[llvm] [AMDGPU] Use alias analysis for SMEM/VMEM WAR waitcnts (PR #206569)

[Matt Arsenault via llvm-commits]

arsenm wrote:

> > > > * `si-triv-disjoint-mem-access.ll`: false positive, for the same MMO TODO in the above test. The extra wait is between an SMRD constant load from `addrspace(4)` and a global/buffer store to `addrspace(1)`.
> > > 
> > > 
> > > AS1 and AS4 can alias to my knowledge, so it is valid. https://llvm.org/doxygen/AMDGPU_8h_source.html#l00617
> > 
> > 
> > They can alias, but as4 can be assumed invariant
> 
> So maybe the MMO for a load from AS4 should always be `isInvariant`?

Yes, long term I want to eliminate as4 and just make sure everything is invariant 

https://github.com/llvm/llvm-project/pull/206569