http://gcc.gnu.org/bugzilla/show_bug.cgi?id=40436
--- Comment #41 from Jan Hubicka <hubicka at gcc dot gnu.org> 2010-11-14 09:06:41 UTC --- Created attachment 22389 --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=22389 preprocessed ext/super.c Hi, this testcase shows that we are no longer able to optimize away ext3_sops in sb->s_magic = (__builtin_constant_p((__u16)((es->s_magic))) ? ({ __u16 __x = (((es->s_magic))); ((__u16)( (((__u16)(__x) & (__u16)0x00ffU) << 8) | (((__u16)(__x) & (__u16)0xff00U) >> 8) )); }) : __fswab16(((es->s_magic)))); if (sb->s_magic != 0xEF53) { if (!silent) printk("<3>" "VFS: Can't find ext3 filesystem on dev %s.\n", bdevname(dev)); goto failed_mount; } The problem is that varpool code now expect that variable accesses are optimized out at the end of gimple queue instead of using TREE_SYMBOL_REFERENCED bit. In this case sb->s_magic shomehow manages to get undefined and in GIMPLE land we keep the conditional around, while in RTL land init-regs initialize it to 0 that consequently makes mount to always fail. I guess it is not real code quality bug since it happens only on undefined behaviour code, but we might consider doing initialization by zero sometime in gimple queue, too. Also I am not quite sure if we are not misoptimizing the code, it is convoluted.
