Rollup merge of #121379 - nnethercote:rm-unchecked_error_guaranteed, …

…r=oli-obk

Remove an `unchecked_error_guaranteed` call.

If we abort immediately after complaining about the obsolete `impl Trait for ..` syntax, then we avoid reaching HIR lowering. This means we can use `TyKind::Dummy` instead of `TyKind::Err`.

r? `@oli-obk`

compiler/rustc_ast_passes/src/ast_validation.rs

@@ -881,9 +881,10 @@ impl<'a> Visitor<'a> for AstValidator<'a> {
                         &item.vis,
                         errors::VisibilityNotPermittedNote::TraitImpl,
                     );
-                    // njn: use Dummy here
-                    if let TyKind::Err(_) = self_ty.kind {
-                        this.dcx().emit_err(errors::ObsoleteAuto { span: item.span });
+                    if let TyKind::Dummy = self_ty.kind {
+                        // Abort immediately otherwise the `TyKind::Dummy` will reach HIR lowering,
+                        // which isn't allowed. Not a problem for this obscure, obsolete syntax.
+                        this.dcx().emit_fatal(errors::ObsoleteAuto { span: item.span });
                     }
                     if let (&Unsafe::Yes(span), &ImplPolarity::Negative(sp)) = (unsafety, polarity)
                     {

compiler/rustc_parse/src/parser/item.rs

@@ -601,22 +601,10 @@ impl<'a> Parser<'a> {
             // We need to report this error after `cfg` expansion for compatibility reasons
             self.bump(); // `..`, do not add it to expected tokens
 
-            // FIXME(nnethercote): AST validation later detects this
-            // `TyKind::Err` and emits an errors. So why the unchecked
-            // ErrorGuaranteed?
-            // - A `span_delayed_bug` doesn't work here, because rustfmt can
-            //   hit this path but then not hit the follow-up path in the AST
-            //   validator that issues the error, which results in ICEs.
-            // - `TyKind::Dummy` doesn't work, because it ends up reaching HIR
-            //   lowering, which results in ICEs. Changing `TyKind::Dummy` to
-            //   `TyKind::Err` during AST validation might fix that, but that's
-            //   not possible because AST validation doesn't allow mutability.
-            //
-            // #121072 will hopefully remove all this special handling of the
-            // obsolete `impl Trait for ..` and then this can go away.
-            #[allow(deprecated)]
-            let guar = rustc_errors::ErrorGuaranteed::unchecked_error_guaranteed();
-            Some(self.mk_ty(self.prev_token.span, TyKind::Err(guar)))
+            // AST validation later detects this `TyKind::Dummy` and emits an
+            // error. (#121072 will hopefully remove all this special handling
+            // of the obsolete `impl Trait for ..` and then this can go away.)
+            Some(self.mk_ty(self.prev_token.span, TyKind::Dummy))
         } else if has_for || self.token.can_begin_type() {
             Some(self.parse_ty()?)
         } else {