From 4a5a25e90c5ab0909a283df8a9cb10d9807f6e7a Mon Sep 17 00:00:00 2001
From: Vicki Pfau <vi@endrift.com>
Date: Fri, 18 Oct 2024 02:58:19 -0700
Subject: [PATCH] GBA DMA: Cache cycle estimation on first DMA

---
 include/mgba/internal/gba/dma.h |  2 ++
 src/gba/audio.c                 |  1 +
 src/gba/dma.c                   | 35 ++++++++++++++++++++++++++++-----
 src/gba/io.c                    |  1 +
 src/gba/memory.c                |  4 ++++
 5 files changed, 38 insertions(+), 5 deletions(-)

diff --git a/include/mgba/internal/gba/dma.h b/include/mgba/internal/gba/dma.h
index 192b3995a..58d590948 100644
--- a/include/mgba/internal/gba/dma.h
+++ b/include/mgba/internal/gba/dma.h
@@ -48,6 +48,7 @@ struct GBADMA {
 	uint32_t nextDest;
 	int32_t nextCount;
 	uint32_t when;
+	int32_t cycles;
 };
 
 struct GBA;
@@ -65,6 +66,7 @@ void GBADMARunHblank(struct GBA* gba, int32_t cycles);
 void GBADMARunVblank(struct GBA* gba, int32_t cycles);
 void GBADMARunDisplayStart(struct GBA* gba, int32_t cycles);
 void GBADMAUpdate(struct GBA* gba);
+void GBADMARecalculateCycles(struct GBA* gba);
 
 CXX_GUARD_END
 
diff --git a/src/gba/audio.c b/src/gba/audio.c
index a6ebafbae..ac394e995 100644
--- a/src/gba/audio.c
+++ b/src/gba/audio.c
@@ -309,6 +309,7 @@ void GBAAudioSampleFIFO(struct GBAAudio* audio, int fifoId, int32_t cycles) {
 		if (GBADMARegisterGetTiming(dma->reg) == GBA_DMA_TIMING_CUSTOM) {
 			dma->when = mTimingCurrentTime(&audio->p->timing) - cycles;
 			dma->nextCount = 4;
+			GBADMARecalculateCycles(audio->p);
 			GBADMASchedule(audio->p, channel->dmaSource, dma);
 		}
 	}
diff --git a/src/gba/dma.c b/src/gba/dma.c
index 290341361..5971b1b03 100644
--- a/src/gba/dma.c
+++ b/src/gba/dma.c
@@ -256,15 +256,21 @@ void GBADMAService(struct GBA* gba, int number, struct GBADMA* info) {
 	if (info->count == info->nextCount) {
 		if (width == 4) {
 			cycles += memory->waitstatesNonseq32[sourceRegion] + memory->waitstatesNonseq32[destRegion];
+			info->cycles = memory->waitstatesSeq32[sourceRegion] + memory->waitstatesSeq32[destRegion];
 		} else {
 			cycles += memory->waitstatesNonseq16[sourceRegion] + memory->waitstatesNonseq16[destRegion];
+			info->cycles = memory->waitstatesSeq16[sourceRegion] + memory->waitstatesSeq16[destRegion];
 		}
 	} else {
-		if (width == 4) {
-			cycles += memory->waitstatesSeq32[sourceRegion] + memory->waitstatesSeq32[destRegion];
-		} else {
-			cycles += memory->waitstatesSeq16[sourceRegion] + memory->waitstatesSeq16[destRegion];
+		// Crossed region boundary; recalculate cached cycles
+		if (UNLIKELY(!(source & 0x00FFFFFC) || !(dest & 0x00FFFFFC))) {
+			if (width == 4) {
+				info->cycles = memory->waitstatesSeq32[sourceRegion] + memory->waitstatesSeq32[destRegion];
+			} else {
+				info->cycles = memory->waitstatesSeq16[sourceRegion] + memory->waitstatesSeq16[destRegion];
+			}
 		}
+		cycles += info->cycles;
 	}
 	info->when += cycles;
 
@@ -281,7 +287,7 @@ void GBADMAService(struct GBA* gba, int number, struct GBADMA* info) {
 			memory->dmaTransferRegister = cpu->memory.load16(cpu, source, 0);
 			memory->dmaTransferRegister |= memory->dmaTransferRegister << 16;
 		}
-		if (destRegion == GBA_REGION_ROM2_EX) {
+		if (UNLIKELY(destRegion == GBA_REGION_ROM2_EX)) {
 			if (memory->savedata.type == GBA_SAVEDATA_AUTODETECT) {
 				mLOG(GBA_MEM, INFO, "Detected EEPROM savegame");
 				GBASavedataInitEEPROM(&memory->savedata);
@@ -327,3 +333,22 @@ void GBADMAService(struct GBA* gba, int number, struct GBADMA* info) {
 	}
 	GBADMAUpdate(gba);
 }
+
+void GBADMARecalculateCycles(struct GBA* gba) {
+	int i;
+	for (i = 0; i < 4; ++i) {
+		struct GBADMA* dma = &gba->memory.dma[i];
+		if (!GBADMARegisterIsEnable(dma->reg)) {
+			continue;
+		}
+
+		uint32_t width = GBADMARegisterGetWidth(dma->reg);
+		uint32_t sourceRegion = dma->nextSource >> BASE_OFFSET;
+		uint32_t destRegion = dma->nextDest >> BASE_OFFSET;
+		if (width) {
+			dma->cycles = gba->memory.waitstatesSeq32[sourceRegion] + gba->memory.waitstatesSeq32[destRegion];
+		} else {
+			dma->cycles = gba->memory.waitstatesSeq16[sourceRegion] + gba->memory.waitstatesSeq16[destRegion];
+		}
+	}
+}
diff --git a/src/gba/io.c b/src/gba/io.c
index 8a4ea7fc2..cb6058f7f 100644
--- a/src/gba/io.c
+++ b/src/gba/io.c
@@ -1079,6 +1079,7 @@ void GBAIODeserialize(struct GBA* gba, const struct GBASerializedState* state) {
 	LOAD_32(gba->dmaPC, 0, &state->dmaBlockPC);
 	LOAD_32(gba->bus, 0, &state->bus);
 
+	GBADMARecalculateCycles(gba);
 	GBADMAUpdate(gba);
 	GBAHardwareDeserialize(&gba->memory.hw, state);
 }
diff --git a/src/gba/memory.c b/src/gba/memory.c
index 3243fc5a1..c818c75c8 100644
--- a/src/gba/memory.c
+++ b/src/gba/memory.c
@@ -1734,6 +1734,10 @@ void GBAAdjustWaitstates(struct GBA* gba, uint16_t parameters) {
 			STORE_32(memory->agbPrintFuncBackup, AGB_PRINT_FLUSH_ADDR | base, memory->rom);
 		}
 	}
+
+	if (gba->performingDMA) {
+		GBADMARecalculateCycles(gba);
+	}
 }
 
 void GBAAdjustEWRAMWaitstates(struct GBA* gba, uint16_t parameters) {