Add flag to disable type family consistency check

Type family consistency checks can become a performance bottleneck (#5224, #19703, #25554, #25555). It would be good to investigate ways to do better here, but it seems fundamentally hard to reconcile the check with separate compilation. Moreover, there are cases in which the user may "know" that their family instances are consistent (or at least be willing to take the soundness risk at some stages of development, knowing that a later compilation or their CI setup will catch any actual problems). So perhaps GHC should have a(n unsafe) flag to disable the check?

Side note: #13099 seems to say that the current check is in some cases not conservative enough. It's also interesting to think about how the check should be managed in one-shot mode; does responsibility for running the check fall to GHC or the build system?

added Tfeature request TypeFamilies compiler perf needs triage labels

Indeed it seems like the consistency check is conceptually closer to a link-time check than a compile-time check. One could imagine a scheme where it could be implemented this way. When performing the final link in make mode the compiler would load the interfaces for all modules invoked in the link and run the check once across their transitive import closure. This would likely already improve things considerably.

Sadly one-shot mode is a bit harder given the current interfaces since we can't necessarily load interface files.

I don't see how one-shot mode is harder in particular? There is still a link phase when the object files are linked together. There is also the --abi-hash major mode which must read all interface files in order to compute the ABI hash for a project.

#13099 seems to be an issue with the recompilation checker rather than the consistency check exactly.

One particular issue is if you are compiling HPT modules then you can end up linking an expression before the final link time (if you are using TH)

As a general point, the language should avoid adding features which go against separate compilation as they do not scale well in the long run.

I don't see how one-shot mode is harder in particular?

When invoking GHC in one-shot mode the user may provide the modules to be linked a paths to object files. Consequently we may not be able to locate the corresponding interface files. Perhaps we just say that this is not a valid invocation but this would be a change; this sort of usage is, IIRC, allowed today.

removed needs triage label