[llvm] [MachineCSE][AArch64] Coalesce subregister copies behind an opt-in target hook (PR #224593)

[Matt Arsenault via llvm-commits]

arsenm wrote:

> Well, I believe the PPC blocker is been pending for 2+ months now. 

So review it? (Though really, the problem is asking for reviews from all the targets on these kind of codegen wide refactors. These kind of patches routinely hold up infrastructure by months and years). 

> Meanwhile, it makes sense to have to this a target hook. Once, the blocker is merged we can enable it unconditionally?

The hook will never be removed in that case. There needs to be incentive to do the yak shaving 

> Having subregs on target instructions always sounded very error prone to me.

This must work, and all code has all the information it needs to make sure this works correctly. This is just a matter of yak shaving all of the places that didn't bother to handle subregs.

IIRC analyzeCompare needs to gain subregister arguments for the last AArch64 regression I looked at. 

https://github.com/llvm/llvm-project/pull/224593