Hi! We ICE on this testcase since forever, because the computed goto's operand needs cleanups and doesn't have corresponding CLEANUP_POINT_EXPR emitted anywhere. Given that the operand is void * and only valid values for it are && LABEL_DECLs, I think it is ok to perform the cleanups before the actual goto.
Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk? 2012-07-03 Jakub Jelinek <ja...@redhat.com> PR c++/53812 * semantics.c (finish_goto_stmt): Surround computed goto argument with CLEANUP_POINT_EXPR if needed. * g++.dg/ext/label14.C: New test. --- gcc/cp/semantics.c.jj 2012-06-27 16:36:03.000000000 +0200 +++ gcc/cp/semantics.c 2012-07-02 10:22:26.377925168 +0200 @@ -571,6 +571,9 @@ finish_goto_stmt (tree destination) tf_warning_or_error); if (error_operand_p (destination)) return NULL_TREE; + destination + = fold_build_cleanup_point_expr (TREE_TYPE (destination), + destination); } } --- gcc/testsuite/g++.dg/ext/label14.C.jj 2012-07-02 10:28:14.305949165 +0200 +++ gcc/testsuite/g++.dg/ext/label14.C 2012-07-02 10:27:11.000000000 +0200 @@ -0,0 +1,17 @@ +// PR c++/53812 +// { dg-do compile } +// { dg-options "" } + +struct T { T () : t(0) {}; int t; ~T (); }; +struct S { void *operator [] (T); }; +void bar (S &, void *, void *); + +void +foo (S &x, T &y) +{ + bar (x, &&l1, &&l2); +l1: + goto *x[y]; +l2: + bar (x, &&l1, &&l2); +} Jakub