On 2025-12-16 09:21, Andrew MacLeod wrote:
The new integrated PTA figures out a couple of things.. first, it knows
* from earlier that buf3 and buf1 points to the same thing,
* and can also figure out that buf3 + 4 == buf1 + 4, so it can
fold away the condition afterwards since arg1 is the result
How does it know of this equivalence? Is PTA somehow ignoring
__noinline__?
No, it earlier in the function we have a few statements like :
if (memcpy (buf3, buf5, 8) != (char *) buf1
abort ();
which translates into
_9 = __builtin___memcpy_chk (buf3_60(D), "a", 1, _5);
if (_9 != &buf1)
goto <bb 13>; [INV]
else
goto <bb 12>; [INV]
where it can establish that buf3_60 == &buf1 since _9 == buf3_60 and _9
== &buf1 from bb12 onward
Aha, ok! It looks like test2_sub needs to be broken up then, to have
each of those tests run independently. That way the compiler won't have
a chance to leak the result of one test into the ones that follow.
Something like:
__attribute__((noinline))
test2_sub_1 (...)
{
if (__builtin_memcpy (buf3, "ABCDEF", 6) != (char *) buf1
|| memcmp (buf1, "ABCDEFghijklmnopq\0", 19))
abort ();
}
test2_sub ()
{
...
test2_sub_1 (buf3, buf4, buf6, n);
...
}
and so on. Basically all of the tests with buf3 based expressions as
the destination need their own noinline function.
Sid
