Fix code for probing availability of AVX-512.
authorNathan Bossart <nathan@postgresql.org>
Tue, 23 Apr 2024 15:54:04 +0000 (10:54 -0500)
committerNathan Bossart <nathan@postgresql.org>
Tue, 23 Apr 2024 15:54:04 +0000 (10:54 -0500)
commit598e0114a3b16bb7edcaba6784b3794f816889ae
tree3c3e335bc66410633c814ab93d132a008c4de189
parentbb3ca232398658291d4b1eaae12cd6e88dfc96a3
Fix code for probing availability of AVX-512.

This commit fixes a few things:
* Instead of checking for CPU support of the "xsave" extension, we
  need to check for OS support of XGETBV instructions via the
  "osxsave" flag.
* We must check that additional XCR0 bits are set to be sure the
  ZMM registers are fully enabled.
* We should use the recommended ordering of steps.  Specifically,
  we need to check that the ZMM registers are enabled prior to
  checking for AVX-512 via CPUID.

In passing, split this code into separate functions to improve
readability.

Reported-by: Andrew Kane
Reviewed-by: Akash Shankaran, Raghuveer Devulapalli
Discussion: https://postgr.es/m/20240418024459.GA3385227%40nathanxps13
src/port/pg_popcount_avx512_choose.c