[llvm] [AA] A conservative fix for atomic store instruction. (PR #155032)

[David Li via llvm-commits]

david-xl wrote:

yes, the same fix should be applied to LoadInst too (with a test case). @nikic can you clarify 'return ModRef for relaxed?'  

The fix can probably be expanded to handle stronger constraints. For instance if the store is 'seq_cst', it imposes order on seq_cst loads of all addresses but not on non seq_cst loads. To handle this, the interface will need to be enhanced to pass the instruction instead of MemoryLocation.

https://github.com/llvm/llvm-project/pull/155032