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

[Jay Foad via Phabricator via llvm-commits]

foad added a comment.

> FMA and MAD formation is kind of complicated and I'm still leaning towards you being better off doing this in a combine where all other fma formation is done

But this isn't really FMA/MAD formation - if the addend is hard coded to 0 then these instructions just do a regular multiply. The "fused"ness or not doesn't have any effect.

> I'm assuming all these instructions internally perform the fma as f32?

Yes the "mix" forms all do a 32-bit mad/fma internally.


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

https://reviews.llvm.org/D153544