M CHANGES → CHANGES

@@ -32,6 +32,7 @@ - 3DS: Use system font for menus
  - PSP2: Use system font for menus
  - All: Faster memory read/write
  - Qt: Make audio channel/video layer options shortcut mappable
+ - GBA Memory: Optimize stalling behavior
 
 0.4.1: (2016-07-11)
 Bugfixes:

M src/gba/memory.c → src/gba/memory.c

@@ -1748,30 +1748,32 @@ // The wait is the stall
 		return wait;
 	}
 
+	int32_t previousLoads = 0;
+
+	// Don't prefetch too much if we're overlapping with a previous prefetch
+	uint32_t dist = (memory->lastPrefetchedPc - cpu->gprs[ARM_PC]) >> 1;
+	if (dist < 8) {
+		previousLoads = dist;
+	}
+
 	int32_t s = cpu->memory.activeSeqCycles16 + 1;
 	int32_t n2s = cpu->memory.activeNonseqCycles16 - cpu->memory.activeSeqCycles16 + 1;
 
 	// Figure out how many sequential loads we can jam in
 	int32_t stall = s;
 	int32_t loads = 1;
-	int32_t previousLoads = 0;
 
-	// Don't prefetch too much if we're overlapping with a previous prefetch
-	uint32_t dist = (memory->lastPrefetchedPc - cpu->gprs[ARM_PC]) >> 1;
-	if (dist < memory->lastPrefetchedLoads) {
-		previousLoads = dist;
-	}
-	while (stall < wait) {
-		stall += s;
-		++loads;
-	}
-	if (loads + previousLoads > 8) {
-		int diff = (loads + previousLoads) - 8;
-		loads -= diff;
-		stall -= s * diff;
-	} else if (stall > wait && loads == 1) {
+	if (stall > wait && !previousLoads) {
 		// We might need to stall a bit extra if we haven't finished the first S cycle
 		wait = stall;
+	} else {
+		while (stall < wait) {
+			stall += s;
+			++loads;
+		}
+		if (loads + previousLoads > 8) {
+			loads = 8 - previousLoads;
+		}
 	}
 	// This instruction used to have an N, convert it to an S.
 	wait -= n2s;