From 0308f136c7e195bd2e245dd3676d559b97595ceb Mon Sep 17 00:00:00 2001
From: Jeffrey Pfau <jeffrey@endrift.com>
Date: Tue, 18 Nov 2014 04:29:16 -0800
Subject: [PATCH] GBA Memory: partially unroll loops for LDM/STM and add
 UNLIKELYs

---
 src/gba/gba-memory.c | 50 ++++++++++++++++++++++++++++++++++++++++----
 1 file changed, 46 insertions(+), 4 deletions(-)

diff --git a/src/gba/gba-memory.c b/src/gba/gba-memory.c
index ad359a413..fe5c65c7c 100644
--- a/src/gba/gba-memory.c
+++ b/src/gba/gba-memory.c
@@ -669,14 +669,35 @@ void GBAStore8(struct ARMCore* cpu, uint32_t address, int8_t value, int* cycleCo
 }
 
 #define LDM_LOOP(LDM) \
-	for (i = 0; i < 16; ++i) { \
-		if (mask & (1 << i)) { \
+	for (i = 0; i < 16; i += 4) { \
+		if (UNLIKELY(mask & (1 << i))) { \
 			LDM; \
 			waitstatesRegion = memory->waitstatesSeq32; \
 			cpu->gprs[i] = value; \
 			++wait; \
 			address += 4; \
 		} \
+		if (UNLIKELY(mask & (2 << i))) { \
+			LDM; \
+			waitstatesRegion = memory->waitstatesSeq32; \
+			cpu->gprs[i + 1] = value; \
+			++wait; \
+			address += 4; \
+		} \
+		if (UNLIKELY(mask & (4 << i))) { \
+			LDM; \
+			waitstatesRegion = memory->waitstatesSeq32; \
+			cpu->gprs[i + 2] = value; \
+			++wait; \
+			address += 4; \
+		} \
+		if (UNLIKELY(mask & (8 << i))) { \
+			LDM; \
+			waitstatesRegion = memory->waitstatesSeq32; \
+			cpu->gprs[i + 3] = value; \
+			++wait; \
+			address += 4; \
+		} \
 	}
 
 uint32_t GBALoadMultiple(struct ARMCore* cpu, uint32_t address, int mask, enum LSMDirection direction, int* cycleCounter) {
@@ -757,14 +778,35 @@ uint32_t GBALoadMultiple(struct ARMCore* cpu, uint32_t address, int mask, enum L
 }
 
 #define STM_LOOP(STM) \
-	for (i = 0; i < 16; ++i) { \
-		if (mask & (1 << i)) { \
+	for (i = 0; i < 16; i += 4) { \
+		if (UNLIKELY(mask & (1 << i))) { \
 			value = cpu->gprs[i]; \
 			STM; \
 			waitstatesRegion = memory->waitstatesSeq32; \
 			++wait; \
 			address += 4; \
 		} \
+		if (UNLIKELY(mask & (2 << i))) { \
+			value = cpu->gprs[i + 1]; \
+			STM; \
+			waitstatesRegion = memory->waitstatesSeq32; \
+			++wait; \
+			address += 4; \
+		} \
+		if (UNLIKELY(mask & (4 << i))) { \
+			value = cpu->gprs[i + 2]; \
+			STM; \
+			waitstatesRegion = memory->waitstatesSeq32; \
+			++wait; \
+			address += 4; \
+		} \
+		if (UNLIKELY(mask & (8 << i))) { \
+			value = cpu->gprs[i + 3]; \
+			STM; \
+			waitstatesRegion = memory->waitstatesSeq32; \
+			++wait; \
+			address += 4; \
+		} \
 	}
 
 uint32_t GBAStoreMultiple(struct ARMCore* cpu, uint32_t address, int mask, enum LSMDirection direction, int* cycleCounter) {