https://gcc.gnu.org/bugzilla/show_bug.cgi?id=105739
--- Comment #9 from Jan Hubicka <hubicka at gcc dot gnu.org> ---
Indeed volatile checks seems to be missing across ipa-prop code. Here is
smaller testcase:
__attribute__((noinline))
static int
test2(int a)
{
if (__builtin_constant_p (a))
__builtin_abort ();
return a;
}
/*__attribute__((noinline))*/
static int
test(int *a)
{
int val = *(volatile int *)a;
if (__builtin_constant_p (val))
__builtin_abort ();
if (val)
return test2(val);
return 0;
}
int a;
int
main()
{
a = 0;
return test (&a);
}
is optimized to
main:
.LFB2:
.cfi_startproc
movl $0, a(%rip)
movl a(%rip), %eax
xorl %eax, %eax
ret
.cfi_endproc
which I don't think is correct. The volatile load can be non-0 and thus we can
return non-0.
It does not trigger unreachable as I hoped for since vrp and fab passes seems
to jump to same conclusion as ipa-cp.