NLL: spurious error emitted on body of (erroneous) const fn body

[pnkfelix]

Spawned off of #54528 (specificially this card in the project).

Consider the following code:

const fn no_dyn_trait_ret() -> &'static dyn std::fmt::Debug { &() }

fn main() { }

With AST-borrowck, you get (play):

error: trait bounds other than `Sized` on const fn parameters are unstable
 --> src/main.rs:1:32
  |
1 | const fn no_dyn_trait_ret() -> &'static dyn std::fmt::Debug { &() }
  |                                ^^^^^^^^^^^^^^^^^^^^^^^^^^^^

error: aborting due to previous error

With NLL (here using migration a la 2018 edition), you get (play):

error: trait bounds other than `Sized` on const fn parameters are unstable
 --> src/main.rs:1:32
  |
1 | const fn no_dyn_trait_ret() -> &'static dyn std::fmt::Debug { &() }
  |                                ^^^^^^^^^^^^^^^^^^^^^^^^^^^^

warning[E0515]: cannot return reference to temporary value
 --> src/main.rs:1:63
  |
1 | const fn no_dyn_trait_ret() -> &'static dyn std::fmt::Debug { &() }
  |                                                               ^--
  |                                                               ||
  |                                                               |temporary value created here
  |                                                               returns a reference to data owned by the current function
  |
  = warning: This error has been downgraded to a warning for backwards compatibility with previous releases.
          It represents potential unsoundness in your code.
          This warning will become a hard error in the future.

error: aborting due to previous error

I have not managed to replicate the above borrow-check error on code that does not exhibit some other const fn error.

And the problem goes away if you opt into #![feature(const_fn)] (play).

So, this is really a very minor diagnostic issue.

[RalfJung]

Cc @oli-obk might be something in (min) const qualification?

[oli-obk]

TLDR: The issue with this is that we have a bunch of separate checks relating to const things, and most of these are spaghetti entangled into one. @eddyb and I are working towards fixing that, but it's nontrivial to say the least.

Details:

We have the following checks that we need to run on MIR (which relate to const):

• min_const_fn compliance
• const_fn compliance
• promotion of runtime values to an anonymous static/constant
• const compliance (valid array length, const item initializer or enum discriminant)
• static compliance (valid static item initializer)
• no interior mutability checks
• no drop in const checks

The current implementation has an early abort for const fns where if the const_fn feature gate is not active, we first run min_const_fn, and if that fails we don't run the entangled other 4 passes. If we ran the other passes in case of min_const_fn failure, you'd get a lot of error duplication.

I don't think we have a tracking issue for the refactoring separating all these passes into distinct dataflow passes, but there's ongoing work.

[Aaron1011]

This now produces the same output both with and without #![feature(nll)].