https://gcc.gnu.org/bugzilla/show_bug.cgi?id=116200
Sergei Trofimovich <slyfox at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |slyfox at gcc dot gnu.org
--- Comment #6 from Sergei Trofimovich <slyfox at gcc dot gnu.org> ---
I think I'm seeing a similar crash on libgcrypt-1.10.3. Here is the minimal
example I extracted from it in case it's easier to explore, fix and add a
regression test:
// $ cat a.c
__attribute__((noipa
//, optimize(0)
))
static unsigned int
KECCAK_F1600_ABSORB_FUNC_NAME(int pos, unsigned long nlanes, int blocklanes)
{
unsigned int burn = 0;
while (nlanes)
{
switch (blocklanes)
{
case 21:
/* SHAKE128 */
while (nlanes >= 21)
{
nlanes -= 21;
}
break;
case 18:
/* SHA3-224 */
while (nlanes >= 18)
{
nlanes -= 18;
}
break;
case 17:
/* SHA3-256 & SHAKE256 */
while (nlanes >= 17)
{
nlanes -= 17;
}
break;
case 13:
/* SHA3-384 */
while (nlanes >= 13)
{
nlanes -= 13;
}
break;
case 9:
/* SHA3-512 */
while (nlanes >= 9)
{
nlanes -= 9;
}
break;
}
while (nlanes)
{
nlanes--;
if (++pos == blocklanes)
{
//sink1(hd);
pos = 0;
break;
}
}
}
return burn;
}
int main() {
KECCAK_F1600_ABSORB_FUNC_NAME(0, 1, -1);
}
$ gcc a.c -o a -O0 && ./a
$ gcc a.c -o a -O1 && ./a
Segmentation fault (core dumped)
I think the sample is still valid: it tries to do nlanes iterations (with some
form of unrolling).