#[test] fns is not type checked when not testing

[nikomatsakis]

If you are not running --test, then #[test] functions are stripped out of the source. This is good, in one sense, but bad in another---the functions are not type checked! It is very easy for them to bitrot this way. I think they ought to stick around but not be translated; we can ensure that #[test] functions are not referenced from other functions in check_fn_usage, or whatever that pass is called.

[brson]

Can we possibly just move the test compiler pass further down the pipeline?

[brson]

There are also #[cfg(test)] functions as well, but it's harder to argue that those should be typechecked - should #[cfg(target_os = "win32")] functions be typechecked on linux?

[nikomatsakis]

I would prefer to type check as much as we can. If we can type check win32 functions, I think we should! But if may turn out that there are cases where two variants are mutually exclusive...? I am thinking of some of the math stuff where it was conditionally compiled such that float == f32 sometimes and sometimes float == f64.

But personally there's nothing I hate more than thinking I have everything good and then finding out that it won't even build on windows... usually it runs after I fix the build error, but the build fails on some trivial thing.

[marijnh]

Can't we just expect people to actually run their tests? Not typechecking is one form of bitrot, not passing is another. We don't want to run tests as part of a normal compilation.

[nikomatsakis]

If we can't compile everything---which I guess we can't---then it's fine. I feel like it's annoying not to know whether tests will at least build, but @marijnh has a point. Anyway it's probably not worth the time to fix. I'll close it for now.