impl_trait_in_bindings fails with Option<impl Trait>

[nikomatsakis]

This example ICEs:

#![feature(nll)]
#![feature(impl_trait_in_bindings)]

use std::fmt::Debug;

fn main() {
    let x: Option<impl Debug> = Some(44_u32);
    println!("{:?}", x);
}

The reason is that the code which chooses when to "reveal" opaque types looks only for opaque types at the top level (and, oddly, it does so only if ordinary unification fails; having code that branches based on whether unification succeeded is in general a bad idea, so we should fix that too):

rust/src/librustc_mir/borrow_check/nll/type_check/mod.rs

Lines 893 to 899 in c7df1f5

	if let Err(terr) = self.sub_types(sub, sup, locations, category) {
	if let TyKind::Opaque(..) = sup.sty {
	return self.eq_opaque_type_and_type(sub, sup, locations, category);
	} else {
	return Err(terr);
	}
	}

More deeply, though, this code is taking a somewhat flawed approach. In particular, it is looking at the results of inference, but -- at least presently -- opaque types can end up in the inferred type in one of two ways. (As shown by #54593.)

For example, one might have a mixture of opaque types that came from a recursive call (and which ought not to be revealed) and opaque types that the user wrote:

#![feature(nll)]
#![feature(impl_trait_in_bindings)]

use std::fmt::Debug;

fn foo() -> impl Copy {
    if false {
        let x: (_, impl Debug) = (foo(), 22);
    }
    ()
}

fn main() { }

The correct way to do this, I suspect, is to leverage the user-type annotations that we are tracking for NLL.

cc @alexreg @cramertj @eddyb

[nikomatsakis]

Another relevant example:

#![feature(nll)]
#![feature(impl_trait_in_bindings)]

use std::fmt::Debug;

fn foo() -> String {
    let x: impl Debug = 22;
    format!("{}", x)
}

fn main() { }

This fails to compile, which seems correct, though I'm not sure precisely what semantics we want.

Alternatively, we may want to encode the impl Debug sort of thing as a kind of cast that "obscures" the underlying type? Have to think about it.

[alexreg]

@nikomatsakis I think the second example works exactly as intended (and as the RFC specifies). It's intuitive to me, at least.

You're right about the first example; that's the wrong way to go about it. I didn't see a better approach at the time though, and I forgot about it while solving other issues. The lowering and type checking is basically okay, that said... it's the NLL borrow checking that's the problem. Which I think is something that requires your expertise, to be honest.

[Centril]

Tagging T-lang due to Niko's notes in #54600 (comment).

[mikeyhew]

@nikomatsakis

another relevant example:

Shouldn't format!("{}", x) should be format!("{:?}", x)?

[alexreg]

@mikeyhew Nope, I think he meant that... since he follows it by saying "This fails to compile, which seems correct". With {:?} it actually compiles, as it should. :-)

[nikomatsakis]

So @alexreg has been asking me to leave some notes on how to do this properly. The high-level strategy that I think we should be pursuing is changing impl Foo into types into a kind of coercion (at least in let position). So this:

let x: impl Foo = y

would sort of compile "as if" the user had written let x = y as impl Foo -- in fact, the latter syntax feels like something we could support. Basically, we're adding an explicit operation that "forgets" the true type of y and coerces it into the "opaque" type. This is sort of analogous to what happens if you do let x: &dyn Foo = y-- there is a coercion that takes place.

In fact, it might be illustrative to see what happens there. Given this example program (playground):

fn main() {
    let x: &dyn std::fmt::Debug = &3;
}

we get the following MIR (excerpted and simplified):

fn main() -> (){
    ...
    let _1: &dyn std::fmt::Debug;    // "x" in scope 2 at src/main.rs:2:9: 2:10
    let mut _2: &i32;
    let mut _3: &i32;

    bb0: {                              
        ...
        _3 = &(promoted[0]: i32);
        _2 = _3;
        _1 = move _2 as &dyn std::fmt::Debug (Unsize);
        ...
    }
}

The key bit is that Unsize coercion. I imagine we'll wind up with something similar, in which we have something like

_1 = hide(move _2, opaque-type)

I'm not sure yet if this should be an "ordinary" cast -- it seems potentially different, since the "opaque type" is treated in rather a particular way. But maybe it's ok to frame it as a cast.

OK, I'll stop there for this comment. I'm not yet sure the best way to implement so I'll probably leave a few comments exploring a few options.

[nikomatsakis]

Hmm before going any further I also want to raise up a few interesting questions. The possibility of ref patterns on the left-hand-side with impl types in the type is sort of "curious" -- I suspect we should just disallow it, so that we can treat the initializer as a "value" (and not a place).

Right now, the initializers can kind of act in two ways. e.g., here, y winds up acting as a place expresion and not a value expression:

let y = 22;
let ref x: u32 = y; // equivalent to `let x = &y;`

In principle you could have:

let ref x: impl Debug = y;

and wind up with x having the type &<opaque Debug>. This may or may not add some complexity — I guess we'll see. If so, we could always disallow ref bindings in the patterns when an impl Trait type is used as a type ascription.

[nikomatsakis]

So what happens when you have a &dyn Trait binding today? The answer is that the you wind up in the "coerce types" code:

rust/src/librustc_typeck/check/coercion.rs

Line 166 in 365b900

fn coerce(&self, a: Ty<'tcx>, b: Ty<'tcx>) -> CoerceResult<'tcx> {

Here a (the "source") would be the type of the initializer and b is some type that includes a dyn Trait. This will wind up considering an "unsize coercion":

rust/src/librustc_typeck/check/coercion.rs

Line 200 in 365b900

let unsize = self.commit_if_ok(|_| self.coerce_unsized(a, b));

This routine is pretty complex and I'll skip over the details (reminds me though that #50753 would be really nice to fix). It will however ultimately produce an Adjustment, which will include an AdjustKind::Unsize step.

This adjustment is stored alongside the source expression (here, y) and is taken into account when producing MIR:

rust/src/librustc_mir/hair/cx/expr.rs

Lines 81 to 85 in 365b900

	fn apply_adjustment<'a, 'gcx, 'tcx>(cx: &mut Cx<'a, 'gcx, 'tcx>,
	hir_expr: &'tcx hir::Expr,
	mut expr: Expr<'tcx>,
	adjustment: &Adjustment<'tcx>)
	-> Expr<'tcx> {

[jyn514]

Triage: the first example still gives an ICE.

error: internal compiler error: broken MIR in DefId(0:6 ~ playground[abce]::main) (_1 = Option::<u32>::Some(const 44_u32)): bad assignment (std::option::Option<impl std::fmt::Debug> = std::option::Option<u32>): NoSolution
 --> src/main.rs:7:33
  |
7 |     let x: Option<impl Debug> = Some(44_u32);
  |                                 ^^^^^^^^^^^^
  |
  = note: delayed at compiler/rustc_mir/src/borrow_check/type_check/mod.rs:253:27

[lqd]

For people following at home, impl-trait-in-bindings is on its way to being removed and reimplemented: #86729