[PATCH] D35839: [X86] SET0 to use XMM registers where possible PR26018 PR32862

[Craig Topper via Phabricator via llvm-commits]

craig.topper added a comment.

I know Simon already accepted this, but shouldn't we be using xmm with VLX?



================
Comment at: lib/Target/X86/X86InstrInfo.cpp:7735
+    if (HasVLX)
+      return Expand2AddrUndef(MIB, get(X86::VPXORDZ256rr));
+    if (TRI->getEncodingValue(SrcReg) < 16) {
----------------
I think we should use VPXORDZ128rr and the equivalent xmm instead to be consistent. That will bring some of the tests back together I think.


================
Comment at: test/CodeGen/X86/vector-shuffle-combining-avx.ll:26
 define <4 x float> @combine_vpermilvar_4f32_identity(<4 x float> %a0) {
 ; X32-LABEL: combine_vpermilvar_4f32_identity:
 ; X32:       # BB#0:
----------------
There's no FileCheck prefix that references this prefix.

If we used xmm instead of ymm for VLX would this test not need to have its filecheck lines split?


https://reviews.llvm.org/D35839