https://godbolt.org/z/Mf8vMfs5r
The improvement is clear in IR, and I don't see any real perf diff on the motivating x86 codegen from bug 51259, so we might as well go with the shorter IR: https://reviews.llvm.org/rG0369714b3168 If there's some enhancement/generalization that would make that better, let me know. Otherwise, we can resolve this.
(In reply to Sanjay Patel from comment #1) > If there's some enhancement/generalization that would make that better, let > me know. Otherwise, we can resolve this. I think we can resolve this, yes.