variant_count intrinsic is evaluated incorrectly in generic code

[tmiasko]

For example:

#![feature(variant_count)]

fn a<T>() -> usize {
    let v = [0u8; std::mem::variant_count::<T>()];
    std::mem::size_of_val(&v)
}

fn b() -> usize {
    let v = [0u8; std::mem::variant_count::<Result<(), ()>>()];
    std::mem::size_of_val(&v)
}

fn main() {
    println!("{}", a::<Result<(), ()>>());
    println!("{}", b());
}

$ rustc a.rs
warning: cannot use constants which depend on generic parameters in types
 --> a.rs:4:19
  |
4 |     let v = [0u8; std::mem::variant_count::<T>()];
  |                   ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  |
  = note: `#[warn(const_evaluatable_unchecked)]` on by default
  = warning: this was previously accepted by the compiler but is being phased out; it will become a hard error in a future release!
  = note: for more information, see issue #76200 <https://github.com/rust-lang/rust/issues/76200>

warning: 1 warning emitted

$ ./a
0
2

Similar to #73976, but with a different intrinsic.

[lcnr]

We can probably just put ensure_monomorphic call in front of that if statement.

rust/compiler/rustc_mir/src/interpret/intrinsics.rs

Line 79 in e0ef0fc

if let ty::Adt(ref adt, _) = tp_ty.kind() {

A more precise way to implement this would be to exhaustively match on tp_ty and forbid ty::Projection, ty::Opaque, ty::Param, ty::Bound, ty::Placeholder and ty::Infer here. This way we the intrinsic can still be evaluated for Option<T> which is probably desirable.

[wusyong]

@rustbot claim

[wusyong]

@lcnr Hello, I'm interested in taking the desirable approach. How is the arm for variants like ty::Projection should be handled? Which macro is more appropriate for this situation?

[lcnr]

I quickly grepped for throw.*TooGeneric and ended up with throw_inval!(TooGeneric) here.
That only works once you know what to search for though.