[PATCH] D124903: [X86][AMX][NFC] Replace PXOR instruction with SET0 in AMX pre config.
LuoYuanke via Phabricator via llvm-commits
llvm-commits at lists.llvm.org
Tue May 3 22:18:45 PDT 2022
LuoYuanke added inline comments.
================
Comment at: llvm/lib/Target/X86/X86PreTileConfig.cpp:393
.addReg(Ymm);
} else {
assert(ST.hasSSE2() && "AMX should assume SSE2 enabled");
----------------
craig.topper wrote:
> LuoYuanke wrote:
> > craig.topper wrote:
> > > Why is there no AVX1 code path to use VMOVUPSmr for 128 bits?
> > I think usaully AMX imply AVX512 enabled, and for the correctness, it can be covered by SSE2 instruction MOVUPSmr. So I think AVX1 code path is nice to have, but not must.
> We usually try not to mix AVX and SSE code due to the impact to the upper bits and how it affects the vzeroupper insertion. But I guess since these are stores they don't dirty the upper bits so maybe it isn't an impact.
Thanks for reminding. Actually I'm going to ask why use VMOVUPSmr for 128 bits given MOVUPSmr works both for SEE and AVX. So AVX-SSE switch penatly is a reason.
Repository:
rG LLVM Github Monorepo
CHANGES SINCE LAST ACTION
https://reviews.llvm.org/D124903/new/
https://reviews.llvm.org/D124903
More information about the llvm-commits
mailing list