caio/clang-p2996
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

[SelectionDAG] Fix NaN regression in fma dag-combine. (#146592)

Browse Source 
After 901e1390c9 (#127770), the DAG
combine would transform `fma(x, 0.0, 1.0)` into `1.0` if
`-fp-contract=fast` was enabled, in addition to when 'x' is marked
nnan/ninf.

It's only valid in the latter case, not the former, so delete the extra
condition.
This commit is contained in:
James Y Knight
2025-07-01 18:10:30 -04:00
committed by GitHub
parent 475cd8dfaf
commit ae2104897c
2 changed files with 13 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

3
llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
View File

@@ -18087,8 +18087,7 @@ template <class MatchContextClass> SDValue DAGCombiner::visitFMA(SDNode *N) {
// FIXME: use fast math flags instead of Options.UnsafeFPMath
// TODO: Finally migrate away from global TargetOptions.
if (Options.AllowFPOpFusion == FPOpFusion::Fast ||
(Options.NoNaNsFPMath && Options.NoInfsFPMath) ||
if ((Options.NoNaNsFPMath && Options.NoInfsFPMath) ||
(N->getFlags().hasNoNaNs() && N->getFlags().hasNoInfs())) {
if (Options.NoSignedZerosFPMath || N->getFlags().hasNoSignedZeros() ||
(N2CFP && !N2CFP->isExactlyValue(-0.0))) {

12
llvm/test/CodeGen/X86/dag-combiner-fma-folding.ll
View File

@@ -1,5 +1,6 @@
; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
; RUN: llc -mtriple=x86_64-- --start-before=x86-isel -mattr=+avx,+fma %s -o - | FileCheck %s
; RUN: llc -mtriple=x86_64-- --start-before=x86-isel -mattr=+avx,+fma %s -o - -fp-contract=fast | FileCheck %s
define double @fma_folding(double %x) {
; CHECK-LABEL: fma_folding:
@@ -20,3 +21,14 @@ define double @fma_no_folding(double %x) {
%fused = call contract nnan ninf double @llvm.fma.f64(double %x, double 0.0, double -0.0)
ret double %fused
}
define double @fma_no_fold_potential_nan(double %x) {
; CHECK-LABEL: fma_no_fold_potential_nan:
; CHECK: # %bb.0:
; CHECK-NEXT: vxorpd %xmm1, %xmm1, %xmm1
; CHECK-NEXT: vfmadd213sd {{.*#+}} xmm0 = (xmm1 * xmm0) + mem
; CHECK-NEXT: retq
%fused = call contract double @llvm.fma.f64(double %x, double 0.0, double 1.0)
ret double %fused
}