[llvm] pr/amdgpu closed world (PR #66488)

[Johannes Doerfert via llvm-commits]

jdoerfert wrote:

> > So, when I put it to the current location (FullLTO last callback), all opt and llc test failed because they don't run that.
> 
> I'm still confused. They don't run it, correct. I expect them to still not run AMDGPUAttributor after this patch (keeping all the test updates I did in [b1bcb7c](https://github.com/llvm/llvm-project/commit/b1bcb7ca460fcd317bbc8309e14c8761bf8394e0) intact.

Options, as far as I can tell:
0) Old PM pipeline, as shown with the first version of this patch: runs always, requires the pipeline to set the closed world assumptions, only then works in a compile job with clang or opt.
1) New PM pipeline, as it is w/o this patch: runs always, never have closed world assumptions in a compile job with clang.
2) New PM pipeline, read LTOMode in the new PM to determine Post(Thin|Full)LTOLink and enable closed world assumption: runs always, has closed world assumption in ThinLTO compile only (when executed via clang), some test might work with opt.
3) New PM pipeline, add to FullLTO pipeline, as it is w/ this patch: runs always with clang (FullLTO or ThinLTO, hence all compiles), doesn't run with opt but tests need llvm-lto(2)

We could add "-attributor-assume-closed-world" to tests, as I did before, but that only works with 1) and 2). It also again requires us to set the flag from the driver/tooling (clang or llc). That's what I had before.

@shiltian offered to look into this. Maybe he has more luck than me.


https://github.com/llvm/llvm-project/pull/66488