[llvm] [SystemZ] Improve handling of memmoves. (PR #196285)

[Jonas Paulsson via llvm-commits]

JonPsson1 wrote:

Performance runs show very little difference between max 2 stores vs 1. I guess this could mean we could keep the 2 limit and skip the implementation of the VLL/VSTL, even though that's not very involved.

Very little performance difference on the few affected benchmarks, although it ofc should be a good win not to do a libcall. Only real difference is a a weird 4% regression on imagick, which has in just one file 8 12-byte memmoves previously done with libcall now made with vll/vstl, or lg+lfh/stg+stfh. There doesn't seem to be any other side-effects (in fact a few less spill/reg-move instructions).



https://github.com/llvm/llvm-project/pull/196285