[PATCH] D32084: AMDGPU/GFX9: Set +fast-fmaf for >=gfx900 unless -cl-denorms-are-zero is set

[Matt Arsenault via Phabricator via cfe-commits]

arsenm added inline comments.


================
Comment at: lib/Basic/Targets.cpp:2208-2210
+      TargetOpts.Features.push_back(
+          (Twine(hasFullSpeedFMAF32(TargetOpts.CPU) &&
+          !CGOpts.FlushDenorm ? '+' : '-') + Twine("fast-fmaf")).str());
----------------
We glue fast-fmaf to the specific subtargets, so I don't think you can actually turn it and off. It has this nonsensical behavior where it assumes if you disable a subtarget feature for the processors you aren't actually targeting the processor at all. I think a dedicated new feature might be needed


https://reviews.llvm.org/D32084