This is an archive of the discontinued LLVM Phabricator instance.

[X86] Disable commuting for the first source operand of zero masked scalar fma intrinsic instructions.
ClosedPublic

Authored by craig.topper on Mar 3 2020, 7:40 AM.

Details

Summary

I believe this is the correct fix for D75506 rather than disabling all commuting. We can still commute the remaining two sources.

Diff Detail

Event Timeline

craig.topper created this revision.Mar 3 2020, 7:40 AM
Herald added a project: Restricted Project. · View Herald TranscriptMar 3 2020, 7:40 AM
Herald added a subscriber: hiraditya. · View Herald Transcript
spatel added inline comments.Mar 3 2020, 8:58 AM
llvm/test/CodeGen/X86/avx512-intrinsics.ll
5816

Not too familiar with this code path, but we can shrink this test a bit and still crash:

define <4 x float> @test_int_x86_avx512_maskz_vfmadd_ss_load0(i1 zeroext %t0, <4 x float>* nocapture readonly %t1, float %t2, float %t3) {
  %t5 = load <4 x float>, <4 x float>* %t1, align 16
  %t6 = extractelement <4 x float> %t5, i64 0
  %t9 = tail call float @llvm.fma.f32(float %t6, float %t2, float %t3) #2
  %t12 = select i1 %t0, float %t9, float 0.0
  %t13 = insertelement <4 x float> %t5, float %t12, i64 0
  ret <4 x float> %t13
}

Simplify test

RKSimon added inline comments.Mar 3 2020, 10:03 AM
llvm/test/CodeGen/X86/avx512-intrinsics.ll
5832

(sidenote) losing the mask register from the asm comment is really bad.......

spatel accepted this revision.Mar 3 2020, 11:08 AM

LGTM

This revision is now accepted and ready to land.Mar 3 2020, 11:08 AM