[PATCH] D74268: [X86] Add back fmaddsub intrinsics to work towards fixing the strict fp implementation

[Craig Topper via Phabricator via llvm-commits]

craig.topper added a comment.

In D74268#1884942 <https://reviews.llvm.org/D74268#1884942>, @RKSimon wrote:

> I agree that its unlikely we won't lose much combine potential - we could add demanded elts support to simplify to fmadd/fmsub but that isn't that likely tbh.
>
> Can the the fmsubadd variants still safely use the fneg ?
>
> Are we missing AVX512 FMADDSUB_RND variants?


I think using neg on the input to represent fmsubadd should be fine. Fneg doesn’t cause exceptions and it shouldn’t affect the rounding behavior if it’s on the input.

We still have rounding intrinsics for avx512 those were never removed.


CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D74268/new/

https://reviews.llvm.org/D74268