rustc_typeck: correctly compute Substs for Res::SelfCtor.

[eddyb]

Fixes #61882.

r? @petrochenkov cc @varkor

[Centril]

Would be good to add the const B: Self = Self(0); case also for some redundancy?

I don't know what an interesting test for A<&'static T> would be... I have:

impl A<&'static u8> {
    fn f() -> Self {
        let x = 0;
        Self(&x)
    }
}

it fails with "cannot return value referencing local variable x".

[eddyb]

You can't return it. Try Self(&x); - that should still error, even if you're not returning it.

[Centril]

Ah good call; it doesn't error:

struct A<T>(T);

impl A<&'static u8> {
    fn f() {
        let x = 0;
        Self(&x);
    }
}

so you can add that to the test file.

[Centril]

Decided to check your branch out to help speed things up... but Self(&x); for impl A<&'static u8> { still does not error.

[eddyb]

It does error, but it needs to be in a separate test because it doesn't borrow-check A::f if A::B doesn't type-check (we should probably remove this limitation and keep going? cc @estebank)

[petrochenkov]

r=me with tests requested by Centril

[eddyb]

@bors r=petrochenkov

[bors]

📌 Commit dedf2ed has been approved by petrochenkov

[pnkfelix]

discussed at T-compiler meeting. accepting for beta-backport.

[emilyalbini]

@eddyb there are conflicts applying this to the beta branch. Is this partial diff for src/librustc_typeck/check/mod.rs correct?

@@ -5255,12 +5209,8 @@ impl<'a, 'gcx, 'tcx> FnCtxt<'a, 'gcx, 'tcx> {
 
         let tcx = self.tcx;
 
-        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(_) | Res::Upvar(..) => Vec::new(),
+            Res::Local(_) | Res::Upvar(..) | Res::SelfCtor(_) => 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),

[eddyb]

@pietroalbini Yes, that seems correct. I could take a look at a backported PR, or open a backport PR myself if you prefer that.