[PATCH] D14909: [X86][FMA] Optimize FNEG(FMUL) Patterns

[Sanjay Patel via llvm-commits]

spatel added a reviewer: scanon.

================
Comment at: lib/Target/X86/X86ISelLowering.cpp:26161
@@ +26160,3 @@
+  // use of a constant by performing (0 - A*B) instead.
+  if (Arg.getOpcode() == ISD::FMUL && DAG.getTarget().Options.UnsafeFPMath &&
+      (SVT == MVT::f32 || SVT == MVT::f64) &&
----------------
Do we need unsafe math for this transform?

================
Comment at: lib/Target/X86/X86ISelLowering.cpp:26163
@@ +26162,3 @@
+      (SVT == MVT::f32 || SVT == MVT::f64) &&
+      (Subtarget->hasFMA() || Subtarget->hasFMA4() || Subtarget->hasAVX512())) {
+    SDLoc DL(N);
----------------
Aren't the FMA checks enough? Is there some target that assumes that AVX512 includes FMA but does not set the FMA attribute?

================
Comment at: lib/Target/X86/X86ISelLowering.cpp:26164
@@ +26163,3 @@
+      (Subtarget->hasFMA() || Subtarget->hasFMA4() || Subtarget->hasAVX512())) {
+    SDLoc DL(N);
+    SDValue Zero = DAG.getConstantFP(0.0, DL, VT);
----------------
Nit: could hoist this and reuse below.


Repository:
  rL LLVM

http://reviews.llvm.org/D14909