[PATCH] D75248: [Arm] Do not lower vmax/vmin to Neon instructions

Tue Mar 3 11:57:46 PST 2020

dmgreen added reviewers: efriedma, t.p.northover, jmolloy, john.brawn.
dmgreen added a comment.

I think this is probably OK, I'm not against it, but you could easily argue it either way. If we have an option called useNEONForSinglePrecisionFP, we should probably stick to it. Even if it's not universally applicable. So long as not-one else objects.

We should keep it consistent for FP16 as well though.

================
Comment at: llvm/lib/Target/ARM/ARMISelLowering.cpp:1432
+    }
     setOperationAction(ISD::FMINIMUM, MVT::f16, Legal);
     setOperationAction(ISD::FMAXIMUM, MVT::f16, Legal);
----------------
If we do this for f32, we should presumably do the same for f16.

Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D75248/new/

https://reviews.llvm.org/D75248