[PATCH] D51706: ARM64: improve non-zero memset isel by ~2x

[JF Bastien via Phabricator via llvm-commits]

jfb added a comment.

In https://reviews.llvm.org/D51706#1225382, @efriedma wrote:

> If I'm understanding correctly, the reason the f128 operations aren't efficient is that float immediate lowering doesn't know how to use movi?  That seems like it would be more straightforward to solve by fixing the float immediate lowering, rather than messing with memset lowering.


I initially looked at this but it wasn't more straightforward. It's also odd to me because we know `memset` is `i8`, so `v2xi64` clearly works, but we don't know that other values will work as `f128` or anything else (and we don't have the value to look at here).


Repository:
  rL LLVM

https://reviews.llvm.org/D51706