------- Comment #29 from ian at airs dot com 2007-05-02 16:57 ------- Created an attachment (id=13497) --> (http://gcc.gnu.org/bugzilla/attachment.cgi?id=13497&action=view) Patch
Here is one approach which fixes the test case. This introduces a new tree code, ALIASING_CONVERT_EXPR. It is conveyed into RTL via a flag on REGS: REG_ALIAS_ALL. I didn't try to really union the alias sets, I just said that the result of placement new can alias anything. This patch is essentially untested. I'm not very happy with this approach because it doesn't fail safe: it's too easy to lose the special aliasing, and then the problem appears again, but only with a more complicated test case. A safer approach might be to change the type returned by placement new and mark it as TYPE_REF_CAN_ALIAS_ALL, but then I'm worried about type conversion and type comparison problems, since it isn't actually a different type. -- http://gcc.gnu.org/bugzilla/show_bug.cgi?id=29286