Yes, I'm using a global IORef.

If we want to avoid that, we need to find a place to store an arbitrary value (in this case, a ParsedModule) and have it make its way all the way to the post-typechecking hook. I was hoping we could use a module annotation (analogously to how you can put already resolved names into a RdrName) , but it looks like those only contain text and an expression.

Does this PR solve your use case? If not, what else would be necessary?

Not fully yet, no. I'll add details to #235, but basically there are two other points where source files are read in by LH. One of them is easily disabled by just setting noannotations = True in the config. The other one is some random place for caching I guess? I'll look it up.

Note that due to the breadcrumb stuff, this PR now includes #2368 as well. I think #2368 has less potential for disruption, and it fixes a manifest bug, so I think it will go in before this anyway.

FWIW, I 100% disagree with the HLint suggestion here:

warning Use tuple-section
Suggestion in swapBreadcrumb in module Language.Haskell.Liquid.GHC.Plugin:

Use tuple-section

• Found: \ old -> (old, new)
• Perhaps: (, new)
• Note: may require {-# LANGUAGE TupleSections #-} adding to the top of the file

In fact, in review I don't like my current code either :)

So what we really want is:

1. First time we parse a module, there is no breadcrumb for the current module so we can store a Parsed breadcrumb. Later on, if we get to the after-parser hook again and do find some existing breadcrumb, that is an assertion failure.

2. First time we typecheck a module, there is a Parsed breadcrumb for the current module. We replace it with a Typechecking breadcrumb, which also makes sure we don't retain all ASTs. Getting to typechecking without a breadcrumb is an assertion failure.

3. Recrusive typecheckers will see that there is already a Typechecking breadcrumb and do nothing.

I'm not sure we should ever remove the Typechecking breadcrumb because then we have no way of later detecting some of the invalid states above. Maybe replace it with a Done breadcrumb that is an error to encounter in both hooks.

I'll redo the implementation based on these ideas.