@@ -96,6 +96,7 @@ - Core: Rework thread state synchronization
  - Core: Improve support for ROM patch cheats, supporting disabling overlapping patches
  - GB: Allow pausing event loop while CPU is blocked
  - GB: Add support for sleep and shutdown callbacks
+ - GB: Redo double speed emulation (closes mgba.io/i/1515)
  - GB Core: Return the current number of banks for ROM/SRAM, not theoretical max
  - GB I/O: Implement preliminary support for PCM12/PCM34 (closes mgba.io/i/1468)
  - GBA: Allow pausing event loop while CPU is blocked

@@ -133,6 +133,7 @@ #pragma pack(pop)
 
 	uint16_t index;
 
+	int tMultiplier;
 	int32_t cycles;
 	int32_t nextEvent;
 	enum SM83ExecutionState executionState;

@@ -65,7 +65,7 @@ audio->style = style;
 	if (style == GB_AUDIO_GBA) {
 		audio->timingFactor = 4;
 	} else {
-		audio->timingFactor = 1;
+		audio->timingFactor = 2;
 	}
 
 	audio->frameEvent.context = audio;
@@ -339,7 +339,7 @@ mTimingDeschedule(audio->timing, &audio->ch3Event);
 	if (audio->playingCh3) {
 		audio->ch3.readable = audio->style != GB_AUDIO_DMG;
 		// TODO: Where does this cycle delay come from?
-		mTimingSchedule(audio->timing, &audio->ch3Event, audio->timingFactor * 4 + 2 * (2048 - audio->ch3.rate));
+		mTimingSchedule(audio->timing, &audio->ch3Event, audio->timingFactor * (4 + 2 * (2048 - audio->ch3.rate)));
 	}
 	*audio->nr52 &= ~0x0004;
 	*audio->nr52 |= audio->playingCh3 << 2;
@@ -477,11 +477,8 @@ } else if (!wasEnable) {
 		audio->skipFrame = false;
 		audio->frame = 7;
 
-		if (audio->p) {
-			unsigned timingFactor = 0x400 >> !audio->p->doubleSpeed;
-			if (audio->p->timer.internalDiv & timingFactor) {
-				audio->skipFrame = true;
-			}
+		if (audio->p && audio->p->timer.internalDiv & 0x400) {
+			audio->skipFrame = true;
 		}
 	}
 }
@@ -914,7 +911,7 @@ ch->sample >>= volume;
 	audio->ch3.readable = true;
 	if (audio->style == GB_AUDIO_DMG) {
 		mTimingDeschedule(audio->timing, &audio->ch3Fade);
-		mTimingSchedule(timing, &audio->ch3Fade, 2 - cyclesLate);
+		mTimingSchedule(timing, &audio->ch3Fade, 4 - cyclesLate);
 	}
 	int cycles = 2 * (2048 - ch->rate);
 	mTimingSchedule(timing, &audio->ch3Event, audio->timingFactor * cycles - cyclesLate);

@@ -742,7 +742,7 @@ if (!enable) {
 		gb->memory.ime = false;
 		GBUpdateIRQs(gb);
 	} else {
-		mTimingSchedule(&gb->timing, &gb->eiPending, 4);
+		mTimingSchedule(&gb->timing, &gb->eiPending, 4 * cpu->tMultiplier);
 	}
 }
 
@@ -796,7 +796,7 @@ void GBStop(struct SM83Core* cpu) {
 	struct GB* gb = (struct GB*) cpu->master;
 	if (gb->model >= GB_MODEL_CGB && gb->memory.io[GB_REG_KEY1] & 1) {
 		gb->doubleSpeed ^= 1;
-		gb->audio.timingFactor = gb->doubleSpeed + 1;
+		gb->cpu->tMultiplier = 2 - gb->doubleSpeed;
 		gb->memory.io[GB_REG_KEY1] = 0;
 		gb->memory.io[GB_REG_KEY1] |= gb->doubleSpeed << 7;
 	} else {

@@ -417,15 +417,15 @@ _writeSGBBits(gb, (value >> 4) & 3);
 		}
 		return;
 	case GB_REG_TIMA:
-		if (value && mTimingUntil(&gb->timing, &gb->timer.irq) > 1) {
+		if (value && mTimingUntil(&gb->timing, &gb->timer.irq) > 2 - (int) gb->doubleSpeed) {
 			mTimingDeschedule(&gb->timing, &gb->timer.irq);
 		}
-		if (mTimingUntil(&gb->timing, &gb->timer.irq) == -1) {
+		if (mTimingUntil(&gb->timing, &gb->timer.irq) == (int) gb->doubleSpeed - 2) {
 			return;
 		}
 		break;
 	case GB_REG_TMA:
-		if (mTimingUntil(&gb->timing, &gb->timer.irq) == -1) {
+		if (mTimingUntil(&gb->timing, &gb->timer.irq) == (int) gb->doubleSpeed - 2) {
 			gb->memory.io[GB_REG_TIMA] = value;
 		}
 		break;

@@ -531,10 +531,7 @@ if (base >= 0xE000) {
 		base &= 0xDFFF;
 	}
 	mTimingDeschedule(&gb->timing, &gb->memory.dmaEvent);
-	mTimingSchedule(&gb->timing, &gb->memory.dmaEvent, 8);
-	if (gb->cpu->cycles + 8 < gb->cpu->nextEvent) {
-		gb->cpu->nextEvent = gb->cpu->cycles + 8;
-	}
+	mTimingSchedule(&gb->timing, &gb->memory.dmaEvent, 8 * (2 - gb->doubleSpeed));
 	gb->memory.dmaSource = base;
 	gb->memory.dmaDest = 0;
 	gb->memory.dmaRemaining = 0xA0;
@@ -580,7 +577,7 @@ ++gb->memory.dmaSource;
 	++gb->memory.dmaDest;
 	gb->memory.dmaRemaining = dmaRemaining - 1;
 	if (gb->memory.dmaRemaining) {
-		mTimingSchedule(timing, &gb->memory.dmaEvent, 4 - cyclesLate);
+		mTimingSchedule(timing, &gb->memory.dmaEvent, 4 * (2 - gb->doubleSpeed) - cyclesLate);
 	}
 }
 
@@ -594,7 +591,7 @@ ++gb->memory.hdmaDest;
 	--gb->memory.hdmaRemaining;
 	if (gb->memory.hdmaRemaining) {
 		mTimingDeschedule(timing, &gb->memory.hdmaEvent);
-		mTimingSchedule(timing, &gb->memory.hdmaEvent, 2 - cyclesLate);
+		mTimingSchedule(timing, &gb->memory.hdmaEvent, 4 - cyclesLate);
 	} else {
 		gb->cpuBlocked = false;
 		gb->memory.io[GB_REG_HDMA1] = gb->memory.hdmaSource >> 8;

@@ -175,8 +175,6 @@ gb->doubleSpeed = GBSerializedCpuFlagsGetDoubleSpeed(flags);
 	gb->cpu->halted = GBSerializedCpuFlagsGetHalted(flags);
 	gb->cpuBlocked = GBSerializedCpuFlagsGetBlocked(flags);
 
-	gb->audio.timingFactor = gb->doubleSpeed + 1;
-
 	LOAD_32LE(gb->cpu->cycles, 0, &state->cpu.cycles);
 	LOAD_32LE(gb->cpu->nextEvent, 0, &state->cpu.nextEvent);
 	gb->timing.root = NULL;

@@ -77,7 +77,7 @@ GBUpdateIRQs(sio->p);
 			sio->pendingSB = 0xFF;
 		}
 	} else {
-		mTimingSchedule(timing, &sio->event, sio->period);
+		mTimingSchedule(timing, &sio->event, sio->period * (2 - sio->p->doubleSpeed));
 	}
 }
 
@@ -93,7 +93,7 @@ sio->period = GBSIOCyclesPerTransfer[GBRegisterSCGetClockSpeed(sc)]; // TODO Shift Clock
 	if (GBRegisterSCIsEnable(sc)) {
 		mTimingDeschedule(&sio->p->timing, &sio->event);
 		if (GBRegisterSCIsShiftClock(sc)) {
-			mTimingSchedule(&sio->p->timing, &sio->event, sio->period);
+			mTimingSchedule(&sio->p->timing, &sio->event, sio->period * (2 - sio->p->doubleSpeed));
 			sio->remainingBits = 8;
 		}
 	}

@@ -128,7 +128,7 @@ break;
 	case TRANSFER_FINISHING:
 		// Finish the transfer
 		// We need to make sure the other GBs catch up so they don't get behind
-		node->nextEvent += node->d.p->period - 8; // Split the cycles to avoid waiting too long
+		node->nextEvent += node->d.p->period * (2 - node->d.p->p->doubleSpeed) - 8; // Split the cycles to avoid waiting too long
 #ifndef NDEBUG
 		ATOMIC_ADD(node->p->d.transferId, 1);
 #endif
@@ -208,7 +208,7 @@ static void _GBSIOLockstepNodeProcessEvents(struct mTiming* timing, void* user, uint32_t cyclesLate) {
 	struct GBSIOLockstepNode* node = user;
 	mLockstepLock(&node->p->d);
 	if (node->p->d.attached < 2) {
-		mTimingSchedule(timing, &node->event, (GBSIOCyclesPerTransfer[0] >> 1) - cyclesLate);
+		mTimingSchedule(timing, &node->event, (GBSIOCyclesPerTransfer[0] >> 1) * (2 - node->d.p->p->doubleSpeed) - cyclesLate);
 		mLockstepUnlock(&node->p->d);
 		return;
 	}

@@ -20,17 +20,18 @@ GBUpdateIRQs(timer->p);
 }
 
 static void _GBTimerDivIncrement(struct GBTimer* timer, uint32_t cyclesLate) {
-	while (timer->nextDiv >= GB_DMG_DIV_PERIOD) {
-		timer->nextDiv -= GB_DMG_DIV_PERIOD;
+	int tMultiplier = 2 - timer->p->doubleSpeed;
+	while (timer->nextDiv >= GB_DMG_DIV_PERIOD * tMultiplier) {
+		timer->nextDiv -= GB_DMG_DIV_PERIOD * tMultiplier;
 
 		// Make sure to trigger when the correct bit is a falling edge
 		if (timer->timaPeriod > 0 && (timer->internalDiv & (timer->timaPeriod - 1)) == timer->timaPeriod - 1) {
 			++timer->p->memory.io[GB_REG_TIMA];
 			if (!timer->p->memory.io[GB_REG_TIMA]) {
-				mTimingSchedule(&timer->p->timing, &timer->irq, 7 - ((timer->p->cpu->executionState - cyclesLate) & 3));
+				mTimingSchedule(&timer->p->timing, &timer->irq, 7 * tMultiplier - ((timer->p->cpu->executionState * tMultiplier - cyclesLate) & (3 * tMultiplier)));
 			}
 		}
-		unsigned timingFactor = 0x3FF >> !timer->p->doubleSpeed;
+		unsigned timingFactor = 0x1FF;
 		if ((timer->internalDiv & timingFactor) == timingFactor) {
 			GBAudioUpdateFrame(&timer->p->audio, &timer->p->timing);
 		}
@@ -52,7 +53,7 @@ }
 	if (timaToGo < divsToGo) {
 		divsToGo = timaToGo;
 	}
-	timer->nextDiv = GB_DMG_DIV_PERIOD * divsToGo;
+	timer->nextDiv = GB_DMG_DIV_PERIOD * divsToGo * (2 - timer->p->doubleSpeed);
 	mTimingSchedule(timing, &timer->event, timer->nextDiv - cyclesLate);
 }
 
@@ -66,7 +67,7 @@ timer->irq.name = "GB Timer IRQ";
 	timer->irq.callback = _GBTimerIRQ;
 	timer->event.priority = 0x21;
 
-	timer->nextDiv = GB_DMG_DIV_PERIOD; // TODO: GBC differences
+	timer->nextDiv = GB_DMG_DIV_PERIOD * 2;
 	timer->timaPeriod = 1024 >> 4;
 }
 
@@ -74,27 +75,27 @@ void GBTimerDivReset(struct GBTimer* timer) {
 	timer->nextDiv -= mTimingUntil(&timer->p->timing, &timer->event);
 	mTimingDeschedule(&timer->p->timing, &timer->event);
 	_GBTimerDivIncrement(timer, 0);
-	if (((timer->internalDiv << 1) | ((timer->nextDiv >> 3) & 1)) & timer->timaPeriod) {
+	int tMultiplier = 2 - timer->p->doubleSpeed;
+	if (((timer->internalDiv << 1) | ((timer->nextDiv >> (4 - timer->p->doubleSpeed)) & 1)) & timer->timaPeriod) {
 		++timer->p->memory.io[GB_REG_TIMA];
 		if (!timer->p->memory.io[GB_REG_TIMA]) {
-			mTimingSchedule(&timer->p->timing, &timer->irq, 7 - (timer->p->cpu->executionState & 3));
+			mTimingSchedule(&timer->p->timing, &timer->irq, (7 - (timer->p->cpu->executionState & 3)) * tMultiplier);
 		}
 	}
-	unsigned timingFactor = 0x400 >> !timer->p->doubleSpeed;
-	if (timer->internalDiv & timingFactor) {
+	if (timer->internalDiv & 0x200) {
 		GBAudioUpdateFrame(&timer->p->audio, &timer->p->timing);
 	}
 	timer->p->memory.io[GB_REG_DIV] = 0;
 	timer->internalDiv = 0;
-	timer->nextDiv = GB_DMG_DIV_PERIOD;
-	mTimingSchedule(&timer->p->timing, &timer->event, timer->nextDiv - ((timer->p->cpu->executionState + 1) & 3));
+	timer->nextDiv = GB_DMG_DIV_PERIOD * (2 - timer->p->doubleSpeed);
+	mTimingSchedule(&timer->p->timing, &timer->event, timer->nextDiv - ((timer->p->cpu->executionState + 1) & 3) * tMultiplier);
 }
 
 uint8_t GBTimerUpdateTAC(struct GBTimer* timer, GBRegisterTAC tac) {
 	if (GBRegisterTACIsRun(tac)) {
 		timer->nextDiv -= mTimingUntil(&timer->p->timing, &timer->event);
 		mTimingDeschedule(&timer->p->timing, &timer->event);
-		_GBTimerDivIncrement(timer, (timer->p->cpu->executionState + 2) & 3);
+		_GBTimerDivIncrement(timer, ((timer->p->cpu->executionState + 2) & 3) * (2 - timer->p->doubleSpeed));
 
 		switch (GBRegisterTACGetClock(tac)) {
 		case 0:
@@ -111,7 +112,7 @@ timer->timaPeriod = 256 >> 4;
 			break;
 		}
 
-		timer->nextDiv += GB_DMG_DIV_PERIOD;
+		timer->nextDiv += GB_DMG_DIV_PERIOD * (2 - timer->p->doubleSpeed);
 		mTimingSchedule(&timer->p->timing, &timer->event, timer->nextDiv);
 	} else {
 		timer->timaPeriod = 0;

@@ -253,7 +253,7 @@ video->p->memory.io[GB_REG_IF] |= (1 << GB_IRQ_VBLANK);
 	GBUpdateIRQs(video->p);
 	video->p->memory.io[GB_REG_STAT] = video->stat;
 	mTimingDeschedule(&video->p->timing, &video->modeEvent);
-	mTimingSchedule(&video->p->timing, &video->modeEvent, next);
+	mTimingSchedule(&video->p->timing, &video->modeEvent, next << 1);
 }
 
 void _endMode0(struct mTiming* timing, void* context, uint32_t cyclesLate) {
@@ -297,7 +297,7 @@ }
 
 	GBUpdateIRQs(video->p);
 	video->p->memory.io[GB_REG_STAT] = video->stat;
-	mTimingSchedule(timing, &video->modeEvent, (next << video->p->doubleSpeed) - cyclesLate);
+	mTimingSchedule(timing, &video->modeEvent, (next << 1) - cyclesLate);
 }
 
 void _endMode1(struct mTiming* timing, void* context, uint32_t cyclesLate) {
@@ -334,14 +334,14 @@ video->p->memory.io[GB_REG_IF] |= (1 << GB_IRQ_LCDSTAT);
 		GBUpdateIRQs(video->p);
 	}
 	video->p->memory.io[GB_REG_STAT] = video->stat;
-	mTimingSchedule(timing, &video->modeEvent, (next << video->p->doubleSpeed) - cyclesLate);
+	mTimingSchedule(timing, &video->modeEvent, (next << 1) - cyclesLate);
 }
 
 void _endMode2(struct mTiming* timing, void* context, uint32_t cyclesLate) {
 	struct GBVideo* video = context;
 	_cleanOAM(video, video->ly);
 	video->x = -(video->p->memory.io[GB_REG_SCX] & 7);
-	video->dotClock = mTimingCurrentTime(timing) - cyclesLate + 5 - (video->x << video->p->doubleSpeed);
+	video->dotClock = mTimingCurrentTime(timing) - cyclesLate + 10 - (video->x << 1);
 	int32_t next = GB_VIDEO_MODE_3_LENGTH_BASE + video->objMax * 6 - video->x;
 	video->mode = 3;
 	video->modeEvent.callback = _endMode3;
@@ -352,7 +352,7 @@ video->p->memory.io[GB_REG_IF] |= (1 << GB_IRQ_LCDSTAT);
 		GBUpdateIRQs(video->p);
 	}
 	video->p->memory.io[GB_REG_STAT] = video->stat;
-	mTimingSchedule(timing, &video->modeEvent, (next << video->p->doubleSpeed) - cyclesLate);
+	mTimingSchedule(timing, &video->modeEvent, (next << 1) - cyclesLate);
 }
 
 void _endMode3(struct mTiming* timing, void* context, uint32_t cyclesLate) {
@@ -375,18 +375,18 @@ }
 	video->p->memory.io[GB_REG_STAT] = video->stat;
 	// TODO: Cache SCX & 7 in case it changes
 	int32_t next = GB_VIDEO_MODE_0_LENGTH_BASE - video->objMax * 6 - (video->p->memory.io[GB_REG_SCX] & 7);
-	mTimingSchedule(timing, &video->modeEvent, (next << video->p->doubleSpeed) - cyclesLate);
+	mTimingSchedule(timing, &video->modeEvent, (next << 1) - cyclesLate);
 }
 
 void _updateFrameCount(struct mTiming* timing, void* context, uint32_t cyclesLate) {
 	UNUSED(cyclesLate);
 	struct GBVideo* video = context;
 	if (video->p->cpu->executionState != SM83_CORE_FETCH) {
-		mTimingSchedule(timing, &video->frameEvent, 4 - ((video->p->cpu->executionState + 1) & 3));
+		mTimingSchedule(timing, &video->frameEvent, (4 - ((video->p->cpu->executionState + 1) & 3)) * (2 - video->p->doubleSpeed));
 		return;
 	}
 	if (!GBRegisterLCDCIsEnable(video->p->memory.io[GB_REG_LCDC])) {
-		mTimingSchedule(timing, &video->frameEvent, GB_VIDEO_TOTAL_LENGTH);
+		mTimingSchedule(timing, &video->frameEvent, GB_VIDEO_TOTAL_LENGTH << 1);
 	}
 
 	--video->frameskipCounter;
@@ -424,7 +424,7 @@ if (video->mode != 3) {
 		return;
 	}
 	int oldX = video->x;
-	video->x = (int32_t) (mTimingCurrentTime(&video->p->timing) - cyclesLate - video->dotClock) >> video->p->doubleSpeed;
+	video->x = ((int32_t) (mTimingCurrentTime(&video->p->timing) - cyclesLate - video->dotClock)) >> 1;
 	if (video->x > GB_VIDEO_HORIZONTAL_PIXELS) {
 		video->x = GB_VIDEO_HORIZONTAL_PIXELS;
 	} else if (video->x < 0) {
@@ -444,7 +444,7 @@ video->mode = 2;
 		video->modeEvent.callback = _endMode2;
 		int32_t next = GB_VIDEO_MODE_2_LENGTH - 5; // TODO: Why is this fudge factor needed? Might be related to T-cycles for load/store differing
 		mTimingDeschedule(&video->p->timing, &video->modeEvent);
-		mTimingSchedule(&video->p->timing, &video->modeEvent, next << video->p->doubleSpeed);
+		mTimingSchedule(&video->p->timing, &video->modeEvent, next << 1);
 
 		video->ly = 0;
 		video->p->memory.io[GB_REG_LY] = 0;
@@ -471,7 +471,7 @@ video->renderer->writePalette(video->renderer, 0, video->dmgPalette[0]);
 	
 		mTimingDeschedule(&video->p->timing, &video->modeEvent);
 		mTimingDeschedule(&video->p->timing, &video->frameEvent);
-		mTimingSchedule(&video->p->timing, &video->frameEvent, GB_VIDEO_TOTAL_LENGTH);
+		mTimingSchedule(&video->p->timing, &video->frameEvent, GB_VIDEO_TOTAL_LENGTH << 1);
 	}
 	video->p->memory.io[GB_REG_STAT] = video->stat;
 }

@@ -61,6 +61,7 @@ cpu->pc = 0;
 
 	cpu->instruction = 0;
 
+	cpu->tMultiplier = 2;
 	cpu->cycles = 0;
 	cpu->nextEvent = 0;
 	cpu->executionState = SM83_CORE_FETCH;
@@ -102,7 +103,7 @@ cpu->instruction = _SM83InstructionIRQDelay;
 }
 
 static void _SM83Step(struct SM83Core* cpu) {
-	++cpu->cycles;
+	cpu->cycles += cpu->tMultiplier;
 	enum SM83ExecutionState state = cpu->executionState;
 	cpu->executionState = SM83_CORE_IDLE_0;
 	switch (state) {
@@ -147,23 +148,31 @@ break;
 	}
 }
 
-void SM83Tick(struct SM83Core* cpu) {
-	while (cpu->cycles >= cpu->nextEvent) {
-		cpu->irqh.processEvents(cpu);
-	}
+static inline bool _SM83TickInternal(struct SM83Core* cpu) {
+	bool running = true;
 	_SM83Step(cpu);
-	if (cpu->cycles + 2 >= cpu->nextEvent) {
+	int t = cpu->tMultiplier;
+	if (cpu->cycles + t * 2 >= cpu->nextEvent) {
 		int32_t diff = cpu->nextEvent - cpu->cycles;
 		cpu->cycles = cpu->nextEvent;
-		cpu->executionState += diff;
+		cpu->executionState += diff >> (t - 1); // NB: This assumes tMultiplier is either 1 or 2
 		cpu->irqh.processEvents(cpu);
-		cpu->cycles += SM83_CORE_EXECUTE - cpu->executionState;
+		cpu->cycles += (SM83_CORE_EXECUTE - cpu->executionState) * t;
+		running = false;
 	} else {
-		cpu->cycles += 2;
+		cpu->cycles += t * 2;
 	}
 	cpu->executionState = SM83_CORE_FETCH;
 	cpu->instruction(cpu);
-	++cpu->cycles;
+	cpu->cycles += t;
+	return running;
+}
+
+void SM83Tick(struct SM83Core* cpu) {
+	while (cpu->cycles >= cpu->nextEvent) {
+		cpu->irqh.processEvents(cpu);
+	}
+	_SM83TickInternal(cpu);
 }
 
 void SM83Run(struct SM83Core* cpu) {
@@ -173,19 +182,6 @@ if (cpu->cycles >= cpu->nextEvent) {
 			cpu->irqh.processEvents(cpu);
 			break;
 		}
-		_SM83Step(cpu);
-		if (cpu->cycles + 2 >= cpu->nextEvent) {
-			int32_t diff = cpu->nextEvent - cpu->cycles;
-			cpu->cycles = cpu->nextEvent;
-			cpu->executionState += diff;
-			cpu->irqh.processEvents(cpu);
-			cpu->cycles += SM83_CORE_EXECUTE - cpu->executionState;
-			running = false;
-		} else {
-			cpu->cycles += 2;
-		}
-		cpu->executionState = SM83_CORE_FETCH;
-		cpu->instruction(cpu);
-		++cpu->cycles;
+		running = _SM83TickInternal(cpu) && running;
 	}
 }