diff --git a/CHANGES b/CHANGES
index 4b80c29dd..83fa6f8da 100644
--- a/CHANGES
+++ b/CHANGES
@@ -10,6 +10,7 @@ Bugfixes:
  - GBA BIOS: Fix HLE Lz77 and RL functions to properly account for width and invalid addresses
  - ARM7: Fix LDM writeback to a register already written
  - GBA Memory: Don't call into GPIO write calls if GPIO devices are absent
+ - ARM7: Extend prefetch by one stage
 Misc:
  - Qt: Disable sync to video by default
  - GBA: Exit cleanly on FATAL if the port supports it
diff --git a/src/arm/arm.c b/src/arm/arm.c
index f2b71bc61..17aa099b8 100644
--- a/src/arm/arm.c
+++ b/src/arm/arm.c
@@ -176,9 +176,10 @@ void ARMRaiseSWI(struct ARMCore* cpu) {
 }
 
 static inline void ARMStep(struct ARMCore* cpu) {
-	uint32_t opcode = cpu->prefetch;
-	LOAD_32(cpu->prefetch, cpu->gprs[ARM_PC] & cpu->memory.activeMask, cpu->memory.activeRegion);
+	uint32_t opcode = cpu->prefetch[0];
+	cpu->prefetch[0] = cpu->prefetch[1];
 	cpu->gprs[ARM_PC] += WORD_SIZE_ARM;
+	LOAD_32(cpu->prefetch[1], cpu->gprs[ARM_PC] & cpu->memory.activeMask, cpu->memory.activeRegion);
 
 	unsigned condition = opcode >> 28;
 	if (condition != 0xE) {
@@ -239,9 +240,10 @@ static inline void ARMStep(struct ARMCore* cpu) {
 }
 
 static inline void ThumbStep(struct ARMCore* cpu) {
-	uint32_t opcode = cpu->prefetch;
-	LOAD_16(cpu->prefetch, cpu->gprs[ARM_PC] & cpu->memory.activeMask, cpu->memory.activeRegion);
+	uint32_t opcode = cpu->prefetch[0];
+	cpu->prefetch[0] = cpu->prefetch[1];
 	cpu->gprs[ARM_PC] += WORD_SIZE_THUMB;
+	LOAD_16(cpu->prefetch[1], cpu->gprs[ARM_PC] & cpu->memory.activeMask, cpu->memory.activeRegion);
 	ThumbInstruction instruction = _thumbTable[opcode >> 6];
 	instruction(cpu, opcode);
 }
diff --git a/src/arm/arm.h b/src/arm/arm.h
index 1140b2237..23efd086a 100644
--- a/src/arm/arm.h
+++ b/src/arm/arm.h
@@ -149,7 +149,7 @@ struct ARMCore {
 	int32_t shifterOperand;
 	int32_t shifterCarryOut;
 
-	uint32_t prefetch;
+	uint32_t prefetch[2];
 	enum ExecutionMode executionMode;
 	enum PrivilegeMode privilegeMode;
 
diff --git a/src/arm/isa-inlines.h b/src/arm/isa-inlines.h
index 1d72cd62f..6c2327a82 100644
--- a/src/arm/isa-inlines.h
+++ b/src/arm/isa-inlines.h
@@ -52,15 +52,17 @@
 #define ARM_WRITE_PC \
 	cpu->gprs[ARM_PC] = (cpu->gprs[ARM_PC] & -WORD_SIZE_ARM); \
 	cpu->memory.setActiveRegion(cpu, cpu->gprs[ARM_PC]); \
-	LOAD_32(cpu->prefetch, cpu->gprs[ARM_PC] & cpu->memory.activeMask, cpu->memory.activeRegion); \
+	LOAD_32(cpu->prefetch[0], cpu->gprs[ARM_PC] & cpu->memory.activeMask, cpu->memory.activeRegion); \
 	cpu->gprs[ARM_PC] += WORD_SIZE_ARM; \
+	LOAD_32(cpu->prefetch[1], cpu->gprs[ARM_PC] & cpu->memory.activeMask, cpu->memory.activeRegion); \
 	currentCycles += 2 + cpu->memory.activeUncachedCycles32 + cpu->memory.activeSeqCycles32;
 
 #define THUMB_WRITE_PC \
 	cpu->gprs[ARM_PC] = (cpu->gprs[ARM_PC] & -WORD_SIZE_THUMB); \
 	cpu->memory.setActiveRegion(cpu, cpu->gprs[ARM_PC]); \
-	LOAD_16(cpu->prefetch, cpu->gprs[ARM_PC] & cpu->memory.activeMask, cpu->memory.activeRegion); \
+	LOAD_16(cpu->prefetch[0], cpu->gprs[ARM_PC] & cpu->memory.activeMask, cpu->memory.activeRegion); \
 	cpu->gprs[ARM_PC] += WORD_SIZE_THUMB; \
+	LOAD_16(cpu->prefetch[1], cpu->gprs[ARM_PC] & cpu->memory.activeMask, cpu->memory.activeRegion); \
 	currentCycles += 2 + cpu->memory.activeUncachedCycles16 + cpu->memory.activeSeqCycles16;
 
 static inline int _ARMModeHasSPSR(enum PrivilegeMode mode) {
diff --git a/src/gba/gba-memory.c b/src/gba/gba-memory.c
index d3b092a1a..b3293879f 100644
--- a/src/gba/gba-memory.c
+++ b/src/gba/gba-memory.c
@@ -14,7 +14,7 @@
 #include "util/memory.h"
 
 static uint32_t _popcount32(unsigned bits);
-static uint32_t _deadbeef = 0xDEADBEEF;
+static uint32_t _deadbeef[2] = { 0xDEADBEEF, 0xFEEDFACE };
 
 static void GBASetActiveRegion(struct ARMCore* cpu, uint32_t region);
 static void GBAMemoryServiceDMA(struct GBA* gba, int number, struct GBADMA* info);
@@ -128,7 +128,7 @@ static void GBASetActiveRegion(struct ARMCore* cpu, uint32_t address) {
 		return;
 	}
 	if (memory->activeRegion == REGION_BIOS) {
-		memory->biosPrefetch = cpu->prefetch;
+		memory->biosPrefetch = cpu->prefetch[0];
 	}
 	memory->activeRegion = newRegion;
 	switch (address & ~OFFSET_MASK) {
@@ -158,7 +158,7 @@ static void GBASetActiveRegion(struct ARMCore* cpu, uint32_t address) {
 		cpu->memory.activeMask = SIZE_CART0 - 1;
 		break;
 	default:
-		cpu->memory.activeRegion = &_deadbeef;
+		cpu->memory.activeRegion = _deadbeef;
 		cpu->memory.activeMask = 0;
 		GBALog(gba, GBA_LOG_FATAL, "Jumped to invalid address");
 		break;
@@ -176,7 +176,7 @@ static void GBASetActiveRegion(struct ARMCore* cpu, uint32_t address) {
 	if (cpu->cycles >= cpu->nextEvent) { \
 		value = gba->bus; \
 	} else { \
-		value = cpu->prefetch; \
+		value = cpu->prefetch[0]; \
 		if (cpu->executionMode == MODE_THUMB) { \
 			value |= value << 16; \
 		} \
@@ -303,7 +303,7 @@ int16_t GBALoad16(struct ARMCore* cpu, uint32_t address, int* cycleCounter) {
 			if (cpu->cycles >= cpu->nextEvent) {
 				value = gba->bus;
 			} else {
-				value = cpu->prefetch;
+				value = cpu->prefetch[0];
 			}
 		}
 		break;
@@ -359,7 +359,7 @@ int16_t GBALoad16(struct ARMCore* cpu, uint32_t address, int* cycleCounter) {
 		if (cpu->cycles >= cpu->nextEvent) {
 			value = gba->bus;
 		} else {
-			value = cpu->prefetch;
+			value = cpu->prefetch[0];
 		}
 		break;
 	}
@@ -396,7 +396,7 @@ int8_t GBALoad8(struct ARMCore* cpu, uint32_t address, int* cycleCounter) {
 			if (cpu->cycles >= cpu->nextEvent) {
 				value = gba->bus;
 			} else {
-				value = cpu->prefetch;
+				value = cpu->prefetch[0];
 			}
 		}
 		break;
@@ -454,7 +454,7 @@ int8_t GBALoad8(struct ARMCore* cpu, uint32_t address, int* cycleCounter) {
 		if (cpu->cycles >= cpu->nextEvent) {
 			value = gba->bus;
 		} else {
-			value = cpu->prefetch;
+			value = cpu->prefetch[0];
 		}
 		break;
 	}
diff --git a/src/gba/gba-serialize.c b/src/gba/gba-serialize.c
index 61c716da3..ccf0b708f 100644
--- a/src/gba/gba-serialize.c
+++ b/src/gba/gba-serialize.c
@@ -82,10 +82,12 @@ void GBADeserialize(struct GBA* gba, struct GBASerializedState* state) {
 	gba->cpu->memory.setActiveRegion(gba->cpu, gba->cpu->gprs[ARM_PC]);
 	if (gba->cpu->cpsr.t) {
 		gba->cpu->executionMode = MODE_THUMB;
-		LOAD_16(gba->cpu->prefetch, (gba->cpu->gprs[ARM_PC] - WORD_SIZE_THUMB) & gba->cpu->memory.activeMask, gba->cpu->memory.activeRegion);
+		LOAD_16(gba->cpu->prefetch[0], (gba->cpu->gprs[ARM_PC] - WORD_SIZE_THUMB) & gba->cpu->memory.activeMask, gba->cpu->memory.activeRegion);
+		LOAD_16(gba->cpu->prefetch[1], (gba->cpu->gprs[ARM_PC]) & gba->cpu->memory.activeMask, gba->cpu->memory.activeRegion);
 	} else {
 		gba->cpu->executionMode = MODE_ARM;
-		LOAD_32(gba->cpu->prefetch, (gba->cpu->gprs[ARM_PC] - WORD_SIZE_ARM) & gba->cpu->memory.activeMask, gba->cpu->memory.activeRegion);
+		LOAD_32(gba->cpu->prefetch[0], (gba->cpu->gprs[ARM_PC] - WORD_SIZE_ARM) & gba->cpu->memory.activeMask, gba->cpu->memory.activeRegion);
+		LOAD_32(gba->cpu->prefetch[1], (gba->cpu->gprs[ARM_PC]) & gba->cpu->memory.activeMask, gba->cpu->memory.activeRegion);
 	}
 
 	GBAMemoryDeserialize(&gba->memory, state);
diff --git a/src/gba/gba.c b/src/gba/gba.c
index 541f1ac2f..643782aae 100644
--- a/src/gba/gba.c
+++ b/src/gba/gba.c
@@ -227,9 +227,9 @@ static void GBAProcessEvents(struct ARMCore* cpu) {
 		int32_t nextEvent = INT_MAX;
 		int32_t testEvent;
 
-		gba->bus = cpu->prefetch;
+		gba->bus = cpu->prefetch[0];
 		if (cpu->executionMode == MODE_THUMB) {
-			gba->bus |= cpu->prefetch << 16;
+			gba->bus |= cpu->prefetch[0] << 16;
 		}
 
 		if (gba->springIRQ) {