Use PUSH/POP to save/restore persistent registers in prologue/epilogie

dstogov · Mar 14, 2024 · 9176609 · 9176609
1 parent b447d62
commit 9176609
Show file tree

Hide file tree

Showing 36 changed files with 310 additions and 92 deletions.
diff --git a/ir_aarch64.dasc b/ir_aarch64.dasc
@@ -11,7 +11,7 @@
 |.globals ir_lb
 |.section code, cold_code, rodata, jmp_table
 
-|.define IR_LOOP_ALIGNMENT, 16
+|.define IR_LOOP_ALIGNMENT, 8
 
 #ifdef IR_DEBUG
 typedef struct _ir_mem {uint64_t v;} ir_mem;

diff --git a/ir_x86.dasc b/ir_x86.dasc
@@ -3229,43 +3229,48 @@ static void ir_emit_prologue(ir_ctx *ctx)
 {
 	ir_backend_data *data = ctx->data;
 	dasm_State **Dst = &data->dasm_state;
+	int offset = ctx->stack_frame_size + ctx->call_stack_size;
 
 	if (ctx->flags & IR_USE_FRAME_POINTER) {
 		|	push Ra(IR_REG_RBP)
 		|	mov Ra(IR_REG_RBP), Ra(IR_REG_RSP)
 	}
+	if (IR_REGSET_INTERSECTION((ir_regset)ctx->used_preserved_regs, IR_REGSET_GP)) {
+		int i;
+		ir_regset used_preserved_regs = IR_REGSET_INTERSECTION((ir_regset)ctx->used_preserved_regs, IR_REGSET_GP);
+
+		for (i = IR_REG_GP_FIRST; i <= IR_REG_GP_LAST; i++) {
+			if (IR_REGSET_IN(used_preserved_regs, i)) {
+				offset -= sizeof(void*);
+				|	push Ra(i)
+			}
+		}
+	}
 	if (ctx->stack_frame_size + ctx->call_stack_size) {
 		if (ctx->fixed_stack_red_zone) {
 			IR_ASSERT(ctx->stack_frame_size + ctx->call_stack_size <= ctx->fixed_stack_red_zone);
-		} else {
-			|	sub Ra(IR_REG_RSP), (ctx->stack_frame_size + ctx->call_stack_size)
+		} else if (offset) {
+			|	sub Ra(IR_REG_RSP), offset
 		}
 	}
-	if (ctx->used_preserved_regs) {
+	if (IR_REGSET_INTERSECTION((ir_regset)ctx->used_preserved_regs, IR_REGSET_FP)) {
 		ir_reg fp;
-		int offset;
-		uint32_t i;
-		ir_regset used_preserved_regs = (ir_regset)ctx->used_preserved_regs;
+		int i;
+		ir_regset used_preserved_regs = IR_REGSET_INTERSECTION((ir_regset)ctx->used_preserved_regs, IR_REGSET_FP);
 
 		if (ctx->flags & IR_USE_FRAME_POINTER) {
 			fp = IR_REG_FRAME_POINTER;
-			offset = 0;
+			offset -= ctx->stack_frame_size + ctx->call_stack_size;
 		} else {
 			fp = IR_REG_STACK_POINTER;
-			offset = ctx->stack_frame_size + ctx->call_stack_size;
 		}
-		for (i = 0; i < IR_REG_NUM; i++) {
+		for (i = IR_REG_FP_FIRST; i <= IR_REG_FP_LAST; i++) {
 			if (IR_REGSET_IN(used_preserved_regs, i)) {
-				if (i < IR_REG_FP_FIRST) {
-					offset -= sizeof(void*);
-					|	mov aword [Ra(fp)+offset], Ra(i)
+				offset -= sizeof(void*);
+				if (ctx->mflags & IR_X86_AVX) {
+					|	vmovsd qword [Ra(fp)+offset], xmm(i-IR_REG_FP_FIRST)
 				} else {
-					offset -= sizeof(void*);
-					if (ctx->mflags & IR_X86_AVX) {
-						|	vmovsd qword [Ra(fp)+offset], xmm(i-IR_REG_FP_FIRST)
-					} else {
-						|	movsd qword [Ra(fp)+offset], xmm(i-IR_REG_FP_FIRST)
-					}
+					|	movsd qword [Ra(fp)+offset], xmm(i-IR_REG_FP_FIRST)
 				}
 			}
 		}
@@ -3332,26 +3337,24 @@ static void ir_emit_epilogue(ir_ctx *ctx)
 	ir_backend_data *data = ctx->data;
 	dasm_State **Dst = &data->dasm_state;
 
-	if (ctx->used_preserved_regs) {
+	if (IR_REGSET_INTERSECTION((ir_regset)ctx->used_preserved_regs, IR_REGSET_FP)) {
+		int i;
 		int offset;
-		uint32_t i;
+		ir_reg fp = (ctx->flags & IR_USE_FRAME_POINTER) ? IR_REG_FRAME_POINTER : IR_REG_STACK_POINTER;
 		ir_regset used_preserved_regs = (ir_regset)ctx->used_preserved_regs;
 
 		if (ctx->flags & IR_USE_FRAME_POINTER) {
+			fp = IR_REG_FRAME_POINTER;
 			offset = 0;
 		} else {
+			fp = IR_REG_STACK_POINTER;
 			offset = ctx->stack_frame_size + ctx->call_stack_size;
 		}
 		for (i = 0; i < IR_REG_NUM; i++) {
 			if (IR_REGSET_IN(used_preserved_regs, i)) {
 				if (i < IR_REG_FP_FIRST) {
-					ir_reg fp = (ctx->flags & IR_USE_FRAME_POINTER) ? IR_REG_FRAME_POINTER : IR_REG_STACK_POINTER;
-
 					offset -= sizeof(void*);
-					|	mov Ra(i), aword [Ra(fp)+offset]
 				} else {
-					ir_reg fp = (ctx->flags & IR_USE_FRAME_POINTER) ? IR_REG_FRAME_POINTER : IR_REG_STACK_POINTER;
-
 					offset -= sizeof(void*);
 					if (ctx->mflags & IR_X86_AVX) {
 						|	vmovsd xmm(i-IR_REG_FP_FIRST), qword [Ra(fp)+offset]
@@ -3363,7 +3366,40 @@ static void ir_emit_epilogue(ir_ctx *ctx)
 		}
 	}
 
-	if (ctx->flags & IR_USE_FRAME_POINTER) {
+	if (IR_REGSET_INTERSECTION((ir_regset)ctx->used_preserved_regs, IR_REGSET_GP)) {
+		int i;
+		ir_regset used_preserved_regs = IR_REGSET_INTERSECTION((ir_regset)ctx->used_preserved_regs, IR_REGSET_GP);
+		int offset;
+
+		if (ctx->flags & IR_USE_FRAME_POINTER) {
+			offset = 0;
+		} else {
+			offset = ctx->stack_frame_size + ctx->call_stack_size;
+		}
+		if (IR_REGSET_INTERSECTION((ir_regset)ctx->used_preserved_regs, IR_REGSET_GP)) {
+			int i;
+			ir_regset used_preserved_regs = IR_REGSET_INTERSECTION((ir_regset)ctx->used_preserved_regs, IR_REGSET_GP);
+
+			for (i = IR_REG_GP_LAST; i >= IR_REG_GP_FIRST; i--) {
+				if (IR_REGSET_IN(used_preserved_regs, i)) {
+					offset -= sizeof(void*);
+				}
+			}
+		}
+		if (ctx->flags & IR_USE_FRAME_POINTER) {
+			|	lea Ra(IR_REG_RSP), [Ra(IR_REG_RBP)+offset]
+		} else if (offset) {
+			|	add Ra(IR_REG_RSP), offset
+		}
+		for (i = IR_REG_GP_LAST; i >= IR_REG_GP_FIRST; i--) {
+			if (IR_REGSET_IN(used_preserved_regs, i)) {
+				|	pop Ra(i)
+			}
+		}
+		if (ctx->flags & IR_USE_FRAME_POINTER) {
+			|	pop Ra(IR_REG_RBP)
+		}
+	} else if (ctx->flags & IR_USE_FRAME_POINTER) {
 		|	mov Ra(IR_REG_RSP), Ra(IR_REG_RBP)
 		|	pop Ra(IR_REG_RBP)
 	} else if (ctx->stack_frame_size + ctx->call_stack_size) {

diff --git a/tests/debug.aarch64/dessa_002.irt b/tests/debug.aarch64/dessa_002.irt
@@ -25,6 +25,7 @@ aarch64
 --EXPECT--
 test:
 	b .L2
+	nop
 .L1:
 	mov w0, w1
 .L2:

diff --git a/tests/debug.aarch64/dessa_003.irt b/tests/debug.aarch64/dessa_003.irt
@@ -35,6 +35,7 @@ test:
 	movz w3, #0xc
 	sdiv w0, w0, w3
 	b .L2
+	nop
 .L1:
 	mov w3, w0
 	mov w0, w1

diff --git a/tests/debug.aarch64/fig.irt b/tests/debug.aarch64/fig.irt
@@ -71,6 +71,7 @@ test:
 	mul w9, w2, w1
 	add w10, w9, #4
 	b .L3
+	nop
 .L1:
 	mul w6, w0, w1
 	add w7, w6, #1

diff --git a/tests/debug.aarch64/loop_002.irt b/tests/debug.aarch64/loop_002.irt
@@ -49,6 +49,7 @@ test:
 	mov w1, wzr
 	mov w0, wzr
 	b .L4
+	nop
 .L1:
 	cmp w2, w1
 	b.eq .L2

diff --git a/tests/debug.aarch64/regset-fib.irt b/tests/debug.aarch64/regset-fib.irt
@@ -81,6 +81,7 @@ main:
 	ldr d0, .L3
 	str d0, [x29, #0x18]
 	b .L2
+	nop
 .L1:
 	ldr d0, [x29, #0x10]
 	fadd d1, d1, d0
@@ -98,7 +99,6 @@ main:
 	ldp x29, x30, [sp], #0x20
 	ret
 .rodata
-	.db 0x1f, 0x20, 0x03, 0xd5
 .L3:
 	.db 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0xf0, 0x3f
 .L4:

diff --git a/tests/debug.aarch64/regset-fib2.irt b/tests/debug.aarch64/regset-fib2.irt
@@ -81,6 +81,7 @@ main:
 	ldr d0, .L3
 	str d0, [x29, #0x18]
 	b .L2
+	nop
 .L1:
 	ldr d0, [x29, #0x10]
 	fadd d1, d1, d0
@@ -101,6 +102,7 @@ main:
 	ldp x29, x30, [sp], #0x20
 	ret
 .rodata
+	.db 0x1f, 0x20, 0x03, 0xd5
 .L3:
 	.db 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0xf0, 0x3f
 .L4:

diff --git a/tests/debug.aarch64/regset-test.irt b/tests/debug.aarch64/regset-test.irt
@@ -160,6 +160,7 @@ test:
 	str d1, [sp, #0x18]
 	mov w0, wzr
 	b .L2
+	nop
 .L1:
 	ldr d1, [sp, #0x18]
 	ldr d0, [sp, #0x10]
@@ -197,6 +198,7 @@ test:
 	add sp, sp, #0x38
 	ret
 .rodata
+	.db 0x1f, 0x20, 0x03, 0xd5
 .L4:
 	.db 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0xe0, 0x3f
 .L5:

diff --git a/tests/debug.aarch64/sccp_002.irt b/tests/debug.aarch64/sccp_002.irt
@@ -68,6 +68,7 @@ test:
 	movz w0, #0x3
 .L4:
 	ret
+	nop
 .L5:
 	cmp w1, #0
 	b.ne .L5

diff --git a/tests/debug.x86/dessa_002.irt b/tests/debug.x86/dessa_002.irt
@@ -26,6 +26,16 @@ x86
 test:
 	movl 4(%esp), %eax
 	jmp .L2
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop
 .L1:
 	movl %ecx, %eax
 .L2:

diff --git a/tests/debug.x86/dessa_003.irt b/tests/debug.x86/dessa_003.irt
@@ -32,22 +32,28 @@ x86
 }
 --EXPECT--
 test:
-	subl $8, %esp
-	movl %ebx, 4(%esp)
-	movl %ebp, (%esp)
+	pushl %ebx
+	pushl %ebp
 	movl 0x10(%esp), %ecx
 	movl 0x14(%esp), %ebx
 	movl 0xc(%esp), %eax
 	movl $0xc, %ebp
 	cltd
 	idivl %ebp
 	jmp .L2
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop
 .L1:
 	xchgl %eax, %ecx
 .L2:
 	testl %ebx, %ebx
 	jne .L1
-	movl 4(%esp), %ebx
-	movl (%esp), %ebp
-	addl $8, %esp
+	popl %ebp
+	popl %ebx
 	retl
diff --git a/tests/debug.x86/fig.irt b/tests/debug.x86/fig.irt
@@ -66,11 +66,11 @@ x86
 }
 --EXPECT--
 test:
-	subl $0x14, %esp
-	movl %ebx, 0x10(%esp)
-	movl %ebp, 0xc(%esp)
-	movl %esi, 8(%esp)
-	movl %edi, 4(%esp)
+	pushl %ebx
+	pushl %ebp
+	pushl %esi
+	pushl %edi
+	subl $4, %esp
 	movl 0x18(%esp), %eax
 	movl 0x1c(%esp), %ecx
 	movl 0x24(%esp), %ebx
@@ -80,6 +80,16 @@ test:
 	movl %edx, (%esp)
 	leal 4(%edx), %edx
 	jmp .L3
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop
 .L1:
 	movl %eax, %esi
 	imull %ecx, %esi
@@ -111,9 +121,9 @@ test:
 	movl 0x34(%esp), %edi
 	addl %edi, %eax
 	leal 1(%eax, %esi), %eax
-	movl 0x10(%esp), %ebx
-	movl 0xc(%esp), %ebp
-	movl 8(%esp), %esi
-	movl 4(%esp), %edi
-	addl $0x14, %esp
+	addl $4, %esp
+	popl %edi
+	popl %esi
+	popl %ebp
+	popl %ebx
 	retl
diff --git a/tests/debug.x86/loop_002.irt b/tests/debug.x86/loop_002.irt
@@ -49,6 +49,16 @@ test:
 	xorl %ecx, %ecx
 	xorl %eax, %eax
 	jmp .L4
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop
 .L1:
 	cmpl %ecx, %edx
 	je .L2

diff --git a/tests/debug.x86/regset-fib.irt b/tests/debug.x86/regset-fib.irt
@@ -77,6 +77,15 @@ main:
 	movsd .L4, %xmm0
 	movsd %xmm0, 0xc(%esp)
 	jmp .L2
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop
+	nop
 .L1:
 	addsd %xmm1, %xmm0
 	movsd %xmm0, 0xc(%esp)
@@ -96,7 +105,7 @@ main:
 	addl $0x1c, %esp
 	retl
 .rodata
-	.db 0x90, 0x90, 0x90, 0x90, 0x90
+	.db 0x90, 0x90, 0x90, 0x90
 .L4:
 	.db 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0xf0, 0x3f
 .L5: