[PATCH] D27692: [x86] use a single shufps when it can save instructions

[Sanjay Patel via Phabricator via llvm-commits]

spatel added a comment.

Thanks, Simon. I see 4 potential outcomes for this patch:

1. Abandon; we should take a different approach starting at a different point in the lowering sequence.
2. Add some limitations/refinements (eg: opt-for-size, check the input/output to see if they can avoid a domain switch?).
3. Accept it, but add fixes for the known problem cases before this lands.
4. Accept it as-is; the wins outweigh the losses. We can convert a shufps back to integer-equivalent shuffles in MachineCombiner as needed.

Any other options/thoughts?


https://reviews.llvm.org/D27692