[PATCH] D48067: [InstCombine] Replacing X86-specific rounding intrinsics with generic floor-ceil

[Craig Topper via Phabricator via llvm-commits]

craig.topper added a comment.

I think where we ultimately want to end up is to remove the masking from the packed intrinsics and replace the scalar intrinsics with versions that use f32/f64 as their types. The IR would then look similar to where we're trying to end up for things like sqrt. But instead of a target independent intrinsic we would have a target specific intrinsic. All the the masking and insert/extract would be completely separate from the operation itself. That would greatly simplify the InstCombine code here because you would just need to trade out the target specific intrinsic for the floor/ceil intrinsic without having to worry about anything else. Thoughts?


Repository:
  rL LLVM

https://reviews.llvm.org/D48067