fix: Don't panic on some weird code

[ChayimFriedman2]

Fixes #19734.

[Veykril]

This seems a bit odd to me, handling it this way. I believe we should be bailing out earlier for a such a broken path already, that is a path solely consisting of a type anchor (<T as Trait>) which i believe is the issue here?.

Also I'd rather have this as a test in hir-ty itself imo (though I can understand if reproducing that without IDE tools might be tricky)

[ChayimFriedman2]

This seems a bit odd to me, handling it this way. I believe we should be bailing out earlier for a such a broken path already, that is a path solely consisting of a type anchor (<T as Trait>) which i believe is the issue here?.

Apparently we don't 🙂 We probably could, and I have no preference to handle it one way or another, but I don't know where we can handle this (parsing must produce some syntax tree anyway, and hir lowering doesn't know the path was malformed).

Also I'd rather have this as a test in hir-ty itself imo (though I can understand if reproducing that without IDE tools might be tricky)

Done.

[Veykril]

Apparently we don't 🙂

Oh no I am aware, I meant that more in a manner of, we should be doing that instead (bad phrasing on my part).

[ChayimFriedman2]

Do you have an idea how?

[Veykril]

No, I am looking a bit at it right now but I think I am wrong either way. We do have to handle this here I think

[Veykril]

Hmm, actually, we can just fail path lowering if we end up with a path whose last segment has has_self_type set I think, as that is always incorrect?

[Veykril]

That is could you check (something along the lines of)

index 629d1f2ada..e664f83f6e 100644
--- a/crates/hir-def/src/expr_store/lower/path.rs
+++ b/crates/hir-def/src/expr_store/lower/path.rs
@@ -110,7 +110,7 @@
                 push_segment(&segment, &mut segments, Name::new_symbol_root(sym::Self_));
             }
             ast::PathSegmentKind::Type { type_ref, trait_ref } => {
-                debug_assert!(path.qualifier().is_none()); // this can only occur at the first segment
+                debug_assert!(path.qualifier().is_none() && type_anchor.is_none()); // this can only occur at the first segment

                 let self_type = collector.lower_type_ref(type_ref?, impl_trait_lower_fn);

@@ -235,6 +235,10 @@
     let mod_path = Interned::new(ModPath::from_segments(kind, segments));
     if type_anchor.is_none() && generic_args.is_empty() {
         return Some(Path::BarePath(mod_path));
+    } else if generic_args.last().and_then(Option::as_ref).is_some_and(|it| it.has_self_type) {
+        // `<T as Trait>` without following segment
+        // FIXME: Report error
+        return None;
     } else {
         return Some(Path::Normal(Box::new(NormalPath {
             type_anchor,

I think that would be a better solution here

[ChayimFriedman2]

I hesitate a bit because maybe there is a valid way for this to happen, but I changed to that, except that I don't return None but only dismiss the generic args, which will cause us to infer them - a better user experience.