Re: timespec test failure on Linux/s390x

[Paul Eggert]

On 8/29/21 6:15 PM, Bruno Haible wrote:

> ../../gltests/test-timespec.c:152: assertion 'eq (timespec_add (a, sumbc), 
> timespec_add (sum, c))' failed
> Aborted (core dumped)
>
> In test-timespec.c:152 the local variables are:
> ...
> ntests = 26
> computed_hz = 1000000000
> i = 1
> a = {tv_sec = -9223372036854775808, tv_nsec = 0}
> roundtrip = {tv_sec = -9223372036854775808, tv_nsec = 0}
> prevroundtrip = {tv_sec = -9223372036854775808, tv_nsec = 0}
> j = 1
> b = {tv_sec = -9223372036854775808, tv_nsec = 0}
> sum = {tv_sec = 0, tv_nsec = 0}

That value of 'sum' is wrong; it should be most-negative value {tv_sec = 
-9223372036854775808, tv_nsec = 0} because A and B are both that value, 
and 'sum = timespec_add (a, b)' is supposed to be saturated addition.

My guess is that timespec-add.c's line 49 'INT_ADD_WRAPV (rs, bs, &rs)' 
is not correctly returning true when RS and BS are both the 
most-negative value. Since you're using GCC, line 49 should be 
equivalent to '__builtin_add_overflow (rs, bs, &rs)' (though you should 
check this), and that suggests a GCC bug. (Yes, I know, everybody at 
first blames the compiler. :-)

Which version of GCC are you using? 
<https://gcc.gnu.org/bugzilla/show_bug.cgi?id=98269> says that 
__builtin_add_overflow does not work in GCC 6.5 on s390x, and that the 
bug is fixed in GCC 7. Could this be the problem?

> sumbc = {tv_sec = -9223372036854775807, tv_nsec = 0}
> timespec_add (a, sumbc) = {tv_sec = -9223372036854775808, tv_nsec = 0}
...> * Questions:
> Is timespec_add (a, sumbc) wrong?

No, it's right.

> Or does this particular triple (a, b, c) need to be exluded from the tests?

I suspect the bug is in the earlier call to timespec_add, as mentioned 
above.