[PATCH] D153544: [AMDGPU] Use V_FMA_MIX* more often

[Mateja Marjanovic via Phabricator via llvm-commits]

matejam added a comment.





================
Comment at: llvm/lib/Target/AMDGPU/VOP3PInstructions.td:187
+                            (f32 (VOP3PMadMixMods f32:$src1, i32:$src1_modifiers))))),
+    (mixlo_inst $src0_modifiers, $src0,
+                $src1_modifiers, $src1,
----------------
matejam wrote:
> foad wrote:
> > arsenm wrote:
> > > If this is using v_mad_mix (i.e >= gfx900 && < gfx906), you can't introduce v_mad* without checking if denormal flushing is enabled
> > Is that a pre-existing problem with the other patterns in this multiclass (I see there is a TODO comment about it at the top)? Or do they check for denormal handling before creating anything that matches `fma_like`?
> If you run llvm/test/CodeGen/AMDGPU/mad-mix-lo.ll with -denormal-fp-math=ieee or preserve-sign, you will get the same result, both will select v_fma/mad_mix* instructions. 
> Or do they check for denormal handling before creating anything that matches fma_like?
It is checked in SITargetLowering::isFMAFasterThanFMulAndFAdd and SITargetLowering::isFMADLegal, I think.


CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D153544/new/

https://reviews.llvm.org/D153544