The program below aborts at runtime when compiled by gcc 4.3.0 with option "-O2
-std=c99" (it runs correctly without optimization or when compiled with "icc
-std=c99 -restrict"). I believe it should have defined behavior according to
the standard (see e.g. PR14192), for p[0] and q[0] are not restricted pointers,
and p[0][0] is based on p[0] but not p. Intuitively, p and q should behave
like two different arrays of pointers, and we may well want to store aliasing
pointers into them.
This is apparently due to the simplistic find_base_decl() in 4.3.0, which is
quite different the standard's definition of "based on". If I understand
correctly, for two pointers p and q, only when p is provably based on some
restricted pointer p0 (along all control paths), and q is provably not based on
it, can we be sure that a modification to *q can never change *p.
>From a casual glance of the 4.3.0 source, it seems that "restrict" is not well
supported by the tree passes, but the alias sets in the generated RTL make some
use of it. Proper support, either at RTL or tree level, may mean a lot of
work, but in the meantime we had better agree on the meaning of the standard
and avoid generating wrong code :)
==============================================
#include <stdlib.h>
/* fail by gcc-4.3.0 */
void f(int **restrict p, int **restrict q)
{
p[0][0] = 1;
q[0][0] = 2;
if (p[0][0] != 2) abort();
}
int main(void)
{
int a;
int *p1 = &a, *p2 = &a;
f(&p1, &p2);
return 0;
}
--
Summary: Wrong code involving restricted pointers to non-
restricted pointers
Product: gcc
Version: 4.3.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: middle-end
AssignedTo: unassigned at gcc dot gnu dot org
ReportedBy: rainy6144 at gmail dot com
GCC build triplet: i686-pc-linux-gnu
GCC host triplet: i686-pc-linux-gnu
GCC target triplet: i686-pc-linux-gnu
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=36013
