[llvm] [SystemZ] Improve shouldCoalesce() for i128. (PR #74942)

[Ulrich Weigand via llvm-commits]

uweigand wrote:

> Sorry, but I am a little worried about that because if we only check the three-free for the coalesced GPR64, those three registers may be different ones than those free over the GPR128 interval.

I don't really think this is a concern - if this happens, then regalloc may have to insert a move at some CFG edge, but that should work fine.  The underlying problem requiring this special handling really was related to local problems with register availability around a call site and/or inline asm, I think.

https://github.com/llvm/llvm-project/pull/74942