]> granicus.if.org Git - llvm/commitdiff
[AArch64] Support for FP FMA when -ffp-contract=fast
authorEvandro Menezes <e.menezes@samsung.com>
Thu, 15 Sep 2016 19:55:23 +0000 (19:55 +0000)
committerEvandro Menezes <e.menezes@samsung.com>
Thu, 15 Sep 2016 19:55:23 +0000 (19:55 +0000)
Currently, the machine combiner can proceed matching when -ffast-math is on.
It should also match when only -ffp-contract=fast is specified as was the
case before when DAGCombiner was doing the job.

Patch by: Abderrazek Zaafrani <a.zaafrani@samsung.com>.

Differential Revision: https://reviews.llvm.org/D24366

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@281649 91177308-0d34-0410-b5e6-96231b3b80d8

lib/Target/AArch64/AArch64InstrInfo.cpp
test/CodeGen/AArch64/arm64-fma-combine-with-fpfusion.ll [new file with mode: 0644]

index 6907e061332e23d8f841771c9c97594516806e4e..859f78289012db18eebf0062df311f347d3d1b49 100644 (file)
@@ -2901,6 +2901,8 @@ static bool isCombineInstrCandidate64(unsigned Opc) {
 // FP Opcodes that can be combined with a FMUL
 static bool isCombineInstrCandidateFP(const MachineInstr &Inst) {
   switch (Inst.getOpcode()) {
+  default:
+    break;
   case AArch64::FADDSrr:
   case AArch64::FADDDrr:
   case AArch64::FADDv2f32:
@@ -2911,9 +2913,9 @@ static bool isCombineInstrCandidateFP(const MachineInstr &Inst) {
   case AArch64::FSUBv2f32:
   case AArch64::FSUBv2f64:
   case AArch64::FSUBv4f32:
-    return Inst.getParent()->getParent()->getTarget().Options.UnsafeFPMath;
-  default:
-    break;
+               TargetOptions Options = Inst.getParent()->getParent()->getTarget().Options; 
+    return (Options.UnsafeFPMath || 
+                                   Options.AllowFPOpFusion == FPOpFusion::Fast);
   }
   return false;
 }
diff --git a/test/CodeGen/AArch64/arm64-fma-combine-with-fpfusion.ll b/test/CodeGen/AArch64/arm64-fma-combine-with-fpfusion.ll
new file mode 100644 (file)
index 0000000..095a0b0
--- /dev/null
@@ -0,0 +1,12 @@
+; RUN: llc < %s -mtriple=aarch64-linux-gnu -fp-contract=fast | FileCheck %s
+define float @mul_add(float %a, float %b, float %c) local_unnamed_addr #0 {
+; CHECK-LABEL: %entry
+; CHECK: fmadd {{s[0-9]+}}, {{s[0-9]+}}, {{s[0-9]+}}
+  entry:
+    %mul = fmul float %a, %b
+    %add = fadd float %mul, %c
+    ret float %add
+}
+
+attributes #0 = { norecurse nounwind readnone "correctly-rounded-divide-sqrt-fp-math"="false" "disable-tail-calls"="false" "less-precise-fpmad"="false" "no-frame-pointer-elim"="true" "no-frame-pointer-elim-non-leaf" "no-infs-fp-math"="false" "no-jump-tables"="false" "no-nans-fp-math"="false" "no-signed-zeros-fp-math"="false" "no-trapping-math"="false" "stack-protector-buffer-size"="8" "target-cpu"="generic" "target-features"="+neon" "unsafe-fp-math"="false" "use-soft-float"="false" }
+