Make bottom type fallback lazy

[edwardw]

PR #17603 introduced bottom type fallback but did it a bit too
eagerly. This patch makes the fallback lazy so that typeck can run
its cause and detect as many type errors as possible with regard to
diverging types.

Closes #21878

[rust-highfive]

r? @Aatch

(rust_highfive has picked a reviewer for you, use r? to override)

[alexcrichton]

cc @nikomatsakis

[edwardw]

r? @nikomatsakis

[nikomatsakis]

Hmm, I'm not sure how I feel about this. This introduces a "second kind" of fallback. I was planning to start work soon on RFC rust-lang/rfcs#213, which would generalize fallback for integer literals to other type parameters, and part of the reason I felt ok with ! falling back to () was that it would fit into this general framework. But then again perhaps it is best to make it special -- I agree that the examples from #21878 are unfortunate. And this change seems pretty reasonable overall. (Have you done any experiments to see how much code breaks if we just remove ! fallback?)

[edwardw]

I just did a quick experiment of removing ! fallback altogether. It failed right away at libcore/num/mod.rs with an error saying "cannot determine a type for this expression: unconstrained type".

[edwardw]

And of course, feel free to close this PR if you have something more general in the making.

[nikomatsakis]

@edwardw well what I had in mind is kind of orthogonal to what you've done here. I'm still tossing it about in my mind.

[steveklabnik]

@nikomatsakis what's your thoughts on this today?

[bors]

☔ The latest upstream changes (presumably #26232) made this pull request unmergeable. Please resolve the merge conflicts.