[PATCH] D72138: [ARM] Fill in FP16 FMA patterns

[Dave Green via Phabricator via llvm-commits]

dmgreen marked an inline comment as done.
dmgreen added inline comments.


================
Comment at: llvm/lib/Target/ARM/ARMInstrVFP.td:2215
       Requires<[HasVFP4]>;
+def : Pat<(f16 (fma (fneg HPR:$Sn), HPR:$Sm, (fneg HPR:$Sdin))),
+          (VFNMAH HPR:$Sdin, HPR:$Sn, HPR:$Sm)>,
----------------
samparker wrote:
> Do you think it would be worth doing some canonicalisation somewhere? With two fnegs I'm assuming this pattern is more expensive than the others if not caught.
I gave this a go, but it appears that the opposite is sometimes true. AMDGPU has instructions that look like `v_fma_f32 v0, -v0, v1, -v1`, where each of the operands can be inverted for free. If we inverted the whole thing, then we would need to add reverse patterns. RISCV seems to have some cases with fnmadd.s that are made worse too.


CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D72138/new/

https://reviews.llvm.org/D72138