llvmbot
commented
11 months ago @llvm/issue-subscribers-backend-aarch64
Author: Sjoerd Meijer (sjoerdmeijer)
We are generating a lot of code with Clang for a loop that contains an if-then statement resulting in predicated instructions, which don't seem to be necessary looking at GCC's codegen. For this kernel s124 in TSVC, we are about 60% behind.
Compile this input with `-O3 -mcpu=neoverse-v2 -ffast-math`:
```
__attribute__((aligned(64))) float a[32000],b[32000],c[32000],d[32000],e[32000],
aa[256][256],bb[256][256],cc[256][256],tt[256][256];
int dummy(float[32000], float[32000], float[32000], float[32000], float[32000], float[256][256], float[256][256], float[256][256], float);
float s124(struct args_t * func_args)
{
int j;
for (int nl = 0; nl < 100000; nl++) {
j = -1;
for (int i = 0; i < 32000; i++) {
if (b[i] > (float)0.) {
j++;
a[j] = b[i] + d[i] * e[i];
} else {
j++;
a[j] = c[i] + d[i] * e[i];
}
}
dummy(a, b, c, d, e, aa, bb, cc, 0.);
}
}
```
Clang's codegen:
```
.LBB0_3: // Parent Loop BB0_2 Depth=1
ld1w { z0.s }, p2/z, [x12, x9, lsl #2]
ld1w { z1.s }, p2/z, [x25, x9, lsl #2]
ld1w { z2.s }, p2/z, [x24, x9, lsl #2]
asr x10, x8, #30
add x8, x8, x20
add x11, x23, x10
ld1w { z3.s }, p2/z, [x26, x9, lsl #2]
ld1w { z4.s }, p2/z, [x22, x9, lsl #2]
ld1w { z5.s }, p2/z, [x19, x9, lsl #2]
fcmgt p0.s, p2/z, z0.s, #0.0
fcmgt p1.s, p2/z, z1.s, #0.0
sel z0.s, p0, z0.s, z2.s
sel z1.s, p1, z1.s, z3.s
ld1w { z2.s }, p2/z, [x21, x9, lsl #2]
ld1w { z3.s }, p2/z, [x27, x9, lsl #2]
add x9, x9, x15
cmp x28, x9
fmla z0.s, p2/m, z4.s, z2.s
fmla z1.s, p2/m, z5.s, z3.s
st1b { z0.b }, p3, [x23, x10]
st1w { z1.s }, p2, [x11, x14, lsl #2]
b.ne .LBB0_3
```
vs. GCC's codegen:
```
.L2:
ldr q30, [x23, x0]
ldr q27, [x22, x0]
ldr q28, [x21, x0]
ldr q31, [x20, x0]
fcmgt v29.4s, v30.4s, 0
fmla v30.4s, v27.4s, v28.4s
fmla v31.4s, v27.4s, v28.4s
bit v31.16b, v30.16b, v29.16b
str q31, [x19, x0]
add x0, x0, 16
cmp x0, x24
bne .L2
```
See also:
https://godbolt.org/z/nb6xYxxKo
We are generating a lot of code with Clang for a loop that contains an if-then statement resulting in predicated instructions, which don't seem to be necessary looking at GCC's codegen. For this kernel s124 in TSVC, we are about 60% behind.
Compile this input with
-O3 -mcpu=neoverse-v2 -ffast-math:Clang's codegen:
vs. GCC's codegen:
See also: https://godbolt.org/z/nb6xYxxKo