ICE on stable in selection #89864

playground link

If I had to guess, this is because in the non-codegen phase, we can't prove <<Dynamic<T> as WrapperTrait>::Assoc as Trait>::Assoc: fmt::Debug, because we're generic over T, but later on during codegen phase, we try to normalize that, see that there's only one applicable impl for WrapperTrait for Dynamic<T>, pick that, then find that ConcreteAssoc: Debug doesn't hold. But by this point we're "too late" and already doing codegen.

Seems like a very lazy-norm thing and this is all speculation.

#90891 changed the format_args expansion slightly so the example no longer ICEs, but it can still be reproduced by calling ArgumentV1::new manually

#![feature(fmt_internals)]

use core::{fmt, marker::PhantomData};

pub trait WrapperTrait {
    type Assoc;
}

pub trait Trait {
    type Assoc;
}

pub struct Ice<T>(<<Dynamic<T> as WrapperTrait>::Assoc as Trait>::Assoc);

fn ice<T>(i: Ice<T>)
where
    <<Dynamic<T> as WrapperTrait>::Assoc as Trait>::Assoc: fmt::Debug,
{
    fmt::ArgumentV1::new(&i.0, fmt::Debug::fmt);
}

pub struct TraitImplementor;

impl Trait for TraitImplementor {
    type Assoc = ConcreteAssoc;
}

pub struct ConcreteAssoc;

impl<T> WrapperTrait for Dynamic<T> {
    type Assoc = TraitImplementor;
}

pub struct Dynamic<T>(PhantomData<T>);

Hmm... this is what I was worried about in the PR. This compiles now, but the trait bound ConcreteAssoc::Debug doesn't hold.

Is it an issue? The function ice is never called, so no code ever has to prove this bound.

oh 🤦 yea, I can't find a way to call it either. Thanks!

This is fixes and we have appropriate tests imo