[llvm] [MLGO] Model selection for models lowered through EmitC (PR #212650)

[Andrzej Warzyński via llvm-commits]

banach-space wrote:

Thanks for working on this @beamandala !

This is not really my area of expertise, so I will defer to folks more qualified than me to review :) I do have three high-level comments though:
* Tests are missing. For example, there are new  CLI flags, but no tests to verify them.
* (Global) CMake flags that include `EMITC` in their name (e.g.`LLVM_HAVE_EMITC_LOWERING`) are IMHO confusing. The new flow is based on MLIR rather than EmitC - EmitC is an implementation detail (a very important one, but still an implementation detail). TOSA and MemRef are other super important dialects involved in this. 
* I recommend removing the hard-coded pass pipeline from `MLGOLower.cmake`. That's for two reasons: (1) That pipeline might not work for every possible model that people come up with and it is likely to evolve in the future. (2) It contains non-trivial logic from MLIR that should be verified in MLIR. Otherwise, when things evolve in MLIR, we may end-up with hard-to-triage test failures in MLIR.

Regarding my last point, the lowering pipeline should be shared with tests added in https://github.com/llvm/llvm-project/pull/211833 so that we always verify the "production" pipeline.

Thank you :)

https://github.com/llvm/llvm-project/pull/212650