[llvm] [DebugInfo] Handle additional types of stores in assignment tracking (PR #129070)

[Stephen Tozer via llvm-commits]

SLTozer wrote:

> Do we need a test covering the poison/undef path, i.e. an untagged unknown store to something where there's not a live dbg.value?

It's unlikely to be a source of errors, but it should be covered.

> Can we use Ìnstruction::mayWriteToMemory` as the test for whether a store is unknown, or is that signing us up to too much maintenance? It might make sense if we allow any intrinsic that writes to memory to adopt this behaviour.

We might use it in the verifier to determine whether a given instruction should be allowed to have a DIAssignID attached; in AssignmentTrackingAnalysis for the "unknown store" case, I'll have to look into it - the important part is that we can identify the base address that is being written to, which looks like it might be possible.

https://github.com/llvm/llvm-project/pull/129070