diff --git a/ir_gcm.c b/ir_gcm.c
index 8a48fed..adbe77a 100644
--- a/ir_gcm.c
+++ b/ir_gcm.c
@@ -15,6 +15,7 @@
 #define IR_GCM_EARLY_BLOCK(b)        ((uint32_t)-((int32_t)(b)))
 
 #define IR_GCM_SPLIT 1
+#define IR_SCHEDULE_SWAP_OPS 1
 
 static uint32_t ir_gcm_schedule_early(ir_ctx *ctx, ir_ref ref, ir_list *queue_late)
 {
@@ -1131,6 +1132,37 @@ int ir_schedule(ir_ctx *ctx)
 				new_insn->op1 = _xlat[insn->op1];
 				new_insn->op2 = _xlat[insn->op2];
 				new_insn->op3 = insn->op3;
+#if IR_SCHEDULE_SWAP_OPS
+				/* Swap operands according to folding rules */
+				if (new_insn->op1 < new_insn->op2) {
+					switch (new_insn->op) {
+						case IR_EQ:
+						case IR_NE:
+						case IR_ADD:
+						case IR_MUL:
+						case IR_ADD_OV:
+						case IR_MUL_OV:
+						case IR_OR:
+						case IR_AND:
+						case IR_XOR:
+						case IR_MIN:
+						case IR_MAX:
+							SWAP_REFS(new_insn->op1, new_insn->op2);
+							break;
+						case IR_LT:
+						case IR_GE:
+						case IR_LE:
+						case IR_GT:
+						case IR_ULT:
+						case IR_UGE:
+						case IR_ULE:
+						case IR_UGT:
+							SWAP_REFS(new_insn->op1, new_insn->op2);
+							new_insn->op ^= 3; /* [U]LT <-> [U]GT, [U]LE <-> [U]GE */
+							break;
+					}
+				}
+#endif
 				break;
 			case 3:
 				new_insn->op1 = _xlat[insn->op1];
diff --git a/tests/014.irt b/tests/014.irt
index 85281e9..96fa8f8 100644
--- a/tests/014.irt
+++ b/tests/014.irt
@@ -38,7 +38,7 @@
 	l_4 = END(l_1);
 	l_5 = LOOP_BEGIN(l_4, l_11);
 	int32_t d_6 = PHI(l_5, c_6, d_7);
-	int32_t d_7 = ADD(d_3, d_6);
+	int32_t d_7 = ADD(d_6, d_3);
 	bool d_8 = LT(d_7, c_4);
 	l_9 = IF(l_5, d_8);
 	l_10 = IF_TRUE(l_9);
diff --git a/tests/023.irt b/tests/023.irt
index 0af490a..60b213b 100644
--- a/tests/023.irt
+++ b/tests/023.irt
@@ -39,7 +39,7 @@
 	l_4 = END(l_1);
 	l_5 = LOOP_BEGIN(l_4, l_11);
 	int32_t d_6 = PHI(l_5, c_6, d_7);
-	int32_t d_7 = ADD(d_3, d_6);
+	int32_t d_7 = ADD(d_6, d_3);
 	bool d_8 = LT(d_7, c_4);
 	l_9 = IF(l_5, d_8);
 	l_10 = IF_TRUE(l_9);
diff --git a/tests/debug.Windows-x86_64/dce_001.irt b/tests/debug.Windows-x86_64/dce_001.irt
index 0587be1..6684a95 100755
--- a/tests/debug.Windows-x86_64/dce_001.irt
+++ b/tests/debug.Windows-x86_64/dce_001.irt
@@ -33,7 +33,7 @@ Windows-x86_64
 test:
 	xorl %eax, %eax
 .L1:
-	leal 1(%rcx, %rax), %eax
+	leal 1(%rax, %rcx), %eax
 	cmpl $0xa, %eax
 	jl .L1
 	addl %eax, %eax
diff --git a/tests/debug.Windows-x86_64/loop_001.irt b/tests/debug.Windows-x86_64/loop_001.irt
index aae3b3b..003bb32 100755
--- a/tests/debug.Windows-x86_64/loop_001.irt
+++ b/tests/debug.Windows-x86_64/loop_001.irt
@@ -32,7 +32,7 @@ Windows-x86_64
 test:
 	xorl %eax, %eax
 .L1:
-	leal 1(%rcx, %rax), %eax
+	leal 1(%rax, %rcx), %eax
 	cmpl $0xa, %eax
 	jl .L1
 	addl %eax, %eax
diff --git a/tests/debug.aarch64/dce_001.irt b/tests/debug.aarch64/dce_001.irt
index 9608c18..01cabce 100644
--- a/tests/debug.aarch64/dce_001.irt
+++ b/tests/debug.aarch64/dce_001.irt
@@ -34,7 +34,7 @@ test:
 	add w0, w0, #1
 	mov w1, wzr
 .L1:
-	add w1, w0, w1
+	add w1, w1, w0
 	cmp w1, #0xa
 	b.lt .L1
 	add w0, w1, w1
diff --git a/tests/debug.aarch64/loop_001.irt b/tests/debug.aarch64/loop_001.irt
index 547639e..39aaf0a 100644
--- a/tests/debug.aarch64/loop_001.irt
+++ b/tests/debug.aarch64/loop_001.irt
@@ -33,7 +33,7 @@ test:
 	add w0, w0, #1
 	mov w1, wzr
 .L1:
-	add w1, w0, w1
+	add w1, w1, w0
 	cmp w1, #0xa
 	b.lt .L1
 	add w0, w1, w1
diff --git a/tests/debug.x86/dce_001.irt b/tests/debug.x86/dce_001.irt
index a0dbf91..5ab06f7 100644
--- a/tests/debug.x86/dce_001.irt
+++ b/tests/debug.x86/dce_001.irt
@@ -34,7 +34,7 @@ test:
 	movl 4(%esp), %eax
 	xorl %ecx, %ecx
 .L1:
-	leal 1(%eax, %ecx), %ecx
+	leal 1(%ecx, %eax), %ecx
 	cmpl $0xa, %ecx
 	jl .L1
 	leal (%ecx, %ecx), %eax
diff --git a/tests/debug.x86/loop_001.irt b/tests/debug.x86/loop_001.irt
index 5ef8b2f..c13944c 100644
--- a/tests/debug.x86/loop_001.irt
+++ b/tests/debug.x86/loop_001.irt
@@ -33,7 +33,7 @@ test:
 	movl 4(%esp), %eax
 	xorl %ecx, %ecx
 .L1:
-	leal 1(%eax, %ecx), %ecx
+	leal 1(%ecx, %eax), %ecx
 	cmpl $0xa, %ecx
 	jl .L1
 	leal (%ecx, %ecx), %eax
diff --git a/tests/debug/dce_001.irt b/tests/debug/dce_001.irt
index 84f856a..8c6b1a7 100644
--- a/tests/debug/dce_001.irt
+++ b/tests/debug/dce_001.irt
@@ -33,7 +33,7 @@ x86_64
 test:
 	xorl %eax, %eax
 .L1:
-	leal 1(%rdi, %rax), %eax
+	leal 1(%rax, %rdi), %eax
 	cmpl $0xa, %eax
 	jl .L1
 	addl %eax, %eax
diff --git a/tests/debug/loop_001.irt b/tests/debug/loop_001.irt
index bd95603..32156e6 100644
--- a/tests/debug/loop_001.irt
+++ b/tests/debug/loop_001.irt
@@ -32,7 +32,7 @@ x86_64
 test:
 	xorl %eax, %eax
 .L1:
-	leal 1(%rdi, %rax), %eax
+	leal 1(%rax, %rdi), %eax
 	cmpl $0xa, %eax
 	jl .L1
 	addl %eax, %eax