[Mlir-commits] [mlir] [mlir] [dataflow] Refactoring the definition of program points in dat… (PR #105656)

[Stella Laurenzo]

stellaraccident wrote:

Flyby comment 1: I am almost certain that all of the features behind InlineGlobalSlots.cpp are very close [to being decommissioned](https://github.com/llvm/torch-mlir/pull/3693), and then this very old code can be outright deleted vs carried forward. We've been clear since January that the day will come soon when the cost of maintaining this path cannot be carried anymore. If this is that day, then delete any remaining tests and I will approve.

Flyby comment 2: While we may get lucky in this case and simply not need to address this, in general, when doing an in-situ upgrade of an API, it is important to not just address where things get to but to not remove semantics that people are relying on along the way. I know this is hard/annoying. For this very specific case of impact to torch-mlir, we may be able to not worry about this at this time, but torch-mlir is not the only user of these APIs, and we need to be evolving them without semantic/representability reduction in intermediate steps. If that isn't possible, then it strongly indicates that an in-situ refactor isn't the appropriate way to go.

https://github.com/llvm/llvm-project/pull/105656