diff --git a/src/arm/isa-arm.c b/src/arm/isa-arm.c
index 530e4e05d..3beab5a13 100644
--- a/src/arm/isa-arm.c
+++ b/src/arm/isa-arm.c
@@ -242,13 +242,13 @@ static inline void _immediate(struct ARMCore* cpu, uint32_t opcode) {
 #define ADDR_MODE_3_WRITEBACK(ADDR) ADDR_MODE_2_WRITEBACK(ADDR)
 
 #define ARM_LOAD_POST_BODY \
+	++currentCycles; \
 	if (rd == ARM_PC) { \
 		ARM_WRITE_PC; \
 	}
 
 #define ARM_STORE_POST_BODY \
-	currentCycles -= ARM_PREFETCH_CYCLES; \
-	currentCycles += 1 + cpu->memory.activeNonseqCycles32;
+	currentCycles += cpu->memory.activeNonseqCycles32 - cpu->memory.activeSeqCycles32;
 
 #define DEFINE_INSTRUCTION_ARM(NAME, BODY) \
 	static void _ARMInstruction ## NAME (struct ARMCore* cpu, uint32_t opcode) { \
diff --git a/src/arm/isa-thumb.c b/src/arm/isa-thumb.c
index 6f054dc91..2e4d22994 100644
--- a/src/arm/isa-thumb.c
+++ b/src/arm/isa-thumb.c
@@ -36,6 +36,8 @@
 
 #define THUMB_PREFETCH_CYCLES (1 + cpu->memory.activeSeqCycles16)
 
+#define THUMB_LOAD_POST_BODY ++currentCycles;
+
 #define THUMB_STORE_POST_BODY \
 	currentCycles += cpu->memory.activeNonseqCycles16 - cpu->memory.activeSeqCycles16;
 
@@ -89,9 +91,9 @@ DEFINE_IMMEDIATE_5_INSTRUCTION_THUMB(ASR1,
 	}
 	THUMB_NEUTRAL_S( , , cpu->gprs[rd]);)
 
-DEFINE_IMMEDIATE_5_INSTRUCTION_THUMB(LDR1, cpu->gprs[rd] = cpu->memory.load32(cpu, cpu->gprs[rm] + immediate * 4, &currentCycles))
-DEFINE_IMMEDIATE_5_INSTRUCTION_THUMB(LDRB1, cpu->gprs[rd] = cpu->memory.loadU8(cpu, cpu->gprs[rm] + immediate, &currentCycles))
-DEFINE_IMMEDIATE_5_INSTRUCTION_THUMB(LDRH1, cpu->gprs[rd] = cpu->memory.loadU16(cpu, cpu->gprs[rm] + immediate * 2, &currentCycles))
+DEFINE_IMMEDIATE_5_INSTRUCTION_THUMB(LDR1, cpu->gprs[rd] = cpu->memory.load32(cpu, cpu->gprs[rm] + immediate * 4, &currentCycles); THUMB_LOAD_POST_BODY;)
+DEFINE_IMMEDIATE_5_INSTRUCTION_THUMB(LDRB1, cpu->gprs[rd] = cpu->memory.loadU8(cpu, cpu->gprs[rm] + immediate, &currentCycles); THUMB_LOAD_POST_BODY;)
+DEFINE_IMMEDIATE_5_INSTRUCTION_THUMB(LDRH1, cpu->gprs[rd] = cpu->memory.loadU16(cpu, cpu->gprs[rm] + immediate * 2, &currentCycles); THUMB_LOAD_POST_BODY;)
 DEFINE_IMMEDIATE_5_INSTRUCTION_THUMB(STR1, cpu->memory.store32(cpu, cpu->gprs[rm] + immediate * 4, cpu->gprs[rd], &currentCycles); THUMB_STORE_POST_BODY;)
 DEFINE_IMMEDIATE_5_INSTRUCTION_THUMB(STRB1, cpu->memory.store8(cpu, cpu->gprs[rm] + immediate, cpu->gprs[rd], &currentCycles); THUMB_STORE_POST_BODY;)
 DEFINE_IMMEDIATE_5_INSTRUCTION_THUMB(STRH1, cpu->memory.store16(cpu, cpu->gprs[rm] + immediate * 2, cpu->gprs[rd], &currentCycles); THUMB_STORE_POST_BODY;)
@@ -261,8 +263,8 @@ DEFINE_INSTRUCTION_WITH_HIGH_THUMB(MOV3,
 #define DEFINE_IMMEDIATE_WITH_REGISTER_THUMB(NAME, BODY) \
 	COUNT_CALL_3(DEFINE_IMMEDIATE_WITH_REGISTER_EX_THUMB, NAME ## _R, BODY)
 
-DEFINE_IMMEDIATE_WITH_REGISTER_THUMB(LDR3, cpu->gprs[rd] = cpu->memory.load32(cpu, (cpu->gprs[ARM_PC] & 0xFFFFFFFC) + immediate, &currentCycles))
-DEFINE_IMMEDIATE_WITH_REGISTER_THUMB(LDR4, cpu->gprs[rd] = cpu->memory.load32(cpu, cpu->gprs[ARM_SP] + immediate, &currentCycles))
+DEFINE_IMMEDIATE_WITH_REGISTER_THUMB(LDR3, cpu->gprs[rd] = cpu->memory.load32(cpu, (cpu->gprs[ARM_PC] & 0xFFFFFFFC) + immediate, &currentCycles); THUMB_LOAD_POST_BODY;)
+DEFINE_IMMEDIATE_WITH_REGISTER_THUMB(LDR4, cpu->gprs[rd] = cpu->memory.load32(cpu, cpu->gprs[ARM_SP] + immediate, &currentCycles); THUMB_LOAD_POST_BODY;)
 DEFINE_IMMEDIATE_WITH_REGISTER_THUMB(STR3, cpu->memory.store32(cpu, cpu->gprs[ARM_SP] + immediate, cpu->gprs[rd], &currentCycles); THUMB_STORE_POST_BODY;)
 
 DEFINE_IMMEDIATE_WITH_REGISTER_THUMB(ADD5, cpu->gprs[rd] = (cpu->gprs[ARM_PC] & 0xFFFFFFFC) + immediate)
@@ -278,11 +280,11 @@ DEFINE_IMMEDIATE_WITH_REGISTER_THUMB(ADD6, cpu->gprs[rd] = cpu->gprs[ARM_SP] + i
 #define DEFINE_LOAD_STORE_WITH_REGISTER_THUMB(NAME, BODY) \
 	COUNT_CALL_3(DEFINE_LOAD_STORE_WITH_REGISTER_EX_THUMB, NAME ## _R, BODY)
 
-DEFINE_LOAD_STORE_WITH_REGISTER_THUMB(LDR2, cpu->gprs[rd] = cpu->memory.load32(cpu, cpu->gprs[rn] + cpu->gprs[rm], &currentCycles))
-DEFINE_LOAD_STORE_WITH_REGISTER_THUMB(LDRB2, cpu->gprs[rd] = cpu->memory.loadU8(cpu, cpu->gprs[rn] + cpu->gprs[rm], &currentCycles))
-DEFINE_LOAD_STORE_WITH_REGISTER_THUMB(LDRH2, cpu->gprs[rd] = cpu->memory.loadU16(cpu, cpu->gprs[rn] + cpu->gprs[rm], &currentCycles))
-DEFINE_LOAD_STORE_WITH_REGISTER_THUMB(LDRSB, cpu->gprs[rd] = cpu->memory.load8(cpu, cpu->gprs[rn] + cpu->gprs[rm], &currentCycles))
-DEFINE_LOAD_STORE_WITH_REGISTER_THUMB(LDRSH, cpu->gprs[rd] = cpu->memory.load16(cpu, cpu->gprs[rn] + cpu->gprs[rm], &currentCycles))
+DEFINE_LOAD_STORE_WITH_REGISTER_THUMB(LDR2, cpu->gprs[rd] = cpu->memory.load32(cpu, cpu->gprs[rn] + cpu->gprs[rm], &currentCycles); THUMB_LOAD_POST_BODY;)
+DEFINE_LOAD_STORE_WITH_REGISTER_THUMB(LDRB2, cpu->gprs[rd] = cpu->memory.loadU8(cpu, cpu->gprs[rn] + cpu->gprs[rm], &currentCycles); THUMB_LOAD_POST_BODY;)
+DEFINE_LOAD_STORE_WITH_REGISTER_THUMB(LDRH2, cpu->gprs[rd] = cpu->memory.loadU16(cpu, cpu->gprs[rn] + cpu->gprs[rm], &currentCycles); THUMB_LOAD_POST_BODY;)
+DEFINE_LOAD_STORE_WITH_REGISTER_THUMB(LDRSB, cpu->gprs[rd] = cpu->memory.load8(cpu, cpu->gprs[rn] + cpu->gprs[rm], &currentCycles); THUMB_LOAD_POST_BODY;)
+DEFINE_LOAD_STORE_WITH_REGISTER_THUMB(LDRSH, cpu->gprs[rd] = cpu->memory.load16(cpu, cpu->gprs[rn] + cpu->gprs[rm], &currentCycles); THUMB_LOAD_POST_BODY;)
 DEFINE_LOAD_STORE_WITH_REGISTER_THUMB(STR2, cpu->memory.store32(cpu, cpu->gprs[rn] + cpu->gprs[rm], cpu->gprs[rd], &currentCycles); THUMB_STORE_POST_BODY;)
 DEFINE_LOAD_STORE_WITH_REGISTER_THUMB(STRB2, cpu->memory.store8(cpu, cpu->gprs[rn] + cpu->gprs[rm], cpu->gprs[rd], &currentCycles); THUMB_STORE_POST_BODY;)
 DEFINE_LOAD_STORE_WITH_REGISTER_THUMB(STRH2, cpu->memory.store16(cpu, cpu->gprs[rn] + cpu->gprs[rm], cpu->gprs[rd], &currentCycles); THUMB_STORE_POST_BODY;)
@@ -313,6 +315,7 @@ DEFINE_LOAD_STORE_WITH_REGISTER_THUMB(STRH2, cpu->memory.store16(cpu, cpu->gprs[
 
 DEFINE_LOAD_STORE_MULTIPLE_THUMB(LDMIA,
 	cpu->gprs[i] = cpu->memory.load32(cpu, address, 0),
+	THUMB_LOAD_POST_BODY;
 	if (!((1 << rn) & rs)) {
 		cpu->gprs[rn] = address;
 	})
@@ -354,7 +357,8 @@ DEFINE_LOAD_STORE_MULTIPLE_EX_THUMB(POP,
 	(m = 0x01, i = 0; i < 8; m <<= 1, ++i),
 	cpu->gprs[i] = cpu->memory.load32(cpu, address, 0),
 	+=,
-	, ,
+	,
+	THUMB_LOAD_POST_BODY;,
 	cpu->gprs[ARM_SP] = address)
 
 DEFINE_LOAD_STORE_MULTIPLE_EX_THUMB(POPR,
@@ -365,7 +369,8 @@ DEFINE_LOAD_STORE_MULTIPLE_EX_THUMB(POPR,
 	+=,
 	,
 	cpu->gprs[ARM_PC] = cpu->memory.load32(cpu, address, 0) & 0xFFFFFFFE;
-	address += 4;,
+	address += 4;
+	THUMB_LOAD_POST_BODY;,
 	cpu->gprs[ARM_SP] = address;
 	THUMB_WRITE_PC;)