[llvm] [SelectionDAG] Require same MMO base in relative-align no-alias proof (PR #217880)

[via llvm-commits]

chenqiant wrote:

You're right — I misdiagnosed this.

The relative-align proof is residue-class arithmetic on BaseAlign; it does
not need a shared MMO Value. Requiring getValue() equality is overly
conservative (and explains the unrelated X86 FileCheck churn).

The RISC-V test also doesn't exercise the patch: dst align 16 vs src
align 8 already fails OrigAlignment0 == OrigAlignment1, so the proof
never fires. The interleaved schedule we saw on our fork came from
getMemmoveLoadsAndStores tagging every load/store with the base
Alignment (and promoting src align up to dst). #177885's
commonAlignment(*SrcAlign, SrcOff) is the actual fix.

I'll close this PR. Thanks.

https://github.com/llvm/llvm-project/pull/217880