`const_evaluatable_unchecked` doesn't correctly deal with non default substs

[lcnr]

While checking whether the fixed crate can be ported from typenum to nightly with generic_const_exprs, I hit this warning. I am not sure whether it's a false positive or whether the behavior I want is meant to be disallowed, but I think it should be allowed when enabling generic_const_exprs.

I've reduced the code though I left some stuff to show why it would be useful.

#![feature(generic_const_exprs)]

pub struct If<const CONDITION: bool>;
pub trait True {}
impl True for If<true> {}

pub struct FixedI8<const FRAC: u32> {
    pub bits: i8,
}

impl<const FRAC_LHS: u32, const FRAC_RHS: u32> PartialEq<FixedI8<FRAC_RHS>> for FixedI8<FRAC_LHS>
where
    If<{ FRAC_RHS <= 8 }>: True,
{
    fn eq(&self, _rhs: &FixedI8<FRAC_RHS>) -> bool {
        unimplemented!()
    }
}

impl<const FRAC: u32> PartialEq<i8> for FixedI8<FRAC> {
    fn eq(&self, rhs: &i8) -> bool {
        let rhs_as_fixed = FixedI8::<0> { bits: *rhs };
        PartialEq::eq(self, &rhs_as_fixed)
    }
}

Originally posted by @tspiteri in #76200 (comment)

[lcnr]

const_evaluatable_unchecked is emitted if we successfully evaluate an anonymous constant with generic const arguments.

Anonymous constants currently inherit all of their parents generics, so for If<{ FRAC_RHS <= 8 }>: True, the constant { FRAC_RHS <= 8 } has both FRAC_RHS and FRAC_LHS in scope, even though it only uses FRAC_RHS.

This would mean that pretty much all anonymous constants would cause that lint, so we add a flag to their MIR to detect whether they actually use any generic parameter: https://doc.rust-lang.org/nightly/nightly-rustc/rustc_middle/mir/struct.Body.html#structfield.is_polymorphic.

{ FRAC_RHS <= 8 } uses FRAC_RHS so that flag is true. The problem is now that FRAC_RHS is concrete, so evaluating the constant is successful, but the constant refers to FRAC_LHS which is still generic.

We should just disable the const_evaluatable_unchecked lint if feature(generic_const_exprs) is enabled 🤷

cc @BoxyUwU

[peter-lyons-kehl]

Another similar. Short example failing on Nightly (thanks to ilyvion#9946 for narrowing it down): https://play.rust-lang.org/?version=nightly&mode=debug&edition=2021&gist=c133c09b2fa155bb3f4b9ee45e6c011e:

#![feature(generic_const_exprs)]

trait Foo<const N: usize>
where
    [(); check_empty_array_size(N)]: Sized,
{
    type Narr: Foo<0>;
}

#[allow(unused_variables)]
pub const fn check_empty_array_size(array_size: usize) -> usize {
    0
}

My use case: https://github.com/ranging-rs/slicing-rs/blob/main/src/slices/mod.rs#L48=. I know it's a semantic/voluntary constraint only, and it's quite recursive, but I believe it deserves its place. (Alternatives I tried are at ranging-rs/slicing-rs#1.)

[lcnr]

this has been fixed, needs test to close