E0493 emitted without drop when destructuring

[CAD97]

Given the following code:

pub struct Wrap(String);

impl Wrap {
    pub const fn into_inner(self) -> String {
        self.0
    }
}

The current output is:

error[E0493]: destructors cannot be evaluated at compile-time
 --> src/lib.rs:4:29
  |
4 |     pub const fn into_inner(self) -> String {
  |                             ^^^^ constant functions cannot evaluate destructors
5 |         self.0
6 |     }
  |     - value is dropped here

This error is clearly wrong: self is not dropped, as it's been moved out of. In fact, no destructors are run, so this should be able to compile. If it shouldn't compile, the error should be changed to not be incorrect.

This seems to be an issue in the drop suppression when moving out field-wise; using e.g. mem::forget works to suppress the destructor and allow the function to compile. I was unable to trick the compiler to accept the function with ManuallyDrop, as extracting the wrapped value requires manifesting a wrapper again, which triggers E0493 "destructors cannot be evaluated at compile-time".

@rustbot label: +A-const-fn +A-destructors

[ghost]

I believe this is fixed by #![feature(const_precise_live_drops)].

[trentj]

Does not seem to be fixed in all cases. While investigating a question raised on Discord:

const fn concat<T>(this: [T; 4], that: [T; 4]) -> [T; 8] {
    let [a, b, c, d] = this;
    let [e, f, g, h] = that;
    [a, b, c, d, e, f, g, h]
}

gives the bogus error for this and that despite const_precise_live_drops (playground).

Curiously, adding std::mem::forget(this); triggers use of partially moved value: `this` which suggests that this is considered only partially moved from for some reason. Perhaps this is a different problem; this is the nearest issue I could find.

[CAD97]

The OP example with just one field can now be straightforwardly worked around with transmute. The array case can currently be worked around with just #![feature(const_refs_to_cell)] and an ugly pile of repetitious unsafe: [playground]

pub const fn concat<T>(this: [T; 4], that: [T; 4]) -> [T; 8] {
    let this = ManuallyDrop::new(this);
    let that = ManuallyDrop::new(that);
    unsafe {
        let a = addr_of!(this).cast::<T>().add(0).read();
        let b = addr_of!(this).cast::<T>().add(1).read();
        let c = addr_of!(this).cast::<T>().add(2).read();
        let d = addr_of!(this).cast::<T>().add(3).read();
        let e = addr_of!(that).cast::<T>().add(0).read();
        let f = addr_of!(that).cast::<T>().add(1).read();
        let g = addr_of!(that).cast::<T>().add(2).read();
        let h = addr_of!(that).cast::<T>().add(3).read();
        [a, b, c, d, e, f, g, h]
    }
}

I believe that the proper fix should be along the lines of that a binding pattern like [a, b, c, d] gets identified as being complete — that it creates a binding covering every place in the scrutinee — and using the fact that the pattern is complete (and in the by-value mode) to mark the scrutinee place as fully moved from instead of only partially moved from.

Unfortunately, it gets more interesting when Copy typed subplaces are involved, since then the scrutinee is not fully moved from (Copy places are not invalidated when moved from). That's where const_precise_live_drops becomes necessary.