Rollup merge of #121047 - compiler-errors:default-impls, r=lcnr

Do not assemble candidates for default impls

There is no reason (as far as I can tell?) that we should assemble an impl candidate for a default impl. This candidate itself does not prove that the impl holds, and any time that it *does* hold, there will be a more specializing non-default impl that also is assembled.

This is because `default impl<T> Foo for T {}` actually expands to `impl<T> Foo for T where T: Foo {}`. The only way to satisfy that where clause (without coinduction) is via *another* implementation that does hold -- precisely an impl that specializes it.

This should fix the specialization related regressions for #116494. That should lead to one root crate regression that doesn't have to do with specialization, which I think we can regress.

r? lcnr cc ``@rust-lang/types``

cc #31844

compiler/rustc_trait_selection/src/traits/select/candidate_assembly.rs

@@ -566,6 +566,14 @@ impl<'cx, 'tcx> SelectionContext<'cx, 'tcx> {
                 {
                     return;
                 }
+
+                // For every `default impl`, there's always a non-default `impl`
+                // that will *also* apply. There's no reason to register a candidate
+                // for this impl, since it is *not* proof that the trait goal holds.
+                if self.tcx().defaultness(impl_def_id).is_default() {
+                    return;
+                }
+
                 if self.reject_fn_ptr_impls(
                     impl_def_id,
                     obligation,