[Bug target/78762] Regression: Splitting unaligned AVX loads also when AVX2 is enabled

venkataramanan.kumar at amd dot com gcc-bugzilla@gcc.gnu.org
Wed Dec 21 10:16:00 GMT 2016


https://gcc.gnu.org/bugzilla/show_bug.cgi?id=78762

--- Comment #9 from Venkataramanan <venkataramanan.kumar at amd dot com> ---
(In reply to Jakub Jelinek from comment #6)
> Sure, the question is (raised several times over the last couple of years)
> is if the generic tuning should not adjust slightly based on the selected
> ISAs.
> AVX2 is right now only supported by Intel CPUs and bdver4/znver1 from the
> AMD CPUs.  The Intel CPUs don't want this kind of tuning, dunno if
> bdver4/znver1 prefers such splitting or not.  If not and it is limited only
> to older AMD chips, the question is whether it shouldn't be removed from the
> generic tuning, or if narrowing the set of CPUs (by adding further ISAs)
> can't change some properties of the generic tuning (say define generic+avx2
> and generic+avx512f tunings that would be enabled for generic tuning if
> MASK_AVX2 or MASK_AVX512F is set).

I did some benchmarking on AMD avx2 target with and without unaligned avx256
load splitting. Noticing regressions when we allow the splitting of avx256
unaligned loads. 

Already for AMD AVX2 target, splitting is turned off.  So I suggest not to do
avx256 unaligned load splitting for generic also.


More information about the Gcc-bugs mailing list