[llvm] r326393 - [X86] Make sure we don't combine (fneg (fma X, Y, Z)) to a target specific node when there are no FMA instructions.
Craig Topper via llvm-commits
llvm-commits at lists.llvm.org
Wed Feb 28 16:08:38 PST 2018
Author: ctopper
Date: Wed Feb 28 16:08:38 2018
New Revision: 326393
URL: http://llvm.org/viewvc/llvm-project?rev=326393&view=rev
Log:
[X86] Make sure we don't combine (fneg (fma X, Y, Z)) to a target specific node when there are no FMA instructions.
This would cause a 'cannot select' error at isel when we should have emitted a lib call and an xor.
Fixes PR36553.
Added:
llvm/trunk/test/CodeGen/X86/pr36553.ll
Modified:
llvm/trunk/lib/Target/X86/X86ISelLowering.cpp
Modified: llvm/trunk/lib/Target/X86/X86ISelLowering.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Target/X86/X86ISelLowering.cpp?rev=326393&r1=326392&r2=326393&view=diff
==============================================================================
--- llvm/trunk/lib/Target/X86/X86ISelLowering.cpp (original)
+++ llvm/trunk/lib/Target/X86/X86ISelLowering.cpp Wed Feb 28 16:08:38 2018
@@ -35737,7 +35737,7 @@ static SDValue combineFneg(SDNode *N, Se
// If we're negating an FMA node, then we can adjust the
// instruction to include the extra negation.
unsigned NewOpcode = 0;
- if (Arg.hasOneUse()) {
+ if (Arg.hasOneUse() && Subtarget.hasAnyFMA()) {
switch (Arg.getOpcode()) {
case ISD::FMA: NewOpcode = X86ISD::FNMSUB; break;
case X86ISD::FMSUB: NewOpcode = X86ISD::FNMADD; break;
Added: llvm/trunk/test/CodeGen/X86/pr36553.ll
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/CodeGen/X86/pr36553.ll?rev=326393&view=auto
==============================================================================
--- llvm/trunk/test/CodeGen/X86/pr36553.ll (added)
+++ llvm/trunk/test/CodeGen/X86/pr36553.ll Wed Feb 28 16:08:38 2018
@@ -0,0 +1,20 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc < %s -mtriple=x86_64-apple-darwin10 | FileCheck %s
+
+; Make sure we don't crash because we negated an fma when we didn't have any fma instructions.
+
+define float @pr36553(float %a, float %b, float %c) nounwind {
+; CHECK-LABEL: pr36553:
+; CHECK: ## %bb.0: ## %entry
+; CHECK-NEXT: pushq %rax
+; CHECK-NEXT: callq _fmaf
+; CHECK-NEXT: xorps {{.*}}(%rip), %xmm0
+; CHECK-NEXT: popq %rax
+; CHECK-NEXT: retq
+entry:
+ %0 = tail call float @llvm.fma.f32(float %a, float %b, float %c)
+ %sub = fsub float -0.000000e+00, %0
+ ret float %sub
+}
+
+declare float @llvm.fma.f32(float, float, float)
More information about the llvm-commits
mailing list