Type inference regression in recent nightly

[Yamakaky]

https://play.rust-lang.org/?gist=188e52dc2ba423a8f78c6e381f278cf1&version=nightly&backtrace=1

source (copied by @pnkfelix from gist linked above):

fn main() {}

fn t() -> Result<(), String> {
    return Err("".into());
    Ok(())
}

Maybe it's because Ok(()) is unreachable?

See rust-lang-deprecated/error-chain#132 for the original report.

The breakage happens between nightlies 668864d 2017-02-16 and 536a900 2017-02-17.

[golddranks]

Seems similar to #39808, in the sense that there is some unreachable code that apparently was used to help inference before?

[TimNN]

Diff between the nightly-2017-02-17 and nightly-2016-02-18.

I expect this was caused by #39485 (and thus is expected breakage?).

[Yamakaky]

Seems like it.

[golddranks]

Is it written somewhere what inference breaking was within range of "expected" and what is not? It started to bother me: the problem doesn't seem to be that some type info isn't available because it would be inferred from dead code. After all, the function signature here is not dead code, and everything's inferrable from there. The problem seems to be that the type of the dead expression isn't fully inferred from the non-dead code. But should the type of E matter since the code path is dead? And if it does, shouldn't it be fully inferrable from the function signature?

[Yamakaky]

Exactly. BTW, it's not convenient if dead code doesn't typecheck. It happens to me a lot when I'm writing new code, along with unused variables for example.

[dhardy]

As @nikomatsakis points out, existing code could be broken by doing proper type-checking in dead code.
However, this code would already have carried a dead_code warning (same as this case).

Ironic that disabling type checking of dead-code is actually causing type errors (this issue).

I stand by my reasoning that dead code should be type checked (even if it breaks some existing code), however it would also make sense if the dead code warning was reported before errors in the dead code.

[golddranks]

So let me check: the recent changes disabled type-checking of dead code? And nevertheless it complains that the type of E in the dead code isn't known?

[dhardy]

I'm not very familiar with how the compiler works, but I suspect the demand_coerce function also causes some type inference to happen, and without that some other part of the code complains that a type parameter never got deduced.
https://github.com/rust-lang/rust/pull/39485/files#diff-1d1b0d29a2e8da97c6bfb6e364d920c7L4165

I'd rather someone who actually understands the compiler answer that question though...

[nikomatsakis]

Hmm. I think what we're doing in #39485 isn't quite right. I'm not sure the best answer here. Before #39485, we mostly but not completely ignored the types that arose from dead code. I'm nominating this for discussion in compiler team meeting; I think we have to address this regression in some way.