[llvm] [AMDGPU] Less-constrained scheduling for ASYNCMARK (PR #197638)

[Krzysztof Drewniak via llvm-commits]

krzysz00 wrote:

> ince we're making wait_asyncmark not hasSideEffects, it and other things should def/use something, but let's not call it ASYNCcnt, that's very confusing. Why don't all the instructions that can be tracked by asyncmark use/def something like "AsyncMarker". Asynccnt is the hardware counter, which is not the same as the modeling of async operations, as clearly demonstrated by tdm

Yeah, that's what I had in mind here

> Can you paste the link to the ticket? I think TENSORcnt can be added to def/use of asyncmark/wait_asyncmark to sequence TDM instrs and them.

https://amd-hub.atlassian.net/browse/LCOMPILER-1619 is the ticket (with apologies to external readers)

https://github.com/llvm/llvm-project/pull/197638