CTFE error shown thrice

[RalfJung]

This program

#![allow(const_err)]

union Foo {
    a: &'static u8,
    b: usize,
}

const BAD: usize = unsafe { Foo { a: &1 }.b * 2};
fn main() {
    let f = BAD;
}

leads to the following errors:

error[E0080]: referenced constant has errors
  --> src/main.rs:10:13
   |
8  | const BAD: usize = unsafe { Foo { a: &1 }.b * 2};
   |                             ------------------- "pointer arithmetic or comparison" needs an rfc before being allowed inside constants
9  | fn main() {
10 |     let f = BAD;
   |             ^^^

error[E0080]: could not evaluate constant
  --> src/main.rs:10:13
   |
10 |     let f = BAD;
   |             ^^^ referenced constant has errors

error[E0080]: constant evaluation error
 --> src/main.rs:8:1
  |
8 | const BAD: usize = unsafe { Foo { a: &1 }.b * 2};
  | ^^^^^^^^^^^^^^^^^^^^^^^^^^^^-------------------^^
  |                             |
  |                             "pointer arithmetic or comparison" needs an rfc before being allowed inside constants

error: aborting due to 3 previous errors

Seems like overkill to say three times that the constant has errors. ;)

Cc @oli-obk

[RalfJung]

It gets even better when having warn(const_err): Then we get if 4 times.

warning: this constant cannot be used
 --> src/main.rs:8:1
  |
8 | const BAD: usize = unsafe { Foo { a: &1 }.b * 2};
  | ^^^^^^^^^^^^^^^^^^^^^^^^^^^^-------------------^^
  |                             |
  |                             "pointer arithmetic or comparison" needs an rfc before being allowed inside constants
  |
note: lint level defined here
 --> src/main.rs:1:9
  |
1 | #![warn(const_err)]
  |         ^^^^^^^^^

error[E0080]: referenced constant has errors
  --> src/main.rs:10:13
   |
8  | const BAD: usize = unsafe { Foo { a: &1 }.b * 2};
   |                             ------------------- "pointer arithmetic or comparison" needs an rfc before being allowed inside constants
9  | fn main() {
10 |     let f = BAD;
   |             ^^^

error[E0080]: could not evaluate constant
  --> src/main.rs:10:13
   |
10 |     let f = BAD;
   |             ^^^ referenced constant has errors

error[E0080]: constant evaluation error
 --> src/main.rs:8:1
  |
8 | const BAD: usize = unsafe { Foo { a: &1 }.b * 2};
  | ^^^^^^^^^^^^^^^^^^^^^^^^^^^^-------------------^^
  |                             |
  |                             "pointer arithmetic or comparison" needs an rfc before being allowed inside constants

[RalfJung]

From conversation with @oli-obk:

The proper way to fix this is to move the errors/warnings into the query, instead of having to do this at all places where the query is made. This is both the right fix for #52849 and avoids duplicate errors.

So the const_eval query would be renamed to const_eval_raw and only used internally. There would be three new queries eval_const, eval_static, eval_promoted with different behavior wrt. showing an error or a warning, and whether to do validation of the final value.

The lint would just trigger all these queries. Thanks to caching, a warning triggered by a query is only ever shown once.

[RalfJung]

Thanks to @oli-obk's const-query-error PR, this is now fixed.

@oli-obk should this get a dedicated regression test? My impression is that we already have tons of tests like this.

[oli-obk]

Yea, there's enough things testing this. The only concern I see now is that if you allow(const_err) you have no idea anymore what's going on with the constant, rustc just tells you

error[E0080]: could not evaluate constant
  --> src/main.rs:10:13
   |
10 |     let f = BAD;
   |             ^^^ referenced constant has errors

without any further info.

[estebank]

Closing as per #53821:

error: any use of this value will cause an error
 --> src/main.rs:8:1
  |
8 | const BAD: usize = unsafe { Foo { a: &1 }.b * 2};
  | ^^^^^^^^^^^^^^^^^^^^^^^^^^^^-------------------^^
  |                             |
  |                             "pointer arithmetic or comparison" needs an rfc before being allowed inside constants
  |
  = note: #[deny(const_err)] on by default