On Fri, Apr 04, 2014 at 04:48:48PM +0100, Jonathan Wakely wrote:
> On 04/04/14 09:17 -0600, Jeff Law wrote:
> >So what are those circumstances? ISTM this deserves some kind of
> >comment at the least.
>
> I found the ICE using ubsan on the libstdc++ testsuite. Including the
> bits/stdc++.h PCH'd header (i.e. the entire library) seemed to trigger
> it. So either related to PCH, or just huge translation units ... not
> sure of the exact cause.
So what happens here is that normally we initialize the builtins via
c_common_nodes_and_builtins, which has:
5711 if (!flag_preprocess_only)
5712 c_define_builtins (va_list_ref_type_node, va_list_arg_type_node);
c_define_builtins defines builtins in builtins.def, that includes even
sanitizer.def. I guess that flag_preprocess_only was in effect due to PCH
and we didn't define the builtins. I still haven't managed to create
some sweet & small testcase, but I've found another ICE with PCH:
$ touch y.h
$ cat y.c
#include "y.h"
int main () {}
$ gcc y.h
$ gcc y.c -fsanitize=undefined -S
y.c: In function ‘main’:
y.c:2:1: internal compiler error: Segmentation fault
int main () {}
^
0x9a20df crash_signal
/home/marek/src/gcc/gcc/toplev.c:337
0x52dfde bind
/home/marek/src/gcc/gcc/c/c-decl.c:646
0x53767f c_builtin_function(tree_node*)
/home/marek/src/gcc/gcc/c/c-decl.c:3664
0x875340 add_builtin_function_common
/home/marek/src/gcc/gcc/langhooks.c:573
0x875ff3 add_builtin_function(char const*, tree_node*, int, built_in_class,
char const*, tree_node*)
/home/marek/src/gcc/gcc/langhooks.c:589
0x9b46ba initialize_sanitizer_builtins()
/home/marek/src/gcc/gcc/sanitizer.def:30
0x9bfca5 ubsan_pass
/home/marek/src/gcc/gcc/ubsan.c:870
0x9bfca5 execute
/home/marek/src/gcc/gcc/ubsan.c:938
Please submit a full bug report,
with preprocessed source if appropriate.
Please include the complete backtrace with any bug report.
See <http://gcc.gnu.org/bugs.html> for instructions.
Marek
