Bounds seem to be ignored in negative impls

[japaric]

#![feature(optin_builtin_traits)]

use std::marker::MarkerTrait;

unsafe trait Pod: MarkerTrait {}
unsafe impl Pod for .. {}

trait Bound: MarkerTrait {}

// it seems that the bound is ignored here, i.e. this gets treated as `impl<T> !Pod for T {}`
impl<T: Bound> !Pod for T {}

fn is_pod<T: Pod>(_: T) {}

fn main() {
    is_pod(0i32);  // `i32` doesn't implement `Bound`, so it should be `Pod`
    is_pod(&0i32);  // likewise here
}

rustc 1.0.0-nightly (3b3bb0e68 2015-03-04) (built 2015-03-05)

cc @nikomatsakis @flaper87

[nikomatsakis]

Yeah, hmm, this is caused by the way that trait dispatch picks a concrete impl if its the only thing available without recursively evaluating the bounds. Hmm. It makes sense for positive impls but for negative impls doesn't seem quite right. 

Niko

-------- Original message --------
From: Jorge Aparicio notifications@github.com
Date:03/05/2015 09:57 (GMT-05:00)
To: rust-lang/rust rust@noreply.github.com
Cc: Niko Matsakis niko@alum.mit.edu
Subject: [rust] Bounds seem to be ignored in negative impls (#23072)

#![feature(optin_builtin_traits)]

use std::marker::MarkerTrait;

unsafe trait Pod: MarkerTrait {}
unsafe impl Pod for .. {}

trait Bound: MarkerTrait {}

// it seems that the bound is ignored here, i.e. this gets treated as impl<T> !Pod for T {}
impl<T: Bound> !Pod for T {}

fn is_pod<T: Pod>(_: T) {}

fn main() {
is_pod(0); // i32 doesn't implement Bound, so it should be Pod
is_pod(&0); // likewise here
}
rustc 1.0.0-nightly (3b3bb0e 2015-03-04) (built 2015-03-05)
cc @nikomatsakis @flaper87

—
Reply to this email directly or view it on GitHub.

[nikomatsakis]

This turns out to be a deeper issue. We should hold off on removing feature-gate for negative impls until it is resolved. The core issue that that negative impls with extra bounds become effectively negative bounds:

trait NegFoo : MarkerTrait { }
impl NegFoo for .. { }
impl<T:!Foo> !NegFoo for T { }

Now T:NegFoo is effectively T:!Foo from rust-lang/rfcs#586. I am very nervous about introducing such unbounded negativity into the trait system (see e.g. #23086). I would like to (for now) limit negative impls in OIBIT so that they can't carry where-clauses -- or at least not more where-clauses than are needed for their types to be well-formed.

[kdeeee]

error still persist in rustc 1.11.0 (9b21dcd6a 2016-08-15)
error message

default trait implementations are experimental and possibly buggy
negative trait bounds are not yet fully implemented; use marker types for now

[Manishearth]

triage: OIBITs cannot have bounds anymore. yay.