Hi! middle-end or back-end created aggregates don't have CLASS_TYPE_P bit set, so build_zero_init_1 ICEs e.g. on va_list if it is array of structures.
Fixed by treating all RECORD_TYPE or UNION_TYPE types that way, rather than just CLASS_TYPE_P ones. Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk? 2013-02-24 Jakub Jelinek <ja...@redhat.com> PR c++/56403 * init.c (build_zero_init_1): Use RECORD_OR_UNION_CODE_P instead of CLASS_TYPE_P. * g++.dg/torture/pr56403.C: New test. --- gcc/cp/init.c.jj 2013-02-07 22:27:09.000000000 +0100 +++ gcc/cp/init.c 2013-02-24 20:07:12.246690837 +0100 @@ -179,7 +179,7 @@ build_zero_init_1 (tree type, tree nelts init = convert (type, nullptr_node); else if (SCALAR_TYPE_P (type)) init = convert (type, integer_zero_node); - else if (CLASS_TYPE_P (type)) + else if (RECORD_OR_UNION_CODE_P (TREE_CODE (type))) { tree field; vec<constructor_elt, va_gc> *v = NULL; --- gcc/testsuite/g++.dg/torture/pr56403.C.jj 2013-02-24 20:05:41.961233765 +0100 +++ gcc/testsuite/g++.dg/torture/pr56403.C 2013-02-24 20:05:14.000000000 +0100 @@ -0,0 +1,12 @@ +// PR c++/56403 +// { dg-do compile } + +#include <stdarg.h> + +struct S { va_list err_args; }; + +void * +foo () +{ + return new S (); +} Jakub