unsized_fn_params should not accept types that don't have a dynamically fixed size (such as `extern` types)

[RalfJung]

The following code should not be accepted:

#![feature(extern_types)]
#![feature(unsized_fn_params)]

extern {
    type E;
}

fn test(e: E) {}

fn calltest(e: Box<E>) {
    test(*e)
}

In terms of MIR semantics I think we want to view a function call as making copies of its arguments. This requires at least dynamically determining the size of the argument, and therefore we shouldn't accept extern type arguments.

I was extremely surprised earlier today when I learned that this code is getting accepted.

There is no generic bound to distinguish extern types from other unsized types, so the options we have are

• require the unsized type to be concrete enough that we can determine its tail, and ensure it is not extern
• raise the error during monomorphization / codegen

Trying to actually build any such code will lead to ICEs such as

error: internal compiler error: compiler/rustc_codegen_llvm/src/type_of.rs:306:13: TyAndLayout::scalar_pair_element_llty(TyAndLayout { ty: *mut E, layout: Layout { size: Size(8 bytes), align: AbiAndPrefAlign { abi: Align(8 bytes), pref: Align(8 bytes) }, abi: Scalar(Initialized { value: Pointer(AddressSpace(0)), valid_range: 0..=18446744073709551615 }), fields: Primitive, largest_niche: None, variants: Single { index: 0 }, max_repr_align: None, unadjusted_abi_align: Align(8 bytes) } }): not applicable

or

thread 'rustc' panicked at compiler/rustc_codegen_llvm/src/builder.rs:491:9:
assertion `left == right` failed
  left: false
 right: true

[SNCPlay42]

See also #79409.

There is no generic bound to distinguish extern types from other unsized types

rust-lang/rfcs#2984 and rust-lang/rfcs#3396 are open RFCs that attempt to address this.

[RalfJung]

Unfortunately just rejecting locals/arguments that are not known to be dyn-sized as a MIR pass does not work: the forget_unsized function in the standard library would stop compiling.

So currently something post-monomorphization is the only option. Is there even some place that check could be put without being duplicated across all codegen backends (and Miri)?