[PATCH] D77152: [SelectionDAG] Better legalization for FSHL and FSHR

Thu Jul 16 01:59:33 PDT 2020

foad marked an inline comment as done.
foad added inline comments.

================
Comment at: llvm/test/CodeGen/PowerPC/funnel-shift.ll:22-24
+; CHECK-NEXT:    clrlwi 5, 5, 27
+; CHECK-NEXT:    srwi 4, 4, 1
+; CHECK-NEXT:    clrlwi 6, 6, 27
----------------
RKSimon wrote:
> foad wrote:
> > RKSimon wrote:
> > > foad wrote:
> > > > This is a slight regression. We need one more instruction overall because of the two clrlwi instructions to mask both normal and the inverted shift amounts.
> > > Any luck with this?
> > No. I think it would need a target-specific lowering to take advantage of the fact that shift by 32 is well defined on PowerPC. Can I leave this for a later patch, and/or could a PowerPC expert volunteer to help with it?
> So we're not just missing a micro optimization we can add to the powerpc backend as part of this patch? @hfinkel Any suggestions?
I can't think of a simple micro optimization that would help. For PowerPC it would be better to expand fshl(X, Y, Z) as `X << (Z & 31) | Y >> (32 - (Z & 31))` (where the right shift supports shifting by 32) which would be better than either the "before" or the "after" code shown in this diff.

Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D77152/new/

https://reviews.llvm.org/D77152