[PATCH] D68651: [InstCombine] Signed saturation patterns

[Dave Green via Phabricator via llvm-commits]

dmgreen added a comment.

Yes. I was going off the prior art for adding sadd_sat and ssub_sat to instcombine. And from the cases this patch is matching, where we are otherwise extending to a higher type, the intrinsic seems to produce equal or better code in most of the cases I've tried now. So this wasn't just for vectorisation, although it does make things a lot simpler there. (On an arm specific note, we have a scalar qadd instruction that can be used, if we can sort out the "q" flag otherwise being visible from C).

If the canonical form for one of these signed saturating adds/subs wasn't an intrinsic, what would it be? Going into a higher type is awkward for us because the i64 add is not legal, and so doesn't look like the kind of instruction that should be vectorised, plus in ISel we'd have to catch the lowering fairly early and do something special.


CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D68651/new/

https://reviews.llvm.org/D68651