[PATCH] D48784: [X86] Prefer to form negate instructions instead of folding a load

[Craig Topper via Phabricator via llvm-commits]

craig.topper added a comment.

I'll run our benchmark list. This was more of an observation that we were different than icc, gcc, and msvc.

There's a related question. Given the option of promoting a (i16 sub 0, load) to 32-bits, should we promote and use neg+movzwl or keep it as 16-bits so we can fold the load.


https://reviews.llvm.org/D48784