Fix stack parameters loading

ir_x86.dasc

@@ -3956,12 +3956,11 @@ static int ir_emit_dessa_move(ir_ctx *ctx, uint8_t type, ir_ref from, ir_ref to)
 	return 1;
 }
 
-static void ir_emit_param_move(ir_ctx *ctx, uint8_t type, ir_reg from_reg, ir_reg to_reg, ir_ref to)
+static void ir_emit_param_move(ir_ctx *ctx, uint8_t type, ir_reg from_reg, ir_reg to_reg, ir_ref to, int32_t offset)
 {
 	ir_backend_data *data = ctx->data;
 	dasm_State **Dst = &data->dasm_state;
 	ir_reg fp = (ctx->flags & IR_USE_FRAME_POINTER) ? IR_REG_RBP : IR_REG_RSP;
-	int32_t offset;
 
 	IR_ASSERT(from_reg != IR_REG_NONE || to_reg != IR_REG_NONE);
 
@@ -3973,7 +3972,6 @@ static void ir_emit_param_move(ir_ctx *ctx, uint8_t type, ir_reg from_reg, ir_re
 				ir_emit_store(ctx, type, to, from_reg);
 			}
 		} else {
-			offset = (ctx->flags & IR_USE_FRAME_POINTER) ? -8 : -8; // TODO:
 			|	ASM_REG_MEM_OP mov, type, to_reg, [Ra(fp)+offset]
 		}
 	} else {
@@ -3984,7 +3982,6 @@ static void ir_emit_param_move(ir_ctx *ctx, uint8_t type, ir_reg from_reg, ir_re
 				ir_emit_store(ctx, type, to, from_reg);
 			}
 		} else {
-			offset = (ctx->flags & IR_USE_FRAME_POINTER) ? -8 : -8; // TODO:
 			|	ASM_FP_REG_MEM_OP movss, movsd, vmovss, vmovsd, type, to_reg, [Ra(fp)+offset]
 		}
 	}
@@ -3992,6 +3989,7 @@ static void ir_emit_param_move(ir_ctx *ctx, uint8_t type, ir_reg from_reg, ir_re
 
 static void ir_emit_load_params(ir_ctx *ctx)
 {
+	ir_backend_data *data = ctx->data;
 	ir_use_list *use_list = &ctx->use_lists[1];
 	ir_insn *insn;
 	ir_ref i, n, *p, use;
@@ -4004,7 +4002,13 @@ static void ir_emit_load_params(ir_ctx *ctx)
 	int fp_reg_params_count = IR_REG_FP_ARGS;
 	const int8_t *int_reg_params = _ir_int_reg_params;
 	const int8_t *fp_reg_params = _ir_fp_reg_params;
+	int32_t stack_offset = 0;
 
+	if (ctx->flags & IR_USE_FRAME_POINTER) {
+		stack_offset = sizeof(void*); /* skip old frame pointer */
+	} else {
+		stack_offset = data->stack_frame_size + data->stack_frame_alignment;
+	}
 	n = use_list->count;
 	for (i = 0, p = &ctx->use_edges[use_list->refs]; i < n; i++, p++) {
 		use = *p;
@@ -4015,23 +4019,22 @@ static void ir_emit_load_params(ir_ctx *ctx)
 					if (int_param_num < int_reg_params_count) {
 						src_reg = int_reg_params[int_param_num];
 					} else {
-						// TODO: replace IR_REG_NONE by stack slot
 						src_reg = IR_REG_NONE;
+						stack_offset += sizeof(void*);
 					}
 					int_param_num++;
 				} else {
 					if (fp_param_num < fp_reg_params_count) {
 						src_reg = fp_reg_params[fp_param_num];
 					} else {
-						// TODO: replace IR_REG_NONE by stack slot
 						src_reg = IR_REG_NONE;
+						stack_offset += sizeof(void*);
 					}
 					fp_param_num++;
 				}
 				dst_reg = IR_REG_NUM(ctx->regs[use][0]);
 				if (src_reg != dst_reg) {
-					// TODO: parallel move
+					ir_emit_param_move(ctx, insn->type, src_reg, dst_reg, use, stack_offset);
-					ir_emit_param_move(ctx, insn->type, src_reg, dst_reg, use);
 				}
 				if (dst_reg != IR_REG_NONE && (ctx->regs[use][0] & IR_REG_SPILL_STORE)) {
 					ir_emit_store(ctx, insn->type, use, dst_reg);

tests/debug/fig-O0.irt

@@ -72,13 +72,13 @@ test:
 	movl %edx, 8(%rsp)
 	movl %ecx, 0xc(%rsp)
 	movl %r9d, 0x10(%rsp)
-	movl -8(%rsp), %eax
+	movl 0xa0(%rsp), %eax
 	movl %eax, 0x14(%rsp)
-	movl -8(%rsp), %eax
+	movl 0xa8(%rsp), %eax
 	movl %eax, 0x18(%rsp)
-	movl -8(%rsp), %eax
+	movl 0xb0(%rsp), %eax
 	movl %eax, 0x1c(%rsp)
-	movl -8(%rsp), %eax
+	movl 0xb8(%rsp), %eax
 	movl %eax, 0x20(%rsp)
 	movl (%rsp), %eax
 	movl %eax, 0x24(%rsp)

tests/debug/fig.irt

@@ -69,10 +69,10 @@ test:
 	subq $0x10, %rsp
 	movq %rbx, 8(%rsp)
 	movq %rbp, (%rsp)
-	movl -8(%rsp), %eax
+	movl 0x18(%rsp), %eax
-	movl -8(%rsp), %r8d
+	movl 0x20(%rsp), %r8d
-	movl -8(%rsp), %r10d
+	movl 0x28(%rsp), %r10d
-	movl -8(%rsp), %r11d
+	movl 0x30(%rsp), %r11d
 	movl %edx, %ebx
 	imull %esi, %ebx
 	leal 4(%rbx), %ebp

tests/debug/params_001.irt

@@ -0,0 +1,48 @@
+--TEST--
+001: Parameter Loading
+--ARGS--
+-S
+--CODE--
+{
+	l_1 = START(l_2);
+	int32_t p_1 = PARAM(l_1, "p_1", 1);
+	int32_t p_2 = PARAM(l_1, "p_2", 1);
+	int32_t p_3 = PARAM(l_1, "p_3", 1);
+	int32_t p_4 = PARAM(l_1, "p_4", 1);
+	int32_t p_5 = PARAM(l_1, "p_5", 1);
+	int32_t p_6 = PARAM(l_1, "p_6", 1);
+	int32_t p_7 = PARAM(l_1, "p_7", 1);
+	int32_t p_8 = PARAM(l_1, "p_8", 1);
+	int32_t p_9 = PARAM(l_1, "p_9", 1);
+	int32_t p_10 = PARAM(l_1, "p_10", 1);
+	int32_t r1 = ADD(p_1, p_2);
+	int32_t r2 = ADD(r1, p_3);
+	int32_t r3 = ADD(r2, p_4);
+	int32_t r4 = ADD(r3, p_5);
+	int32_t r5 = ADD(r4, p_6);
+	int32_t r6 = ADD(r5, p_7);
+	int32_t r7 = ADD(r6, p_8);
+	int32_t r8 = ADD(r7, p_9);
+	int32_t ret = SUB(r8, p_10);
+	l_2 = RETURN(l_1, ret);
+}
+--EXPECT--
+test:
+	subq $8, %rsp
+	movq %rbx, (%rsp)
+	movl 0x10(%rsp), %eax
+	movl 0x18(%rsp), %r10d
+	movl 0x20(%rsp), %r11d
+	movl 0x28(%rsp), %ebx
+	leal (%rsi, %rdi), %esi
+	leal (%rsi, %rdx), %edx
+	leal (%rdx, %rcx), %ecx
+	leal (%rcx, %r8), %ecx
+	leal (%rcx, %r9), %ecx
+	leal (%rcx, %rax), %eax
+	leal (%rax, %r10), %eax
+	leal (%rax, %r11), %eax
+	subl %ebx, %eax
+	movq (%rsp), %rbx
+	addq $8, %rsp
+	retq

tests/debug/params_002.irt

@@ -0,0 +1,43 @@
+--TEST--
+002: Parameter Loading
+--ARGS--
+-S
+--CODE--
+{
+	l_1 = START(l_2);
+	double p_1 = PARAM(l_1, "p_1", 1);
+	double p_2 = PARAM(l_1, "p_2", 1);
+	double p_3 = PARAM(l_1, "p_3", 1);
+	double p_4 = PARAM(l_1, "p_4", 1);
+	double p_5 = PARAM(l_1, "p_5", 1);
+	double p_6 = PARAM(l_1, "p_6", 1);
+	double p_7 = PARAM(l_1, "p_7", 1);
+	double p_8 = PARAM(l_1, "p_8", 1);
+	double p_9 = PARAM(l_1, "p_9", 1);
+	double p_10 = PARAM(l_1, "p_10", 1);
+	double r1 = ADD(p_1, p_2);
+	double r2 = ADD(r1, p_3);
+	double r3 = ADD(r2, p_4);
+	double r4 = ADD(r3, p_5);
+	double r5 = ADD(r4, p_6);
+	double r6 = ADD(r5, p_7);
+	double r7 = ADD(r6, p_8);
+	double r8 = ADD(r7, p_9);
+	double ret = SUB(r8, p_10);
+	l_2 = RETURN(l_1, ret);
+}
+--EXPECT--
+test:
+	movsd 8(%rsp), %xmm8
+	movsd 0x10(%rsp), %xmm9
+	addsd %xmm0, %xmm1
+	addsd %xmm2, %xmm1
+	addsd %xmm3, %xmm1
+	addsd %xmm4, %xmm1
+	addsd %xmm5, %xmm1
+	addsd %xmm6, %xmm1
+	addsd %xmm7, %xmm1
+	movapd %xmm1, %xmm0
+	addsd %xmm8, %xmm0
+	subsd %xmm9, %xmm0
+	retq