[Mlir-commits] [mlir] [mlir][CMake] enable disabling MLIR_ENABLE_EXECUTION_ENGINE (PR #171060)

[Mehdi Amini]

joker-eph wrote:

One high-level note: please don't merge PR before ensuring that reviewers are OK with it when there are on-going discussions, like here.

> It's a top-level CMake variable - I don't understand what the issue is with giving people the choice to disable it? 

There is no "issue" per-se: it's all a question of motivation and tradeoff.
MLIR is very modular in that you only need to link what you need and we try to not "pollute" your final binary with useless code, but we don't go the extra mile with a gazillion CMake option to avoid building some components by default. It increases the support surface and makes incredibly difficult to support the combination of these. So in the end it is a "ROI" kind of concern.

The existing logic is present because we detect an "incorrect" configuration where the JIT cannot work without the host backend available: we're trying to avoid people having weird errors. 
Hence the angle I took in my answer: the motivation you had express with your change was because of a configuration that leads to errors that could be "surprising" to the user, and hard for them to figure out what to do to unblock themselves.
In general my take in MLIR has been to try to catch these scenario earlier and diagnose the problem close to the source, where we can indicate which flag to tweak for example.

For a different motivation, I would answer differently, and right now the PR description does not include any unfortunately.

> I don't see the difference between that variable and these which are fully user settable

There is a fundamental difference in one that is "hermetic" to the LLVM build system (the same invocation gives the same results) and one where depending on whether cmake detects the right cuda or python  you end up with a different build configuration.


https://github.com/llvm/llvm-project/pull/171060