revert stabilization of `!` and `TryFrom`

[nikomatsakis]

Per #49593, we need to rollback the stabilization of ! (and a few dependent things) from 1.26. That's sad, but hopefully only temporary -- we'll bring it back once we can fix the coercion, but we don't want to be tinkering with the core type system in a beta backport. I'm opening this P-high issue specifically to track the backport, not the response.

[nikomatsakis]

triage: P-high

[nikomatsakis]

cc @canndrew -- this is sorta sad :(

[mitsuhiko]

Since ! is hard to search for I just left a comment on probably not the best issue. However I came across a case which now fails to compile I want to ensure does not fall under the radar: #48950 (comment)

[nikomatsakis]

@mitsuhiko

Interesting.

[nikomatsakis]

So @alexcrichton and I were saying that it would make sense to revert the TryFrom stabilization PRs first, and then the ! changes. I think he said @SimonSapin was going to do that?

[SimonSapin]

FWIW I think it sucks if we can’t find a satisfying solution to #49593 in time. If we must revert it should’t be in the opposite order but there’s no need to separate them either. I could prepare a commit that changes some #[stable] attributes back to #[unstable] and it can land together with the rest.

[nikomatsakis]

Argh. I meant to do this today, but I failed to get it done. It's going to be a bit tricky to revert !. If you look at the actual PR (#47630), you'll see there was a lot of complicated stuff.

I do not want to just revert back to what we had -- in particular I do not want to keep that flag we had on () type (we used to have two flavors of tuples, bad). However, that was the future compat warning mechanism we were using. We could use a coarser mechanism, where we check for any trait impls that concern () -- or just omit warnings for now. I'm inclined to do the latter at least to start (which in fact sidesteps the majority of the diff from #47630).

[nikomatsakis]

Hmm, ping @pnkfelix -- do you think you could take on reverting this next week?

[nikomatsakis]

FWIW I think it sucks if we can’t find a satisfying solution to #49593 in time

@SimonSapin I agree, but I feel reluctant to make changes to inference in a beta backport. But maybe it is worth expending some effort to see if we can find a minimal delta that we'd feel comfortable with.