| Issue |
202756
|
| Summary |
[X86] Suboptimal generation of same-argument cmpunord
|
| Labels |
|
| Assignees |
|
| Reporter |
dzaima
|
The code:
```c
#include<immintrin.h>
__m256d f(__m256d a) {
return _mm256_cmp_pd(a, a, 3);
}
```
gets compiled to:
```asm
vxorpd xmm1, xmm1, xmm1
vcmpunordpd ymm0, ymm0, ymm1
ret
```
despite being trivially better as explicitly written:
```asm
vcmpunordpd ymm0, ymm0, ymm0
ret
```
https://godbolt.org/z/z7WsY1Yb6
IR from middle-end:
```llvm
define noundef <4 x double> @f(<4 x double> noundef %a) {
%0 = fcmp uno <4 x double> %a, zeroinitializer
%1 = sext <4 x i1> %0 to <4 x i64>
%2 = bitcast <4 x i64> %1 to <4 x double>
ret <4 x double> %2
}
```
![]()
_______________________________________________
llvm-bugs mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs
