On 11/01/2022 19:04, Jakub Jelinek wrote:
On Tue, Jan 11, 2022 at 06:40:44PM +0530, Siddhesh Poyarekar wrote:
Never try to compute size for offset when the object size is -1, which
is either unknown maximum or uninitialized minimum irrespective of the
osi->pass number.
gcc/ChangeLog:
PR tree-optimization/pr103961
* tree-object-size.c (plus_stmt_object_size): Always avoid
computing offset for -1 size.
gcc/testsuite/ChangeLog:
PR tree-optimization/pr103961
* gcc.dg/pr103961.c: New test case.
Ok. Martin's executable testcase would work too but only if
it would be limited to targets with glibc with __sprintf_chk in or
it would need to be linked against -lssp, so I think it is fine as is.
Thanks I'll push this then.
+void
+cap_to_text (int c)
+{
+ char buf[1572];
+ char *p;
+ int n, t;
+ p = 20 + buf;
+ for (t = 8; t--; )
+ {
+ for (n = 0; n < c; n++)
+ p += sprintf (p, "a,");
+ p--;
+ if (__builtin_object_size (p, 1) == 0)
+ abort ();
+ }
Just curious, does your PR77608 patch change this such that __bos (p, 1)
is not ~(size_t)0 but 1572?
We don't know if c isn't 0, so can't count on p += sprintf (p, "a,");
actually incrementing the pointer (and unless we try to understand more
what exactly it is doing we'd need to also assume it could e.g. return -1,
and we aren't too smart about loops anyway, but if the PR77608 patch assumes
that variable bounds aren't out of bounds, then a pointer that started
as 20 + buf can be minimum buf + 0 and maximum buf + 1572.
Note, as __bos is trying to prevent exploiting UB, perhaps assuming the
pointer arithmetics or array refs aren't out of bounds isn't perfect, but
then I'd say a __bos (p, 1) == 1572 means bigger security than __bos (p, 1)
== -1 where we punt and don't check anything.
That's a good catch; the 77608 patch actually breaks this again, but for
a different reason that the patch itself introduces. The "safe"
assumption of wholesize for p_4 is not safe enough because p_17 (which
is p_4 - 1) is assumed to have an underflow, resulting in a zero size again.
I need to rethink the 77608 fix; it won't always be as simple as
returning the wholesize.
Thanks,
Siddhesh
