@@ -22,6 +22,7 @@ - GBA: Fix timing advancing too quickly in rare cases
  - GBA Audio: Fix deserializing SOUNDCNT_L
  - GBA Audio: Fix stereo in XQ audio
  - GBA Audio: Fix volume/mute in XQ audio (fixes mgba.io/i/1864)
+ - GBA Audio: Revamp FIFO emulation (fixes mgba.io/i/356, mgba.io/i/875, mgba.io/i/1847)
  - GBA BIOS: Implement dummy sound driver calls
  - GBA BIOS: Improve HLE BIOS timing
  - GBA BIOS: Fix reloading video registers after reset (fixes mgba.io/i/1808)

@@ -15,6 +15,8 @@ #include <mgba/core/log.h>
 #include <mgba/internal/gb/audio.h>
 #include <mgba-util/circle-buffer.h>
 
+#define GBA_AUDIO_FIFO_SIZE 8
+
 #define MP2K_MAGIC 0x68736D53
 #define MP2K_MAX_SOUND_CHANNELS 12
 
@@ -26,7 +28,11 @@ extern const unsigned GBA_AUDIO_SAMPLES;
 extern const int GBA_AUDIO_VOLUME_MAX;
 
 struct GBAAudioFIFO {
-	struct CircleBuffer fifo;
+	uint32_t fifo[GBA_AUDIO_FIFO_SIZE];
+	int fifoWrite;
+	int fifoRead;
+	uint32_t internalSample;
+	int internalRemaining;
 	int dmaSource;
 	int8_t sample;
 };
@@ -298,7 +304,7 @@ void GBAAudioWriteSOUNDCNT_X(struct GBAAudio* audio, uint16_t value);
 void GBAAudioWriteSOUNDBIAS(struct GBAAudio* audio, uint16_t value);
 
 void GBAAudioWriteWaveRAM(struct GBAAudio* audio, int address, uint32_t value);
-void GBAAudioWriteFIFO(struct GBAAudio* audio, int address, uint32_t value);
+uint32_t GBAAudioWriteFIFO(struct GBAAudio* audio, int address, uint32_t value);
 void GBAAudioSampleFIFO(struct GBAAudio* audio, int fifoId, int32_t cycles);
 
 struct GBASerializedState;

@@ -20,7 +20,7 @@
 mLOG_DECLARE_CATEGORY(GBA_STATE);
 
 /* Savestate format:
- * 0x00000 - 0x00003: Version Magic (0x01000001)
+ * 0x00000 - 0x00003: Version Magic (0x01000004)
  * 0x00004 - 0x00007: BIOS checksum (e.g. 0xBAAE187F for official BIOS)
  * 0x00008 - 0x0000B: ROM CRC32
  * 0x0000C - 0x0000F: Master cycles
@@ -69,10 +69,16 @@ * | 0x00188 - 0x0018B: Next event
  * 0x0018C - 0x001AB: Audio FIFO 1
  * 0x001AC - 0x001CB: Audio FIFO 2
  * 0x001CC - 0x001DF: Audio miscellaneous state
- * | 0x001CC - 0x001CF: FIFO 1 size
- * | 0x001D0 - 0x001D3: Reserved
+ * | 0x001CC - 0x001CF: Channel A internal audio samples
+ * | 0x001D0 - 0x001D3: Channel B internal audio samples
  * | 0x001D4 - 0x001D7: Next sample
- * | 0x001D8 - 0x001DB: FIFO 2 size
+ * | 0x001D8: Channel A current sample
+ * | 0x001D9: Channel B current sample
+ * | 0x001DA - 0x001DB: Flags
+ *   | bits 0 - 1: Channel B internal samples remaining
+ *   | bits 2 - 4: Channel B readable words
+ *   | bits 5 - 6: Channel A internal samples remaining
+ *   | bits 7 - 9: Channel A readable words
  * | TODO: Fix this, they're in big-endian order, but field is little-endian
  * | 0x001DC - 0x001DC: Channel 1 envelope state
  *   | bits 0 - 3: Current volume
@@ -217,6 +223,12 @@ * 0x21000 - 0x60FFF: WRAM
  * Total size: 0x61000 (397,312) bytes
  */
 
+DECL_BITFIELD(GBASerializedAudioFlags, uint16_t);
+DECL_BITS(GBASerializedAudioFlags, FIFOInternalSamplesB, 0, 2);
+DECL_BITS(GBASerializedAudioFlags, FIFOSamplesB, 2, 3); // Yay legacy?
+DECL_BITS(GBASerializedAudioFlags, FIFOInternalSamplesA, 5, 2);
+DECL_BITS(GBASerializedAudioFlags, FIFOSamplesA, 7, 3);
+
 DECL_BITFIELD(GBASerializedVideoFlags, uint32_t);
 DECL_BITS(GBASerializedVideoFlags, Mode, 0, 2);
 
@@ -267,12 +279,14 @@ } cpu;
 
 	struct {
 		struct GBSerializedPSGState psg;
-		uint8_t fifoA[32];
-		uint8_t fifoB[32];
-		uint32_t fifoSizeA;
-		int32_t reserved;
+		uint32_t fifoA[8];
+		uint32_t fifoB[8];
+		uint32_t internalA;
+		uint32_t internalB;
 		int32_t nextSample;
-		uint32_t fifoSizeB;
+		int8_t sampleA;
+		int8_t sampleB;
+		GBASerializedAudioFlags gbaFlags;
 		GBSerializedAudioFlags flags;
 	} audio;
 

@@ -23,7 +23,6 @@
 mLOG_DEFINE_CATEGORY(GBA_AUDIO, "GBA Audio", "gba.audio");
 
 const unsigned GBA_AUDIO_SAMPLES = 2048;
-const unsigned GBA_AUDIO_FIFO_SIZE = 8 * sizeof(int32_t);
 const int GBA_AUDIO_VOLUME_MAX = 0x100;
 
 static const int CLOCKS_PER_FRAME = 0x800;
@@ -48,8 +47,6 @@ audio->samples = samples;
 	// Guess too large; we hang producing extra samples if we guess too low
 	blip_set_rates(audio->psg.left, GBA_ARM7TDMI_FREQUENCY, 96000);
 	blip_set_rates(audio->psg.right, GBA_ARM7TDMI_FREQUENCY, 96000);
-	CircleBufferInit(&audio->chA.fifo, GBA_AUDIO_FIFO_SIZE);
-	CircleBufferInit(&audio->chB.fifo, GBA_AUDIO_FIFO_SIZE);
 
 	audio->externalMixing = false;
 	audio->forceDisableChA = false;
@@ -64,7 +61,17 @@ mTimingDeschedule(&audio->p->timing, &audio->sampleEvent);
 	mTimingSchedule(&audio->p->timing, &audio->sampleEvent, 0);
 	audio->chA.dmaSource = 1;
 	audio->chB.dmaSource = 2;
+	audio->chA.fifoWrite = 0;
+	audio->chA.fifoRead = 0;
+	audio->chA.internalSample = 0;
+	audio->chA.internalRemaining = 0;
+	memset(audio->chA.fifo, 0, sizeof(audio->chA.fifo));
 	audio->chA.sample = 0;
+	audio->chB.fifoWrite = 0;
+	audio->chB.fifoRead = 0;
+	audio->chB.internalSample = 0;
+	audio->chB.internalRemaining = 0;
+	memset(audio->chB.fifo, 0, sizeof(audio->chB.fifo));
 	audio->chB.sample = 0;
 	audio->sampleRate = 0x8000;
 	audio->soundbias = 0x200;
@@ -84,14 +91,10 @@
 	blip_clear(audio->psg.left);
 	blip_clear(audio->psg.right);
 	audio->clock = 0;
-	CircleBufferClear(&audio->chA.fifo);
-	CircleBufferClear(&audio->chB.fifo);
 }
 
 void GBAAudioDeinit(struct GBAAudio* audio) {
 	GBAudioDeinit(&audio->psg);
-	CircleBufferDeinit(&audio->chA.fifo);
-	CircleBufferDeinit(&audio->chB.fifo);
 }
 
 void GBAAudioResizeBuffer(struct GBAAudio* audio, size_t samples) {
@@ -199,10 +202,12 @@ audio->chBRight = GBARegisterSOUNDCNT_HIGetChBRight(value);
 	audio->chBLeft = GBARegisterSOUNDCNT_HIGetChBLeft(value);
 	audio->chBTimer = GBARegisterSOUNDCNT_HIGetChBTimer(value);
 	if (GBARegisterSOUNDCNT_HIIsChAReset(value)) {
-		CircleBufferClear(&audio->chA.fifo);
+		audio->chA.fifoWrite = 0;
+		audio->chA.fifoRead = 0;
 	}
 	if (GBARegisterSOUNDCNT_HIIsChBReset(value)) {
-		CircleBufferClear(&audio->chB.fifo);
+		audio->chB.fifoWrite = 0;
+		audio->chB.fifoRead = 0;
 	}
 }
 
@@ -219,26 +224,25 @@ void GBAAudioWriteWaveRAM(struct GBAAudio* audio, int address, uint32_t value) {
 	audio->psg.ch3.wavedata32[address | (!audio->psg.ch3.bank * 4)] = value;
 }
 
-void GBAAudioWriteFIFO(struct GBAAudio* audio, int address, uint32_t value) {
-	struct CircleBuffer* fifo;
+uint32_t GBAAudioWriteFIFO(struct GBAAudio* audio, int address, uint32_t value) {
+	struct GBAAudioFIFO* channel;
 	switch (address) {
 	case REG_FIFO_A_LO:
-		fifo = &audio->chA.fifo;
+		channel = &audio->chA;
 		break;
 	case REG_FIFO_B_LO:
-		fifo = &audio->chB.fifo;
+		channel = &audio->chB;
 		break;
 	default:
 		mLOG(GBA_AUDIO, ERROR, "Bad FIFO write to address 0x%03x", address);
-		return;
+		return value;
 	}
-	int i;
-	for (i = 0; i < 4; ++i) {
-		while (!CircleBufferWrite8(fifo, value >> (8 * i))) {
-			int8_t dummy;
-			CircleBufferRead8(fifo, &dummy);
-		}
+	channel->fifo[channel->fifoWrite] = value;
+	++channel->fifoWrite;
+	if (channel->fifoWrite == GBA_AUDIO_FIFO_SIZE) {
+		channel->fifoWrite = 0;
 	}
+	return channel->fifo[channel->fifoWrite];
 }
 
 void GBAAudioSampleFIFO(struct GBAAudio* audio, int fifoId, int32_t cycles) {
@@ -251,17 +255,33 @@ } else {
 		mLOG(GBA_AUDIO, ERROR, "Bad FIFO write to address 0x%03x", fifoId);
 		return;
 	}
-	if (CircleBufferSize(&channel->fifo) <= 4 * sizeof(int32_t) && channel->dmaSource > 0) {
+	int fifoSize;
+	if (channel->fifoWrite >= channel->fifoRead) {
+		fifoSize = channel->fifoWrite - channel->fifoRead;
+	} else {
+		fifoSize = GBA_AUDIO_FIFO_SIZE - channel->fifoRead + channel->fifoWrite;
+	}
+	if (GBA_AUDIO_FIFO_SIZE - fifoSize > 4 && channel->dmaSource > 0) {
 		struct GBADMA* dma = &audio->p->memory.dma[channel->dmaSource];
 		if (GBADMARegisterGetTiming(dma->reg) == GBA_DMA_TIMING_CUSTOM) {
 			dma->when = mTimingCurrentTime(&audio->p->timing) - cycles;
 			dma->nextCount = 4;
 			GBADMASchedule(audio->p, channel->dmaSource, dma);
-		} else {
-			channel->dmaSource = 0;
+		}
+	}
+	if (!channel->internalRemaining && fifoSize) {
+		channel->internalSample = channel->fifo[channel->fifoRead];
+		channel->internalRemaining = 4;
+		++channel->fifoRead;
+		if (channel->fifoRead == GBA_AUDIO_FIFO_SIZE) {
+			channel->fifoRead = 0;
 		}
 	}
-	CircleBufferRead8(&channel->fifo, (int8_t*) &channel->sample);
+	channel->sample = channel->internalSample;
+	if (channel->internalRemaining) {
+		channel->internalSample >>= 8;
+		--channel->internalRemaining;
+	}
 }
 
 static int _applyBias(struct GBAAudio* audio, int sample) {
@@ -345,37 +365,76 @@
 void GBAAudioSerialize(const struct GBAAudio* audio, struct GBASerializedState* state) {
 	GBAudioPSGSerialize(&audio->psg, &state->audio.psg, &state->audio.flags);
 
-	CircleBufferDump(&audio->chA.fifo, state->audio.fifoA, sizeof(state->audio.fifoA));
-	CircleBufferDump(&audio->chB.fifo, state->audio.fifoB, sizeof(state->audio.fifoB));
-	uint32_t fifoSize = CircleBufferSize(&audio->chA.fifo);
-	STORE_32(fifoSize, 0, &state->audio.fifoSizeA);
-	fifoSize = CircleBufferSize(&audio->chB.fifo);
-	STORE_32(fifoSize, 0, &state->audio.fifoSizeB);
+	STORE_32(audio->chA.internalSample, 0, &state->audio.internalA);
+	STORE_32(audio->chB.internalSample, 0, &state->audio.internalB);
+	state->audio.sampleA = audio->chA.sample;
+	state->audio.sampleB = audio->chB.sample;
+
+	int readA = audio->chA.fifoRead;
+	int readB = audio->chB.fifoRead;
+	size_t i;
+	for (i = 0; i < GBA_AUDIO_FIFO_SIZE; ++i) {
+		STORE_32(audio->chA.fifo[readA], i << 2, state->audio.fifoA);
+		STORE_32(audio->chB.fifo[readB], i << 2, state->audio.fifoB);
+		++readA;
+		if (readA == GBA_AUDIO_FIFO_SIZE) {
+			readA = 0;
+		}
+		++readB;
+		if (readB == GBA_AUDIO_FIFO_SIZE) {
+			readB = 0;
+		}
+	}
+
+	int fifoSizeA;
+	if (audio->chA.fifoWrite >= audio->chA.fifoRead) {
+		fifoSizeA = audio->chA.fifoWrite - audio->chA.fifoRead;
+	} else {
+		fifoSizeA = GBA_AUDIO_FIFO_SIZE - audio->chA.fifoRead + audio->chA.fifoWrite;
+	}
+
+	int fifoSizeB;
+	if (audio->chB.fifoWrite >= audio->chB.fifoRead) {
+		fifoSizeB = audio->chB.fifoWrite - audio->chB.fifoRead;
+	} else {
+		fifoSizeB = GBA_AUDIO_FIFO_SIZE - audio->chB.fifoRead + audio->chB.fifoWrite;
+	}
+
+	GBASerializedAudioFlags flags = 0;
+	flags = GBASerializedAudioFlagsSetFIFOSamplesA(flags, fifoSizeA);
+	flags = GBASerializedAudioFlagsSetFIFOSamplesB(flags, fifoSizeB);
+	flags = GBASerializedAudioFlagsSetFIFOInternalSamplesA(flags, audio->chA.internalRemaining);
+	flags = GBASerializedAudioFlagsSetFIFOInternalSamplesB(flags, audio->chB.internalRemaining);
+	STORE_32(flags, 0, &state->audio.gbaFlags);
 	STORE_32(audio->sampleEvent.when - mTimingCurrentTime(&audio->p->timing), 0, &state->audio.nextSample);
 }
 
 void GBAAudioDeserialize(struct GBAAudio* audio, const struct GBASerializedState* state) {
 	GBAudioPSGDeserialize(&audio->psg, &state->audio.psg, &state->audio.flags);
 
-	CircleBufferClear(&audio->chA.fifo);
-	CircleBufferClear(&audio->chB.fifo);
-	uint32_t fifoSize;
-	LOAD_32(fifoSize, 0, &state->audio.fifoSizeA);
-	if (fifoSize > CircleBufferCapacity(&audio->chA.fifo)) {
-		fifoSize = CircleBufferCapacity(&audio->chA.fifo);
-	}
+	LOAD_32(audio->chA.internalSample, 0, &state->audio.internalA);
+	LOAD_32(audio->chB.internalSample, 0, &state->audio.internalB);
+	audio->chA.sample = state->audio.sampleA;
+	audio->chB.sample = state->audio.sampleB;
+
+	int readA = 0;
+	int readB = 0;
 	size_t i;
-	for (i = 0; i < fifoSize; ++i) {
-		CircleBufferWrite8(&audio->chA.fifo, state->audio.fifoA[i]);
+	for (i = 0; i < GBA_AUDIO_FIFO_SIZE; ++i) {
+		LOAD_32(audio->chA.fifo[readA], i << 2, state->audio.fifoA);
+		LOAD_32(audio->chB.fifo[readB], i << 2, state->audio.fifoB);
+		++readA;
+		++readB;
 	}
+	audio->chA.fifoRead = 0;
+	audio->chB.fifoRead = 0;
 
-	LOAD_32(fifoSize, 0, &state->audio.fifoSizeB);
-	if (fifoSize > CircleBufferCapacity(&audio->chB.fifo)) {
-		fifoSize = CircleBufferCapacity(&audio->chB.fifo);
-	}
-	for (i = 0; i < fifoSize; ++i) {
-		CircleBufferWrite8(&audio->chB.fifo, state->audio.fifoB[i]);
-	}
+	GBASerializedAudioFlags flags;
+	LOAD_32(flags, 0, &state->audio.gbaFlags);
+	audio->chA.fifoWrite = GBASerializedAudioFlagsGetFIFOSamplesA(flags);
+	audio->chB.fifoWrite = GBASerializedAudioFlagsGetFIFOSamplesB(flags);
+	audio->chA.internalRemaining = GBASerializedAudioFlagsGetFIFOInternalSamplesA(flags);
+	audio->chB.internalRemaining = GBASerializedAudioFlagsGetFIFOInternalSamplesB(flags);
 
 	uint32_t when;
 	LOAD_32(when, 0, &state->audio.nextSample);

@@ -429,12 +429,12 @@
 		case REG_FIFO_A_LO:
 		case REG_FIFO_B_LO:
 			GBAIOWrite32(gba, address, (gba->memory.io[(address >> 1) + 1] << 16) | value);
-			break;
+			return;
 
 		case REG_FIFO_A_HI:
 		case REG_FIFO_B_HI:
 			GBAIOWrite32(gba, address - 2, gba->memory.io[(address >> 1) - 1] | (value << 16));
-			break;
+			return;
 
 		// DMA
 		case REG_DMA0SAD_LO:
@@ -630,7 +630,7 @@ GBAAudioWriteWaveRAM(&gba->audio, 3, value);
 		break;
 	case REG_FIFO_A_LO:
 	case REG_FIFO_B_LO:
-		GBAAudioWriteFIFO(&gba->audio, address, value);
+		value = GBAAudioWriteFIFO(&gba->audio, address, value);
 		break;
 	case REG_DMA0SAD_LO:
 		value = GBADMAWriteSAD(gba, 0, value);

@@ -15,7 +15,7 @@
 #include <fcntl.h>
 
 const uint32_t GBA_SAVESTATE_MAGIC = 0x01000000;
-const uint32_t GBA_SAVESTATE_VERSION = 0x00000003;
+const uint32_t GBA_SAVESTATE_VERSION = 0x00000004;
 
 mLOG_DEFINE_CATEGORY(GBA_STATE, "GBA Savestate", "gba.serialize");