rematch_impl should be infallible

[lcnr]

rust/compiler/rustc_trait_selection/src/traits/select/mod.rs

Lines 2117 to 2144 in f90a4ff

	fn rematch_impl(
	&mut self,
	impl_def_id: DefId,
	obligation: &TraitObligation<'tcx>,
	) -> Normalized<'tcx, SubstsRef<'tcx>> {
	let impl_trait_ref = self.tcx().bound_impl_trait_ref(impl_def_id).unwrap();
	match self.match_impl(impl_def_id, impl_trait_ref, obligation) {
	Ok(substs) => substs,
	Err(()) => {
	self.infcx.tcx.sess.delay_span_bug(
	obligation.cause.span,
	&format!(
	"Impl {:?} was matchable against {:?} but now is not",
	impl_def_id, obligation
	),
	);
	let value = self.infcx.fresh_substs_for_item(obligation.cause.span, impl_def_id);
	let err = self.tcx().ty_error();
	let value = value.fold_with(&mut BottomUpFolder {
	tcx: self.tcx(),
	ty_op: |_| err,
	lt_op: |l| l,
	ct_op: |c| c,
	});
	Normalized { value, obligations: vec![] }
	}
	}
	}

We currently emit a delay_span_bug if match_impl in rematch_impl fails. Considering that we should only call rematch_impl for impls for which we've already called match_impl before, rematch_impl should always succeed.

There are currently 2 tests which hit that error case:

src/test/ui/impl-trait/issues/issue-62742.rs
src/test/ui/impl-trait/issues/issue-84073.rs

We should figure out why they fail and either fix that, or add a comment to rematch_impl explaining what's going on.