With a generic struct, `Self(x)` works where `x` is of the wrong type argument

[Centril]

In particular, we can write:

struct A<T>(T);

impl A<bool> {
    const B: A<u8> = Self(0); // OK but shouldn't be?

    const C: A<u8> = Self { 0: 0 }; // Correctly rejected.
}

This can lead to the following interesting situation:

struct A<T>(T);

impl A<bool> {
    const B: Self = Self(0); // Rejected, so `Self(..)` is not of type `Self`?
}

cc @petrochenkov, @varkor, @eddyb, and @alexreg

This does seem like a bug to me... In particular, type_alias_enum_variants does not allow Self::Some(transform(x)) where Self == Option<T> but where Option<U> is expected.

cc @rust-lang/lang

[petrochenkov]

Yes, that's a bug, Self in the example should be A<bool> and A::<bool>(0) should report an error.

We missed a test for this in #53751 somehow (for Self types this property is tested in struct-path-self-type-mismatch.rs).

[petrochenkov]

Self constructors were stabilized not too long ago, so this should preferably be fixed soon and backported to beta.

[eddyb]

Ouch, this explains why we were able to remove the Ty result from rewrite_self_ctor as redundant in #61085 / #61189 - there was a subst call missing on that type.
But that's not the end of it, this is incorrectly allowed right now, as well:

struct A<T>(T);

impl A<bool> {
    const B: A<u8> = Self::<u8>(0);
}

So the order here is wrong:

rust/src/librustc_typeck/check/mod.rs

Lines 5368 to 5377 in 374c63e

	let res = match self.rewrite_self_ctor(res, span) {
	Ok(res) => res,
	Err(ErrorReported) => return (tcx.types.err, res),
	};
	let path_segs = match res {
	Res::Local(_) => vec![],
	Res::Def(kind, def_id) =>
	AstConv::def_ids_for_value_path_segments(self, segments, self_ty, kind, def_id),
	_ => bug!("instantiate_value_path on {:?}", res),
	};

In that Res::SelfCtor really needs to be treated like Res::Local.

In general, I think instead of rewrite_self_ctor we need a method that only returns the right information (such as the modified Res and substituted Ty) given the impl DefId inside Res::SelfCtor.
And every call should be audited. EDIT: there's only one call now, huh, so it can be inlined.
I guess I could prepare a master and beta PR?

[eddyb]

We shouldn't forget tests like A<&'static T> being created with Self(non_static_ref).

[eddyb]

Looks like the test added in #59025 for #57924 has an explicit generic arg: Self::<E>(e), but somehow @varkor and I didn't spot that during the review, whoops.