Regression: implementing `Clone` for trait objects

[ryzhyk]

The following code used to work with rustc 1.32 and earlier, but it crashes with stack overflow due to infinite recursion in 1.35

pub trait CBFn: Fn() {
   fn clone_boxed(&self) -> Box<dyn CBFn>;
}

impl<T> CBFn for T
where
   T: 'static + Clone + Fn()
{
   fn clone_boxed(&self) -> Box<dyn CBFn> {
       Box::new(T::clone(self))
   }
}

impl Clone for Box<dyn CBFn> {
   fn clone(&self) -> Self {
       self.clone_boxed()
   }
}

fn main() { 
    let f = || {};
    let b: Box<dyn CBFn> = Box::new(f);
    let bb = b.clone();
}

(Playground)

Errors:

   Compiling playground v0.0.1 (/playground)
warning: unused variable: `bb`
  --> src/main.rs:23:9
   |
23 |     let bb = b.clone();
   |         ^^ help: consider prefixing with an underscore: `_bb`
   |
   = note: #[warn(unused_variables)] on by default

    Finished dev [unoptimized + debuginfo] target(s) in 0.81s
     Running `target/debug/playground`

thread 'main' has overflowed its stack
fatal runtime error: stack overflow
timeout: the monitored command dumped core
/root/entrypoint.sh: line 8:     8 Aborted                 timeout --signal=KILL ${timeout} "$@"

Update

It does work correctly with a slightly modified main using as_ref() to explicitly extract reference out of the box.

fn main() { 
    let f = || {};
    let b: Box<dyn CBFn> = Box::new(f);
    let bb = b.as_ref().clone();
}

[RustyYato]

This also works if you extract the reference out of the box inside the Clone impl

impl Clone for Box<dyn CBFn> {
   fn clone(&self) -> Self {
       self.as_ref().clone_boxed()
   }
}

https://play.rust-lang.org/?version=stable&mode=debug&edition=2018&gist=a2d576b485eee34dc1cecdb06e78a0f6

It looks like Rust sees that Box<dyn CBFn>: 'static + Clone + Fn() then applies the blanket impl for CBFn to Box<dyn CBFn>. This in turn leads self.clone_box() using <Box<dyn CBFn> as CBFn>::clone_boxed(&self), which leads to an infinite recursion problem.
This can be fixed by explicitly getting the reference to the trait object using as_ref.

[ryzhyk]

Thanks, @KrishnaSannasi. This makes sense, but I find it disturbing that the behavior changed in one of the recent compiler releases, potentially breaking software that relied on the old behavior. I wonder if this is deliberate or a side effect of some other change.

[RustyYato]

This is likely a side effect of changes/optimizations to how blanket trait impls interact with each other.

[ollie27]

This looks like fallout from #59500.

[pnkfelix]

pre-triage: P-high (for bisection to verify hypothesis that this was injected by #59500), unnominating, assigning to self.