[PATCH] D55163: AMDGPU: Add optimization patterns to combine fp32->fp16 conversions
Matt Arsenault via Phabricator via llvm-commits
llvm-commits at lists.llvm.org
Mon Jan 21 15:01:00 PST 2019
arsenm added inline comments.
================
Comment at: lib/Target/AMDGPU/SIInstructions.td:1597
+let SubtargetPredicate = isGCN in {
+
----------------
pendingchaos wrote:
> arsenm wrote:
> > Should use isVI, or maybe these should be distinguished by GCN3Encoding? Needs a comment for why these are separated
> Since all GCN versions support the VOP3a form, shouldn't it use isGCN (to combine the modifiers into the instruction on SI/CI)?
>
> The VOP2 form is only supported on SI/CI, so isSICI is used. IIRC VOP2 ended up being used when no modifiers could be folded.
isGCN is obsolete and should be removed anywhere it's used.
Since we try to shrink instructions later, we only want to select the _e64 version when possible
Repository:
rL LLVM
CHANGES SINCE LAST ACTION
https://reviews.llvm.org/D55163/new/
https://reviews.llvm.org/D55163
More information about the llvm-commits
mailing list