[X86] llvm-mca reports wrong data for vpscatterqd, vscatterqps with -mcpu=skylake-avx512

[mikolaj-pirog]

llvm-mca with -mcpu=skx (or skylake-avx512 or common-avx512) reports wrong data (1==latency==uops==throughtput) for theses asm instructions:

vpscatterqd ymmword ptr [rsp + zmm0 + 16] {k1}, ymm1
vscatterqps ymmword ptr [rsp + zmm0 + 16] {k1}, ymm1

For their xmmword and qword the numbers reported are fine. For other newer -mcpu options the numbers are also fine. The llvm-mca version is 20.0.0git.

[llvmbot]

@llvm/issue-subscribers-tools-llvm-mca

Author: Mikołaj Piróg (mikolaj-pirog)

llvm-mca with -mcpu=skx (or skylake-avx512 or common-avx512) reports wrong data (1==latency==uops==throughtput) for theses asm instructions:

vpscatterqd ymmword ptr [rsp + zmm0 + 16] {k1}, ymm1
vscatterqps ymmword ptr [rsp + zmm0 + 16] {k1}, ymm1

For their xmmword and qword the numbers reported are fine. For other newer -mcpu options the numbers are also fine. The llvm-mca version is 20.0.0git.

[llvmbot]

@llvm/issue-subscribers-backend-x86

Author: Mikołaj Piróg (mikolaj-pirog)

llvm-mca with -mcpu=skx (or skylake-avx512 or common-avx512) reports wrong data (1==latency==uops==throughtput) for theses asm instructions:

vpscatterqd ymmword ptr [rsp + zmm0 + 16] {k1}, ymm1
vscatterqps ymmword ptr [rsp + zmm0 + 16] {k1}, ymm1

For their xmmword and qword the numbers reported are fine. For other newer -mcpu options the numbers are also fine. The llvm-mca version is 20.0.0git.

[boomanaiden154]

Seems like the information is just missing in the scheduling model which should be a relatively simple fix to add in. uops.info has (seemingly) correct port data that should not be too difficult to move into the LLVM skylake-avx512 scheduling model.

[RKSimon]

we haven't gotten much llvm-mca test coverage for scatter - I'll get that added first