| Issue |
180695
|
| Summary |
[AVX10.2] Move out mask from bias convert (VCVTBIAS)* instruction intrinsic arguments
|
| Labels |
new issue
|
| Assignees |
|
| Reporter |
mahesh-attarde
|
In 1 of motivating case, we compute bias twice since mask combine did went through.
```
__m128i foo(__m128i __W, __mmask8 __U, __m128i __A, __m128h __B) {
return _mm_mask_cvtbiasph_bf8(__W, 0xf0, __A, __B) + _mm_mask_cvtbiasph_bf8(__W, 0x0f, __A, __B);
}
__m128i bar(__m128i __W, __mmask8 __U, __m128i __A, __m128h __B) {
return _mm_mask_dpph_ps(__W, 0xf0, __A, __B) + _mm_mask_dpph_ps(__W, 0x0f, __A, __B);
}
```
```
foo(long long vector[2], unsigned char, long long vector[2], _Float16 vector[8]):
mov al, -16
kmovd k1, eax
vmovaps xmm3, xmm0
vcvtbiasph2bf8 xmm3 {k1}, xmm1, xmm2
mov al, 15
kmovd k1, eax
vcvtbiasph2bf8 xmm0 {k1}, xmm1, xmm2
vpaddq xmm0, xmm0, xmm3
ret
bar(long long vector[2], unsigned char, long long vector[2], _Float16 vector[8]):
vmovaps xmm3, xmm0
vdpphps xmm3, xmm1, xmm2
vaddps xmm0, xmm3, xmm0
ret
```
We may need to use __builtin_select in bias intrinsic functions to let transform succeed.
Ref : https://godbolt.org/z/s51W33Ph1
![]()
_______________________________________________
llvm-bugs mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs
