Issue 242205
Summary [LV] failed to use correct VF to vectorize small const trip loops
Labels missed-optimization, llvm:loopvectorize
Assignees
Reporter Andarwinux
    https://godbolt.org/z/zaPfs57Wj
```c
void foo(float * restrict a, float * restrict b) {
    for (int i = 0; i < 27; ++i) {
        a[i]=a[i]*b[i];
    }
}
```
-g0 -O3 -march=novalake -mllvm -unroll-full-max-count=0
```asm
foo:
        vmovups ymm0, ymmword ptr [rdi]
        vmovups ymm1, ymmword ptr [rdi + 32]
        vmovups ymm2, ymmword ptr [rdi + 64]
        vmulps  ymm0, ymm0, ymmword ptr [rsi]
 vmovups ymmword ptr [rdi], ymm0
        vmulps  ymm0, ymm1, ymmword ptr [rsi + 32]
        vmovups ymmword ptr [rdi + 32], ymm0
        vmulps ymm0, ymm2, ymmword ptr [rsi + 64]
        vmovups ymmword ptr [rdi + 64], ymm0
        vmovss  xmm0, dword ptr [rdi + 96]
        vmulss  xmm0, xmm0, dword ptr [rsi + 96]
        vmovss  dword ptr [rdi + 96], xmm0
 vmovss  xmm0, dword ptr [rdi + 100]
        vmulss  xmm0, xmm0, dword ptr [rsi + 100]
        vmovss  dword ptr [rdi + 100], xmm0
        vmovss xmm0, dword ptr [rdi + 104]
        vmulss  xmm0, xmm0, dword ptr [rsi + 104]
        vmovss  dword ptr [rdi + 104], xmm0
        vzeroupper
 ret
```
```
Cost for VF 8: 6 (Estimated cost per lane: 0.75)
Cost for VF 16: 6 (Estimated cost per lane: 0.375)
LV: VF 8 has lower cost than VF 16 when taking the cost of the remaining scalar loop iterations into consideration for a maximum trip count of 27.
LV: Selecting VF: 8.
LEV: Epilogue vectorization is not profitable for this loop
```
 -O3 -march=novalake -mllvm -vectorizer-min-trip-count=32 -mllvm -unroll-full-max-count=0
```asm
foo:
        vmovups zmm0, zmmword ptr [rdi]
        vmulps  zmm0, zmm0, zmmword ptr [rsi]
        vmovups zmmword ptr [rdi], zmm0
        mov     ax, 2047
        kmovd   k1, eax
 vmovups zmm0 {k1} {z}, zmmword ptr [rdi + 64]
        vmovups zmm1 {k1} {z}, zmmword ptr [rsi + 64]
        vmulps  zmm0, zmm0, zmm1
        vmovups zmmword ptr [rdi + 64] {k1}, zmm0
        vzeroupper
        ret
```
As long as epilogue can be tail-folded, the widest VF should be selected.

_______________________________________________
llvm-bugs mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs

Reply via email to