[PATCH] D23583: [AArch64] Add feature has-fast-fma

[Matt Arsenault via llvm-commits]

arsenm added a subscriber: arsenm.
arsenm added a comment.

In https://reviews.llvm.org/D23583#517358, @MatzeB wrote:

> Can you explain in more detail what this is about? I would expect FMA to be as fast or faster than FMUL+FADD pretty much everywhere (oterhwise why would we bother with an extra instruction).


We have this problem in AMDGPU, the higher precision is slower for f32, I don't know if any AArch64 variants have this issue also


================
Comment at: llvm/lib/Target/AArch64/AArch64ISelLowering.cpp:7378
@@ +7377,3 @@
+bool AArch64TargetLowering::enableAggressiveFMAFusion(EVT VT) const {
+  return Subtarget->hasFastFMA() && isFMAFasterThanFMulAndFAdd(VT);
+}
----------------
Why is this check not part of isFMAFasterThanFMulAndFAdd?


Repository:
  rL LLVM

https://reviews.llvm.org/D23583