_assert(!info->cff.size);
-
-   if (glyph_index >= info->numGlyphs) return -1; // glyph index out of range
-   if (info->indexToLocFormat >= 2)    return -1; // unknown index->glyph map format
-
-   if (info->indexToLocFormat == 0) {
-      g1 = info->glyf + ttUSHORT(info->data + info->loca + glyph_index * 2) * 2;
-      g2 = info->glyf + ttUSHORT(info->data + info->loca + glyph_index * 2 + 2) * 2;
-   } else {
-      g1 = info->glyf + ttULONG (info->data + info->loca + glyph_index * 4);
-      g2 = info->glyf + ttULONG (info->data + info->loca + glyph_index * 4 + 4);
-   }
-
-   return g1==g2 ? -1 : g1; // if length is 0, return -1
-}
-
-static int stbtt__GetGlyphInfoT2(const stbtt_fontinfo *info, int glyph_index, int *x0, int *y0, int *x1, int *y1);
-
-STBTT_DEF int stbtt_GetGlyphBox(const stbtt_fontinfo *info, int glyph_index, int *x0, int *y0, int *x1, int *y1)
-{
-   if (info->cff.size) {
-      stbtt__GetGlyphInfoT2(info, glyph_index, x0, y0, x1, y1);
-   } else {
-      int g = stbtt__GetGlyfOffset(info, glyph_index);
-      if (g < 0) return 0;
-
-      if (x0) *x0 = ttSHORT(info->data + g + 2);
-      if (y0) *y0 = ttSHORT(info->data + g + 4);
-      if (x1) *x1 = ttSHORT(info->data + g + 6);
-      if (y1) *y1 = ttSHORT(info->data + g + 8);
-   }
-   return 1;
-}
-
-STBTT_DEF int stbtt_GetCodepointBox(const stbtt_fontinfo *info, int codepoint, int *x0, int *y0, int *x1, int *y1)
-{
-   return stbtt_GetGlyphBox(info, stbtt_FindGlyphIndex(info,codepoint), x0,y0,x1,y1);
-}
-
-STBTT_DEF int stbtt_IsGlyphEmpty(const stbtt_fontinfo *info, int glyph_index)
-{
-   stbtt_int16 numberOfContours;
-   int g;
-   if (info->cff.size)
-      return stbtt__GetGlyphInfoT2(info, glyph_index, NULL, NULL, NULL, NULL) == 0;
-   g = stbtt__GetGlyfOffset(info, glyph_index);
-   if (g < 0) return 1;
-   numberOfContours = ttSHORT(info->data + g);
-   return numberOfContours == 0;
-}
-
-static int stbtt__close_shape(stbtt_vertex *vertices, int num_vertices, int was_off, int start_off,
-    stbtt_int32 sx, stbtt_int32 sy, stbtt_int32 scx, stbtt_int32 scy, stbtt_int32 cx, stbtt_int32 cy)
-{
-   if (start_off) {
-      if (was_off)
-         stbtt_setvertex(&vertices[num_vertices++], STBTT_vcurve, (cx+scx)>>1, (cy+scy)>>1, cx,cy);
-      stbtt_setvertex(&vertices[num_vertices++], STBTT_vcurve, sx,sy,scx,scy);
-   } else {
-      if (was_off)
-         stbtt_setvertex(&vertices[num_vertices++], STBTT_vcurve,sx,sy,cx,cy);
-      else
-         stbtt_setvertex(&vertices[num_vertices++], STBTT_vline,sx,sy,0,0);
-   }
-   return num_vertices;
-}
-
-static int stbtt__GetGlyphShapeTT(const stbtt_fontinfo *info, int glyph_index, stbtt_vertex **pvertices)
-{
-   stbtt_int16 numberOfContours;
-   stbtt_uint8 *endPtsOfContours;
-   stbtt_uint8 *data = info->data;
-   stbtt_vertex *vertices=0;
-   int num_vertices=0;
-   int g = stbtt__GetGlyfOffset(info, glyph_index);
-
-   *pvertices = NULL;
-
-   if (g < 0) return 0;
-
-   numberOfContours = ttSHORT(data + g);
-
-   if (numberOfContours > 0) {
-      stbtt_uint8 flags=0,flagcount;
-      stbtt_int32 ins, i,j=0,m,n, next_move, was_off=0, off, start_off=0;
-      stbtt_int32 x,y,cx,cy,sx,sy, scx,scy;
-      stbtt_uint8 *points;
-      endPtsOfContours = (data + g + 10);
-      ins = ttUSHORT(data + g + 10 + numberOfContours * 2);
-      points = data + g + 10 + numberOfContours * 2 + 2 + ins;
-
-      n = 1+ttUSHORT(endPtsOfContours + numberOfContours*2-2);
-
-      m = n + 2*numberOfContours;  // a loose bound on how many vertices we might need
-      vertices = (stbtt_vertex *) STBTT_malloc(m * sizeof(vertices[0]), info->userdata);
-      if (vertices == 0)
-         return 0;
-
-      next_move = 0;
-      flagcount=0;
-
-      // in first pass, we load uninterpreted data into the allocated array
-      // above, shifted to the end of the array so we won't overwrite it when
-      // we create our final data starting from the front
-
-      off = m - n; // starting offset for uninterpreted data, regardless of how m ends up being calculated
-
-      // first load flags
-
-      for (i=0; i < n; ++i) {
-         if (flagcount == 0) {
-            flags = *points++;
-            if (flags & 8)
-               flagcount = *points++;
-         } else
-            --flagcount;
-         vertices[off+i].type = flags;
-      }
-
-      // now load x coordinates
-      x=0;
-      for (i=0; i < n; ++i) {
-         flags = vertices[off+i].type;
-         if (flags & 2) {
-            stbtt_int16 dx = *points++;
-            x += (flags & 16) ? dx : -dx; // ???
-         } else {
-            if (!(flags & 16)) {
-               x = x + (stbtt_int16) (points[0]*256 + points[1]);
-               points += 2;
-            }
-         }
-         vertices[off+i].x = (stbtt_int16) x;
-      }
-
-      // now load y coordinates
-      y=0;
-      for (i=0; i < n; ++i) {
-         flags = vertices[off+i].type;
-         if (flags & 4) {
-            stbtt_int16 dy = *points++;
-            y += (flags & 32) ? dy : -dy; // ???
-         } else {
-            if (!(flags & 32)) {
-               y = y + (stbtt_int16) (points[0]*256 + points[1]);
-               points += 2;
-            }
-         }
-         vertices[off+i].y = (stbtt_int16) y;
-      }
-
-      // now convert them to our format
-      num_vertices=0;
-      sx = sy = cx = cy = scx = scy = 0;
-      for (i=0; i < n; ++i) {
-         flags = vertices[off+i].type;
-         x     = (stbtt_int16) vertices[off+i].x;
-         y     = (stbtt_int16) vertices[off+i].y;
-
-         if (next_move == i) {
-            if (i != 0)
-               num_vertices = stbtt__close_shape(vertices, num_vertices, was_off, start_off, sx,sy,scx,scy,cx,cy);
-
-            // now start the new one
-            start_off = !(flags & 1);
-            if (start_off) {
-               // if we start off with an off-curve point, then when we need to find a point on the curve
-               // where we can start, and we need to save some state for when we wraparound.
-               scx = x;
-               scy = y;
-               if (!(vertices[off+i+1].type & 1)) {
-                  // next point is also a curve point, so interpolate an on-point curve
-                  sx = (x + (stbtt_int32) vertices[off+i+1].x) >> 1;
-                  sy = (y + (stbtt_int32) vertices[off+i+1].y) >> 1;
-               } else {
-                  // otherwise just use the next point as our start point
-                  sx = (stbtt_int32) vertices[off+i+1].x;
-                  sy = (stbtt_int32) vertices[off+i+1].y;
-                  ++i; // we're using point i+1 as the starting point, so skip it
-               }
-            } else {
-               sx = x;
-               sy = y;
-            }
-            stbtt_setvertex(&vertices[num_vertices++], STBTT_vmove,sx,sy,0,0);
-            was_off = 0;
-            next_move = 1 + ttUSHORT(endPtsOfContours+j*2);
-            ++j;
-         } else {
-            if (!(flags & 1)) { // if it's a curve
-               if (was_off) // two off-curve control points in a row means interpolate an on-curve midpoint
-                  stbtt_setvertex(&vertices[num_vertices++], STBTT_vcurve, (cx+x)>>1, (cy+y)>>1, cx, cy);
-               cx = x;
-               cy = y;
-               was_off = 1;
-            } else {
-               if (was_off)
-                  stbtt_setvertex(&vertices[num_vertices++], STBTT_vcurve, x,y, cx, cy);
-               else
-                  stbtt_setvertex(&vertices[num_vertices++], STBTT_vline, x,y,0,0);
-               was_off = 0;
-            }
-         }
-      }
-      num_vertices = stbtt__close_shape(vertices, num_vertices, was_off, start_off, sx,sy,scx,scy,cx,cy);
-   } else if (numberOfContours < 0) {
-      // Compound shapes.
-      int more = 1;
-      stbtt_uint8 *comp = data + g + 10;
-      num_vertices = 0;
-      vertices = 0;
-      while (more) {
-         stbtt_uint16 flags, gidx;
-         int comp_num_verts = 0, i;
-         stbtt_vertex *comp_verts = 0, *tmp = 0;
-         float mtx[6] = {1,0,0,1,0,0}, m, n;
-
-         flags = ttSHORT(comp); comp+=2;
-         gidx = ttSHORT(comp); comp+=2;
-
-         if (flags & 2) { // XY values
-            if (flags & 1) { // shorts
-               mtx[4] = ttSHORT(comp); comp+=2;
-               mtx[5] = ttSHORT(comp); comp+=2;
-            } else {
-               mtx[4] = ttCHAR(comp); comp+=1;
-               mtx[5] = ttCHAR(comp); comp+=1;
-            }
-         }
-         else {
-            // @TODO handle matching point
-            STBTT_assert(0);
-         }
-         if (flags & (1<<3)) { // WE_HAVE_A_SCALE
-            mtx[0] = mtx[3] = ttSHORT(comp)/16384.0f; comp+=2;
-            mtx[1] = mtx[2] = 0;
-         } else if (flags & (1<<6)) { // WE_HAVE_AN_X_AND_YSCALE
-            mtx[0] = ttSHORT(comp)/16384.0f; comp+=2;
-            mtx[1] = mtx[2] = 0;
-            mtx[3] = ttSHORT(comp)/16384.0f; comp+=2;
-         } else if (flags & (1<<7)) { // WE_HAVE_A_TWO_BY_TWO
-            mtx[0] = ttSHORT(comp)/16384.0f; comp+=2;
-            mtx[1] = ttSHORT(comp)/16384.0f; comp+=2;
-            mtx[2] = ttSHORT(comp)/16384.0f; comp+=2;
-            mtx[3] = ttSHORT(comp)/16384.0f; comp+=2;
-         }
-
-         // Find transformation scales.
-         m = (float) STBTT_sqrt(mtx[0]*mtx[0] + mtx[1]*mtx[1]);
-         n = (float) STBTT_sqrt(mtx[2]*mtx[2] + mtx[3]*mtx[3]);
-
-         // Get indexed glyph.
-         comp_num_verts = stbtt_GetGlyphShape(info, gidx, &comp_verts);
-         if (comp_num_verts > 0) {
-            // Transform vertices.
-            for (i = 0; i < comp_num_verts; ++i) {
-               stbtt_vertex* v = &comp_verts[i];
-               stbtt_vertex_type x,y;
-               x=v->x; y=v->y;
-               v->x = (stbtt_vertex_type)(m * (mtx[0]*x + mtx[2]*y + mtx[4]));
-               v->y = (stbtt_vertex_type)(n * (mtx[1]*x + mtx[3]*y + mtx[5]));
-               x=v->cx; y=v->cy;
-               v->cx = (stbtt_vertex_type)(m * (mtx[0]*x + mtx[2]*y + mtx[4]));
-               v->cy = (stbtt_vertex_type)(n * (mtx[1]*x + mtx[3]*y + mtx[5]));
-            }
-            // Append vertices.
-            tmp = (stbtt_vertex*)STBTT_malloc((num_vertices+comp_num_verts)*sizeof(stbtt_vertex), info->userdata);
-            if (!tmp) {
-               if (vertices) STBTT_free(vertices, info->userdata);
-               if (comp_verts) STBTT_free(comp_verts, info->userdata);
-               return 0;
-            }
-            if (num_vertices > 0 && vertices) STBTT_memcpy(tmp, vertices, num_vertices*sizeof(stbtt_vertex));
-            STBTT_memcpy(tmp+num_vertices, comp_verts, comp_num_verts*sizeof(stbtt_vertex));
-            if (vertices) STBTT_free(vertices, info->userdata);
-            vertices = tmp;
-            STBTT_free(comp_verts, info->userdata);
-            num_vertices += comp_num_verts;
-         }
-         // More components ?
-         more = flags & (1<<5);
-      }
-   } else {
-      // numberOfCounters == 0, do nothing
-   }
-
-   *pvertices = vertices;
-   return num_vertices;
-}
-
-typedef struct
-{
-   int bounds;
-   int started;
-   float first_x, first_y;
-   float x, y;
-   stbtt_int32 min_x, max_x, min_y, max_y;
-
-   stbtt_vertex *pvertices;
-   int num_vertices;
-} stbtt__csctx;
-
-#define STBTT__CSCTX_INIT(bounds) {bounds,0, 0,0, 0,0, 0,0,0,0, NULL, 0}
-
-static void stbtt__track_vertex(stbtt__csctx *c, stbtt_int32 x, stbtt_int32 y)
-{
-   if (x > c->max_x || !c->started) c->max_x = x;
-   if (y > c->max_y || !c->started) c->max_y = y;
-   if (x < c->min_x || !c->started) c->min_x = x;
-   if (y < c->min_y || !c->started) c->min_y = y;
-   c->started = 1;
-}
-
-static void stbtt__csctx_v(stbtt__csctx *c, stbtt_uint8 type, stbtt_int32 x, stbtt_int32 y, stbtt_int32 cx, stbtt_int32 cy, stbtt_int32 cx1, stbtt_int32 cy1)
-{
-   if (c->bounds) {
-      stbtt__track_vertex(c, x, y);
-      if (type == STBTT_vcubic) {
-         stbtt__track_vertex(c, cx, cy);
-         stbtt__track_vertex(c, cx1, cy1);
-      }
-   } else {
-      stbtt_setvertex(&c->pvertices[c->num_vertices], type, x, y, cx, cy);
-      c->pvertices[c->num_vertices].cx1 = (stbtt_int16) cx1;
-      c->pvertices[c->num_vertices].cy1 = (stbtt_int16) cy1;
-   }
-   c->num_vertices++;
-}
-
-static void stbtt__csctx_close_shape(stbtt__csctx *ctx)
-{
-   if (ctx->first_x != ctx->x || ctx->first_y != ctx->y)
-      stbtt__csctx_v(ctx, STBTT_vline, (int)ctx->first_x, (int)ctx->first_y, 0, 0, 0, 0);
-}
-
-static void stbtt__csctx_rmove_to(stbtt__csctx *ctx, float dx, float dy)
-{
-   stbtt__csctx_close_shape(ctx);
-   ctx->first_x = ctx->x = ctx->x + dx;
-   ctx->first_y = ctx->y = ctx->y + dy;
-   stbtt__csctx_v(ctx, STBTT_vmove, (int)ctx->x, (int)ctx->y, 0, 0, 0, 0);
-}
-
-static void stbtt__csctx_rline_to(stbtt__csctx *ctx, float dx, float dy)
-{
-   ctx->x += dx;
-   ctx->y += dy;
-   stbtt__csctx_v(ctx, STBTT_vline, (int)ctx->x, (int)ctx->y, 0, 0, 0, 0);
-}
-
-static void stbtt__csctx_rccurve_to(stbtt__csctx *ctx, float dx1, float dy1, float dx2, float dy2, float dx3, float dy3)
-{
-   float cx1 = ctx->x + dx1;
-   float cy1 = ctx->y + dy1;
-   float cx2 = cx1 + dx2;
-   float cy2 = cy1 + dy2;
-   ctx->x = cx2 + dx3;
-   ctx->y = cy2 + dy3;
-   stbtt__csctx_v(ctx, STBTT_vcubic, (int)ctx->x, (int)ctx->y, (int)cx1, (int)cy1, (int)cx2, (int)cy2);
-}
-
-static stbtt__buf stbtt__get_subr(stbtt__buf idx, int n)
-{
-   int count = stbtt__cff_index_count(&idx);
-   int bias = 107;
-   if (count >= 33900)
-      bias = 32768;
-   else if (count >= 1240)
-      bias = 1131;
-   n += bias;
-   if (n < 0 || n >= count)
-      return stbtt__new_buf(NULL, 0);
-   return stbtt__cff_index_get(idx, n);
-}
-
-static stbtt__buf stbtt__cid_get_glyph_subrs(const stbtt_fontinfo *info, int glyph_index)
-{
-   stbtt__buf fdselect = info->fdselect;
-   int nranges, start, end, v, fmt, fdselector = -1, i;
-
-   stbtt__buf_seek(&fdselect, 0);
-   fmt = stbtt__buf_get8(&fdselect);
-   if (fmt == 0) {
-      // untested
-      stbtt__buf_skip(&fdselect, glyph_index);
-      fdselector = stbtt__buf_get8(&fdselect);
-   } else if (fmt == 3) {
-      nranges = stbtt__buf_get16(&fdselect);
-      start = stbtt__buf_get16(&fdselect);
-      for (i = 0; i < nranges; i++) {
-         v = stbtt__buf_get8(&fdselect);
-         end = stbtt__buf_get16(&fdselect);
-         if (glyph_index >= start && glyph_index < end) {
-            fdselector = v;
-            break;
-         }
-         start = end;
-      }
-   }
-   if (fdselector == -1) stbtt__new_buf(NULL, 0);
-   return stbtt__get_subrs(info->cff, stbtt__cff_index_get(info->fontdicts, fdselector));
-}
-
-static int stbtt__run_charstring(const stbtt_fontinfo *info, int glyph_index, stbtt__csctx *c)
-{
-   int in_header = 1, maskbits = 0, subr_stack_height = 0, sp = 0, v, i, b0;
-   int has_subrs = 0, clear_stack;
-   float s[48];
-   stbtt__buf subr_stack[10], subrs = info->subrs, b;
-   float f;
-
-#define STBTT__CSERR(s) (0)
-
-   // this currently ignores the initial width value, which isn't needed if we have hmtx
-   b = stbtt__cff_index_get(info->charstrings, glyph_index);
-   while (b.cursor < b.size) {
-      i = 0;
-      clear_stack = 1;
-      b0 = stbtt__buf_get8(&b);
-      switch (b0) {
-      // @TODO implement hinting
-      case 0x13: // hintmask
-      case 0x14: // cntrmask
-         if (in_header)
-            maskbits += (sp / 2); // implicit "vstem"
-         in_header = 0;
-         stbtt__buf_skip(&b, (maskbits + 7) / 8);
-         break;
-
-      case 0x01: // hstem
-      case 0x03: // vstem
-      case 0x12: // hstemhm
-      case 0x17: // vstemhm
-         maskbits += (sp / 2);
-         break;
-
-      case 0x15: // rmoveto
-         in_header = 0;
-         if (sp < 2) return STBTT__CSERR("rmoveto stack");
-         stbtt__csctx_rmove_to(c, s[sp-2], s[sp-1]);
-         break;
-      case 0x04: // vmoveto
-         in_header = 0;
-         if (sp < 1) return STBTT__CSERR("vmoveto stack");
-         stbtt__csctx_rmove_to(c, 0, s[sp-1]);
-         break;
-      case 0x16: // hmoveto
-         in_header = 0;
-         if (sp < 1) return STBTT__CSERR("hmoveto stack");
-         stbtt__csctx_rmove_to(c, s[sp-1], 0);
-         break;
-
-      case 0x05: // rlineto
-         if (sp < 2) return STBTT__CSERR("rlineto stack");
-         for (; i + 1 < sp; i += 2)
-            stbtt__csctx_rline_to(c, s[i], s[i+1]);
-         break;
-
-      // hlineto/vlineto and vhcurveto/hvcurveto alternate horizontal and vertical
-      // starting from a different place.
-
-      case 0x07: // vlineto
-         if (sp < 1) return STBTT__CSERR("vlineto stack");
-         goto vlineto;
-      case 0x06: // hlineto
-         if (sp < 1) return STBTT__CSERR("hlineto stack");
-         for (;;) {
-            if (i >= sp) break;
-            stbtt__csctx_rline_to(c, s[i], 0);
-            i++;
-      vlineto:
-            if (i >= sp) break;
-            stbtt__csctx_rline_to(c, 0, s[i]);
-            i++;
-         }
-         break;
-
-      case 0x1F: // hvcurveto
-         if (sp < 4) return STBTT__CSERR("hvcurveto stack");
-         goto hvcurveto;
-      case 0x1E: // vhcurveto
-         if (sp < 4) return STBTT__CSERR("vhcurveto stack");
-         for (;;) {
-            if (i + 3 >= sp) break;
-            stbtt__csctx_rccurve_to(c, 0, s[i], s[i+1], s[i+2], s[i+3], (sp - i == 5) ? s[i + 4] : 0.0f);
-            i += 4;
-      hvcurveto:
-            if (i + 3 >= sp) break;
-            stbtt__csctx_rccurve_to(c, s[i], 0, s[i+1], s[i+2], (sp - i == 5) ? s[i+4] : 0.0f, s[i+3]);
-            i += 4;
-         }
-         break;
-
-      case 0x08: // rrcurveto
-         if (sp < 6) return STBTT__CSERR("rcurveline stack");
-         for (; i + 5 < sp; i += 6)
-            stbtt__csctx_rccurve_to(c, s[i], s[i+1], s[i+2], s[i+3], s[i+4], s[i+5]);
-         break;
-
-      case 0x18: // rcurveline
-         if (sp < 8) return STBTT__CSERR("rcurveline stack");
-         for (; i + 5 < sp - 2; i += 6)
-            stbtt__csctx_rccurve_to(c, s[i], s[i+1], s[i+2], s[i+3], s[i+4], s[i+5]);
-         if (i + 1 >= sp) return STBTT__CSERR("rcurveline stack");
-         stbtt__csctx_rline_to(c, s[i], s[i+1]);
-         break;
-
-      case 0x19: // rlinecurve
-         if (sp < 8) return STBTT__CSERR("rlinecurve stack");
-         for (; i + 1 < sp - 6; i += 2)
-            stbtt__csctx_rline_to(c, s[i], s[i+1]);
-         if (i + 5 >= sp) return STBTT__CSERR("rlinecurve stack");
-         stbtt__csctx_rccurve_to(c, s[i], s[i+1], s[i+2], s[i+3], s[i+4], s[i+5]);
-         break;
-
-      case 0x1A: // vvcurveto
-      case 0x1B: // hhcurveto
-         if (sp < 4) return STBTT__CSERR("(vv|hh)curveto stack");
-         f = 0.0;
-         if (sp & 1) { f = s[i]; i++; }
-         for (; i + 3 < sp; i += 4) {
-            if (b0 == 0x1B)
-               stbtt__csctx_rccurve_to(c, s[i], f, s[i+1], s[i+2], s[i+3], 0.0);
-            else
-               stbtt__csctx_rccurve_to(c, f, s[i], s[i+1], s[i+2], 0.0, s[i+3]);
-            f = 0.0;
-         }
-         break;
-
-      case 0x0A: // callsubr
-         if (!has_subrs) {
-            if (info->fdselect.size)
-               subrs = stbtt__cid_get_glyph_subrs(info, glyph_index);
-            has_subrs = 1;
-         }
-         // FALLTHROUGH
-      case 0x1D: // callgsubr
-         if (sp < 1) return STBTT__CSERR("call(g|)subr stack");
-         v = (int) s[--sp];
-         if (subr_stack_height >= 10) return STBTT__CSERR("recursion limit");
-         subr_stack[subr_stack_height++] = b;
-         b = stbtt__get_subr(b0 == 0x0A ? subrs : info->gsubrs, v);
-         if (b.size == 0) return STBTT__CSERR("subr not found");
-         b.cursor = 0;
-         clear_stack = 0;
-         break;
-
-      case 0x0B: // return
-         if (subr_stack_height <= 0) return STBTT__CSERR("return outside subr");
-         b = subr_stack[--subr_stack_height];
-         clear_stack = 0;
-         break;
-
-      case 0x0E: // endchar
-         stbtt__csctx_close_shape(c);
-         return 1;
-
-      case 0x0C: { // two-byte escape
-         float dx1, dx2, dx3, dx4, dx5, dx6, dy1, dy2, dy3, dy4, dy5, dy6;
-         float dx, dy;
-         int b1 = stbtt__buf_get8(&b);
-         switch (b1) {
-         // @TODO These "flex" implementations ignore the flex-depth and resolution,
-         // and always draw beziers.
-         case 0x22: // hflex
-            if (sp < 7) return STBTT__CSERR("hflex stack");
-            dx1 = s[0];
-            dx2 = s[1];
-            dy2 = s[2];
-            dx3 = s[3];
-            dx4 = s[4];
-            dx5 = s[5];
-            dx6 = s[6];
-            stbtt__csctx_rccurve_to(c, dx1, 0, dx2, dy2, dx3, 0);
-            stbtt__csctx_rccurve_to(c, dx4, 0, dx5, -dy2, dx6, 0);
-            break;
-
-         case 0x23: // flex
-            if (sp < 13) return STBTT__CSERR("flex stack");
-            dx1 = s[0];
-            dy1 = s[1];
-            dx2 = s[2];
-            dy2 = s[3];
-            dx3 = s[4];
-            dy3 = s[5];
-            dx4 = s[6];
-            dy4 = s[7];
-            dx5 = s[8];
-            dy5 = s[9];
-            dx6 = s[10];
-            dy6 = s[11];
-            //fd is s[12]
-            stbtt__csctx_rccurve_to(c, dx1, dy1, dx2, dy2, dx3, dy3);
-            stbtt__csctx_rccurve_to(c, dx4, dy4, dx5, dy5, dx6, dy6);
-            break;
-
-         case 0x24: // hflex1
-            if (sp < 9) return STBTT__CSERR("hflex1 stack");
-            dx1 = s[0];
-            dy1 = s[1];
-            dx2 = s[2];
-            dy2 = s[3];
-            dx3 = s[4];
-            dx4 = s[5];
-            dx5 = s[6];
-            dy5 = s[7];
-            dx6 = s[8];
-            stbtt__csctx_rccurve_to(c, dx1, dy1, dx2, dy2, dx3, 0);
-            stbtt__csctx_rccurve_to(c, dx4, 0, dx5, dy5, dx6, -(dy1+dy2+dy5));
-            break;
-
-         case 0x25: // flex1
-            if (sp < 11) return STBTT__CSERR("flex1 stack");
-            dx1 = s[0];
-            dy1 = s[1];
-            dx2 = s[2];
-            dy2 = s[3];
-            dx3 = s[4];
-            dy3 = s[5];
-            dx4 = s[6];
-            dy4 = s[7];
-            dx5 = s[8];
-            dy5 = s[9];
-            dx6 = dy6 = s[10];
-            dx = dx1+dx2+dx3+dx4+dx5;
-            dy = dy1+dy2+dy3+dy4+dy5;
-            if (STBTT_fabs(dx) > STBTT_fabs(dy))
-               dy6 = -dy;
-            else
-               dx6 = -dx;
-            stbtt__csctx_rccurve_to(c, dx1, dy1, dx2, dy2, dx3, dy3);
-            stbtt__csctx_rccurve_to(c, dx4, dy4, dx5, dy5, dx6, dy6);
-            break;
-
-         default:
-            return STBTT__CSERR("unimplemented");
-         }
-      } break;
-
-      default:
-         if (b0 != 255 && b0 != 28 && b0 < 32)
-            return STBTT__CSERR("reserved operator");
-
-         // push immediate
-         if (b0 == 255) {
-            f = (float)(stbtt_int32)stbtt__buf_get32(&b) / 0x10000;
-         } else {
-            stbtt__buf_skip(&b, -1);
-            f = (float)(stbtt_int16)stbtt__cff_int(&b);
-         }
-         if (sp >= 48) return STBTT__CSERR("push stack overflow");
-         s[sp++] = f;
-         clear_stack = 0;
-         break;
-      }
-      if (clear_stack) sp = 0;
-   }
-   return STBTT__CSERR("no endchar");
-
-#undef STBTT__CSERR
-}
-
-static int stbtt__GetGlyphShapeT2(const stbtt_fontinfo *info, int glyph_index, stbtt_vertex **pvertices)
-{
-   // runs the charstring twice, once to count and once to output (to avoid realloc)
-   stbtt__csctx count_ctx = STBTT__CSCTX_INIT(1);
-   stbtt__csctx output_ctx = STBTT__CSCTX_INIT(0);
-   if (stbtt__run_charstring(info, glyph_index, &count_ctx)) {
-      *pvertices = (stbtt_vertex*)STBTT_malloc(count_ctx.num_vertices*sizeof(stbtt_vertex), info->userdata);
-      output_ctx.pvertices = *pvertices;
-      if (stbtt__run_charstring(info, glyph_index, &output_ctx)) {
-         STBTT_assert(output_ctx.num_vertices == count_ctx.num_vertices);
-         return output_ctx.num_vertices;
-      }
-   }
-   *pvertices = NULL;
-   return 0;
-}
-
-static int stbtt__GetGlyphInfoT2(const stbtt_fontinfo *info, int glyph_index, int *x0, int *y0, int *x1, int *y1)
-{
-   stbtt__csctx c = STBTT__CSCTX_INIT(1);
-   int r = stbtt__run_charstring(info, glyph_index, &c);
-   if (x0)  *x0 = r ? c.min_x : 0;
-   if (y0)  *y0 = r ? c.min_y : 0;
-   if (x1)  *x1 = r ? c.max_x : 0;
-   if (y1)  *y1 = r ? c.max_y : 0;
-   return r ? c.num_vertices : 0;
-}
-
-STBTT_DEF int stbtt_GetGlyphShape(const stbtt_fontinfo *info, int glyph_index, stbtt_vertex **pvertices)
-{
-   if (!info->cff.size)
-      return stbtt__GetGlyphShapeTT(info, glyph_index, pvertices);
-   else
-      return stbtt__GetGlyphShapeT2(info, glyph_index, pvertices);
-}
-
-STBTT_DEF void stbtt_GetGlyphHMetrics(const stbtt_fontinfo *info, int glyph_index, int *advanceWidth, int *leftSideBearing)
-{
-   stbtt_uint16 numOfLongHorMetrics = ttUSHORT(info->data+info->hhea + 34);
-   if (glyph_index < numOfLongHorMetrics) {
-      if (advanceWidth)     *advanceWidth    = ttSHORT(info->data + info->hmtx + 4*glyph_index);
-      if (leftSideBearing)  *leftSideBearing = ttSHORT(info->data + info->hmtx + 4*glyph_index + 2);
-   } else {
-      if (advanceWidth)     *advanceWidth    = ttSHORT(info->data + info->hmtx + 4*(numOfLongHorMetrics-1));
-      if (leftSideBearing)  *leftSideBearing = ttSHORT(info->data + info->hmtx + 4*numOfLongHorMetrics + 2*(glyph_index - numOfLongHorMetrics));
-   }
-}
-
-STBTT_DEF int  stbtt_GetKerningTableLength(const stbtt_fontinfo *info)
-{
-   stbtt_uint8 *data = info->data + info->kern;
-
-   // we only look at the first table. it must be 'horizontal' and format 0.
-   if (!info->kern)
-      return 0;
-   if (ttUSHORT(data+2) < 1) // number of tables, need at least 1
-      return 0;
-   if (ttUSHORT(data+8) != 1) // horizontal flag must be set in format
-      return 0;
-
-   return ttUSHORT(data+10);
-}
-
-STBTT_DEF int stbtt_GetKerningTable(const stbtt_fontinfo *info, stbtt_kerningentry* table, int table_length)
-{
-   stbtt_uint8 *data = info->data + info->kern;
-   int k, length;
-
-   // we only look at the first table. it must be 'horizontal' and format 0.
-   if (!info->kern)
-      return 0;
-   if (ttUSHORT(data+2) < 1) // number of tables, need at least 1
-      return 0;
-   if (ttUSHORT(data+8) != 1) // horizontal flag must be set in format
-      return 0;
-
-   length = ttUSHORT(data+10);
-   if (table_length < length)
-      length = table_length;
-
-   for (k = 0; k < length; k++)
-   {
-      table[k].glyph1 = ttUSHORT(data+18+(k*6));
-      table[k].glyph2 = ttUSHORT(data+20+(k*6));
-      table[k].advance = ttSHORT(data+22+(k*6));
-   }
-
-   return length;
-}
-
-static int stbtt__GetGlyphKernInfoAdvance(const stbtt_fontinfo *info, int glyph1, int glyph2)
-{
-   stbtt_uint8 *data = info->data + info->kern;
-   stbtt_uint32 needle, straw;
-   int l, r, m;
-
-   // we only look at the first table. it must be 'horizontal' and format 0.
-   if (!info->kern)
-      return 0;
-   if (ttUSHORT(data+2) < 1) // number of tables, need at least 1
-      return 0;
-   if (ttUSHORT(data+8) != 1) // horizontal flag must be set in format
-      return 0;
-
-   l = 0;
-   r = ttUSHORT(data+10) - 1;
-   needle = glyph1 << 16 | glyph2;
-   while (l <= r) {
-      m = (l + r) >> 1;
-      straw = ttULONG(data+18+(m*6)); // note: unaligned read
-      if (needle < straw)
-         r = m - 1;
-      else if (needle > straw)
-         l = m + 1;
-      else
-         return ttSHORT(data+22+(m*6));
-   }
-   return 0;
-}
-
-static stbtt_int32 stbtt__GetCoverageIndex(stbtt_uint8 *coverageTable, int glyph)
-{
-   stbtt_uint16 coverageFormat = ttUSHORT(coverageTable);
-   switch (coverageFormat) {
-      case 1: {
-         stbtt_uint16 glyphCount = ttUSHORT(coverageTable + 2);
-
-         // Binary search.
-         stbtt_int32 l=0, r=glyphCount-1, m;
-         int straw, needle=glyph;
-         while (l <= r) {
-            stbtt_uint8 *glyphArray = coverageTable + 4;
-            stbtt_uint16 glyphID;
-            m = (l + r) >> 1;
-            glyphID = ttUSHORT(glyphArray + 2 * m);
-            straw = glyphID;
-            if (needle < straw)
-               r = m - 1;
-            else if (needle > straw)
-               l = m + 1;
-            else {
-               return m;
-            }
-         }
-         break;
-      }
-
-      case 2: {
-         stbtt_uint16 rangeCount = ttUSHORT(coverageTable + 2);
-         stbtt_uint8 *rangeArray = coverageTable + 4;
-
-         // Binary search.
-         stbtt_int32 l=0, r=rangeCount-1, m;
-         int strawStart, strawEnd, needle=glyph;
-         while (l <= r) {
-            stbtt_uint8 *rangeRecord;
-            m = (l + r) >> 1;
-            rangeRecord = rangeArray + 6 * m;
-            strawStart = ttUSHORT(rangeRecord);
-            strawEnd = ttUSHORT(rangeRecord + 2);
-            if (needle < strawStart)
-               r = m - 1;
-            else if (needle > strawEnd)
-               l = m + 1;
-            else {
-               stbtt_uint16 startCoverageIndex = ttUSHORT(rangeRecord + 4);
-               return startCoverageIndex + glyph - strawStart;
-            }
-         }
-         break;
-      }
-
-      default: return -1; // unsupported
-   }
-
-   return -1;
-}
-
-static stbtt_int32  stbtt__GetGlyphClass(stbtt_uint8 *classDefTable, int glyph)
-{
-   stbtt_uint16 classDefFormat = ttUSHORT(classDefTable);
-   switch (classDefFormat)
-   {
-      case 1: {
-         stbtt_uint16 startGlyphID = ttUSHORT(classDefTable + 2);
-         stbtt_uint16 glyphCount = ttUSHORT(classDefTable + 4);
-         stbtt_uint8 *classDef1ValueArray = classDefTable + 6;
-
-         if (glyph >= startGlyphID && glyph < startGlyphID + glyphCount)
-            return (stbtt_int32)ttUSHORT(classDef1ValueArray + 2 * (glyph - startGlyphID));
-         break;
-      }
-
-      case 2: {
-         stbtt_uint16 classRangeCount = ttUSHORT(classDefTable + 2);
-         stbtt_uint8 *classRangeRecords = classDefTable + 4;
-
-         // Binary search.
-         stbtt_int32 l=0, r=classRangeCount-1, m;
-         int strawStart, strawEnd, needle=glyph;
-         while (l <= r) {
-            stbtt_uint8 *classRangeRecord;
-            m = (l + r) >> 1;
-            classRangeRecord = classRangeRecords + 6 * m;
-            strawStart = ttUSHORT(classRangeRecord);
-            strawEnd = ttUSHORT(classRangeRecord + 2);
-            if (needle < strawStart)
-               r = m - 1;
-            else if (needle > strawEnd)
-               l = m + 1;
-            else
-               return (stbtt_int32)ttUSHORT(classRangeRecord + 4);
-         }
-         break;
-      }
-
-      default:
-         return -1; // Unsupported definition type, return an error.
-   }
-
-   // "All glyphs not assigned to a class fall into class 0". (OpenType spec)
-   return 0;
-}
-
-// Define to STBTT_assert(x) if you want to break on unimplemented formats.
-#define STBTT_GPOS_TODO_assert(x)
-
-static stbtt_int32 stbtt__GetGlyphGPOSInfoAdvance(const stbtt_fontinfo *info, int glyph1, int glyph2)
-{
-   stbtt_uint16 lookupListOffset;
-   stbtt_uint8 *lookupList;
-   stbtt_uint16 lookupCount;
-   stbtt_uint8 *data;
-   stbtt_int32 i, sti;
-
-   if (!info->gpos) return 0;
-
-   data = info->data + info->gpos;
-
-   if (ttUSHORT(data+0) != 1) return 0; // Major version 1
-   if (ttUSHORT(data+2) != 0) return 0; // Minor version 0
-
-   lookupListOffset = ttUSHORT(data+8);
-   lookupList = data + lookupListOffset;
-   lookupCount = ttUSHORT(lookupList);
-
-   for (i=0; i<lookupCount; ++i) {
-      stbtt_uint16 lookupOffset = ttUSHORT(lookupList + 2 + 2 * i);
-      stbtt_uint8 *lookupTable = lookupList + lookupOffset;
-
-      stbtt_uint16 lookupType = ttUSHORT(lookupTable);
-      stbtt_uint16 subTableCount = ttUSHORT(lookupTable + 4);
-      stbtt_uint8 *subTableOffsets = lookupTable + 6;
-      if (lookupType != 2) // Pair Adjustment Positioning Subtable
-         continue;
-
-      for (sti=0; sti<subTableCount; sti++) {
-         stbtt_uint16 subtableOffset = ttUSHORT(subTableOffsets + 2 * sti);
-         stbtt_uint8 *table = lookupTable + subtableOffset;
-         stbtt_uint16 posFormat = ttUSHORT(table);
-         stbtt_uint16 coverageOffset = ttUSHORT(table + 2);
-         stbtt_int32 coverageIndex = stbtt__GetCoverageIndex(table + coverageOffset, glyph1);
-         if (coverageIndex == -1) continue;
-
-         switch (posFormat) {
-            case 1: {
-               stbtt_int32 l, r, m;
-               int straw, needle;
-               stbtt_uint16 valueFormat1 = ttUSHORT(table + 4);
-               stbtt_uint16 valueFormat2 = ttUSHORT(table + 6);
-               if (valueFormat1 == 4 && valueFormat2 == 0) { // Support more formats?
-                  stbtt_int32 valueRecordPairSizeInBytes = 2;
-                  stbtt_uint16 pairSetCount = ttUSHORT(table + 8);
-                  stbtt_uint16 pairPosOffset = ttUSHORT(table + 10 + 2 * coverageIndex);
-                  stbtt_uint8 *pairValueTable = table + pairPosOffset;
-                  stbtt_uint16 pairValueCount = ttUSHORT(pairValueTable);
-                  stbtt_uint8 *pairValueArray = pairValueTable + 2;
-
-                  if (coverageIndex >= pairSetCount) return 0;
-
-                  needle=glyph2;
-                  r=pairValueCount-1;
-                  l=0;
-
-                  // Binary search.
-                  while (l <= r) {
-                     stbtt_uint16 secondGlyph;
-                     stbtt_uint8 *pairValue;
-                     m = (l + r) >> 1;
-                     pairValue = pairValueArray + (2 + valueRecordPairSizeInBytes) * m;
-                     secondGlyph = ttUSHORT(pairValue);
-                     straw = secondGlyph;
-                     if (needle < straw)
-                        r = m - 1;
-                     else if (needle > straw)
-                        l = m + 1;
-                     else {
-                        stbtt_int16 xAdvance = ttSHORT(pairValue + 2);
-                        return xAdvance;
-                     }
-                  }
-               } else
-                  return 0;
-               break;
-            }
-
-            case 2: {
-               stbtt_uint16 valueFormat1 = ttUSHORT(table + 4);
-               stbtt_uint16 valueFormat2 = ttUSHORT(table + 6);
-               if (valueFormat1 == 4 && valueFormat2 == 0) { // Support more formats?
-                  stbtt_uint16 classDef1Offset = ttUSHORT(table + 8);
-                  stbtt_uint16 classDef2Offset = ttUSHORT(table + 10);
-                  int glyph1class = stbtt__GetGlyphClass(table + classDef1Offset, glyph1);
-                  int glyph2class = stbtt__GetGlyphClass(table + classDef2Offset, glyph2);
-
-                  stbtt_uint16 class1Count = ttUSHORT(table + 12);
-                  stbtt_uint16 class2Count = ttUSHORT(table + 14);
-                  stbtt_uint8 *class1Records, *class2Records;
-                  stbtt_int16 xAdvance;
-
-                  if (glyph1class < 0 || glyph1class >= class1Count) return 0; // malformed
-                  if (glyph2class < 0 || glyph2class >= class2Count) return 0; // malformed
-
-                  class1Records = table + 16;
-                  class2Records = class1Records + 2 * (glyph1class * class2Count);
-                  xAdvance = ttSHORT(class2Records + 2 * glyph2class);
-                  return xAdvance;
-               } else
-                  return 0;
-               break;
-            }
-
-            default:
-               return 0; // Unsupported position format
-         }
-      }
-   }
-
-   return 0;
-}
-
-STBTT_DEF int  stbtt_GetGlyphKernAdvance(const stbtt_fontinfo *info, int g1, int g2)
-{
-   int xAdvance = 0;
-
-   if (info->gpos)
-      xAdvance += stbtt__GetGlyphGPOSInfoAdvance(info, g1, g2);
-   else if (info->kern)
-      xAdvance += stbtt__GetGlyphKernInfoAdvance(info, g1, g2);
-
-   return xAdvance;
-}
-
-STBTT_DEF int  stbtt_GetCodepointKernAdvance(const stbtt_fontinfo *info, int ch1, int ch2)
-{
-   if (!info->kern && !info->gpos) // if no kerning table, don't waste time looking up both codepoint->glyphs
-      return 0;
-   return stbtt_GetGlyphKernAdvance(info, stbtt_FindGlyphIndex(info,ch1), stbtt_FindGlyphIndex(info,ch2));
-}
-
-STBTT_DEF void stbtt_GetCodepointHMetrics(const stbtt_fontinfo *info, int codepoint, int *advanceWidth, int *leftSideBearing)
-{
-   stbtt_GetGlyphHMetrics(info, stbtt_FindGlyphIndex(info,codepoint), advanceWidth, leftSideBearing);
-}
-
-STBTT_DEF void stbtt_GetFontVMetrics(const stbtt_fontinfo *info, int *ascent, int *descent, int *lineGap)
-{
-   if (ascent ) *ascent  = ttSHORT(info->data+info->hhea + 4);
-   if (descent) *descent = ttSHORT(info->data+info->hhea + 6);
-   if (lineGap) *lineGap = ttSHORT(info->data+info->hhea + 8);
-}
-
-STBTT_DEF int  stbtt_GetFontVMetricsOS2(const stbtt_fontinfo *info, int *typoAscent, int *typoDescent, int *typoLineGap)
-{
-   int tab = stbtt__find_table(info->data, info->fontstart, "OS/2");
-   if (!tab)
-      return 0;
-   if (typoAscent ) *typoAscent  = ttSHORT(info->data+tab + 68);
-   if (typoDescent) *typoDescent = ttSHORT(info->data+tab + 70);
-   if (typoLineGap) *typoLineGap = ttSHORT(info->data+tab + 72);
-   return 1;
-}
-
-STBTT_DEF void stbtt_GetFontBoundingBox(const stbtt_fontinfo *info, int *x0, int *y0, int *x1, int *y1)
-{
-   *x0 = ttSHORT(info->data + info->head + 36);
-   *y0 = ttSHORT(info->data + info->head + 38);
-   *x1 = ttSHORT(info->data + info->head + 40);
-   *y1 = ttSHORT(info->data + info->head + 42);
-}
-
-STBTT_DEF float stbtt_ScaleForPixelHeight(const stbtt_fontinfo *info, float height)
-{
-   int fheight = ttSHORT(info->data + info->hhea + 4) - ttSHORT(info->data + info->hhea + 6);
-   return (float) height / fheight;
-}
-
-STBTT_DEF float stbtt_ScaleForMappingEmToPixels(const stbtt_fontinfo *info, float pixels)
-{
-   int unitsPerEm = ttUSHORT(info->data + info->head + 18);
-   return pixels / unitsPerEm;
-}
-
-STBTT_DEF void stbtt_FreeShape(const stbtt_fontinfo *info, stbtt_vertex *v)
-{
-   STBTT_free(v, info->userdata);
-}
-
-STBTT_DEF stbtt_uint8 *stbtt_FindSVGDoc(const stbtt_fontinfo *info, int gl)
-{
-   int i;
-   stbtt_uint8 *data = info->data;
-   stbtt_uint8 *svg_doc_list = data + stbtt__get_svg((stbtt_fontinfo *) info);
-
-   int numEntries = ttUSHORT(svg_doc_list);
-   stbtt_uint8 *svg_docs = svg_doc_list + 2;
-
-   for(i=0; i<numEntries; i++) {
-      stbtt_uint8 *svg_doc = svg_docs + (12 * i);
-      if ((gl >= ttUSHORT(svg_doc)) && (gl <= ttUSHORT(svg_doc + 2)))
-         return svg_doc;
-   }
-   return 0;
-}
-
-STBTT_DEF int stbtt_GetGlyphSVG(const stbtt_fontinfo *info, int gl, const char **svg)
-{
-   stbtt_uint8 *data = info->data;
-   stbtt_uint8 *svg_doc;
-
-   if (info->svg == 0)
-      return 0;
-
-   svg_doc = stbtt_FindSVGDoc(info, gl);
-   if (svg_doc != NULL) {
-      *svg = (char *) data + info->svg + ttULONG(svg_doc + 4);
-      return ttULONG(svg_doc + 8);
-   } else {
-      return 0;
-   }
-}
-
-STBTT_DEF int stbtt_GetCodepointSVG(const stbtt_fontinfo *info, int unicode_codepoint, const char **svg)
-{
-   return stbtt_GetGlyphSVG(info, stbtt_FindGlyphIndex(info, unicode_codepoint), svg);
-}
-
-//////////////////////////////////////////////////////////////////////////////
-//
-// antialiasing software rasterizer
-//
-
-STBTT_DEF void stbtt_GetGlyphBitmapBoxSubpixel(const stbtt_fontinfo *font, int glyph, float scale_x, float scale_y,float shift_x, float shift_y, int *ix0, int *iy0, int *ix1, int *iy1)
-{
-   int x0=0,y0=0,x1,y1; // =0 suppresses compiler warning
-   if (!stbtt_GetGlyphBox(font, glyph, &x0,&y0,&x1,&y1)) {
-      // e.g. space character
-      if (ix0) *ix0 = 0;
-      if (iy0) *iy0 = 0;
-      if (ix1) *ix1 = 0;
-      if (iy1) *iy1 = 0;
-   } else {
-      // move to integral bboxes (treating pixels as little squares, what pixels get touched)?
-      if (ix0) *ix0 = STBTT_ifloor( x0 * scale_x + shift_x);
-      if (iy0) *iy0 = STBTT_ifloor(-y1 * scale_y + shift_y);
-      if (ix1) *ix1 = STBTT_iceil ( x1 * scale_x + shift_x);
-      if (iy1) *iy1 = STBTT_iceil (-y0 * scale_y + shift_y);
-   }
-}
-
-STBTT_DEF void stbtt_GetGlyphBitmapBox(const stbtt_fontinfo *font, int glyph, float scale_x, float scale_y, int *ix0, int *iy0, int *ix1, int *iy1)
-{
-   stbtt_GetGlyphBitmapBoxSubpixel(font, glyph, scale_x, scale_y,0.0f,0.0f, ix0, iy0, ix1, iy1);
-}
-
-STBTT_DEF void stbtt_GetCodepointBitmapBoxSubpixel(const stbtt_fontinfo *font, int codepoint, float scale_x, float scale_y, float shift_x, float shift_y, int *ix0, int *iy0, int *ix1, int *iy1)
-{
-   stbtt_GetGlyphBitmapBoxSubpixel(font, stbtt_FindGlyphIndex(font,codepoint), scale_x, scale_y,shift_x,shift_y, ix0,iy0,ix1,iy1);
-}
-
-STBTT_DEF void stbtt_GetCodepointBitmapBox(const stbtt_fontinfo *font, int codepoint, float scale_x, float scale_y, int *ix0, int *iy0, int *ix1, int *iy1)
-{
-   stbtt_GetCodepointBitmapBoxSubpixel(font, codepoint, scale_x, scale_y,0.0f,0.0f, ix0,iy0,ix1,iy1);
-}
-
-//////////////////////////////////////////////////////////////////////////////
-//
-//  Rasterizer
-
-typedef struct stbtt__hheap_chunk
-{
-   struct stbtt__hheap_chunk *next;
-} stbtt__hheap_chunk;
-
-typedef struct stbtt__hheap
-{
-   struct stbtt__hheap_chunk *head;
-   void   *first_free;
-   int    num_remaining_in_head_chunk;
-} stbtt__hheap;
-
-static void *stbtt__hheap_alloc(stbtt__hheap *hh, size_t size, void *userdata)
-{
-   if (hh->first_free) {
-      void *p = hh->first_free;
-      hh->first_free = * (void **) p;
-      return p;
-   } else {
-      if (hh->num_remaining_in_head_chunk == 0) {
-         int count = (size < 32 ? 2000 : size < 128 ? 800 : 100);
-         stbtt__hheap_chunk *c = (stbtt__hheap_chunk *) STBTT_malloc(sizeof(stbtt__hheap_chunk) + size * count, userdata);
-         if (c == NULL)
-            return NULL;
-         c->next = hh->head;
-         hh->head = c;
-         hh->num_remaining_in_head_chunk = count;
-      }
-      --hh->num_remaining_in_head_chunk;
-      return (char *) (hh->head) + sizeof(stbtt__hheap_chunk) + size * hh->num_remaining_in_head_chunk;
-   }
-}
-
-static void stbtt__hheap_free(stbtt__hheap *hh, void *p)
-{
-   *(void **) p = hh->first_free;
-   hh->first_free = p;
-}
-
-static void stbtt__hheap_cleanup(stbtt__hheap *hh, void *userdata)
-{
-   stbtt__hheap_chunk *c = hh->head;
-   while (c) {
-      stbtt__hheap_chunk *n = c->next;
-      STBTT_free(c, userdata);
-      c = n;
-   }
-}
-
-typedef struct stbtt__edge {
-   float x0,y0, x1,y1;
-   int invert;
-} stbtt__edge;
-
-
-typedef struct stbtt__active_edge
-{
-   struct stbtt__active_edge *next;
-   #if STBTT_RASTERIZER_VERSION==1
-   int x,dx;
-   float ey;
-   int direction;
-   #elif STBTT_RASTERIZER_VERSION==2
-   float fx,fdx,fdy;
-   float direction;
-   float sy;
-   float ey;
-   #else
-   #error "Unrecognized value of STBTT_RASTERIZER_VERSION"
-   #endif
-} stbtt__active_edge;
-
-#if STBTT_RASTERIZER_VERSION == 1
-#define STBTT_FIXSHIFT   10
-#define STBTT_FIX        (1 << STBTT_FIXSHIFT)
-#define STBTT_FIXMASK    (STBTT_FIX-1)
-
-static stbtt__active_edge *stbtt__new_active(stbtt__hheap *hh, stbtt__edge *e, int off_x, float start_point, void *userdata)
-{
-   stbtt__active_edge *z = (stbtt__active_edge *) stbtt__hheap_alloc(hh, sizeof(*z), userdata);
-   float dxdy = (e->x1 - e->x0) / (e->y1 - e->y0);
-   STBTT_assert(z != NULL);
-   if (!z) return z;
-
-   // round dx down to avoid overshooting
-   if (dxdy < 0)
-      z->dx = -STBTT_ifloor(STBTT_FIX * -dxdy);
-   else
-      z->dx = STBTT_ifloor(STBTT_FIX * dxdy);
-
-   z->x = STBTT_ifloor(STBTT_FIX * e->x0 + z->dx * (start_point - e->y0)); // use z->dx so when we offset later it's by the same amount
-   z->x -= off_x * STBTT_FIX;
-
-   z->ey = e->y1;
-   z->next = 0;
-   z->direction = e->invert ? 1 : -1;
-   return z;
-}
-#elif STBTT_RASTERIZER_VERSION == 2
-static stbtt__active_edge *stbtt__new_active(stbtt__hheap *hh, stbtt__edge *e, int off_x, float start_point, void *userdata)
-{
-   stbtt__active_edge *z = (stbtt__active_edge *) stbtt__hheap_alloc(hh, sizeof(*z), userdata);
-   float dxdy = (e->x1 - e->x0) / (e->y1 - e->y0);
-   STBTT_assert(z != NULL);
-   //STBTT_assert(e->y0 <= start_point);
-   if (!z) return z;
-   z->fdx = dxdy;
-   z->fdy = dxdy != 0.0f ? (1.0f/dxdy) : 0.0f;
-   z->fx = e->x0 + dxdy * (start_point - e->y0);
-   z->fx -= off_x;
-   z->direction = e->invert ? 1.0f : -1.0f;
-   z->sy = e->y0;
-   z->ey = e->y1;
-   z->next = 0;
-   return z;
-}
-#else
-#error "Unrecognized value of STBTT_RASTERIZER_VERSION"
-#endif
-
-#if STBTT_RASTERIZER_VERSION == 1
-// note: this routine clips fills that extend off the edges... ideally this
-// wouldn't happen, but it could happen if the truetype glyph bounding boxes
-// are wrong, or if the user supplies a too-small bitmap
-static void stbtt__fill_active_edges(unsigned char *scanline, int len, stbtt__active_edge *e, int max_weight)
-{
-   // non-zero winding fill
-   int x0=0, w=0;
-
-   while (e) {
-      if (w == 0) {
-         // if we're currently at zero, we need to record the edge start point
-         x0 = e->x; w += e->direction;
-      } else {
-         int x1 = e->x; w += e->direction;
-         // if we went to zero, we need to draw
-         if (w == 0) {
-            int i = x0 >> STBTT_FIXSHIFT;
-            int j = x1 >> STBTT_FIXSHIFT;
-
-            if (i < len && j >= 0) {
-               if (i == j) {
-                  // x0,x1 are the same pixel, so compute combined coverage
-                  scanline[i] = scanline[i] + (stbtt_uint8) ((x1 - x0) * max_weight >> STBTT_FIXSHIFT);
-               } else {
-                  if (i >= 0) // add antialiasing for x0
-                     scanline[i] = scanline[i] + (stbtt_uint8) (((STBTT_FIX - (x0 & STBTT_FIXMASK)) * max_weight) >> STBTT_FIXSHIFT);
-                  else
-                     i = -1; // clip
-
-                  if (j < len) // add antialiasing for x1
-                     scanline[j] = scanline[j] + (stbtt_uint8) (((x1 & STBTT_FIXMASK) * max_weight) >> STBTT_FIXSHIFT);
-                  else
-                     j = len; // clip
-
-                  for (++i; i < j; ++i) // fill pixels between x0 and x1
-                     scanline[i] = scanline[i] + (stbtt_uint8) max_weight;
-               }
-            }
-         }
-      }
-
-      e = e->next;
-   }
-}
-
-static void stbtt__rasterize_sorted_edges(stbtt__bitmap *result, stbtt__edge *e, int n, int vsubsample, int off_x, int off_y, void *userdata)
-{
-   stbtt__hheap hh = { 0, 0, 0 };
-   stbtt__active_edge *active = NULL;
-   int y,j=0;
-   int max_weight = (255 / vsubsample);  // weight per vertical scanline
-   int s; // vertical subsample index
-   unsigned char scanline_data[512], *scanline;
-
-   if (result->w > 512)
-      scanline = (unsigned char *) STBTT_malloc(result->w, userdata);
-   else
-      scanline = scanline_data;
-
-   y = off_y * vsubsample;
-   e[n].y0 = (off_y + result->h) * (float) vsubsample + 1;
-
-   while (j < result->h) {
-      STBTT_memset(scanline, 0, result->w);
-      for (s=0; s < vsubsample; ++s) {
-         // find center of pixel for this scanline
-         float scan_y = y + 0.5f;
-         stbtt__active_edge **step = &active;
-
-         // update all active edges;
-         // remove all active edges that terminate before the center of this scanline
-         while (*step) {
-            stbtt__active_edge * z = *step;
-            if (z->ey <= scan_y) {
-               *step = z->next; // delete from list
-               STBTT_assert(z->direction);
-               z->direction = 0;
-               stbtt__hheap_free(&hh, z);
-            } else {
-               z->x += z->dx; // advance to position for current scanline
-               step = &((*step)->next); // advance through list
-            }
-         }
-
-         // resort the list if needed
-         for(;;) {
-            int changed=0;
-            step = &active;
-            while (*step && (*step)->next) {
-               if ((*step)->x > (*step)->next->x) {
-                  stbtt__active_edge *t = *step;
-                  stbtt__active_edge *q = t->next;
-
-                  t->next = q->next;
-                  q->next = t;
-                  *step = q;
-                  changed = 1;
-               }
-               step = &(*step)->next;
-            }
-            if (!changed) break;
-         }
-
-         // insert all edges that start before the center of this scanline -- omit ones that also end on this scanline
-         while (e->y0 <= scan_y) {
-            if (e->y1 > scan_y) {
-               stbtt__active_edge *z = stbtt__new_active(&hh, e, off_x, scan_y, userdata);
-               if (z != NULL) {
-                  // find insertion point
-                  if (active == NULL)
-                     active = z;
-                  else if (z->x < active->x) {
-                     // insert at front
-                     z->next = active;
-                     active = z;
-                  } else {
-                     // find thing to insert AFTER
-                     stbtt__active_edge *p = active;
-                     while (p->next && p->next->x < z->x)
-                        p = p->next;
-                     // at this point, p->next->x is NOT < z->x
-                     z->next = p->next;
-                     p->next = z;
-                  }
-               }
-            }
-            ++e;
-         }
-
-         // now process all active edges in XOR fashion
-         if (active)
-            stbtt__fill_active_edges(scanline, result->w, active, max_weight);
-
-         ++y;
-      }
-      STBTT_memcpy(result->pixels + j * result->stride, scanline, result->w);
-      ++j;
-   }
-
-   stbtt__hheap_cleanup(&hh, userdata);
-
-   if (scanline != scanline_data)
-      STBTT_free(scanline, userdata);
-}
-
-#elif STBTT_RASTERIZER_VERSION == 2
-
-// the edge passed in here does not cross the vertical line at x or the vertical line at x+1
-// (i.e. it has already been clipped to those)
-static void stbtt__handle_clipped_edge(float *scanline, int x, stbtt__active_edge *e, float x0, float y0, float x1, float y1)
-{
-   if (y0 == y1) return;
-   STBTT_assert(y0 < y1);
-   STBTT_assert(e->sy <= e->ey);
-   if (y0 > e->ey) return;
-   if (y1 < e->sy) return;
-   if (y0 < e->sy) {
-      x0 += (x1-x0) * (e->sy - y0) / (y1-y0);
-      y0 = e->sy;
-   }
-   if (y1 > e->ey) {
-      x1 += (x1-x0) * (e->ey - y1) / (y1-y0);
-      y1 = e->ey;
-   }
-
-   if (x0 == x)
-      STBTT_assert(x1 <= x+1);
-   else if (x0 == x+1)
-      STBTT_assert(x1 >= x);
-   else if (x0 <= x)
-      STBTT_assert(x1 <= x);
-   else if (x0 >= x+1)
-      STBTT_assert(x1 >= x+1);
-   else
-      STBTT_assert(x1 >= x && x1 <= x+1);
-
-   if (x0 <= x && x1 <= x)
-      scanline[x] += e->direction * (y1-y0);
-   else if (x0 >= x+1 && x1 >= x+1)
-      ;
-   else {
-      STBTT_assert(x0 >= x && x0 <= x+1 && x1 >= x && x1 <= x+1);
-      scanline[x] += e->direction * (y1-y0) * (1-((x0-x)+(x1-x))/2); // coverage = 1 - average x position
-   }
-}
-
-static float stbtt__sized_trapezoid_area(float height, float top_width, float bottom_width)
-{
-   STBTT_assert(top_width >= 0);
-   STBTT_assert(bottom_width >= 0);
-   return (top_width + bottom_width) / 2.0f * height;
-}
-
-static float stbtt__position_trapezoid_area(float height, float tx0, float tx1, float bx0, float bx1)
-{
-   return stbtt__sized_trapezoid_area(height, tx1 - tx0, bx1 - bx0);
-}
-
-static float stbtt__sized_triangle_area(float height, float width)
-{
-   return height * width / 2;
-}
-
-static void stbtt__fill_active_edges_new(float *scanline, float *scanline_fill, int len, stbtt__active_edge *e, float y_top)
-{
-   float y_bottom = y_top+1;
-
-   while (e) {
-      // brute force every pixel
-
-      // compute intersection points with top & bottom
-      STBTT_assert(e->ey >= y_top);
-
-      if (e->fdx == 0) {
-         float x0 = e->fx;
-         if (x0 < len) {
-            if (x0 >= 0) {
-               stbtt__handle_clipped_edge(scanline,(int) x0,e, x0,y_top, x0,y_bottom);
-               stbtt__handle_clipped_edge(scanline_fill-1,(int) x0+1,e, x0,y_top, x0,y_bottom);
-            } else {
-               stbtt__handle_clipped_edge(scanline_fill-1,0,e, x0,y_top, x0,y_bottom);
-            }
-         }
-      } else {
-         float x0 = e->fx;
-         float dx = e->fdx;
-         float xb = x0 + dx;
-         float x_top, x_bottom;
-         float sy0,sy1;
-         float dy = e->fdy;
-         STBTT_assert(e->sy <= y_bottom && e->ey >= y_top);
-
-         // compute endpoints of line segment clipped to this scanline (if the
-         // line segment starts on this scanline. x0 is the intersection of the
-         // line with y_top, but that may be off the line segment.
-         if (e->sy > y_top) {
-            x_top = x0 + dx * (e->sy - y_top);
-            sy0 = e->sy;
-         } else {
-            x_top = x0;
-            sy0 = y_top;
-         }
-         if (e->ey < y_bottom) {
-            x_bottom = x0 + dx * (e->ey - y_top);
-            sy1 = e->ey;
-         } else {
-            x_bottom = xb;
-            sy1 = y_bottom;
-         }
-
-         if (x_top >= 0 && x_bottom >= 0 && x_top < len && x_bottom < len) {
-            // from here on, we don't have to range check x values
-
-            if ((int) x_top == (int) x_bottom) {
-               float height;
-               // simple case, only spans one pixel
-               int x = (int) x_top;
-               height = (sy1 - sy0) * e->direction;
-               STBTT_assert(x >= 0 && x < len);
-               scanline[x]      += stbtt__position_trapezoid_area(height, x_top, x+1.0f, x_bottom, x+1.0f);
-               scanline_fill[x] += height; // everything right of this pixel is filled
-            } else {
-               int x,x1,x2;
-               float y_crossing, y_final, step, sign, area;
-               // covers 2+ pixels
-               if (x_top > x_bottom) {
-                  // flip scanline vertically; signed area is the same
-                  float t;
-                  sy0 = y_bottom - (sy0 - y_top);
-                  sy1 = y_bottom - (sy1 - y_top);
-                  t = sy0, sy0 = sy1, sy1 = t;
-                  t = x_bottom, x_bottom = x_top, x_top = t;
-                  dx = -dx;
-                  dy = -dy;
-                  t = x0, x0 = xb, xb = t;
-               }
-               STBTT_assert(dy >= 0);
-               STBTT_assert(dx >= 0);
-
-               x1 = (int) x_top;
-               x2 = (int) x_bottom;
-               // compute intersection with y axis at x1+1
-               y_crossing = y_top + dy * (x1+1 - x0);
-
-               // compute intersection with y axis at x2
-               y_final = y_top + dy * (x2 - x0);
-
-               //           x1    x_top                            x2    x_bottom
-               //     y_top  +------|-----+------------+------------+--------|---+------------+
-               //            |            |            |            |            |            |
-               //            |            |            |            |            |            |
-               //       sy0  |      Txxxxx|............|............|............|............|
-               // y_crossing |            *xxxxx.......|............|............|............|
-               //            |            |     xxxxx..|............|............|............|
-               //            |            |     /-   xx*xxxx........|............|............|
-               //            |            | dy <       |    xxxxxx..|............|............|
-               //   y_final  |            |     \-     |          xx*xxx.........|............|
-               //       sy1  |            |            |            |   xxxxxB...|............|
-               //            |            |            |            |            |            |
-               //            |            |            |            |            |            |
-               //  y_bottom  +------------+------------+------------+------------+------------+
-               //
-               // goal is to measure the area covered by '.' in each pixel
-
-               // if x2 is right at the right edge of x1, y_crossing can blow up, github #1057
-               // @TODO: maybe test against sy1 rather than y_bottom?
-               if (y_crossing > y_bottom)
-                  y_crossing = y_bottom;
-
-               sign = e->direction;
-
-               // area of the rectangle covered from sy0..y_crossing
-               area = sign * (y_crossing-sy0);
-
-               // area of the triangle (x_top,sy0), (x1+1,sy0), (x1+1,y_crossing)
-               scanline[x1] += stbtt__sized_triangle_area(area, x1+1 - x_top);
-
-               // check if final y_crossing is blown up; no test case for this
-               if (y_final > y_bottom) {
-                  y_final = y_bottom;
-                  dy = (y_final - y_crossing ) / (x2 - (x1+1)); // if denom=0, y_final = y_crossing, so y_final <= y_bottom
-               }
-
-               // in second pixel, area covered by line segment found in first pixel
-               // is always a rectangle 1 wide * the height of that line segment; this
-               // is exactly what the variable 'area' stores. it also gets a contribution
-               // from the line segment within it. the THIRD pixel will get the first
-               // pixel's rectangle contribution, the second pixel's rectangle contribution,
-               // and its own contribution. the 'own contribution' is the same in every pixel except
-               // the leftmost and rightmost, a trapezoid that slides down in each pixel.
-               // the second pixel's contribution to the third pixel will be the
-               // rectangle 1 wide times the height change in the second pixel, which is dy.
-
-               step = sign * dy * 1; // dy is dy/dx, change in y for every 1 change in x,
-               // which multiplied by 1-pixel-width is how much pixel area changes for each step in x
-               // so the area advances by 'step' every time
-
-               for (x = x1+1; x < x2; ++x) {
-                  scanline[x] += area + step/2; // area of trapezoid is 1*step/2
-                  area += step;
-               }
-               STBTT_assert(STBTT_fabs(area) <= 1.01f); // accumulated error from area += step unless we round step down
-               STBTT_assert(sy1 > y_final-0.01f);
-
-               // area covered in the last pixel is the rectangle from all the pixels to the left,
-               // plus the trapezoid filled by the line segment in this pixel all the way to the right edge
-               scanline[x2] += area + sign * stbtt__position_trapezoid_area(sy1-y_final, (float) x2, x2+1.0f, x_bottom, x2+1.0f);
-
-               // the rest of the line is filled based on the total height of the line segment in this pixel
-               scanline_fill[x2] += sign * (sy1-sy0);
-            }
-         } else {
-            // if edge goes outside of box we're drawing, we require
-            // clipping logic. since this does not match the intended use
-            // of this library, we use a different, very slow brute
-            // force implementation
-            // note though that this does happen some of the time because
-            // x_top and x_bottom can be extrapolated at the top & bottom of
-            // the shape and actually lie outside the bounding box
-            int x;
-            for (x=0; x < len; ++x) {
-               // cases:
-               //
-               // there can be up to two intersections with the pixel. any intersection
-               // with left or right edges can be handled by splitting into two (or three)
-               // regions. intersections with top & bottom do not necessitate case-wise logic.
-               //
-               // the old way of doing this found the intersections with the left & right edges,
-               // then used some simple logic to produce up to three segments in sorted order
-               // from top-to-bottom. however, this had a problem: if an x edge was epsilon
-               // across the x border, then the corresponding y position might not be distinct
-               // from the other y segment, and it might ignored as an empty segment. to avoid
-               // that, we need to explicitly produce segments based on x positions.
-
-               // rename variables to clearly-defined pairs
-               float y0 = y_top;
-               float x1 = (float) (x);
-               float x2 = (float) (x+1);
-               float x3 = xb;
-               float y3 = y_bottom;
-
-               // x = e->x + e->dx * (y-y_top)
-               // (y-y_top) = (x - e->x) / e->dx
-               // y = (x - e->x) / e->dx + y_top
-               float y1 = (x - x0) / dx + y_top;
-               float y2 = (x+1 - x0) / dx + y_top;
-
-               if (x0 < x1 && x3 > x2) {         // three segments descending down-right
-                  stbtt__handle_clipped_edge(scanline,x,e, x0,y0, x1,y1);
-                  stbtt__handle_clipped_edge(scanline,x,e, x1,y1, x2,y2);
-                  stbtt__handle_clipped_edge(scanline,x,e, x2,y2, x3,y3);
-               } else if (x3 < x1 && x0 > x2) {  // three segments descending down-left
-                  stbtt__handle_clipped_edge(scanline,x,e, x0,y0, x2,y2);
-                  stbtt__handle_clipped_edge(scanline,x,e, x2,y2, x1,y1);
-                  stbtt__handle_clipped_edge(scanline,x,e, x1,y1, x3,y3);
-               } else if (x0 < x1 && x3 > x1) {  // two segments across x, down-right
-                  stbtt__handle_clipped_edge(scanline,x,e, x0,y0, x1,y1);
-                  stbtt__handle_clipped_edge(scanline,x,e, x1,y1, x3,y3);
-               } else if (x3 < x1 && x0 > x1) {  // two segments across x, down-left
-                  stbtt__handle_clipped_edge(scanline,x,e, x0,y0, x1,y1);
-                  stbtt__handle_clipped_edge(scanline,x,e, x1,y1, x3,y3);
-               } else if (x0 < x2 && x3 > x2) {  // two segments across x+1, down-right
-                  stbtt__handle_clipped_edge(scanline,x,e, x0,y0, x2,y2);
-                  stbtt__handle_clipped_edge(scanline,x,e, x2,y2, x3,y3);
-               } else if (x3 < x2 && x0 > x2) {  // two segments across x+1, down-left
-                  stbtt__handle_clipped_edge(scanline,x,e, x0,y0, x2,y2);
-                  stbtt__handle_clipped_edge(scanline,x,e, x2,y2, x3,y3);
-               } else {  // one segment
-                  stbtt__handle_clipped_edge(scanline,x,e, x0,y0, x3,y3);
-               }
-            }
-         }
-      }
-      e = e->next;
-   }
-}
-
-// directly AA rasterize edges w/o supersampling
-static void stbtt__rasterize_sorted_edges(stbtt__bitmap *result, stbtt__edge *e, int n, int vsubsample, int off_x, int off_y, void *userdata)
-{
-   stbtt__hheap hh = { 0, 0, 0 };
-   stbtt__active_edge *active = NULL;
-   int y,j=0, i;
-   float scanline_data[129], *scanline, *scanline2;
-
-   STBTT__NOTUSED(vsubsample);
-
-   if (result->w > 64)
-      scanline = (float *) STBTT_malloc((result->w*2+1) * sizeof(float), userdata);
-   else
-      scanline = scanline_data;
-
-   scanline2 = scanline + result->w;
-
-   y = off_y;
-   e[n].y0 = (float) (off_y + result->h) + 1;
-
-   while (j < result->h) {
-      // find center of pixel for this scanline
-      float scan_y_top    = y + 0.0f;
-      float scan_y_bottom = y + 1.0f;
-      stbtt__active_edge **step = &active;
-
-      STBTT_memset(scanline , 0, result->w*sizeof(scanline[0]));
-      STBTT_memset(scanline2, 0, (result->w+1)*sizeof(scanline[0]));
-
-      // update all active edges;
-      // remove all active edges that terminate before the top of this scanline
-      while (*step) {
-         stbtt__active_edge * z = *step;
-         if (z->ey <= scan_y_top) {
-            *step = z->next; // delete from list
-            STBTT_assert(z->direction);
-            z->direction = 0;
-            stbtt__hheap_free(&hh, z);
-         } else {
-            step = &((*step)->next); // advance through list
-         }
-      }
-
-      // insert all edges that start before the bottom of this scanline
-      while (e->y0 <= scan_y_bottom) {
-         if (e->y0 != e->y1) {
-            stbtt__active_edge *z = stbtt__new_active(&hh, e, off_x, scan_y_top, userdata);
-            if (z != NULL) {
-               if (j == 0 && off_y != 0) {
-                  if (z->ey < scan_y_top) {
-                     // this can happen due to subpixel positioning and some kind of fp rounding error i think
-                     z->ey = scan_y_top;
-                  }
-               }
-               STBTT_assert(z->ey >= scan_y_top); // if we get really unlucky a tiny bit of an edge can be out of bounds
-               // insert at front
-               z->next = active;
-               active = z;
-            }
-         }
-         ++e;
-      }
-
-      // now process all active edges
-      if (active)
-         stbtt__fill_active_edges_new(scanline, scanline2+1, result->w, active, scan_y_top);
-
-      {
-         float sum = 0;
-         for (i=0; i < result->w; ++i) {
-            float k;
-            int m;
-            sum += scanline2[i];
-            k = scanline[i] + sum;
-            k = (float) STBTT_fabs(k)*255 + 0.5f;
-            m = (int) k;
-            if (m > 255) m = 255;
-            result->pixels[j*result->stride + i] = (unsigned char) m;
-         }
-      }
-      // advance all the edges
-      step = &active;
-      while (*step) {
-         stbtt__active_edge *z = *step;
-         z->fx += z->fdx; // advance to position for current scanline
-         step = &((*step)->next); // advance through list
-      }
-
-      ++y;
-      ++j;
-   }
-
-   stbtt__hheap_cleanup(&hh, userdata);
-
-   if (scanline != scanline_data)
-      STBTT_free(scanline, userdata);
-}
-#else
-#error "Unrecognized value of STBTT_RASTERIZER_VERSION"
-#endif
-
-#define STBTT__COMPARE(a,b)  ((a)->y0 < (b)->y0)
-
-static void stbtt__sort_edges_ins_sort(stbtt__edge *p, int n)
-{
-   int i,j;
-   for (i=1; i < n; ++i) {
-      stbtt__edge t = p[i], *a = &t;
-      j = i;
-      while (j > 0) {
-         stbtt__edge *b = &p[j-1];
-         int c = STBTT__COMPARE(a,b);
-         if (!c) break;
-         p[j] = p[j-1];
-         --j;
-      }
-      if (i != j)
-         p[j] = t;
-   }
-}
-
-static void stbtt__sort_edges_quicksort(stbtt__edge *p, int n)
-{
-   /* threshold for transitioning to insertion sort */
-   while (n > 12) {
-      stbtt__edge t;
-      int c01,c12,c,m,i,j;
-
-      /* compute median of three */
-      m = n >> 1;
-      c01 = STBTT__COMPARE(&p[0],&p[m]);
-      c12 = STBTT__COMPARE(&p[m],&p[n-1]);
-      /* if 0 >= mid >= end, or 0 < mid < end, then use mid */
-      if (c01 != c12) {
-         /* otherwise, we'll need to swap something else to middle */
-         int z;
-         c = STBTT__COMPARE(&p[0],&p[n-1]);
-         /* 0>mid && mid<n:  0>n => n; 0<n => 0 */
-         /* 0<mid && mid>n:  0>n => 0; 0<n => n */
-         z = (c == c12) ? 0 : n-1;
-         t = p[z];
-         p[z] = p[m];
-         p[m] = t;
-      }
-      /* now p[m] is the median-of-three */
-      /* swap it to the beginning so it won't move around */
-      t = p[0];
-      p[0] = p[m];
-      p[m] = t;
-
-      /* partition loop */
-      i=1;
-      j=n-1;
-      for(;;) {
-         /* handling of equality is crucial here */
-         /* for sentinels & efficiency with duplicates */
-         for (;;++i) {
-            if (!STBTT__COMPARE(&p[i], &p[0])) break;
-         }
-         for (;;--j) {
-            if (!STBTT__COMPARE(&p[0], &p[j])) break;
-         }
-         /* make sure we haven't crossed */
-         if (i >= j) break;
-         t = p[i];
-         p[i] = p[j];
-         p[j] = t;
-
-         ++i;
-         --j;
-      }
-      /* recurse on smaller side, iterate on larger */
-      if (j < (n-i)) {
-         stbtt__sort_edges_quicksort(p,j);
-         p = p+i;
-         n = n-i;
-      } else {
-         stbtt__sort_edges_quicksort(p+i, n-i);
-         n = j;
-      }
-   }
-}
-
-static void stbtt__sort_edges(stbtt__edge *p, int n)
-{
-   stbtt__sort_edges_quicksort(p, n);
-   stbtt__sort_edges_ins_sort(p, n);
-}
-
-typedef struct
-{
-   float x,y;
-} stbtt__point;
-
-static void stbtt__rasterize(stbtt__bitmap *result, stbtt__point *pts, int *wcount, int windings, float scale_x, float scale_y, float shift_x, float shift_y, int off_x, int off_y, int invert, void *userdata)
-{
-   float y_scale_inv = invert ? -scale_y : scale_y;
-   stbtt__edge *e;
-   int n,i,j,k,m;
-#if STBTT_RASTERIZER_VERSION == 1
-   int vsubsample = result->h < 8 ? 15 : 5;
-#elif STBTT_RASTERIZER_VERSION == 2
-   int vsubsample = 1;
-#else
-   #error "Unrecognized value of STBTT_RASTERIZER_VERSION"
-#endif
-   // vsubsample should divide 255 evenly; otherwise we won't reach full opacity
-
-   // now we have to blow out the windings into explicit edge lists
-   n = 0;
-   for (i=0; i < windings; ++i)
-      n += wcount[i];
-
-   e = (stbtt__edge *) STBTT_malloc(sizeof(*e) * (n+1), userdata); // add an extra one as a sentinel
-   if (e == 0) return;
-   n = 0;
-
-   m=0;
-   for (i=0; i < windings; ++i) {
-      stbtt__point *p = pts + m;
-      m += wcount[i];
-      j = wcount[i]-1;
-      for (k=0; k < wcount[i]; j=k++) {
-         int a=k,b=j;
-         // skip the edge if horizontal
-         if (p[j].y == p[k].y)
-            continue;
-         // add edge from j to k to the list
-         e[n].invert = 0;
-         if (invert ? p[j].y > p[k].y : p[j].y < p[k].y) {
-            e[n].invert = 1;
-            a=j,b=k;
-         }
-         e[n].x0 = p[a].x * scale_x + shift_x;
-         e[n].y0 = (p[a].y * y_scale_inv + shift_y) * vsubsample;
-         e[n].x1 = p[b].x * scale_x + shift_x;
-         e[n].y1 = (p[b].y * y_scale_inv + shift_y) * vsubsample;
-         ++n;
-      }
-   }
-
-   // now sort the edges by their highest point (should snap to integer, and then by x)
-   //STBTT_sort(e, n, sizeof(e[0]), stbtt__edge_compare);
-   stbtt__sort_edges(e, n);
-
-   // now, traverse the scanlines and find the intersections on each scanline, use xor winding rule
-   stbtt__rasterize_sorted_edges(result, e, n, vsubsample, off_x, off_y, userdata);
-
-   STBTT_free(e, userdata);
-}
-
-static void stbtt__add_point(stbtt__point *points, int n, float x, float y)
-{
-   if (!points) return; // during first pass, it's unallocated
-   points[n].x = x;
-   points[n].y = y;
-}
-
-// tessellate until threshold p is happy... @TODO warped to compensate for non-linear stretching
-static int stbtt__tesselate_curve(stbtt__point *points, int *num_points, float x0, float y0, float x1, float y1, float x2, float y2, float objspace_flatness_squared, int n)
-{
-   // midpoint
-   float mx = (x0 + 2*x1 + x2)/4;
-   float my = (y0 + 2*y1 + y2)/4;
-   // versus directly drawn line
-   float dx = (x0+x2)/2 - mx;
-   float dy = (y0+y2)/2 - my;
-   if (n > 16) // 65536 segments on one curve better be enough!
-      return 1;
-   if (dx*dx+dy*dy > objspace_flatness_squared) { // half-pixel error allowed... need to be smaller if AA
-      stbtt__tesselate_curve(points, num_points, x0,y0, (x0+x1)/2.0f,(y0+y1)/2.0f, mx,my, objspace_flatness_squared,n+1);
-      stbtt__tesselate_curve(points, num_points, mx,my, (x1+x2)/2.0f,(y1+y2)/2.0f, x2,y2, objspace_flatness_squared,n+1);
-   } else {
-      stbtt__add_point(points, *num_points,x2,y2);
-      *num_points = *num_points+1;
-   }
-   return 1;
-}
-
-static void stbtt__tesselate_cubic(stbtt__point *points, int *num_points, float x0, float y0, float x1, float y1, float x2, float y2, float x3, float y3, float objspace_flatness_squared, int n)
-{
-   // @TODO this "flatness" calculation is just made-up nonsense that seems to work well enough
-   float dx0 = x1-x0;
-   float dy0 = y1-y0;
-   float dx1 = x2-x1;
-   float dy1 = y2-y1;
-   float dx2 = x3-x2;
-   float dy2 = y3-y2;
-   float dx = x3-x0;
-   float dy = y3-y0;
-   float longlen = (float) (STBTT_sqrt(dx0*dx0+dy0*dy0)+STBTT_sqrt(dx1*dx1+dy1*dy1)+STBTT_sqrt(dx2*dx2+dy2*dy2));
-   float shortlen = (float) STBTT_sqrt(dx*dx+dy*dy);
-   float flatness_squared = longlen*longlen-shortlen*shortlen;
-
-   if (n > 16) // 65536 segments on one curve better be enough!
-      return;
-
-   if (flatness_squared > objspace_flatness_squared) {
-      float x01 = (x0+x1)/2;
-      float y01 = (y0+y1)/2;
-      float x12 = (x1+x2)/2;
-      float y12 = (y1+y2)/2;
-      float x23 = (x2+x3)/2;
-      float y23 = (y2+y3)/2;
-
-      float xa = (x01+x12)/2;
-      float ya = (y01+y12)/2;
-      float xb = (x12+x23)/2;
-      float yb = (y12+y23)/2;
-
-      float mx = (xa+xb)/2;
-      float my = (ya+yb)/2;
-
-      stbtt__tesselate_cubic(points, num_points, x0,y0, x01,y01, xa,ya, mx,my, objspace_flatness_squared,n+1);
-      stbtt__tesselate_cubic(points, num_points, mx,my, xb,yb, x23,y23, x3,y3, objspace_flatness_squared,n+1);
-   } else {
-      stbtt__add_point(points, *num_points,x3,y3);
-      *num_points = *num_points+1;
-   }
-}
-
-// returns number of contours
-static stbtt__point *stbtt_FlattenCurves(stbtt_vertex *vertices, int num_verts, float objspace_flatness, int **contour_lengths, int *num_contours, void *userdata)
-{
-   stbtt__point *points=0;
-   int num_points=0;
-
-   float objspace_flatness_squared = objspace_flatness * objspace_flatness;
-   int i,n=0,start=0, pass;
-
-   // count how many "moves" there are to get the contour count
-   for (i=0; i < num_verts; ++i)
-      if (vertices[i].type == STBTT_vmove)
-         ++n;
-
-   *num_contours = n;
-   if (n == 0) return 0;
-
-   *contour_lengths = (int *) STBTT_malloc(sizeof(**contour_lengths) * n, userdata);
-
-   if (*contour_lengths == 0) {
-      *num_contours = 0;
-      return 0;
-   }
-
-   // make two passes through the points so we don't need to realloc
-   for (pass=0; pass < 2; ++pass) {
-      float x=0,y=0;
-      if (pass == 1) {
-         points = (stbtt__point *) STBTT_malloc(num_points * sizeof(points[0]), userdata);
-         if (points == NULL) goto error;
-      }
-      num_points = 0;
-      n= -1;
-      for (i=0; i < num_verts; ++i) {
-         switch (vertices[i].type) {
-            case STBTT_vmove:
-               // start the next contour
-               if (n >= 0)
-                  (*contour_lengths)[n] = num_points - start;
-               ++n;
-               start = num_points;
-
-               x = vertices[i].x, y = vertices[i].y;
-               stbtt__add_point(points, num_points++, x,y);
-               break;
-            case STBTT_vline:
-               x = vertices[i].x, y = vertices[i].y;
-               stbtt__add_point(points, num_points++, x, y);
-               break;
-            case STBTT_vcurve:
-               stbtt__tesselate_curve(points, &num_points, x,y,
-                                        vertices[i].cx, vertices[i].cy,
-                                        vertices[i].x,  vertices[i].y,
-                                        objspace_flatness_squared, 0);
-               x = vertices[i].x, y = vertices[i].y;
-               break;
-            case STBTT_vcubic:
-               stbtt__tesselate_cubic(points, &num_points, x,y,
-                                        vertices[i].cx, vertices[i].cy,
-                                        vertices[i].cx1, vertices[i].cy1,
-                                        vertices[i].x,  vertices[i].y,
-                                        objspace_flatness_squared, 0);
-               x = vertices[i].x, y = vertices[i].y;
-               break;
-         }
-      }
-      (*contour_lengths)[n] = num_points - start;
-   }
-
-   return points;
-error:
-   STBTT_free(points, userdata);
-   STBTT_free(*contour_lengths, userdata);
-   *contour_lengths = 0;
-   *num_contours = 0;
-   return NULL;
-}
-
-STBTT_DEF void stbtt_Rasterize(stbtt__bitmap *result, float flatness_in_pixels, stbtt_vertex *vertices, int num_verts, float scale_x, float scale_y, float shift_x, float shift_y, int x_off, int y_off, int invert, void *userdata)
-{
-   float scale            = scale_x > scale_y ? scale_y : scale_x;
-   int winding_count      = 0;
-   int *winding_lengths   = NULL;
-   stbtt__point *windings = stbtt_FlattenCurves(vertices, num_verts, flatness_in_pixels / scale, &winding_lengths, &winding_count, userdata);
-   if (windings) {
-      stbtt__rasterize(result, windings, winding_lengths, winding_count, scale_x, scale_y, shift_x, shift_y, x_off, y_off, invert, userdata);
-      STBTT_free(winding_lengths, userdata);
-      STBTT_free(windings, userdata);
-   }
-}
-
-STBTT_DEF void stbtt_FreeBitmap(unsigned char *bitmap, void *userdata)
-{
-   STBTT_free(bitmap, userdata);
-}
-
-STBTT_DEF unsigned char *stbtt_GetGlyphBitmapSubpixel(const stbtt_fontinfo *info, float scale_x, float scale_y, float shift_x, float shift_y, int glyph, int *width, int *height, int *xoff, int *yoff)
-{
-   int ix0,iy0,ix1,iy1;
-   stbtt__bitmap gbm;
-   stbtt_vertex *vertices;
-   int num_verts = stbtt_GetGlyphShape(info, glyph, &vertices);
-
-   if (scale_x == 0) scale_x = scale_y;
-   if (scale_y == 0) {
-      if (scale_x == 0) {
-         STBTT_free(vertices, info->userdata);
-         return NULL;
-      }
-      scale_y = scale_x;
-   }
-
-   stbtt_GetGlyphBitmapBoxSubpixel(info, glyph, scale_x, scale_y, shift_x, shift_y, &ix0,&iy0,&ix1,&iy1);
-
-   // now we get the size
-   gbm.w = (ix1 - ix0);
-   gbm.h = (iy1 - iy0);
-   gbm.pixels = NULL; // in case we error
-
-   if (width ) *width  = gbm.w;
-   if (height) *height = gbm.h;
-   if (xoff  ) *xoff   = ix0;
-   if (yoff  ) *yoff   = iy0;
-
-   if (gbm.w && gbm.h) {
-      gbm.pixels = (unsigned char *) STBTT_malloc(gbm.w * gbm.h, info->userdata);
-      if (gbm.pixels) {
-         gbm.stride = gbm.w;
-
-         stbtt_Rasterize(&gbm, 0.35f, vertices, num_verts, scale_x, scale_y, shift_x, shift_y, ix0, iy0, 1, info->userdata);
-      }
-   }
-   STBTT_free(vertices, info->userdata);
-   return gbm.pixels;
-}
-
-STBTT_DEF unsigned char *stbtt_GetGlyphBitmap(const stbtt_fontinfo *info, float scale_x, float scale_y, int glyph, int *width, int *height, int *xoff, int *yoff)
-{
-   return stbtt_GetGlyphBitmapSubpixel(info, scale_x, scale_y, 0.0f, 0.0f, glyph, width, height, xoff, yoff);
-}
-
-STBTT_DEF void stbtt_MakeGlyphBitmapSubpixel(const stbtt_fontinfo *info, unsigned char *output, int out_w, int out_h, int out_stride, float scale_x, float scale_y, float shift_x, float shift_y, int glyph)
-{
-   int ix0,iy0;
-   stbtt_vertex *vertices;
-   int num_verts = stbtt_GetGlyphShape(info, glyph, &vertices);
-   stbtt__bitmap gbm;
-
-   stbtt_GetGlyphBitmapBoxSubpixel(info, glyph, scale_x, scale_y, shift_x, shift_y, &ix0,&iy0,0,0);
-   gbm.pixels = output;
-   gbm.w = out_w;
-   gbm.h = out_h;
-   gbm.stride = out_stride;
-
-   if (gbm.w && gbm.h)
-      stbtt_Rasterize(&gbm, 0.35f, vertices, num_verts, scale_x, scale_y, shift_x, shift_y, ix0,iy0, 1, info->userdata);
-
-   STBTT_free(vertices, info->userdata);
-}
-
-STBTT_DEF void stbtt_MakeGlyphBitmap(const stbtt_fontinfo *info, unsigned char *output, int out_w, int out_h, int out_stride, float scale_x, float scale_y, int glyph)
-{
-   stbtt_MakeGlyphBitmapSubpixel(info, output, out_w, out_h, out_stride, scale_x, scale_y, 0.0f,0.0f, glyph);
-}
-
-STBTT_DEF unsigned char *stbtt_GetCodepointBitmapSubpixel(const stbtt_fontinfo *info, float scale_x, float scale_y, float shift_x, float shift_y, int codepoint, int *width, int *height, int *xoff, int *yoff)
-{
-   return stbtt_GetGlyphBitmapSubpixel(info, scale_x, scale_y,shift_x,shift_y, stbtt_FindGlyphIndex(info,codepoint), width,height,xoff,yoff);
-}
-
-STBTT_DEF void stbtt_MakeCodepointBitmapSubpixelPrefilter(const stbtt_fontinfo *info, unsigned char *output, int out_w, int out_h, int out_stride, float scale_x, float scale_y, float shift_x, float shift_y, int oversample_x, int oversample_y, float *sub_x, float *sub_y, int codepoint)
-{
-   stbtt_MakeGlyphBitmapSubpixelPrefilter(info, output, out_w, out_h, out_stride, scale_x, scale_y, shift_x, shift_y, oversample_x, oversample_y, sub_x, sub_y, stbtt_FindGlyphIndex(info,codepoint));
-}
-
-STBTT_DEF void stbtt_MakeCodepointBitmapSubpixel(const stbtt_fontinfo *info, unsigned char *output, int out_w, int out_h, int out_stride, float scale_x, float scale_y, float shift_x, float shift_y, int codepoint)
-{
-   stbtt_MakeGlyphBitmapSubpixel(info, output, out_w, out_h, out_stride, scale_x, scale_y, shift_x, shift_y, stbtt_FindGlyphIndex(info,codepoint));
-}
-
-STBTT_DEF unsigned char *stbtt_GetCodepointBitmap(const stbtt_fontinfo *info, float scale_x, float scale_y, int codepoint, int *width, int *height, int *xoff, int *yoff)
-{
-   return stbtt_GetCodepointBitmapSubpixel(info, scale_x, scale_y, 0.0f,0.0f, codepoint, width,height,xoff,yoff);
-}
-
-STBTT_DEF void stbtt_MakeCodepointBitmap(const stbtt_fontinfo *info, unsigned char *output, int out_w, int out_h, int out_stride, float scale_x, float scale_y, int codepoint)
-{
-   stbtt_MakeCodepointBitmapSubpixel(info, output, out_w, out_h, out_stride, scale_x, scale_y, 0.0f,0.0f, codepoint);
-}
-
-//////////////////////////////////////////////////////////////////////////////
-//
-// bitmap baking
-//
-// This is SUPER-CRAPPY packing to keep source code small
-
-static int stbtt_BakeFontBitmap_internal(unsigned char *data, int offset,  // font location (use offset=0 for plain .ttf)
-                                float pixel_height,                     // height of font in pixels
-                                unsigned char *pixels, int pw, int ph,  // bitmap to be filled in
-                                int first_char, int num_chars,          // characters to bake
-                                stbtt_bakedchar *chardata)
-{
-   float scale;
-   int x,y,bottom_y, i;
-   stbtt_fontinfo f;
-   f.userdata = NULL;
-   if (!stbtt_InitFont(&f, data, offset))
-      return -1;
-   STBTT_memset(pixels, 0, pw*ph); // background of 0 around pixels
-   x=y=1;
-   bottom_y = 1;
-
-   scale = stbtt_ScaleForPixelHeight(&f, pixel_height);
-
-   for (i=0; i < num_chars; ++i) {
-      int advance, lsb, x0,y0,x1,y1,gw,gh;
-      int g = stbtt_FindGlyphIndex(&f, first_char + i);
-      stbtt_GetGlyphHMetrics(&f, g, &advance, &lsb);
-      stbtt_GetGlyphBitmapBox(&f, g, scale,scale, &x0,&y0,&x1,&y1);
-      gw = x1-x0;
-      gh = y1-y0;
-      if (x + gw + 1 >= pw)
-         y = bottom_y, x = 1; // advance to next row
-      if (y + gh + 1 >= ph) // check if it fits vertically AFTER potentially moving to next row
-         return -i;
-      STBTT_assert(x+gw < pw);
-      STBTT_assert(y+gh < ph);
-      stbtt_MakeGlyphBitmap(&f, pixels+x+y*pw, gw,gh,pw, scale,scale, g);
-      chardata[i].x0 = (stbtt_int16) x;
-      chardata[i].y0 = (stbtt_int16) y;
-      chardata[i].x1 = (stbtt_int16) (x + gw);
-      chardata[i].y1 = (stbtt_int16) (y + gh);
-      chardata[i].xadvance = scale * advance;
-      chardata[i].xoff     = (float) x0;
-      chardata[i].yoff     = (float) y0;
-      x = x + gw + 1;
-      if (y+gh+1 > bottom_y)
-         bottom_y = y+gh+1;
-   }
-   return bottom_y;
-}
-
-STBTT_DEF void stbtt_GetBakedQuad(const stbtt_bakedchar *chardata, int pw, int ph, int char_index, float *xpos, float *ypos, stbtt_aligned_quad *q, int opengl_fillrule)
-{
-   float d3d_bias = opengl_fillrule ? 0 : -0.5f;
-   float ipw = 1.0f / pw, iph = 1.0f / ph;
-   const stbtt_bakedchar *b = chardata + char_index;
-   int round_x = STBTT_ifloor((*xpos + b->xoff) + 0.5f);
-   int round_y = STBTT_ifloor((*ypos + b->yoff) + 0.5f);
-
-   q->x0 = round_x + d3d_bias;
-   q->y0 = round_y + d3d_bias;
-   q->x1 = round_x + b->x1 - b->x0 + d3d_bias;
-   q->y1 = round_y + b->y1 - b->y0 + d3d_bias;
-
-   q->s0 = b->x0 * ipw;
-   q->t0 = b->y0 * iph;
-   q->s1 = b->x1 * ipw;
-   q->t1 = b->y1 * iph;
-
-   *xpos += b->xadvance;
-}
-
-//////////////////////////////////////////////////////////////////////////////
-//
-// rectangle packing replacement routines if you don't have stb_rect_pack.h
-//
-
-#ifndef STB_RECT_PACK_VERSION
-
-typedef int stbrp_coord;
-
-////////////////////////////////////////////////////////////////////////////////////
-//                                                                                //
-//                                                                                //
-// COMPILER WARNING ?!?!?                                                         //
-//                                                                                //
-//                                                                                //
-// if you get a compile warning due to these symbols being defined more than      //
-// once, move #include "stb_rect_pack.h" before #include "stb_truetype.h"         //
-//                                                                                //
-////////////////////////////////////////////////////////////////////////////////////
-
-typedef struct
-{
-   int width,height;
-   int x,y,bottom_y;
-} stbrp_context;
-
-typedef struct
-{
-   unsigned char x;
-} stbrp_node;
-
-struct stbrp_rect
-{
-   stbrp_coord x,y;
-   int id,w,h,was_packed;
-};
-
-static void stbrp_init_target(stbrp_context *con, int pw, int ph, stbrp_node *nodes, int num_nodes)
-{
-   con->width  = pw;
-   con->height = ph;
-   con->x = 0;
-   con->y = 0;
-   con->bottom_y = 0;
-   STBTT__NOTUSED(nodes);
-   STBTT__NOTUSED(num_nodes);
-}
-
-static void stbrp_pack_rects(stbrp_context *con, stbrp_rect *rects, int num_rects)
-{
-   int i;
-   for (i=0; i < num_rects; ++i) {
-      if (con->x + rects[i].w > con->width) {
-         con->x = 0;
-         con->y = con->bottom_y;
-      }
-      if (con->y + rects[i].h > con->height)
-         break;
-      rects[i].x = con->x;
-      rects[i].y = con->y;
-      rects[i].was_packed = 1;
-      con->x += rects[i].w;
-      if (con->y + rects[i].h > con->bottom_y)
-         con->bottom_y = con->y + rects[i].h;
-   }
-   for (   ; i < num_rects; ++i)
-      rects[i].was_packed = 0;
-}
-#endif
-
-//////////////////////////////////////////////////////////////////////////////
-//
-// bitmap baking
-//
-// This is SUPER-AWESOME (tm Ryan Gordon) packing using stb_rect_pack.h. If
-// stb_rect_pack.h isn't available, it uses the BakeFontBitmap strategy.
-
-STBTT_DEF int stbtt_PackBegin(stbtt_pack_context *spc, unsigned char *pixels, int pw, int ph, int stride_in_bytes, int padding, void *alloc_context)
-{
-   stbrp_context *context = (stbrp_context *) STBTT_malloc(sizeof(*context)            ,alloc_context);
-   int            num_nodes = pw - padding;
-   stbrp_node    *nodes   = (stbrp_node    *) STBTT_malloc(sizeof(*nodes  ) * num_nodes,alloc_context);
-
-   if (context == NULL || nodes == NULL) {
-      if (context != NULL) STBTT_free(context, alloc_context);
-      if (nodes   != NULL) STBTT_free(nodes  , alloc_context);
-      return 0;
-   }
-
-   spc->user_allocator_context = alloc_context;
-   spc->width = pw;
-   spc->height = ph;
-   spc->pixels = pixels;
-   spc->pack_info = context;
-   spc->nodes = nodes;
-   spc->padding = padding;
-   spc->stride_in_bytes = stride_in_bytes != 0 ? stride_in_bytes : pw;
-   spc->h_oversample = 1;
-   spc->v_oversample = 1;
-   spc->skip_missing = 0;
-
-   stbrp_init_target(context, pw-padding, ph-padding, nodes, num_nodes);
-
-   if (pixels)
-      STBTT_memset(pixels, 0, pw*ph); // background of 0 around pixels
-
-   return 1;
-}
-
-STBTT_DEF void stbtt_PackEnd  (stbtt_pack_context *spc)
-{
-   STBTT_free(spc->nodes    , spc->user_allocator_context);
-   STBTT_free(spc->pack_info, spc->user_allocator_context);
-}
-
-STBTT_DEF void stbtt_PackSetOversampling(stbtt_pack_context *spc, unsigned int h_oversample, unsigned int v_oversample)
-{
-   STBTT_assert(h_oversample <= STBTT_MAX_OVERSAMPLE);
-   STBTT_assert(v_oversample <= STBTT_MAX_OVERSAMPLE);
-   if (h_oversample <= STBTT_MAX_OVERSAMPLE)
-      spc->h_oversample = h_oversample;
-   if (v_oversample <= STBTT_MAX_OVERSAMPLE)
-      spc->v_oversample = v_oversample;
-}
-
-STBTT_DEF void stbtt_PackSetSkipMissingCodepoints(stbtt_pack_context *spc, int skip)
-{
-   spc->skip_missing = skip;
-}
-
-#define STBTT__OVER_MASK  (STBTT_MAX_OVERSAMPLE-1)
-
-static void stbtt__h_prefilter(unsigned char *pixels, int w, int h, int stride_in_bytes, unsigned int kernel_width)
-{
-   unsigned char buffer[STBTT_MAX_OVERSAMPLE];
-   int safe_w = w - kernel_width;
-   int j;
-   STBTT_memset(buffer, 0, STBTT_MAX_OVERSAMPLE); // suppress bogus warning from VS2013 -analyze
-   for (j=0; j < h; ++j) {
-      int i;
-      unsigned int total;
-      STBTT_memset(buffer, 0, kernel_width);
-
-      total = 0;
-
-      // make kernel_width a constant in common cases so compiler can optimize out the divide
-      switch (kernel_width) {
-         case 2:
-            for (i=0; i <= safe_w; ++i) {
-               total += pixels[i] - buffer[i & STBTT__OVER_MASK];
-               buffer[(i+kernel_width) & STBTT__OVER_MASK] = pixels[i];
-               pixels[i] = (unsigned char) (total / 2);
-            }
-            break;
-         case 3:
-            for (i=0; i <= safe_w; ++i) {
-               total += pixels[i] - buffer[i & STBTT__OVER_MASK];
-               buffer[(i+kernel_width) & STBTT__OVER_MASK] = pixels[i];
-               pixels[i] = (unsigned char) (total / 3);
-            }
-            break;
-         case 4:
-            for (i=0; i <= safe_w; ++i) {
-               total += pixels[i] - buffer[i & STBTT__OVER_MASK];
-               buffer[(i+kernel_width) & STBTT__OVER_MASK] = pixels[i];
-               pixels[i] = (unsigned char) (total / 4);
-            }
-            break;
-         case 5:
-            for (i=0; i <= safe_w; ++i) {
-               total += pixels[i] - buffer[i & STBTT__OVER_MASK];
-               buffer[(i+kernel_width) & STBTT__OVER_MASK] = pixels[i];
-               pixels[i] = (unsigned char) (total / 5);
-            }
-            break;
-         default:
-            for (i=0; i <= safe_w; ++i) {
-               total += pixels[i] - buffer[i & STBTT__OVER_MASK];
-               buffer[(i+kernel_width) & STBTT__OVER_MASK] = pixels[i];
-               pixels[i] = (unsigned char) (total / kernel_width);
-            }
-            break;
-      }
-
-      for (; i < w; ++i) {
-         STBTT_assert(pixels[i] == 0);
-         total -= buffer[i & STBTT__OVER_MASK];
-         pixels[i] = (unsigned char) (total / kernel_width);
-      }
-
-      pixels += stride_in_bytes;
-   }
-}
-
-static void stbtt__v_prefilter(unsigned char *pixels, int w, int h, int stride_in_bytes, unsigned int kernel_width)
-{
-   unsigned char buffer[STBTT_MAX_OVERSAMPLE];
-   int safe_h = h - kernel_width;
-   int j;
-   STBTT_memset(buffer, 0, STBTT_MAX_OVERSAMPLE); // suppress bogus warning from VS2013 -analyze
-   for (j=0; j < w; ++j) {
-      int i;
-      unsigned int total;
-      STBTT_memset(buffer, 0, kernel_width);
-
-      total = 0;
-
-      // make kernel_width a constant in common cases so compiler can optimize out the divide
-      switch (kernel_width) {
-         case 2:
-            for (i=0; i <= safe_h; ++i) {
-               total += pixels[i*stride_in_bytes] - buffer[i & STBTT__OVER_MASK];
-               buffer[(i+kernel_width) & STBTT__OVER_MASK] = pixels[i*stride_in_bytes];
-               pixels[i*stride_in_bytes] = (unsigned char) (total / 2);
-            }
-            break;
-         case 3:
-            for (i=0; i <= safe_h; ++i) {
-               total += pixels[i*stride_in_bytes] - buffer[i & STBTT__OVER_MASK];
-               buffer[(i+kernel_width) & STBTT__OVER_MASK] = pixels[i*stride_in_bytes];
-               pixels[i*stride_in_bytes] = (unsigned char) (total / 3);
-            }
-            break;
-         case 4:
-            for (i=0; i <= safe_h; ++i) {
-               total += pixels[i*stride_in_bytes] - buffer[i & STBTT__OVER_MASK];
-               buffer[(i+kernel_width) & STBTT__OVER_MASK] = pixels[i*stride_in_bytes];
-               pixels[i*stride_in_bytes] = (unsigned char) (total / 4);
-            }
-            break;
-         case 5:
-            for (i=0; i <= safe_h; ++i) {
-               total += pixels[i*stride_in_bytes] - buffer[i & STBTT__OVER_MASK];
-               buffer[(i+kernel_width) & STBTT__OVER_MASK] = pixels[i*stride_in_bytes];
-               pixels[i*stride_in_bytes] = (unsigned char) (total / 5);
-            }
-            break;
-         default:
-            for (i=0; i <= safe_h; ++i) {
-               total += pixels[i*stride_in_bytes] - buffer[i & STBTT__OVER_MASK];
-               buffer[(i+kernel_width) & STBTT__OVER_MASK] = pixels[i*stride_in_bytes];
-               pixels[i*stride_in_bytes] = (unsigned char) (total / kernel_width);
-            }
-            break;
-      }
-
-      for (; i < h; ++i) {
-         STBTT_assert(pixels[i*stride_in_bytes] == 0);
-         total -= buffer[i & STBTT__OVER_MASK];
-         pixels[i*stride_in_bytes] = (unsigned char) (total / kernel_width);
-      }
-
-      pixels += 1;
-   }
-}
-
-static float stbtt__oversample_shift(int oversample)
-{
-   if (!oversample)
-      return 0.0f;
-
-   // The prefilter is a box filter of width "oversample",
-   // which shifts phase by (oversample - 1)/2 pixels in
-   // oversampled space. We want to shift in the opposite
-   // direction to counter this.
-   return (float)-(oversample - 1) / (2.0f * (float)oversample);
-}
-
-// rects array must be big enough to accommodate all characters in the given ranges
-STBTT_DEF int stbtt_PackFontRangesGatherRects(stbtt_pack_context *spc, const stbtt_fontinfo *info, stbtt_pack_range *ranges, int num_ranges, stbrp_rect *rects)
-{
-   int i,j,k;
-   int missing_glyph_added = 0;
-
-   k=0;
-   for (i=0; i < num_ranges; ++i) {
-      float fh = ranges[i].font_size;
-      float scale = fh > 0 ? stbtt_ScaleForPixelHeight(info, fh) : stbtt_ScaleForMappingEmToPixels(info, -fh);
-      ranges[i].h_oversample = (unsigned char) spc->h_oversample;
-      ranges[i].v_oversample = (unsigned char) spc->v_oversample;
-      for (j=0; j < ranges[i].num_chars; ++j) {
-         int x0,y0,x1,y1;
-         int codepoint = ranges[i].array_of_unicode_codepoints == NULL ? ranges[i].first_unicode_codepoint_in_range + j : ranges[i].array_of_unicode_codepoints[j];
-         int glyph = stbtt_FindGlyphIndex(info, codepoint);
-         if (glyph == 0 && (spc->skip_missing || missing_glyph_added)) {
-            rects[k].w = rects[k].h = 0;
-         } else {
-            stbtt_GetGlyphBitmapBoxSubpixel(info,glyph,
-                                            scale * spc->h_oversample,
-                                            scale * spc->v_oversample,
-                                            0,0,
-                                            &x0,&y0,&x1,&y1);
-            rects[k].w = (stbrp_coord) (x1-x0 + spc->padding + spc->h_oversample-1);
-            rects[k].h = (stbrp_coord) (y1-y0 + spc->padding + spc->v_oversample-1);
-            if (glyph == 0)
-               missing_glyph_added = 1;
-         }
-         ++k;
-      }
-   }
-
-   return k;
-}
-
-STBTT_DEF void stbtt_MakeGlyphBitmapSubpixelPrefilter(const stbtt_fontinfo *info, unsigned char *output, int out_w, int out_h, int out_stride, float scale_x, float scale_y, float shift_x, float shift_y, int prefilter_x, int prefilter_y, float *sub_x, float *sub_y, int glyph)
-{
-   stbtt_MakeGlyphBitmapSubpixel(info,
-                                 output,
-                                 out_w - (prefilter_x - 1),
-                                 out_h - (prefilter_y - 1),
-                                 out_stride,
-                                 scale_x,
-                                 scale_y,
-                                 shift_x,
-                                 shift_y,
-                                 glyph);
-
-   if (prefilter_x > 1)
-      stbtt__h_prefilter(output, out_w, out_h, out_stride, prefilter_x);
-
-   if (prefilter_y > 1)
-      stbtt__v_prefilter(output, out_w, out_h, out_stride, prefilter_y);
-
-   *sub_x = stbtt__oversample_shift(prefilter_x);
-   *sub_y = stbtt__oversample_shift(prefilter_y);
-}
-
-// rects array must be big enough to accommodate all characters in the given ranges
-STBTT_DEF int stbtt_PackFontRangesRenderIntoRects(stbtt_pack_context *spc, const stbtt_fontinfo *info, stbtt_pack_range *ranges, int num_ranges, stbrp_rect *rects)
-{
-   int i,j,k, missing_glyph = -1, return_value = 1;
-
-   // save current values
-   int old_h_over = spc->h_oversample;
-   int old_v_over = spc->v_oversample;
-
-   k = 0;
-   for (i=0; i < num_ranges; ++i) {
-      float fh = ranges[i].font_size;
-      float scale = fh > 0 ? stbtt_ScaleForPixelHeight(info, fh) : stbtt_ScaleForMappingEmToPixels(info, -fh);
-      float recip_h,recip_v,sub_x,sub_y;
-      spc->h_oversample = ranges[i].h_oversample;
-      spc->v_oversample = ranges[i].v_oversample;
-      recip_h = 1.0f / spc->h_oversample;
-      recip_v = 1.0f / spc->v_oversample;
-      sub_x = stbtt__oversample_shift(spc->h_oversample);
-      sub_y = stbtt__oversample_shift(spc->v_oversample);
-      for (j=0; j < ranges[i].num_chars; ++j) {
-         stbrp_rect *r = &rects[k];
-         if (r->was_packed && r->w != 0 && r->h != 0) {
-            stbtt_packedchar *bc = &ranges[i].chardata_for_range[j];
-            int advance, lsb, x0,y0,x1,y1;
-            int codepoint = ranges[i].array_of_unicode_codepoints == NULL ? ranges[i].first_unicode_codepoint_in_range + j : ranges[i].array_of_unicode_codepoints[j];
-            int glyph = stbtt_FindGlyphIndex(info, codepoint);
-            stbrp_coord pad = (stbrp_coord) spc->padding;
-
-            // pad on left and top
-            r->x += pad;
-            r->y += pad;
-            r->w -= pad;
-            r->h -= pad;
-            stbtt_GetGlyphHMetrics(info, glyph, &advance, &lsb);
-            stbtt_GetGlyphBitmapBox(info, glyph,
-                                    scale * spc->h_oversample,
-                                    scale * spc->v_oversample,
-                                    &x0,&y0,&x1,&y1);
-            stbtt_MakeGlyphBitmapSubpixel(info,
-                                          spc->pixels + r->x + r->y*spc->stride_in_bytes,
-                                          r->w - spc->h_oversample+1,
-                                          r->h - spc->v_oversample+1,
-                                          spc->stride_in_bytes,
-                                          scale * spc->h_oversample,
-                                          scale * spc->v_oversample,
-                                          0,0,
-                                          glyph);
-
-            if (spc->h_oversample > 1)
-               stbtt__h_prefilter(spc->pixels + r->x + r->y*spc->stride_in_bytes,
-                                  r->w, r->h, spc->stride_in_bytes,
-                                  spc->h_oversample);
-
-            if (spc->v_oversample > 1)
-               stbtt__v_prefilter(spc->pixels + r->x + r->y*spc->stride_in_bytes,
-                                  r->w, r->h, spc->stride_in_bytes,
-                                  spc->v_oversample);
-
-            bc->x0       = (stbtt_int16)  r->x;
-            bc->y0       = (stbtt_int16)  r->y;
-            bc->x1       = (stbtt_int16) (r->x + r->w);
-            bc->y1       = (stbtt_int16) (r->y + r->h);
-            bc->xadvance =                scale * advance;
-            bc->xoff     =       (float)  x0 * recip_h + sub_x;
-            bc->yoff     =       (float)  y0 * recip_v + sub_y;
-            bc->xoff2    =                (x0 + r->w) * recip_h + sub_x;
-            bc->yoff2    =                (y0 + r->h) * recip_v + sub_y;
-
-            if (glyph == 0)
-               missing_glyph = j;
-         } else if (spc->skip_missing) {
-            return_value = 0;
-         } else if (r->was_packed && r->w == 0 && r->h == 0 && missing_glyph >= 0) {
-            ranges[i].chardata_for_range[j] = ranges[i].chardata_for_range[missing_glyph];
-         } else {
-            return_value = 0; // if any fail, report failure
-         }
-
-         ++k;
-      }
-   }
-
-   // restore original values
-   spc->h_oversample = old_h_over;
-   spc->v_oversample = old_v_over;
-
-   return return_value;
-}
-
-STBTT_DEF void stbtt_PackFontRangesPackRects(stbtt_pack_context *spc, stbrp_rect *rects, int num_rects)
-{
-   stbrp_pack_rects((stbrp_context *) spc->pack_info, rects, num_rects);
-}
-
-STBTT_DEF int stbtt_PackFontRanges(stbtt_pack_context *spc, const unsigned char *fontdata, int font_index, stbtt_pack_range *ranges, int num_ranges)
-{
-   stbtt_fontinfo info;
-   int i,j,n, return_value = 1;
-   //stbrp_context *context = (stbrp_context *) spc->pack_info;
-   stbrp_rect    *rects;
-
-   // flag all characters as NOT packed
-   for (i=0; i < num_ranges; ++i)
-      for (j=0; j < ranges[i].num_chars; ++j)
-         ranges[i].chardata_for_range[j].x0 =
-         ranges[i].chardata_for_range[j].y0 =
-         ranges[i].chardata_for_range[j].x1 =
-         ranges[i].chardata_for_range[j].y1 = 0;
-
-   n = 0;
-   for (i=0; i < num_ranges; ++i)
-      n += ranges[i].num_chars;
-
-   rects = (stbrp_rect *) STBTT_malloc(sizeof(*rects) * n, spc->user_allocator_context);
-   if (rects == NULL)
-      return 0;
-
-   info.userdata = spc->user_allocator_context;
-   stbtt_InitFont(&info, fontdata, stbtt_GetFontOffsetForIndex(fontdata,font_index));
-
-   n = stbtt_PackFontRangesGatherRects(spc, &info, ranges, num_ranges, rects);
-
-   stbtt_PackFontRangesPackRects(spc, rects, n);
-
-   return_value = stbtt_PackFontRangesRenderIntoRects(spc, &info, ranges, num_ranges, rects);
-
-   STBTT_free(rects, spc->user_allocator_context);
-   return return_value;
-}
-
-STBTT_DEF int stbtt_PackFontRange(stbtt_pack_context *spc, const unsigned char *fontdata, int font_index, float font_size,
-            int first_unicode_codepoint_in_range, int num_chars_in_range, stbtt_packedchar *chardata_for_range)
-{
-   stbtt_pack_range range;
-   range.first_unicode_codepoint_in_range = first_unicode_codepoint_in_range;
-   range.array_of_unicode_codepoints = NULL;
-   range.num_chars                   = num_chars_in_range;
-   range.chardata_for_range          = chardata_for_range;
-   range.font_size                   = font_size;
-   return stbtt_PackFontRanges(spc, fontdata, font_index, &range, 1);
-}
-
-STBTT_DEF void stbtt_GetScaledFontVMetrics(const unsigned char *fontdata, int index, float size, float *ascent, float *descent, float *lineGap)
-{
-   int i_ascent, i_descent, i_lineGap;
-   float scale;
-   stbtt_fontinfo info;
-   stbtt_InitFont(&info, fontdata, stbtt_GetFontOffsetForIndex(fontdata, index));
-   scale = size > 0 ? stbtt_ScaleForPixelHeight(&info, size) : stbtt_ScaleForMappingEmToPixels(&info, -size);
-   stbtt_GetFontVMetrics(&info, &i_ascent, &i_descent, &i_lineGap);
-   *ascent  = (float) i_ascent  * scale;
-   *descent = (float) i_descent * scale;
-   *lineGap = (float) i_lineGap * scale;
-}
-
-STBTT_DEF void stbtt_GetPackedQuad(const stbtt_packedchar *chardata, int pw, int ph, int char_index, float *xpos, float *ypos, stbtt_aligned_quad *q, int align_to_integer)
-{
-   float ipw = 1.0f / pw, iph = 1.0f / ph;
-   const stbtt_packedchar *b = chardata + char_index;
-
-   if (align_to_integer) {
-      float x = (float) STBTT_ifloor((*xpos + b->xoff) + 0.5f);
-      float y = (float) STBTT_ifloor((*ypos + b->yoff) + 0.5f);
-      q->x0 = x;
-      q->y0 = y;
-      q->x1 = x + b->xoff2 - b->xoff;
-      q->y1 = y + b->yoff2 - b->yoff;
-   } else {
-      q->x0 = *xpos + b->xoff;
-      q->y0 = *ypos + b->yoff;
-      q->x1 = *xpos + b->xoff2;
-      q->y1 = *ypos + b->yoff2;
-   }
-
-   q->s0 = b->x0 * ipw;
-   q->t0 = b->y0 * iph;
-   q->s1 = b->x1 * ipw;
-   q->t1 = b->y1 * iph;
-
-   *xpos += b->xadvance;
-}
-
-//////////////////////////////////////////////////////////////////////////////
-//
-// sdf computation
-//
-
-#define STBTT_min(a,b)  ((a) < (b) ? (a) : (b))
-#define STBTT_max(a,b)  ((a) < (b) ? (b) : (a))
-
-static int stbtt__ray_intersect_bezier(float orig[2], float ray[2], float q0[2], float q1[2], float q2[2], float hits[2][2])
-{
-   float q0perp = q0[1]*ray[0] - q0[0]*ray[1];
-   float q1perp = q1[1]*ray[0] - q1[0]*ray[1];
-   float q2perp = q2[1]*ray[0] - q2[0]*ray[1];
-   float roperp = orig[1]*ray[0] - orig[0]*ray[1];
-
-   float a = q0perp - 2*q1perp + q2perp;
-   float b = q1perp - q0perp;
-   float c = q0perp - roperp;
-
-   float s0 = 0., s1 = 0.;
-   int num_s = 0;
-
-   if (a != 0.0) {
-      float discr = b*b - a*c;
-      if (discr > 0.0) {
-         float rcpna = -1 / a;
-         float d = (float) STBTT_sqrt(discr);
-         s0 = (b+d) * rcpna;
-         s1 = (b-d) * rcpna;
-         if (s0 >= 0.0 && s0 <= 1.0)
-            num_s = 1;
-         if (d > 0.0 && s1 >= 0.0 && s1 <= 1.0) {
-            if (num_s == 0) s0 = s1;
-            ++num_s;
-         }
-      }
-   } else {
-      // 2*b*s + c = 0
-      // s = -c / (2*b)
-      s0 = c / (-2 * b);
-      if (s0 >= 0.0 && s0 <= 1.0)
-         num_s = 1;
-   }
-
-   if (num_s == 0)
-      return 0;
-   else {
-      float rcp_len2 = 1 / (ray[0]*ray[0] + ray[1]*ray[1]);
-      float rayn_x = ray[0] * rcp_len2, rayn_y = ray[1] * rcp_len2;
-
-      float q0d =   q0[0]*rayn_x +   q0[1]*rayn_y;
-      float q1d =   q1[0]*rayn_x +   q1[1]*rayn_y;
-      float q2d =   q2[0]*rayn_x +   q2[1]*rayn_y;
-      float rod = orig[0]*rayn_x + orig[1]*rayn_y;
-
-      float q10d = q1d - q0d;
-      float q20d = q2d - q0d;
-      float q0rd = q0d - rod;
-
-      hits[0][0] = q0rd + s0*(2.0f - 2.0f*s0)*q10d + s0*s0*q20d;
-      hits[0][1] = a*s0+b;
-
-      if (num_s > 1) {
-         hits[1][0] = q0rd + s1*(2.0f - 2.0f*s1)*q10d + s1*s1*q20d;
-         hits[1][1] = a*s1+b;
-         return 2;
-      } else {
-         return 1;
-      }
-   }
-}
-
-static int equal(float *a, float *b)
-{
-   return (a[0] == b[0] && a[1] == b[1]);
-}
-
-static int stbtt__compute_crossings_x(float x, float y, int nverts, stbtt_vertex *verts)
-{
-   int i;
-   float orig[2], ray[2] = { 1, 0 };
-   float y_frac;
-   int winding = 0;
-
-   // make sure y never passes through a vertex of the shape
-   y_frac = (float) STBTT_fmod(y, 1.0f);
-   if (y_frac < 0.01f)
-      y += 0.01f;
-   else if (y_frac > 0.99f)
-      y -= 0.01f;
-
-   orig[0] = x;
-   orig[1] = y;
-
-   // test a ray from (-infinity,y) to (x,y)
-   for (i=0; i < nverts; ++i) {
-      if (verts[i].type == STBTT_vline) {
-         int x0 = (int) verts[i-1].x, y0 = (int) verts[i-1].y;
-         int x1 = (int) verts[i  ].x, y1 = (int) verts[i  ].y;
-         if (y > STBTT_min(y0,y1) && y < STBTT_max(y0,y1) && x > STBTT_min(x0,x1)) {
-            float x_inter = (y - y0) / (y1 - y0) * (x1-x0) + x0;
-            if (x_inter < x)
-               winding += (y0 < y1) ? 1 : -1;
-         }
-      }
-      if (verts[i].type == STBTT_vcurve) {
-         int x0 = (int) verts[i-1].x , y0 = (int) verts[i-1].y ;
-         int x1 = (int) verts[i  ].cx, y1 = (int) verts[i  ].cy;
-         int x2 = (int) verts[i  ].x , y2 = (int) verts[i  ].y ;
-         int ax = STBTT_min(x0,STBTT_min(x1,x2)), ay = STBTT_min(y0,STBTT_min(y1,y2));
-         int by = STBTT_max(y0,STBTT_max(y1,y2));
-         if (y > ay && y < by && x > ax) {
-            float q0[2],q1[2],q2[2];
-            float hits[2][2];
-            q0[0] = (float)x0;
-            q0[1] = (float)y0;
-            q1[0] = (float)x1;
-            q1[1] = (float)y1;
-            q2[0] = (float)x2;
-            q2[1] = (float)y2;
-            if (equal(q0,q1) || equal(q1,q2)) {
-               x0 = (int)verts[i-1].x;
-               y0 = (int)verts[i-1].y;
-               x1 = (int)verts[i  ].x;
-               y1 = (int)verts[i  ].y;
-               if (y > STBTT_min(y0,y1) && y < STBTT_max(y0,y1) && x > STBTT_min(x0,x1)) {
-                  float x_inter = (y - y0) / (y1 - y0) * (x1-x0) + x0;
-                  if (x_inter < x)
-                     winding += (y0 < y1) ? 1 : -1;
-               }
-            } else {
-               int num_hits = stbtt__ray_intersect_bezier(orig, ray, q0, q1, q2, hits);
-               if (num_hits >= 1)
-                  if (hits[0][0] < 0)
-                     winding += (hits[0][1] < 0 ? -1 : 1);
-               if (num_hits >= 2)
-                  if (hits[1][0] < 0)
-                     winding += (hits[1][1] < 0 ? -1 : 1);
-            }
-         }
-      }
-   }
-   return winding;
-}
-
-static float stbtt__cuberoot( float x )
-{
-   if (x<0)
-      return -(float) STBTT_pow(-x,1.0f/3.0f);
-   else
-      return  (float) STBTT_pow( x,1.0f/3.0f);
-}
-
-// x^3 + a*x^2 + b*x + c = 0
-static int stbtt__solve_cubic(float a, float b, float c, float* r)
-{
-   float s = -a / 3;
-   float p = b - a*a / 3;
-   float q = a * (2*a*a - 9*b) / 27 + c;
-   float p3 = p*p*p;
-   float d = q*q + 4*p3 / 27;
-   if (d >= 0) {
-      float z = (float) STBTT_sqrt(d);
-      float u = (-q + z) / 2;
-      float v = (-q - z) / 2;
-      u = stbtt__cuberoot(u);
-      v = stbtt__cuberoot(v);
-      r[0] = s + u + v;
-      return 1;
-   } else {
-      float u = (float) STBTT_sqrt(-p/3);
-      float v = (float) STBTT_acos(-STBTT_sqrt(-27/p3) * q / 2) / 3; // p3 must be negative, since d is negative
-      float m = (float) STBTT_cos(v);
-      float n = (float) STBTT_cos(v-3.141592/2)*1.732050808f;
-      r[0] = s + u * 2 * m;
-      r[1] = s - u * (m + n);
-      r[2] = s - u * (m - n);
-
-      //STBTT_assert( STBTT_fabs(((r[0]+a)*r[0]+b)*r[0]+c) < 0.05f);  // these asserts may not be safe at all scales, though they're in bezier t parameter units so maybe?
-      //STBTT_assert( STBTT_fabs(((r[1]+a)*r[1]+b)*r[1]+c) < 0.05f);
-      //STBTT_assert( STBTT_fabs(((r[2]+a)*r[2]+b)*r[2]+c) < 0.05f);
-      return 3;
-   }
-}
-
-STBTT_DEF unsigned char * stbtt_GetGlyphSDF(const stbtt_fontinfo *info, float scale, int glyph, int padding, unsigned char onedge_value, float pixel_dist_scale, int *width, int *height, int *xoff, int *yoff)
-{
-   float scale_x = scale, scale_y = scale;
-   int ix0,iy0,ix1,iy1;
-   int w,h;
-   unsigned char *data;
-
-   if (scale == 0) return NULL;
-
-   stbtt_GetGlyphBitmapBoxSubpixel(info, glyph, scale, scale, 0.0f,0.0f, &ix0,&iy0,&ix1,&iy1);
-
-   // if empty, return NULL
-   if (ix0 == ix1 || iy0 == iy1)
-      return NULL;
-
-   ix0 -= padding;
-   iy0 -= padding;
-   ix1 += padding;
-   iy1 += padding;
-
-   w = (ix1 - ix0);
-   h = (iy1 - iy0);
-
-   if (width ) *width  = w;
-   if (height) *height = h;
-   if (xoff  ) *xoff   = ix0;
-   if (yoff  ) *yoff   = iy0;
-
-   // invert for y-downwards bitmaps
-   scale_y = -scale_y;
-
-   {
-      int x,y,i,j;
-      float *precompute;
-      stbtt_vertex *verts;
-      int num_verts = stbtt_GetGlyphShape(info, glyph, &verts);
-      data = (unsigned char *) STBTT_malloc(w * h, info->userdata);
-      precompute = (float *) STBTT_malloc(num_verts * sizeof(float), info->userdata);
-
-      for (i=0,j=num_verts-1; i < num_verts; j=i++) {
-         if (verts[i].type == STBTT_vline) {
-            float x0 = verts[i].x*scale_x, y0 = verts[i].y*scale_y;
-            float x1 = verts[j].x*scale_x, y1 = verts[j].y*scale_y;
-            float dist = (float) STBTT_sqrt((x1-x0)*(x1-x0) + (y1-y0)*(y1-y0));
-            precompute[i] = (dist == 0) ? 0.0f : 1.0f / dist;
-         } else if (verts[i].type == STBTT_vcurve) {
-            float x2 = verts[j].x *scale_x, y2 = verts[j].y *scale_y;
-            float x1 = verts[i].cx*scale_x, y1 = verts[i].cy*scale_y;
-            float x0 = verts[i].x *scale_x, y0 = verts[i].y *scale_y;
-            float bx = x0 - 2*x1 + x2, by = y0 - 2*y1 + y2;
-            float len2 = bx*bx + by*by;
-            if (len2 != 0.0f)
-               precompute[i] = 1.0f / (bx*bx + by*by);
-            else
-               precompute[i] = 0.0f;
-         } else
-            precompute[i] = 0.0f;
-      }
-
-      for (y=iy0; y < iy1; ++y) {
-         for (x=ix0; x < ix1; ++x) {
-            float val;
-            float min_dist = 999999.0f;
-            float sx = (float) x + 0.5f;
-            float sy = (float) y + 0.5f;
-            float x_gspace = (sx / scale_x);
-            float y_gspace = (sy / scale_y);
-
-            int winding = stbtt__compute_crossings_x(x_gspace, y_gspace, num_verts, verts); // @OPTIMIZE: this could just be a rasterization, but needs to be line vs. non-tesselated curves so a new path
-
-            for (i=0; i < num_verts; ++i) {
-               float x0 = verts[i].x*scale_x, y0 = verts[i].y*scale_y;
-
-               if (verts[i].type == STBTT_vline && precompute[i] != 0.0f) {
-                  float x1 = verts[i-1].x*scale_x, y1 = verts[i-1].y*scale_y;
-
-                  float dist,dist2 = (x0-sx)*(x0-sx) + (y0-sy)*(y0-sy);
-                  if (dist2 < min_dist*min_dist)
-                     min_dist = (float) STBTT_sqrt(dist2);
-
-                  // coarse culling against bbox
-                  //if (sx > STBTT_min(x0,x1)-min_dist && sx < STBTT_max(x0,x1)+min_dist &&
-                  //    sy > STBTT_min(y0,y1)-min_dist && sy < STBTT_max(y0,y1)+min_dist)
-                  dist = (float) STBTT_fabs((x1-x0)*(y0-sy) - (y1-y0)*(x0-sx)) * precompute[i];
-                  STBTT_assert(i != 0);
-                  if (dist < min_dist) {
-                     // check position along line
-                     // x' = x0 + t*(x1-x0), y' = y0 + t*(y1-y0)
-                     // minimize (x'-sx)*(x'-sx)+(y'-sy)*(y'-sy)
-                     float dx = x1-x0, dy = y1-y0;
-                     float px = x0-sx, py = y0-sy;
-                     // minimize (px+t*dx)^2 + (py+t*dy)^2 = px*px + 2*px*dx*t + t^2*dx*dx + py*py + 2*py*dy*t + t^2*dy*dy
-                     // derivative: 2*px*dx + 2*py*dy + (2*dx*dx+2*dy*dy)*t, set to 0 and solve
-                     float t = -(px*dx + py*dy) / (dx*dx + dy*dy);
-                     if (t >= 0.0f && t <= 1.0f)
-                        min_dist = dist;
-                  }
-               } else if (verts[i].type == STBTT_vcurve) {
-                  float x2 = verts[i-1].x *scale_x, y2 = verts[i-1].y *scale_y;
-                  float x1 = verts[i  ].cx*scale_x, y1 = verts[i  ].cy*scale_y;
-                  float box_x0 = STBTT_min(STBTT_min(x0,x1),x2);
-                  float box_y0 = STBTT_min(STBTT_min(y0,y1),y2);
-                  float box_x1 = STBTT_max(STBTT_max(x0,x1),x2);
-                  float box_y1 = STBTT_max(STBTT_max(y0,y1),y2);
-                  // coarse culling against bbox to avoid computing cubic unnecessarily
-                  if (sx > box_x0-min_dist && sx < box_x1+min_dist && sy > box_y0-min_dist && sy < box_y1+min_dist) {
-                     int num=0;
-                     float ax = x1-x0, ay = y1-y0;
-                     float bx = x0 - 2*x1 + x2, by = y0 - 2*y1 + y2;
-                     float mx = x0 - sx, my = y0 - sy;
-                     float res[3] = {0.f,0.f,0.f};
-                     float px,py,t,it,dist2;
-                     float a_inv = precompute[i];
-                     if (a_inv == 0.0) { // if a_inv is 0, it's 2nd degree so use quadratic formula
-                        float a = 3*(ax*bx + ay*by);
-                        float b = 2*(ax*ax + ay*ay) + (mx*bx+my*by);
-                        float c = mx*ax+my*ay;
-                        if (a == 0.0) { // if a is 0, it's linear
-                           if (b != 0.0) {
-                              res[num++] = -c/b;
-                           }
-                        } else {
-                           float discriminant = b*b - 4*a*c;
-                           if (discriminant < 0)
-                              num = 0;
-                           else {
-                              float root = (float) STBTT_sqrt(discriminant);
-                              res[0] = (-b - root)/(2*a);
-                              res[1] = (-b + root)/(2*a);
-                              num = 2; // don't bother distinguishing 1-solution case, as code below will still work
-                           }
-                        }
-                     } else {
-                        float b = 3*(ax*bx + ay*by) * a_inv; // could precompute this as it doesn't depend on sample point
-                        float c = (2*(ax*ax + ay*ay) + (mx*bx+my*by)) * a_inv;
-                        float d = (mx*ax+my*ay) * a_inv;
-                        num = stbtt__solve_cubic(b, c, d, res);
-                     }
-                     dist2 = (x0-sx)*(x0-sx) + (y0-sy)*(y0-sy);
-                     if (dist2 < min_dist*min_dist)
-                        min_dist = (float) STBTT_sqrt(dist2);
-
-                     if (num >= 1 && res[0] >= 0.0f && res[0] <= 1.0f) {
-                        t = res[0], it = 1.0f - t;
-                        px = it*it*x0 + 2*t*it*x1 + t*t*x2;
-                        py = it*it*y0 + 2*t*it*y1 + t*t*y2;
-                        dist2 = (px-sx)*(px-sx) + (py-sy)*(py-sy);
-                        if (dist2 < min_dist * min_dist)
-                           min_dist = (float) STBTT_sqrt(dist2);
-                     }
-                     if (num >= 2 && res[1] >= 0.0f && res[1] <= 1.0f) {
-                        t = res[1], it = 1.0f - t;
-                        px = it*it*x0 + 2*t*it*x1 + t*t*x2;
-                        py = it*it*y0 + 2*t*it*y1 + t*t*y2;
-                        dist2 = (px-sx)*(px-sx) + (py-sy)*(py-sy);
-                        if (dist2 < min_dist * min_dist)
-                           min_dist = (float) STBTT_sqrt(dist2);
-                     }
-                     if (num >= 3 && res[2] >= 0.0f && res[2] <= 1.0f) {
-                        t = res[2], it = 1.0f - t;
-                        px = it*it*x0 + 2*t*it*x1 + t*t*x2;
-                        py = it*it*y0 + 2*t*it*y1 + t*t*y2;
-                        dist2 = (px-sx)*(px-sx) + (py-sy)*(py-sy);
-                        if (dist2 < min_dist * min_dist)
-                           min_dist = (float) STBTT_sqrt(dist2);
-                     }
-                  }
-               }
-            }
-            if (winding == 0)
-               min_dist = -min_dist;  // if outside the shape, value is negative
-            val = onedge_value + pixel_dist_scale * min_dist;
-            if (val < 0)
-               val = 0;
-            else if (val > 255)
-               val = 255;
-            data[(y-iy0)*w+(x-ix0)] = (unsigned char) val;
-         }
-      }
-      STBTT_free(precompute, info->userdata);
-      STBTT_free(verts, info->userdata);
-   }
-   return data;
-}
-
-STBTT_DEF unsigned char * stbtt_GetCodepointSDF(const stbtt_fontinfo *info, float scale, int codepoint, int padding, unsigned char onedge_value, float pixel_dist_scale, int *width, int *height, int *xoff, int *yoff)
-{
-   return stbtt_GetGlyphSDF(info, scale, stbtt_FindGlyphIndex(info, codepoint), padding, onedge_value, pixel_dist_scale, width, height, xoff, yoff);
-}
-
-STBTT_DEF void stbtt_FreeSDF(unsigned char *bitmap, void *userdata)
-{
-   STBTT_free(bitmap, userdata);
-}
-
-//////////////////////////////////////////////////////////////////////////////
-//
-// font name matching -- recommended not to use this
-//
-
-// check if a utf8 string contains a prefix which is the utf16 string; if so return length of matching utf8 string
-static stbtt_int32 stbtt__CompareUTF8toUTF16_bigendian_prefix(stbtt_uint8 *s1, stbtt_int32 len1, stbtt_uint8 *s2, stbtt_int32 len2)
-{
-   stbtt_int32 i=0;
-
-   // convert utf16 to utf8 and compare the results while converting
-   while (len2) {
-      stbtt_uint16 ch = s2[0]*256 + s2[1];
-      if (ch < 0x80) {
-         if (i >= len1) return -1;
-         if (s1[i++] != ch) return -1;
-      } else if (ch < 0x800) {
-         if (i+1 >= len1) return -1;
-         if (s1[i++] != 0xc0 + (ch >> 6)) return -1;
-         if (s1[i++] != 0x80 + (ch & 0x3f)) return -1;
-      } else if (ch >= 0xd800 && ch < 0xdc00) {
-         stbtt_uint32 c;
-         stbtt_uint16 ch2 = s2[2]*256 + s2[3];
-         if (i+3 >= len1) return -1;
-         c = ((ch - 0xd800) << 10) + (ch2 - 0xdc00) + 0x10000;
-         if (s1[i++] != 0xf0 + (c >> 18)) return -1;
-         if (s1[i++] != 0x80 + ((c >> 12) & 0x3f)) return -1;
-         if (s1[i++] != 0x80 + ((c >>  6) & 0x3f)) return -1;
-         if (s1[i++] != 0x80 + ((c      ) & 0x3f)) return -1;
-         s2 += 2; // plus another 2 below
-         len2 -= 2;
-      } else if (ch >= 0xdc00 && ch < 0xe000) {
-         return -1;
-      } else {
-         if (i+2 >= len1) return -1;
-         if (s1[i++] != 0xe0 + (ch >> 12)) return -1;
-         if (s1[i++] != 0x80 + ((ch >> 6) & 0x3f)) return -1;
-         if (s1[i++] != 0x80 + ((ch     ) & 0x3f)) return -1;
-      }
-      s2 += 2;
-      len2 -= 2;
-   }
-   return i;
-}
-
-static int stbtt_CompareUTF8toUTF16_bigendian_internal(char *s1, int len1, char *s2, int len2)
-{
-   return len1 == stbtt__CompareUTF8toUTF16_bigendian_prefix((stbtt_uint8*) s1, len1, (stbtt_uint8*) s2, len2);
-}
-
-// returns results in whatever encoding you request... but note that 2-byte encodings
-// will be BIG-ENDIAN... use stbtt_CompareUTF8toUTF16_bigendian() to compare
-STBTT_DEF const char *stbtt_GetFontNameString(const stbtt_fontinfo *font, int *length, int platformID, int encodingID, int languageID, int nameID)
-{
-   stbtt_int32 i,count,stringOffset;
-   stbtt_uint8 *fc = font->data;
-   stbtt_uint32 offset = font->fontstart;
-   stbtt_uint32 nm = stbtt__find_table(fc, offset, "name");
-   if (!nm) return NULL;
-
-   count = ttUSHORT(fc+nm+2);
-   stringOffset = nm + ttUSHORT(fc+nm+4);
-   for (i=0; i < count; ++i) {
-      stbtt_uint32 loc = nm + 6 + 12 * i;
-      if (platformID == ttUSHORT(fc+loc+0) && encodingID == ttUSHORT(fc+loc+2)
-          && languageID == ttUSHORT(fc+loc+4) && nameID == ttUSHORT(fc+loc+6)) {
-         *length = ttUSHORT(fc+loc+8);
-         return (const char *) (fc+stringOffset+ttUSHORT(fc+loc+10));
-      }
-   }
-   return NULL;
-}
-
-static int stbtt__matchpair(stbtt_uint8 *fc, stbtt_uint32 nm, stbtt_uint8 *name, stbtt_int32 nlen, stbtt_int32 target_id, stbtt_int32 next_id)
-{
-   stbtt_int32 i;
-   stbtt_int32 count = ttUSHORT(fc+nm+2);
-   stbtt_int32 stringOffset = nm + ttUSHORT(fc+nm+4);
-
-   for (i=0; i < count; ++i) {
-      stbtt_uint32 loc = nm + 6 + 12 * i;
-      stbtt_int32 id = ttUSHORT(fc+loc+6);
-      if (id == target_id) {
-         // find the encoding
-         stbtt_int32 platform = ttUSHORT(fc+loc+0), encoding = ttUSHORT(fc+loc+2), language = ttUSHORT(fc+loc+4);
-
-         // is this a Unicode encoding?
-         if (platform == 0 || (platform == 3 && encoding == 1) || (platform == 3 && encoding == 10)) {
-            stbtt_int32 slen = ttUSHORT(fc+loc+8);
-            stbtt_int32 off = ttUSHORT(fc+loc+10);
-
-            // check if there's a prefix match
-            stbtt_int32 matchlen = stbtt__CompareUTF8toUTF16_bigendian_prefix(name, nlen, fc+stringOffset+off,slen);
-            if (matchlen >= 0) {
-               // check for target_id+1 immediately following, with same encoding & language
-               if (i+1 < count && ttUSHORT(fc+loc+12+6) == next_id && ttUSHORT(fc+loc+12) == platform && ttUSHORT(fc+loc+12+2) == encoding && ttUSHORT(fc+loc+12+4) == language) {
-                  slen = ttUSHORT(fc+loc+12+8);
-                  off = ttUSHORT(fc+loc+12+10);
-                  if (slen == 0) {
-                     if (matchlen == nlen)
-                        return 1;
-                  } else if (matchlen < nlen && name[matchlen] == ' ') {
-                     ++matchlen;
-                     if (stbtt_CompareUTF8toUTF16_bigendian_internal((char*) (name+matchlen), nlen-matchlen, (char*)(fc+stringOffset+off),slen))
-                        return 1;
-                  }
-               } else {
-                  // if nothing immediately following
-                  if (matchlen == nlen)
-                     return 1;
-               }
-            }
-         }
-
-         // @TODO handle other encodings
-      }
-   }
-   return 0;
-}
-
-static int stbtt__matches(stbtt_uint8 *fc, stbtt_uint32 offset, stbtt_uint8 *name, stbtt_int32 flags)
-{
-   stbtt_int32 nlen = (stbtt_int32) STBTT_strlen((char *) name);
-   stbtt_uint32 nm,hd;
-   if (!stbtt__isfont(fc+offset)) return 0;
-
-   // check italics/bold/underline flags in macStyle...
-   if (flags) {
-      hd = stbtt__find_table(fc, offset, "head");
-      if ((ttUSHORT(fc+hd+44) & 7) != (flags & 7)) return 0;
-   }
-
-   nm = stbtt__find_table(fc, offset, "name");
-   if (!nm) return 0;
-
-   if (flags) {
-      // if we checked the macStyle flags, then just check the family and ignore the subfamily
-      if (stbtt__matchpair(fc, nm, name, nlen, 16, -1))  return 1;
-      if (stbtt__matchpair(fc, nm, name, nlen,  1, -1))  return 1;
-      if (stbtt__matchpair(fc, nm, name, nlen,  3, -1))  return 1;
-   } else {
-      if (stbtt__matchpair(fc, nm, name, nlen, 16, 17))  return 1;
-      if (stbtt__matchpair(fc, nm, name, nlen,  1,  2))  return 1;
-      if (stbtt__matchpair(fc, nm, name, nlen,  3, -1))  return 1;
-   }
-
-   return 0;
-}
-
-static int stbtt_FindMatchingFont_internal(unsigned char *font_collection, char *name_utf8, stbtt_int32 flags)
-{
-   stbtt_int32 i;
-   for (i=0;;++i) {
-      stbtt_int32 off = stbtt_GetFontOffsetForIndex(font_collection, i);
-      if (off < 0) return off;
-      if (stbtt__matches((stbtt_uint8 *) font_collection, off, (stbtt_uint8*) name_utf8, flags))
-         return off;
-   }
-}
-
-#if defined(__GNUC__) || defined(__clang__)
-#pragma GCC diagnostic push
-#pragma GCC diagnostic ignored "-Wcast-qual"
-#endif
-
-STBTT_DEF int stbtt_BakeFontBitmap(const unsigned char *data, int offset,
-                                float pixel_height, unsigned char *pixels, int pw, int ph,
-                                int first_char, int num_chars, stbtt_bakedchar *chardata)
-{
-   return stbtt_BakeFontBitmap_internal((unsigned char *) data, offset, pixel_height, pixels, pw, ph, first_char, num_chars, chardata);
-}
-
-STBTT_DEF int stbtt_GetFontOffsetForIndex(const unsigned char *data, int index)
-{
-   return stbtt_GetFontOffsetForIndex_internal((unsigned char *) data, index);
-}
-
-STBTT_DEF int stbtt_GetNumberOfFonts(const unsigned char *data)
-{
-   return stbtt_GetNumberOfFonts_internal((unsigned char *) data);
-}
-
-STBTT_DEF int stbtt_InitFont(stbtt_fontinfo *info, const unsigned char *data, int offset)
-{
-   return stbtt_InitFont_internal(info, (unsigned char *) data, offset);
-}
-
-STBTT_DEF int stbtt_FindMatchingFont(const unsigned char *fontdata, const char *name, int flags)
-{
-   return stbtt_FindMatchingFont_internal((unsigned char *) fontdata, (char *) name, flags);
-}
-
-STBTT_DEF int stbtt_CompareUTF8toUTF16_bigendian(const char *s1, int len1, const char *s2, int len2)
-{
-   return stbtt_CompareUTF8toUTF16_bigendian_internal((char *) s1, len1, (char *) s2, len2);
-}
-
-#if defined(__GNUC__) || defined(__clang__)
-#pragma GCC diagnostic pop
-#endif
-
-#endif // STB_TRUETYPE_IMPLEMENTATION
-
-
-// FULL VERSION HISTORY
-//
-//   1.25 (2021-07-11) many fixes
-//   1.24 (2020-02-05) fix warning
-//   1.23 (2020-02-02) query SVG data for glyphs; query whole kerning table (but only kern not GPOS)
-//   1.22 (2019-08-11) minimize missing-glyph duplication; fix kerning if both 'GPOS' and 'kern' are defined
-//   1.21 (2019-02-25) fix warning
-//   1.20 (2019-02-07) PackFontRange skips missing codepoints; GetScaleFontVMetrics()
-//   1.19 (2018-02-11) OpenType GPOS kerning (horizontal only), STBTT_fmod
-//   1.18 (2018-01-29) add missing function
-//   1.17 (2017-07-23) make more arguments const; doc fix
-//   1.16 (2017-07-12) SDF support
-//   1.15 (2017-03-03) make more arguments const
-//   1.14 (2017-01-16) num-fonts-in-TTC function
-//   1.13 (2017-01-02) support OpenType fonts, certain Apple fonts
-//   1.12 (2016-10-25) suppress warnings about casting away const with -Wcast-qual
-//   1.11 (2016-04-02) fix unused-variable warning
-//   1.10 (2016-04-02) allow user-defined fabs() replacement
-//                     fix memory leak if fontsize=0.0
-//                     fix warning from duplicate typedef
-//   1.09 (2016-01-16) warning fix; avoid crash on outofmem; use alloc userdata for PackFontRanges
-//   1.08 (2015-09-13) document stbtt_Rasterize(); fixes for vertical & horizontal edges
-//   1.07 (2015-08-01) allow PackFontRanges to accept arrays of sparse codepoints;
-//                     allow PackFontRanges to pack and render in separate phases;
-//                     fix stbtt_GetFontOFfsetForIndex (never worked for non-0 input?);
-//                     fixed an assert() bug in the new rasterizer
-//                     replace assert() with STBTT_assert() in new rasterizer
-//   1.06 (2015-07-14) performance improvements (~35% faster on x86 and x64 on test machine)
-//                     also more precise AA rasterizer, except if shapes overlap
-//                     remove need for STBTT_sort
-//   1.05 (2015-04-15) fix misplaced definitions for STBTT_STATIC
-//   1.04 (2015-04-15) typo in example
-//   1.03 (2015-04-12) STBTT_STATIC, fix memory leak in new packing, various fixes
-//   1.02 (2014-12-10) fix various warnings & compile issues w/ stb_rect_pack, C++
-//   1.01 (2014-12-08) fix subpixel position when oversampling to exactly match
-//                        non-oversampled; STBTT_POINT_SIZE for packed case only
-//   1.00 (2014-12-06) add new PackBegin etc. API, w/ support for oversampling
-//   0.99 (2014-09-18) fix multiple bugs with subpixel rendering (ryg)
-//   0.9  (2014-08-07) support certain mac/iOS fonts without an MS platformID
-//   0.8b (2014-07-07) fix a warning
-//   0.8  (2014-05-25) fix a few more warnings
-//   0.7  (2013-09-25) bugfix: subpixel glyph bug fixed in 0.5 had come back
-//   0.6c (2012-07-24) improve documentation
-//   0.6b (2012-07-20) fix a few more warnings
-//   0.6  (2012-07-17) fix warnings; added stbtt_ScaleForMappingEmToPixels,
-//                        stbtt_GetFontBoundingBox, stbtt_IsGlyphEmpty
-//   0.5  (2011-12-09) bugfixes:
-//                        subpixel glyph renderer computed wrong bounding box
-//                        first vertex of shape can be off-curve (FreeSans)
-//   0.4b (2011-12-03) fixed an error in the font baking example
-//   0.4  (2011-12-01) kerning, subpixel rendering (tor)
-//                    bugfixes for:
-//                        codepoint-to-glyph conversion using table fmt=12
-//                        codepoint-to-glyph conversion using table fmt=4
-//                        stbtt_GetBakedQuad with non-square texture (Zer)
-//                    updated Hello World! sample to use kerning and subpixel
-//                    fixed some warnings
-//   0.3  (2009-06-24) cmap fmt=12, compound shapes (MM)
-//                    userdata, malloc-from-userdata, non-zero fill (stb)
-//   0.2  (2009-03-11) Fix unsigned/signed char warnings
-//   0.1  (2009-03-09) First public release
-//
-
-/*
-------------------------------------------------------------------------------
-This software is available under 2 licenses -- choose whichever you prefer.
-------------------------------------------------------------------------------
-ALTERNATIVE A - MIT License
-Copyright (c) 2017 Sean Barrett
-Permission is hereby granted, free of charge, to any person obtaining a copy of
-this software and associated documentation files (the "Software"), to deal in
-the Software without restriction, including without limitation the rights to
-use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies
-of the Software, and to permit persons to whom the Software is furnished to do
-so, subject to the following conditions:
-The above copyright notice and this permission notice shall be included in all
-copies or substantial portions of the Software.
-THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
-IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
-FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
-AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
-LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
-OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
-SOFTWARE.
-------------------------------------------------------------------------------
-ALTERNATIVE B - Public Domain (www.unlicense.org)
-This is free and unencumbered software released into the public domain.
-Anyone is free to copy, modify, publish, use, compile, sell, or distribute this
-software, either in source code form or as a compiled binary, for any purpose,
-commercial or non-commercial, and by any means.
-In jurisdictions that recognize copyright laws, the author or authors of this
-software dedicate any and all copyright interest in the software to the public
-domain. We make this dedication for the benefit of the public at large and to
-the detriment of our heirs and successors. We intend this dedication to be an
-overt act of relinquishment in perpetuity of all present and future rights to
-this software under copyright law.
-THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
-IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
-FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
-AUTHORS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
-ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION
-WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
-------------------------------------------------------------------------------
-*/
diff --git a/std_printers.c b/std_printers.c
deleted file mode 100644
index abb111a..0000000
--- a/std_printers.c
+++ /dev/null
@@ -1,56 +0,0 @@
-#include <stdarg.h>
-#include <stdio.h>
-#include <unistd.h>
-
-extern int fileno(FILE*);
-
-void print(const char* fmt, ...) {
-	va_list args;
-	va_start(args, fmt);
-	vfprintf(stdout, fmt, args);
-	va_end(args);
-}
-
-void print_err(const char* fmt, ...) {
-	va_list args;
-	va_start(args, fmt);
-
-	if (isatty(fileno(stderr))) {
-		fprintf(stderr, "\033[31;31m");
-	}
-
-	vfprintf(stderr, fmt, args);
-
-	if (isatty(fileno(stderr))) {
-		fprintf(stderr, "\033[0m");
-	}
-
-	va_end(args);
-}
-
-void print_war(const char* fmt, ...) {
-	va_list args;
-	va_start(args, fmt);
-
-	if (isatty(fileno(stderr))) {
-		fprintf(stderr, "\033[31;35m");
-	}
-
-	vfprintf(stderr, fmt, args);
-
-	if (isatty(fileno(stderr))) {
-		fprintf(stderr, "\033[0m");
-	}
-
-	va_end(args);
-}
-
-void pbreak(Error code) {
-#if defined(DEBUG) && defined(plat_x86)
-	__asm__("int3;");
-	(void)code;
-#else
-	exit(code);
-#endif
-}
-
diff --git a/ui.c b/ui.c
deleted file mode 100644
index a7d1070..0000000
--- a/ui.c
+++ /dev/null
@@ -1,152 +0,0 @@
-#include "config.h"
-#include "render.h"
-#include "plat.h"
-#include "rcache.h"
-#include "ui.h"
-
-Rectangle rectcut_left(Rectangle* l, int a) {
-	Rectangle r;
-	r.x = l->x;
-	r.y = l->y;
-	r.w = a;
-	r.h = l->h;
-	l->x += a;
-	l->w -= a;
-	return r;
-}
-
-Rectangle rectcut_right(Rectangle* l, int a) {
-	Rectangle r;
-	r.x = l->x + l->w - a;
-	r.y = l->y;
-	r.w = a;
-	r.h = l->h;
-	l->w -= a;
-	return r;
-}
-
-Rectangle rectcut_top(Rectangle* l, int a) {
-	Rectangle r;
-	r.x = l->x;
-	r.y = l->y;
-	r.w = l->w;
-	r.h = a;
-	l->y += a;
-	l->h -= a;
-	return r;
-}
-
-Rectangle rectcut_bottom(Rectangle* l, int a) {
-	Rectangle r;
-	r.x = l->x;
-	r.y = l->y + l->h - a;
-	r.w = l->w;
-	r.h = a;
-	l->h -= a;
-	return r;
-}
-
-Rectangle shrink_rect(const Rectangle* l, int a) {
-	Rectangle r;
-	int a2 = a * 2;
-	r.x = l->x + a;
-	r.y = l->y + a;
-	r.w = l->w - a2;
-	r.h = l->h - a2;
-	return r;
-}
-
-Rectangle centre_rect(
-	const Rectangle* l,
-	const Rectangle* t
-) {
-	Rectangle r;
-	r.x = l->x + l->w / 2 - t->w / 2;
-	r.y = l->y + l->h / 2 - t->h / 2;
-	r.w = t->w;
-	r.h = t->h;
-	return r;
-}
-
-void init_ui(UI* u, Font* f) {
-	u->font = f;
-}
-
-void ui_begin(UI* u, Rectangle* l) {
-	(void)u;
-	l->x = 0;
-	l->y = 0;
-	l->w = get_render_w();
-	l->h = get_render_h();
-}
-
-void ui_end(UI* u) {
-	(void)u;
-}
-
-void ui_label(
-	UI* u,
-	const Rectangle* l,
-	const char* text
-) {
-	rc_add_cmd_rfont_text_col(
-		u->font,
-		l->x,
-		l->y,
-		text,
-		theme_text_colour
-	);
-}
-
-int ui_button(
-	UI* u,
-	const Rectangle* r,
-	const char* label
-) {
-	Rectangle o, t;
-	Colour c;
-	int h;
-	o = shrink_rect(r, theme_outline_width);
-	h = point_rect_overlap(
-		&o,
-		get_mouse_x(),
-		get_mouse_y()
-	);
-	if (h) {
-		if (mbtn_pressed(mbtn_left))
-			c = theme_active_colour;
-		else
-			c = theme_hovered_colour;
-	} else
-		c = theme_background_colour;
-	rc_add_cmd_rect(r, theme_outline_colour);
-	rc_add_cmd_rect(&o, c);
-	t = text_rect(u->font, label);
-	o = centre_rect(&o, &t);
-	rc_add_cmd_rfont_text_col(
-		u->font,
-		o.x,
-		o.y,
-		label,
-		theme_text_colour
-	);
-	return h && mbtn_just_released(mbtn_left);
-}
-
-void ui_container(
-	UI* u,
-	const Rectangle* l,
-	Rectangle* r,
-	const Rectangle* c
-) {
-	Rectangle o, cl;
-	(void)u;
-	rc_add_cmd_rect(l, theme_outline_colour);
-	o = shrink_rect(l, theme_outline_width);
-	rc_add_cmd_rect(&o, theme_background_colour);
-	cl = o;
-	cl.w += cl.x;
-	cl.h += cl.y;
-	rc_add_cmd_clip(&cl);
-	*r = shrink_rect(&o, theme_padding);
-}
diff --git a/ui.h b/ui.h
deleted file mode 100644
index 6a36566..0000000
--- a/ui.h
+++ /dev/null
@@ -1,41 +0,0 @@
-#ifndef ui_h
-#define ui_h
-
-#include "render.h"
-#include "rect.h"
-
-typedef struct {
-	Font* font;
-} UI;
-
-Rectangle rectcut_left  (Rectangle* l, int a);
-Rectangle rectcut_right (Rectangle* l, int a);
-Rectangle rectcut_top   (Rectangle* l, int a);
-Rectangle rectcut_bottom(Rectangle* l, int a);
-Rectangle shrink_rect   (const Rectangle* l, int a);
-Rectangle centre_rect   (
-	const Rectangle* l,
-	const Rectangle* t
-);
-
-void init_ui(UI* u, Font* f);
-void ui_begin(UI* u, Rectangle* l);
-void ui_end(UI* u);
-int ui_button(
-	UI* u,
-	const Rectangle* l,
-	const char* label
-);
-void ui_label(
-	UI* u,
-	const Rectangle* l,
-	const char* text
-);
-void ui_container(
-	UI* u,
-	const Rectangle* l,
-	Rectangle* r,
-	const Rectangle* c
-);
-
-#endif