Ensure that GT_CNS_VEC is handled in LinearScan::isMatchingConstant #70171
Conversation
|
Tagging subscribers to this area: @JulieLeeMSFT Issue DetailsThis is a small follow up to #68874 but provides decent improvements, particularly where the same constant (Zero and AllBitsSet in particular) is repeatedly used.
|
|
Example diffs Note that vpcmpeqd ymm7, ymm7, ymm7
vxorps ymm6, ymm7, ymm6
vcmpps ymm4, ymm4, ymm0, 0
vorps ymm4, ymm4, ymm6
- vpcmpeqd ymm6, ymm6, ymm6
- vpcmpeqd ymm4, ymm4, ymm6
+ vpcmpeqd ymm4, ymm4, ymm7Note how vmovupd ymm2, ymmword ptr[reloc @RWD00]
vpshufb ymm0, ymm0, ymm2
vperm2i128 ymm0, ymm0, ymm0, 1
- vmovupd ymm2, ymmword ptr[reloc @RWD00]
vpshufb ymm1, ymm1, ymm2 |
| case GT_CNS_INT: | ||
| case GT_CNS_LNG: | ||
| case GT_CNS_DBL: | ||
| case GT_CNS_VEC: |
There was a problem hiding this comment.
Noting ARM64 already had a path here since it needed special handling for Zero and AllBitsSet, it just wasn't getting used in LSRA
|
@kunalspathak could you take a look? Its super small (8 lines) but touches LSRA so I'd appreciate your sign-off as well. |
Co-authored-by: Egor Bogatov <[email protected]>
This is a small follow up to #68874 but provides decent improvements, particularly where the same constant (Zero and AllBitsSet in particular) is repeatedly used.