diff --git a/CHANGES b/CHANGES
index 09eb2956d..72e59f0d8 100644
--- a/CHANGES
+++ b/CHANGES
@@ -41,6 +41,7 @@ Bugfixes:
  - GBA Memory: Fix copy-on-write memory leak
  - Core: Fix ROM patches not being unloaded when disabled (fixes mgba.io/i/962)
  - GBA I/O: Fix writing to DISPCNT CGB flag (fixes mgba.io/i/902)
+ - GBA Memory: Partially revert prefetch changes (fixes mgba.io/i/840)
 Misc:
  - GBA Timer: Use global cycles for timers
  - GBA: Extend oddly-sized ROMs to full address space (fixes mgba.io/i/722)
diff --git a/src/gba/memory.c b/src/gba/memory.c
index 704a660e7..54e8ef8fd 100644
--- a/src/gba/memory.c
+++ b/src/gba/memory.c
@@ -1589,8 +1589,8 @@ int32_t GBAMemoryStall(struct ARMCore* cpu, int32_t wait) {
 		maxLoads -= previousLoads;
 	}
 
-	int32_t s = cpu->memory.activeSeqCycles16;
-	int32_t n2s = cpu->memory.activeNonseqCycles16 - 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;
@@ -1611,7 +1611,7 @@ int32_t GBAMemoryStall(struct ARMCore* cpu, int32_t wait) {
 	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
-	cpu->cycles -= stall;
+	cpu->cycles -= (s - 1) * loads;
 	return wait;
 }