https://gcc.gnu.org/bugzilla/show_bug.cgi?id=70035
--- Comment #5 from Jakub Jelinek <jakub at gcc dot gnu.org> --- Created attachment 37867 --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=37867&action=edit gcc6-pr70035.patch As for the crash in libubsan, the problem is that -fsanitize=vptr instrumentation pretty much assumes that the vtable pointer in the object is either NULL, or valid table (perhaps of some other class). But in this case it is totally uninitialized, random garbage. The attached patch is an attempt to just clear those vtable pointers in objects for -fsanitize=vptr early in the ctor (right after the spot where -flifetime-dse=2 emits the clobber of the object), so that you get nice diagnostics instead of a crash. It would be desirable if the libubsan code assumed that the pointer in the object could be random garbage and deal with it more carefully (e.g. through trying a syscall and see if it ended up with EFAULT etc.). Though of course, both this patch, or being more careful in libubsan, will slow -fsanitize=vptr a little bit.
