Help with underflows on Fuchsia about cctz HOT 3 CLOSED

It looks like the expectation you are referring to is ...

EXPECT_EQ("-2147481748-01-01T00:00:00+00:00", format(RFC3339, tp, cut));

around line 1025. If so, my first concern is that the test has been changed because that line uses the timezone cut, whereas your output says utc.

That said, that part of the test is about conversions in the libc:UTC timezone, which means it mostly concerns the behavior of gmtime_r(3) from your C library.

It is difficult to say more without being able to run the code, but my first question would be "how does gmtime_r() behave on a time_t value of -67768040609740800?" My guess is that it fails. It should return {0, 0, 0, 1, 0, -2147483648}.

from cctz.

Did you discover anything about the behavior of your gmtime_r(), or is this no longer an issue?

from cctz.

I apologize for not getting back to this earlier.

Fuchsia uses musl for its libc implementation:

• gmtime_r.c
• __secs_to_tm.c

gmtime_r() for -67768040609740800 does indeed fail; it yields a tm with all fields set to 0.

These are the two underflow conditions in __secs_to_tm:

  if (t < INT_MIN * 31622400LL || t > INT_MAX * 31622400LL)
    return -1;

  if (years + 100 > INT_MAX || years + 100 < INT_MIN)
    return -1;

The second is the one that's triggered, with an off-by-one error. (years + 100 ends up being -2147483649.)

It appears there's a patch attempting to fix this that was never merged: https://www.openwall.com/lists/musl/2016/11/02/1
However, the patch is incomplete, because with it, the month ends up off by 2:

{0, 0, 0, 1, 2, -2147483648}

from cctz.