[llvm] [LoopIdiom] Select llvm.experimental.memset.pattern intrinsic rather than memset_pattern16 libcall (PR #126736)

[Philip Reames via llvm-commits]

preames wrote:

JFYI, @asb and I discussed this patch offline.  I asked if any of the remaining differences were likely to be material performance wise, and the tentative answer was no.  Given that, he's going to update the patch and I intend to approve (after normal review).  We're at the point where enabling and watching for real regressions is likely the best path forward.

Once this goes in, I intend to implement some of the obvious combines on the new intrinsic.  We've been deliberately waiting on that to have a fair comparison baseline.  

https://github.com/llvm/llvm-project/pull/126736