diff --git a/src/gba/gba.c b/src/gba/gba.c
index 97b4668fe..e3023ebc8 100644
--- a/src/gba/gba.c
+++ b/src/gba/gba.c
@@ -87,6 +87,7 @@ static void GBAInit(struct ARMCore* cpu, struct ARMComponent* component) {
 	gba->idleOptimization = IDLE_LOOP_REMOVE;
 	gba->idleLoop = IDLE_LOOP_NONE;
 	gba->lastJump = 0;
+	gba->haltPending = false;
 	gba->idleDetectionStep = 0;
 	gba->idleDetectionFailures = 0;
 	gba->performingDMA = false;
diff --git a/src/gba/gba.h b/src/gba/gba.h
index 3594b4ab8..1366fd1e6 100644
--- a/src/gba/gba.h
+++ b/src/gba/gba.h
@@ -157,6 +157,7 @@ struct GBA {
 	enum GBAIdleLoopOptimization idleOptimization;
 	uint32_t idleLoop;
 	uint32_t lastJump;
+	bool haltPending;
 	int idleDetectionStep;
 	int idleDetectionFailures;
 	int32_t cachedRegisters[16];
diff --git a/src/gba/io.c b/src/gba/io.c
index 76305cb1d..65f178fde 100644
--- a/src/gba/io.c
+++ b/src/gba/io.c
@@ -567,7 +567,7 @@ void GBAIOWrite32(struct GBA* gba, uint32_t address, uint32_t value) {
 }
 
 uint16_t GBAIORead(struct GBA* gba, uint32_t address) {
-	gba->lastJump = -1; // IO reads need to invalidate detected idle loops
+	gba->haltPending = false; // IO reads need to invalidate detected idle loops
 	switch (address) {
 	case REG_TM0CNT_LO:
 		GBATimerUpdateRegister(gba, 0);
diff --git a/src/gba/memory.c b/src/gba/memory.c
index bc5a32e20..eddc33601 100644
--- a/src/gba/memory.c
+++ b/src/gba/memory.c
@@ -198,7 +198,12 @@ static void GBASetActiveRegion(struct ARMCore* cpu, uint32_t address) {
 	int newRegion = address >> BASE_OFFSET;
 	if (gba->idleOptimization >= IDLE_LOOP_REMOVE && memory->activeRegion != REGION_BIOS) {
 		if (address == gba->idleLoop) {
-			GBAHalt(gba);
+			if (gba->haltPending) {
+				gba->haltPending = false;
+				GBAHalt(gba);
+			} else {
+				gba->haltPending = true;
+			}
 		} else if (gba->idleOptimization >= IDLE_LOOP_DETECT && newRegion == memory->activeRegion) {
 			if (address == gba->lastJump) {
 				switch (gba->idleDetectionStep) {