diff --git a/ir_x86.dasc b/ir_x86.dasc
index 3abab8e..916717f 100644
--- a/ir_x86.dasc
+++ b/ir_x86.dasc
@@ -1611,7 +1611,7 @@ static void ir_emit_prologue(ir_ctx *ctx)
 		|	push rbp
 		|	mov rbp, rsp
 	}
-	if (data->stack_frame_size) {
+	if (data->stack_frame_size + data->call_stack_size) {
 		|	sub rsp, (data->stack_frame_size + data->call_stack_size)
 	}
 	if (data->used_preserved_regs) {
@@ -1671,7 +1671,7 @@ static void ir_emit_epilogue(ir_ctx *ctx)
 	if (ctx->flags & IR_USE_FRAME_POINTER) {
 		|	mov rsp, rbp
 		|	pop rbp
-	} else if (data->stack_frame_size) {
+	} else if (data->stack_frame_size + data->call_stack_size) {
 		|	add rsp, (data->stack_frame_size + data->call_stack_size)
 	}
 }
@@ -3565,7 +3565,7 @@ static int ir_parallel_copy(ir_ctx *ctx, ir_copy *copies, int count, ir_reg tmp_
 	return 1;
 }
 
-static int32_t ir_call_used_stack(ir_ctx *ctx, ir_insn *insn)
+static int32_t ir_call_used_stack(ir_ctx *ctx, ir_insn *insn, bool align)
 {
 	int j, n;
 	ir_type type;
@@ -3593,7 +3593,7 @@ static int32_t ir_call_used_stack(ir_ctx *ctx, ir_insn *insn)
 		}
 	}
 
-	if (used_stack % 16 != 0) {
+	if (align && used_stack % 16 != 0) {
 		/* Stack must be 16 byte aligned */
 		used_stack += 8;
 	}
@@ -3626,7 +3626,7 @@ static int32_t ir_emit_arguments(ir_ctx *ctx, ir_ref def, ir_insn *insn, ir_reg
 		// TODO: support for preallocated stack
 		used_stack = 0;
 	} else {
-		used_stack = ir_call_used_stack(ctx, insn);
+		used_stack = ir_call_used_stack(ctx, insn, 1);
 		data->call_stack_size += used_stack;
 		if (used_stack) {
 			|	sub rsp, used_stack
@@ -4395,7 +4395,7 @@ static void ir_preallocate_call_stack(ir_ctx *ctx, ir_backend_data *data)
 
 	for (i = 1, insn = ctx->ir_base + 1; i < ctx->insns_count;) {
 		if (insn->op == IR_CALL) {
-			call_stack_size = ir_call_used_stack(ctx, insn);
+			call_stack_size = ir_call_used_stack(ctx, insn, 0);
 			if (call_stack_size > peak_call_stack_size) {
 				peak_call_stack_size = call_stack_size;
 			}
@@ -4443,11 +4443,12 @@ static void ir_calc_stack_frame_size(ir_ctx *ctx, ir_backend_data *data)
 				data->stack_frame_alignment += 8;
 			}
 		} else {
-			while (IR_ALIGNED_SIZE(data->stack_frame_size + sizeof(void*), 16) != data->stack_frame_size + sizeof(void*)) {
+			ir_preallocate_call_stack(ctx, data);
+			while (IR_ALIGNED_SIZE(data->stack_frame_size + data->call_stack_size + sizeof(void*), 16) !=
+					data->stack_frame_size + data->call_stack_size + sizeof(void*)) {
 				data->stack_frame_size += 8;
 				data->stack_frame_alignment += 8;
 			}
-			ir_preallocate_call_stack(ctx, data);
 		}
 	}
 }
diff --git a/tests/debug/args_001.irt b/tests/debug/args_001.irt
index 5806981..2311387 100644
--- a/tests/debug/args_001.irt
+++ b/tests/debug/args_001.irt
@@ -22,7 +22,7 @@
 }
 --EXPECT--
 test:
-	subq $0x38, %rsp
+	subq $0x28, %rsp
 	movl $6, (%rsp)
 	movl $7, 8(%rsp)
 	movl $8, 0x10(%rsp)
@@ -37,7 +37,7 @@ test:
 	movabsq $0x100000000, %r9
 	movabsq $_IO_printf, %rax
 	callq *%rax
-	addq $0x38, %rsp
+	addq $0x28, %rsp
 	retq
 
 1 2 3 4 0x100000000 6 7 8 9 0x100000000