rustc panic with "assertion failed: !value.needs_subst()"

[tuxmark5]

The below example panics with: thread 'main' panicked at 'assertion failed: !value.needs_subst()', librustc/traits/query/normalize_erasing_regions.rs:69:9

I've tried minimizing the example as much as possible and that's the best I can do for now.

Version: rustc 1.27.0-nightly (428ea5f 2018-05-06)

use std::marker::PhantomData;

struct Meta<A> {
    value: i32,
    type_: PhantomData<A>
}

trait MetaTrait {
    fn get_value(&self) -> i32;
}

impl<A> MetaTrait for Meta<A> {
    fn get_value(&self) -> i32 { self.value }
}

trait Bar {
    fn get_const(&self) -> &dyn MetaTrait;
}

struct Foo<A> {
    value: A
}

impl<A: 'static> Foo<A> {
    const CONST: &'static dyn MetaTrait = &Meta::<Self> {
        value: 10,
        type_: PhantomData
    };
}

impl<A: 'static> Bar for Foo<A> {
    fn get_const(&self) -> &dyn MetaTrait { Self::CONST }
}

fn main() {
    let foo = Foo::<i32> { value: 10 };
    let bar: &dyn Bar = &foo;
    println!("const {}", bar.get_const().get_value());
}

[simon-auch]

I have the same error, interesstingly the error seems to have something to do with PhantomData:

use std::marker::PhantomData;

struct Foo<T>(PhantomData<T>);

impl<T> Foo<T>{
  const OP: fn(&T) = |_t| {};
  
  fn op(_t: &T){}
  
  fn foo(&mut self, t: T){
    //does not work
    Self::OP(&t);
    //does work
    Self::op(&t);
  }
}

fn main(){
  Foo(PhantomData).foo(1);
}

Edit 1 & 2: Smaller example

[simon-auch]

After some testing i found out it has nothing to do with the PhantomData, since the following code still gives the same error:

struct Foo<T>(T);

impl<T> Foo<T>{
  const OP: fn(T) = |_t| {};
  
  fn foo(self){
    //does not work
    Self::OP(self.0);
  }
}

fn main(){
  Foo(1).foo();
}

edit: further simplification

[tuxmark5]

I think it's related to promoting references to fat pointers in const contexts. In your example you promote a closure |_t| {} to dyn FnOnce(T) (?). In my example, I promote Meta to dyn MetaTrait.

[simon-auch]

It seems that the Problem only occurs, when using the associated constant from inside the implementation block:
http://play.rust-lang.org/?gist=e7f9f10cde5ca4b0c52588c26b6388cf&version=stable&mode=debug&edition=2015

[nikolads]

I found a different example which gives the same error:

struct MyBox<T>(*mut T);

impl<T> MyBox<T> {
    fn get(&self) -> Option<&T> {
        match self.0 {
            Self::NULL => None,
            ptr => unsafe { Some(&*ptr) },
        }
    }

    const NULL: *mut T = ::std::ptr::null_mut();
}

thread 'main' panicked at 'assertion failed: !value.needs_subst()', librustc/traits/query/normalize_erasing_regions.rs:69:9
error: internal compiler error: unexpected panic
note: rustc 1.27.0 (3eda71b00 2018-06-19) running on x86_64-unknown-linux-gnu

It only ICEs if I initialize NULL with a const fn. It compiles after replacing the line with

    const NULL: *mut T = 0 as *mut T;

Moving the function outside the impl block still ICEs, but with a different error:

struct MyBox<T>(*mut T);

impl<T> MyBox<T> {
    const NULL: *mut T = ::std::ptr::null_mut();
}

fn get<'a, T>(ptr: *mut T) -> Option<&'a T> {
    match ptr {
        MyBox::<T>::NULL => None,
        ptr => unsafe { Some(&*ptr) },
    }
}

error: internal compiler error: librustc/ty/subst.rs:493: Type parameter `T/#1` (T/1) out of range when substituting (root type=Some(T)) substs=[T]

thread 'main' panicked at 'Box<Any>', librustc_errors/lib.rs:499:9

[matklad]

Hit this in https://github.com/matklad/once_cell, in stable(1.28.0), beta(1.29.0-beta.1) and nightly ( 2018-08-03).

Here's the commit which exemplifies the issue: matklad/once_cell@ea1c14b (run cargo test --no-default-features to trigger). Basically, I cant newtype-wrap a struct with INIT: Self, which requires me to duplicate doc comments :(

[matklad]

Triage: both the original example and the once_cell bug no longer reproduce for me on 1.31.1 (2018 release) and 1.34.1 (stable at the time of writing).

The issue should probably be closed. Or do we want to add run-pass test just in case?

[estebank]

Lets add a regression test.