http://gcc.gnu.org/bugzilla/show_bug.cgi?id=54038
--- Comment #4 from Jakub Jelinek <jakub at gcc dot gnu.org> 2012-07-20 11:08:49 UTC --- Created attachment 27844 --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=27844 gcc48-pr54038.patch Untested fix. On array23.C testcase, we have Foo as main variant of Foo2 and we have: t: const Foo2 [] c: const Foo [] m: Foo [] and in that case we want to change c's TYPE_MAIN_VARIANT to m. On this new testcase we have char as TYPE_MAIN_VARIANT of gchar, and have: t: const gchar *const [] c: const char *const [] m: const gchar * [] and trying to change TYPE_MAIN_VARIANT of c in that case (whose TYPE_MAIN_VARIANT is const char * [] at that point) is fatal. I believe the important difference between the two is that in array23.C TYPE_MAIN_VARIANT (TREE_TYPE (c)) is the same as TREE_TYPE (m) (aka TYPE_MAIN_VARIANT (elt_type)), but on this PR it is different and as the element type's of c vs. m + t aren't variants of the same type, neither should the array types be variants of one another.
