diff --git a/src/gba/gba.c b/src/gba/gba.c
index f639e6063..35d5d88fd 100644
--- a/src/gba/gba.c
+++ b/src/gba/gba.c
@@ -452,8 +452,8 @@ void GBATimerUpdateRegister(struct GBA* gba, int timer) {
 	struct GBATimer* currentTimer = &gba->timers[timer];
 	if (currentTimer->enable && !currentTimer->countUp) {
 		int32_t prefetchSkew = 0;
-		if ((gba->memory.lastPrefetchedPc - gba->cpu->gprs[ARM_PC]) < gba->memory.lastPrefetchedLoads * WORD_SIZE_THUMB) {
-			prefetchSkew = (gba->memory.lastPrefetchedPc - gba->cpu->gprs[ARM_PC]) / gba->cpu->memory.activeSeqCycles16;
+		if (gba->memory.lastPrefetchedPc - gba->memory.lastPrefetchedLoads * WORD_SIZE_THUMB >= (uint32_t) gba->cpu->gprs[ARM_PC]) {
+			prefetchSkew = (gba->memory.lastPrefetchedPc - gba->cpu->gprs[ARM_PC]) * (gba->cpu->memory.activeSeqCycles16 + 1) / WORD_SIZE_THUMB;
 		}
 		// Reading this takes two cycles (1N+1I), so let's remove them preemptively
 		gba->memory.io[(REG_TM0CNT_LO + (timer << 2)) >> 1] = currentTimer->oldReload + ((gba->cpu->cycles - currentTimer->lastEvent - 2 + prefetchSkew) >> currentTimer->prescaleBits);
diff --git a/src/gba/memory.c b/src/gba/memory.c
index 6f5fa9fbb..704f217c0 100644
--- a/src/gba/memory.c
+++ b/src/gba/memory.c
@@ -235,6 +235,8 @@ static void GBASetActiveRegion(struct ARMCore* cpu, uint32_t address) {
 	}
 
 	gba->lastJump = address;
+	memory->lastPrefetchedPc = 0;
+	memory->lastPrefetchedLoads = 0;
 	if (newRegion == memory->activeRegion && (newRegion < REGION_CART0 || (address & (SIZE_CART0 - 1)) < memory->romSize)) {
 		return;
 	}
@@ -1541,10 +1543,10 @@ int32_t GBAMemoryStall(struct ARMCore* cpu, int32_t wait) {
 	int32_t previousLoads = 0;
 
 	// Don't prefetch too much if we're overlapping with a previous prefetch
-	if ((memory->lastPrefetchedPc - cpu->gprs[ARM_PC]) < memory->lastPrefetchedLoads * WORD_SIZE_THUMB) {
+	if (UNLIKELY((memory->lastPrefetchedPc - cpu->gprs[ARM_PC]) < memory->lastPrefetchedLoads * WORD_SIZE_THUMB)) {
 		previousLoads = (memory->lastPrefetchedPc - cpu->gprs[ARM_PC]) >> 1;
 	}
-	while (stall < wait && loads + previousLoads < 8) {
+	while (stall < wait && LIKELY(loads + previousLoads < 8)) {
 		stall += s;
 		++loads;
 	}