Ignore expected type in diverging blocks

[canndrew]

As per comment: #39297 (comment)

[rust-highfive]

r? @alexcrichton

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

[eddyb]

r? @nikomatsakis

Also, please use a more descriptive title (e.g. "Allow diverging blocks to have a type different than that of their trailing expression." but that's maybe too long).

[nikomatsakis]

@canndrew instead of removing the affected tests, can you move at least one of them into a run-pass test, with a comment explaining why the code now type-checks?

[nikomatsakis]

cc @eddyb

[nikomatsakis]

Oh, I see he already commented =)

[nikomatsakis]

It occurs to me that we perhaps ought to crater this too. That is, I guess that this can lead to types becoming constrained that were not otherwise (as we saw in #39009).

[canndrew]

Is there a way to skip debuginfo tests when testing rustc? They always fail for me and it throws me off because sometimes I think the other tests have succeeded when I've gotten that far.

As for the failing run-pass tests. This now fails (unable to infer enough type information for E)

fnfoo() -> Result<Foo,isize>{returnOk(Foo{x:Bar{x:22},a:returnErr(32)});}

Which is unfortunate. Shouldn't inference be able to infer that the Ok being returned has the type of the return type?

Also, this now fails (unable to infer enough type information for _):

pubfnindex_colors<Pix>(image:&ImageBuffer<Pix,Vec<u8>>)
-> ImageBuffer<Luma<u8>,Vec<u8>>wherePix:Pixel<Subpixel=u8> + 'static,{letmut indices:ImageBuffer<_,Vec<_>> = loop{};for(pixel, idx)in image.pixels().zip(indices.pixels_mut()){// failured occurred here ^^ because we were requiring that we// could project Pixel or Subpixel from `T_indices` (type of// `indices`), but the type is insufficiently constrained// until we reach the return below.}
indices
}

Which makes sense, but is also unfortunate. (Because the function body is diverging, the type of indices as the final expression isn't unified with the return type).

[nikomatsakis]

@canndrew

Shouldn't inference be able to infer that the Ok being returned has the type of the return type?

Well, the whole change here is basically that it doesn't have to. (Since it is dead-code.)

cc @eddyb -- it makes sense that this would be causing some back-compat issues, but it's a drag. Definitely have to do a crater run I guess.

[nikomatsakis]

Started a crater run.

[nikomatsakis]

Crater run results.

Root regressions

• libimagstore-0.2.0 (before) (after)
• imag-0.2.0 (before) (after)
• imag-counter-0.2.0 (before) (after)
• dotfiler-0.1.0 (before) (after)
• lion-0.0.0 (before) (after)
• mdblog-0.1.8 (before) (after)
• fiber-0.1.0 (before) (after)
• twig-0.1.0 (before) (after)

[canndrew]

So, according to that crater run there's 4 directly-broken packages: libimagestore and tera have already been fixed upstream, they just need to publish new versions or get people using those new versions. lion has been yanked from crates.io and twig can be fixed easily enough.

Of course if we did default to ! in these cases there wouldn't be any regressions. I'm not saying I prefer that solution just that it's a possibility.

What's everyone's take on all this?

Well, the whole change here is basically that it doesn't have to. (Since it is dead-code.)

Sure, I was just surprised that it didn't still force that particular bit of unification to happen.

[canndrew]

I've fixed the two failing tests.

[nikomatsakis]

@canndrew ok, well, seems like we might as well go through with this. I feel like the current behavior is pretty inconsistent.

Of course if we did default to ! in these cases there wouldn't be any regressions. I'm not saying I prefer that solution just that it's a possibility.

I'm not sure I understand what you mean here by "default to !".

[canndrew]

@canndrew ok, well, seems like we might as well go through with this. I feel like the current behavior is pretty inconsistent.

Sounds good to me.

I'm not sure I understand what you mean here by "default to !".

I just meant set the type variable which can't be inferred to !.

[nikomatsakis]

@canndrew I see. Well, I'm not a big fan of that idea, for a couple of reasons. For one thing, it would re-create the problem we are trying to fix here. Consider e.g. something like this:

letmut v = vec![];if condition {return;
v.push(vec![]);// this type variable is created in dead code}

Now the type of v becomes Vec<Vec<_0>> where _0 is "from dead code", and hence defaults to !. But we don't want that type to default to !.

In general, the rule of "dead code halts information flow up the tree (i.e., from expression to where expression is used)" but not "side-to-side" seems pretty good to me. Among other things, since the code is dead, there is (indeed) no data or information flow up the tree, and we can't know if you intended to add conversion and so forth.

[nikomatsakis]

cc @Nercury -- twig-rs may be affected by this change. The testing of the PR got this error:

error[E0282]: unable to infer enough type information about `E`
--> src/nodes/parser/expr.rs:379:5
|
379 | Ok(Expr::new_at(
| ^^ cannot infer type for `E`
|
= note: type annotations or generic parameter binding required

This PR affects information flow around dead-code. The problem seems to be that the type of this variable arg is "inferred" from an unimplemented! macro. The simplest fix is probably to remove this code but some type hints would help.

[nikomatsakis]

I'm marking this for relnotes, since it does affect code. The relnotes could read:

"Fixed bugs around how type inference interacts with dead-code. The existing code generally ignores the type of dead-code unless a type-hint is provided; this can cause surprising inference interactions particularly around defaulting. The new code uniformly ignores the result type of dead-code."

[nikomatsakis]

@bors r+

[bors]

📌 Commit 18be42c has been approved by nikomatsakis

[bors]

⌛ Testing commit 18be42c with merge 8725c90...

[bors]

💔 Test failed - status-appveyor

[alexcrichton]

@bors: retry * appveyor timeout

…

On Thu, Feb 16, 2017 at 1:47 PM, bors ***@***.***> wrote: 💔 Test failed - status-appveyor <https://ci.appveyor.com/project/rust-lang/rust/build/1.0.1981> — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#39485 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AAD95OQ-avTCsRkLda_veYnN4AtQf28Aks5rdKftgaJpZM4L2UC7> .

[bors]

⌛ Testing commit 18be42c with merge 82756e4...

[bors]

💔 Test failed - status-travis

[alexcrichton]

@bors: retry * network failure

…

On Thu, Feb 16, 2017 at 9:26 PM, bors ***@***.***> wrote: 💔 Test failed - status-travis <https://travis-ci.org/rust-lang/rust/builds/202486669> — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#39485 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AAD95C-9golHpDkFDDG6f4ZHN6ed11vWks5rdRNQgaJpZM4L2UC7> .

[Nercury]

@nikomatsakis twig was my early Rust project that is not finished and is paused now, so the breakage won't cause any issues

[bors]

⌛ Testing commit 18be42c with merge dc0bb3f...

[bors]

☀️ Test successful - status-appveyor, status-travis
Approved by: nikomatsakis
Pushing dc0bb3f to master...

[golddranks]

There was some concerns about what is expected breakage and what is not in #39984 but I suspect they are more likely to be answered here.

There is this code that returns an Ok(()) in dead code: https://play.rust-lang.org/?gist=188e52dc2ba423a8f78c6e381f278cf1&version=nightly&backtrace=1 It's broken, because the inference can't infer the Err(_) variant anymore. The variant is directly inferrable from the signature of the function that contains the dead code.

So, is it expected breakage that the dead branch of code is 1) required to have fully known types but 2) is not type-inferred, unlike live branches, so these dead branches must have additional type annotations, which they wouldn't need if they were live branches?