M CHANGES → CHANGES

@@ -92,7 +92,6 @@ - ARM: Partially fix LDM/STM writeback with empty register list
  - ARM: Fix stepping when events are pending
  - GBA DMA: Fix case where DMAs could get misaligned (fixes mgba.io/i/1092)
  - GBA Memory: Fix open bus from IWRAM (fixes mgba.io/i/1575)
- - GBA Memory: Improve gamepak prefetch timing
  - GBA Timers: Improve timer startup emulation
  - GBA Video: Fix OpenGL renderer 512x512 backgrounds (fixes mgba.io/i/1572)
  - GBA Video: Fix BLDY for semitransparent sprite on non-target-2 backgrounds

M src/gba/memory.c → src/gba/memory.c

@@ -1622,19 +1622,17 @@ previousLoads = dist >> 1;
 		maxLoads -= previousLoads;
 	}
 
-	int32_t s = cpu->memory.activeSeqCycles16;
+	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 + 1;
+	int32_t stall = s;
 	int32_t loads = 1;
 
 	while (stall < wait && loads < maxLoads) {
 		stall += s;
 		++loads;
 	}
-	memory->lastPrefetchedPc = cpu->gprs[ARM_PC] + WORD_SIZE_THUMB * (loads + previousLoads - 1);
-
 	if (stall > wait) {
 		// The wait cannot take less time than the prefetch stalls
 		wait = stall;
@@ -1643,9 +1641,10 @@
 	// This instruction used to have an N, convert it to an S.
 	wait -= n2s;
 
+	memory->lastPrefetchedPc = cpu->gprs[ARM_PC] + WORD_SIZE_THUMB * (loads + previousLoads - 1);
+
 	// The next |loads|S waitstates disappear entirely, so long as they're all in a row
-	wait -= stall - 1;
-
+	cpu->cycles -= (s - 1) * loads;
 	return wait;
 }