[llvm] [AMDGPU] Preserve resource requirements through call-graph cycles (PR #221625)

[Mian Miftah via llvm-commits]

mmiftahx wrote:

> I think it would be generally helpful for reviewers if you chart out what the 6 patches are going to be about. If you could use Github Stacked PRs / Graphite / spr, that may be helpful too.

@chinmaydd, sorry for the slow answer. The goal is finite resource bounds for AMDGPU indirect calls that list their callees. The series:

1. #221550: define empty and malformed `!callees` semantics
2. #225307 (merged): keep `!callees` when merging calls
3. carry the callee lists through call site info and MIR (needs 1)
4. record them in AMDGPU SelectionDAG and GlobalISel (needs 3)
5. **this pr** #221625
6. use the lists for finite AMDGPU resource bounds (needs 4 and 5)

3, 4 and 6 aren't posted yet. I agree 100% that stacking them would help, but GitHub stacked PRs, Graphite and spr all need branches in llvm/llvm-project, which needs commit access. I'll post them as separate PRs with a dependency note, as described in https://llvm.org/docs/GitHub.html#stacked-pull-requests which should hopefully help.

https://github.com/llvm/llvm-project/pull/221625