[llvm] [SystemZ] Avoid unaligned/overlapping accesses with memcpy/memset. (PR #187100)

[Ulrich Weigand via llvm-commits]

uweigand wrote:

> Added full testing of memcpy/memmove/memset (however no volatile versions), with the current trunk output on HEAD^. This shows the unfortunate changes for memmove as they all became libcalls with the recent change in common-code. It also shows that memset-zero of 16 bytes results in an unaligned VST even though 'false' is returned for it here. Other than that, all VL/VSTs are now aligned in these tests.

I agree this is a good first step (assuming performance verification goes well).   Can you maybe pre-commit the tests, so we clearly see what actually changes with the patch itself?

The memmove change is indeed unfortunate, we need to fix this as a follow-on.

https://github.com/llvm/llvm-project/pull/187100