[Mlir-commits] [mlir] [MLIR] Add option to postpone remark emission (PR #157434)

[Razvan Lupusoru]

razvanlupusoru wrote:

Idea 1 does not make sense to me because of my impression that users of MLIR will build their compilers with different pass pipelines. So the disadvantage you captured seems to be an intractable point:

> Disadvantage: Pass developers need to think what to postpone, and they should know the entire pipeline to design diagnostics correctly.

Back to OpenACC - we will support both Clang (CIR) and Flang. They will each have their own pass pipelines with a different order and use of MLIR OpenACC passes. It is not realistic that a pass itself knows anything about the pipeline; Mehdi captured this concern well: "That would be coupling a utility trying to do something to a specific pass pipeline!"

Thus Idea 2 seems like the better option. I don't think I grasp what the command line options has to do with the postponing aspect - but at least from my perspective I don't see idea 2 as being disruptive to how I am hoping to use the remarks framework. And I am indeed interested in seeing something implemented that will allow me to collect all Remarks instead of printing them when InFlightRemark gets destroyed.


https://github.com/llvm/llvm-project/pull/157434