Ineffectual bitwise or with constant emitted for mask operand of vperm(b|w|d|q|ps|pd)

[cvijdea-bd]

Same thing as #106256, but also happens for the (avx2/avs512) permute[x]var intrinsics, while the PR #106377 seems to only fix it for (v)pshufb specifically.

Godbolt examples: https://godbolt.org/z/MsTcx7qYc

The vector permute intrinsics ignore all bits except the ones that match the required index size, e.g.:

• vpermb only uses 4, 5, 6 bits out of each mask byte element for 128, 256, 512 bit sized vectors respectively
• vpermw only uses 3, 4, 5 bits out of each 16-bit element in the mask
• etc.

The OR operations with unrelated bits should be optimzied out.

Probably applies to vpermt2 (e.g. _mm512_permutex2var_epi16) also, with 1 more bit used since they selected from two concatenated vectors.

cc @RKSimon

[llvmbot]

@llvm/issue-subscribers-backend-x86

Author: Cristian Vîjdea (cvijdea-bd)

Same thing as https://github.com//issues/106256, but also happens for the (avx2/avs512) permute[x]var intrinsics, while the PR https://github.com//pull/106377 seems to only fix it for (v)pshufb specifically.

Godbolt examples: https://godbolt.org/z/MsTcx7qYc

The vector permute intrinsics ignore all bits except the ones that match the required index size, e.g.:

• vpermb only uses 4, 5, 6 bits out of each mask byte element for 128, 256, 512 bit sized vectors respectively
• vpermw only uses 3, 4, 5 bits out of each 16-bit element in the mask
• etc.

The OR operations with unrelated bits should be optimzied out.

Probably applies to vpermt2 (e.g. _mm512_permutex2var_epi16) also, with 1 more bit used since they selected from two concatenated vectors.

cc @RKSimon

[RKSimon]

vpermilpd/vpermilps will need support as well - vpermilpd is annoying as it doesn't use the lsb for the index