SR-add opt性能分析

以如下case为例：
```
void  add(int n) {
  for (int i = 0;i < n; i++) {
    int c = a[i];
    a[i] = c+1;
    b[i] = c;
  }
}
```
gcc O2 shows: (从结果看没有做sr)
```
add:
        cmp     w0, 0
        ble     .L1
        adrp    x3, .LANCHOR0
        mov     x1, 0
        add     x3, x3, :lo12:.LANCHOR0
        add     x5, x3, 24
        .p2align 2
.L3:
        ldr     w2, [x3, x1, lsl 2]
        add     w4, w2, 1
        str     w2, [x5, x1, lsl 2]
        str     w4, [x3, x1, lsl 2]
        add     x1, x1, 1
        cmp     w0, w1
        bgt     .L3
.L1:
        ret
        .size   add, .-add
```
maplec O2 shows: （只做mul的sr）
```
        .text
        .align 3
        .globl  add
        .hidden add
        .type   add, %function
add:
.L.107__5:
        cmp     w0, #0
        ble     .L.107__1
        sxtw    x1, w0
        mov     x0, #0
        lsl     x4, x1, #2
        adrp    x1, b
        add     x5, x1, #:lo12:b
        adrp    x1, a
        add     x2, x1, #:lo12:a
.L.107__2:
        ldr     w1, [x2,x0]
        add     w3, w1, #1
        str     w3, [x2,x0]
        str     w1, [x5,x0]
        add     x0, x0, #4
        cmp     x0, x4
        blo     .L.107__2
.L.107__1:
        ret
.L.107__7:
        .size   add, .-add
```
maplec O2 + sradd shows:
```
add:
.L.107__5:
        cmp     w0, #0
        ble     .L.107__1
        sxtw    x2, w0
        adrp    x0, a
        add     x0, x0, #:lo12:a
        adrp    x1, b
        add     x1, x1, #:lo12:b
        add     x4, x0, x2, LSL #2
.L.107__2:
        ldr     w2, [x0]
        add     w3, w2, #1
        str     w3, [x0]
        add     x0, x0, #4      <====== 可向上合并
        str     w2, [x1]
        add     x1, x1, #4       <====  可向上合并，合并后循环内可以少一条add指令。
        cmp     x0, x4
        blo     .L.107__2
.L.107__1:
        ret
.L.107__7:
        .size   add, .-add
```
llvm O2 shows:
```
add:                                    // @add
// %bb.0:                               // %entry
	cmp	w0, #1                          // =1
	b.lt	.LBB0_3
// %bb.1:                               // %for.body.preheader
	adrp	x9, b
	adrp	x10, a
	mov	w8, w0
	add	x9, x9, :lo12:b
	add	x10, x10, :lo12:a
.LBB0_2:                                // %for.body
                                        // =>This Inner Loop Header: Depth=1
	ldr	w11, [x10]
	subs	x8, x8, #1                      // =1
	add	w12, w11, #1                    // =1
	str	w12, [x10], #4
	str	w11, [x9], #4
	b.ne	.LBB0_2
.LBB0_3:                                // %for.cond.cleanup
	ret
.Lfunc_end0:
	.size	add, .Lfunc_end0-add
```

GVP 方舟编译器/OpenArkCompiler

Content Risk Flag

Comments (0)

GVP方舟编译器/OpenArkCompiler

Content Risk Flag