http://gcc.gnu.org/bugzilla/show_bug.cgi?id=55679
--- Comment #16 from Dominique d'Humieres <dominiq at lps dot ens.fr>
2012-12-16 18:23:24 UTC ---
(In reply to comments #12 and #13)
With -O1, the invalid tests c-c++-common/asan/global-overflow-1.c and
c-c++-common/asan/stack-overflow-1.c are optimized to a valid "return 0" as
shown by the following differences between the outputs of -dump-tree-optimized
with and without -fno-tree-fre:
--- stack-overflow-1.c.164t.optimized 2012-12-16 17:26:13.000000000 +0100
+++ stack-overflow-1.c.164t_f.optimized 2012-12-16 17:25:13.000000000 +0100
@@ -3,18 +3,14 @@
main ()
{
- int res;
char x[10];
int ten.1;
- char _4;
<bb 2>:
__builtin_memset (&x, 0, 10);
ten.1_3 ={v} ten;
- _4 = x[ten.1_3];
- res_5 = (int) _4;
x ={v} {CLOBBER};
- return res_5;
+ return 0;
}
So the problem for these two tests has nothing to do with asan. Note that they
regress cleanly if I add '-fno-tree-fre -fno-tree-pre' to '-fno-builtin-memset'
in the dg-options. The optimization is also defeated If I add a line such as
x[1] = 1;
Although I doubt this behavior is a bug, I'll open a new PR for it tomorrow
unless there is some objection.
