[PATCH] D64710: [x86] try to keep FP casted+truncated+extracted vector element out of GPRs

Mon Jul 15 05:56:15 PDT 2019

RKSimon accepted this revision.
RKSimon added a comment.
This revision is now accepted and ready to land.

LGTM

================
Comment at: llvm/lib/Target/X86/X86ISelLowering.cpp:42410
+/// the sequence on XMM registers rather than moving between vector and GPRs.
+static SDValue combineToFPTruncExtElt(SDNode *N, SelectionDAG &DAG) {
+  // TODO: This is currently only used by combineSIntToFP, but it is generalized
----------------
spatel wrote:
> RKSimon wrote:
> > Could this be folded into vectorizeExtractedCast (which is done at lowering)?
> I did an experiment that just moved this code as-is to be called at lowering rather than combining, and we miss all but the last 2 test diffs in this patch. I didn't check to see how the other tests mutated, but clearly we'd have to match more patterns if we wait until lowering if we wanted to get those other cases. I can investigate more if that seems like the better option. 
> 
> My reason for doing the transform earlier is that we know vector bitcasts are generally free if produced early. If created later, they tend to interfere with other pattern matching.
OK - for now adding a TODO suggesting we merge this with vectorizeExtractedCast is acceptable - we already have a TODO there about supporting SIGN/ZERO_EXTEND so TRUNCATE as well would make sense.

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D64710/new/

https://reviews.llvm.org/D64710