add initial reference invalidation

+ Makes vec example actually memory safe, which is cool

+ Specify owner > sub relationships with ">" in closure parameter
  lists, uses the same group idea as closure calls

+ Relies on users implementing functions in a consistent manner,
  since you can kind of do whatever with pointers. Presumably
  there would be a stdlib of vec/map/set etc. which applications
  could then use and by proxy be memory safe. Although some more
  checks wouldn't hurt, I suppose?

+ Not sure I like having reference invalidation be 'just a move',
  seems to work alright but the semantics of it are a bit muddy.

1 files changed, 13 insertions, 1 deletions

diff --git a/src/move.c b/src/move.c
index 5137719..a76ffd5 100644
--- a/src/move.c
+++ b/src/move.c
@@ -372,6 +372,19 @@ static int mvcheck_id(struct state *state, struct ast *node)
 	if (def->k != AST_VAR_DEF)
 		return 0;
 
+	struct ast_pair *prev = find_move(state, def);
+
+	/* a reference invalidation is represented as a 'moved'
+	 * reference, which is not expressible within the
+	 * language but is constructed as part of
+	 * opt_group_left/opt_group_right forcing a move to
+	 * happen. Hack? */
+	if (def->t->k == TYPE_REF && prev) {
+		/** @todo a more fitting error message? */
+		move_error(node, prev->use);
+		return -1;
+	}
+
 	if (is_trivially_copyable(def->t))
 		return 0;
 
@@ -381,7 +394,6 @@ static int mvcheck_id(struct state *state, struct ast *node)
 		return -1;
 	}
 
-	struct ast_pair *prev = find_move(state, def);
 	if (prev) {
 		/* error messages for opt groups could be improved */
 		move_error(node, prev->use);