[llvm] [SystemZ] Use LCGR/AGHI for i64 XOR with -1 (PR #74882)

[Ulrich Weigand via llvm-commits]

uweigand wrote:

If we want to do that, the patch looks good.  However, I'm still not 100% sure we *do* want to do it.  The sequence is more compact, but we should also verify it has the same execution speed.  (On some CPUs, the XILF and XIHF *might* be able to execute in parallel as they touch different halves of the GPR.  The LCGR / AGHI always must be serialized.)  We should check the performance on a small loop on actual machines.

https://github.com/llvm/llvm-project/pull/74882