timespec test failure on Linux/s390x

timespec test failure on Linux/s390x

[Bruno Haible]

In a testdir of all of Gnulib, on Linux/s390x, test-timespec fails:

../../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:

test = {{tv_sec = -9223372036854775808, tv_nsec = -1},
        {tv_sec = -9223372036854775808, tv_nsec = 0},
        {tv_sec = -9223372036854775808, tv_nsec = 1},
        {tv_sec = -9223372036854775808, tv_nsec = 999999999},
        {tv_sec = -9223372036854775807, tv_nsec = 0},
        {tv_sec = -9223372036854775807, tv_nsec = 1},
        {tv_sec = -9223372036854775807, tv_nsec = 999999999},
        {tv_sec = -1, tv_nsec = 0},
        {tv_sec = -1, tv_nsec = 1},
        {tv_sec = -1, tv_nsec = 999999999},
        {tv_sec = 0, tv_nsec = 0},
        {tv_sec = 0, tv_nsec = 1},
        {tv_sec = 0, tv_nsec = 999999999},
        {tv_sec = 1, tv_nsec = 0},
        {tv_sec = 1, tv_nsec = 1},
        {tv_sec = 1, tv_nsec = 999999999},
        {tv_sec = 1234567890, tv_nsec = 0},
        {tv_sec = 1234567890, tv_nsec = 1},
        {tv_sec = 1234567890, tv_nsec = 999999999},
        {tv_sec = 9223372036854775806, tv_nsec = 0},
        {tv_sec = 9223372036854775806, tv_nsec = 1},
        {tv_sec = 9223372036854775806, tv_nsec = 999999999},
        {tv_sec = 9223372036854775807, tv_nsec = 0},
        {tv_sec = 9223372036854775807, tv_nsec = 1},
        {tv_sec = 9223372036854775807, tv_nsec = 999999999},
        {tv_sec = 9223372036854775807, tv_nsec = 2000000000}}
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}
diff = {tv_sec = 0, tv_nsec = 0}
rdiff = {tv_sec = 0, tv_nsec = 0}
k = 13
c = {tv_sec = 1, tv_nsec = 0}
sumbc = {tv_sec = -9223372036854775807, tv_nsec = 0}
timespec_add (a, sumbc) = {tv_sec = -9223372036854775808, tv_nsec = 0}
timespec_add (sum, c) = {tv_sec = 1, tv_nsec = 0}

* Questions:
Is timespec_add (a, sumbc) wrong?
Or does this particular triple (a, b, c) need to be exluded from the tests?

* struct timespec is defined as:

struct timespec
{
  __time_t tv_sec;              /* Seconds.  */
  ...
};

and apparently __time_t is 64 bits wide and signed.

* Debugging the relevant invocation of timespec_add (a, sumbc):

#0  timespec_add (a=..., b=...) at ../../gllib/timespec-add.c:30
30      {
(gdb) next
31        time_t rs = a.tv_sec;
(gdb) 
32        time_t bs = b.tv_sec;
(gdb) 
33        int ns = a.tv_nsec + b.tv_nsec;
(gdb) print rs
$4 = -9223372036854775808
(gdb) print bs
$5 = -9223372036854775807
(gdb) next
34        int nsd = ns - TIMESPEC_HZ;
(gdb) 
35        int rns = ns;
(gdb) 
37        if (0 <= nsd)
(gdb) 
49        if (INT_ADD_WRAPV (rs, bs, &rs))
(gdb) 
51            if (bs < 0)
(gdb) 
53                rs = TYPE_MINIMUM (time_t);
(gdb) 
54                rns = 0;
(gdb) 
64        return make_timespec (rs, rns);
(gdb) print rs
$6 = -9223372036854775808


Bruno

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.

Re: timespec test failure on Linux/s390x

[Bruno Haible]

Paul Eggert wrote:
> Which version of GCC are you using?

s390x-linux-gnu-gcc-10 (Ubuntu 10.3.0-1ubuntu1~20.04) 10.3.0

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

I'll debug that call...

Bruno

Re: timespec test failure on Linux/s390x

[Bruno Haible]

Paul Eggert wrote:
> 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.

Yes, this is what happens.

Breakpoint 4, main () at ../../gltests/test-timespec.c:135
135                     struct timespec sum = timespec_add (a, b);
(gdb) print a
$11 = {tv_sec = -9223372036854775808, tv_nsec = 0}
(gdb) print b
$12 = {tv_sec = -9223372036854775808, tv_nsec = 0}
(gdb) step
timespec_add (a=..., b=...) at ../../gllib/timespec-add.c:30
30      {
(gdb) next
31        time_t rs = a.tv_sec;
(gdb) 
32        time_t bs = b.tv_sec;
(gdb) 
33        int ns = a.tv_nsec + b.tv_nsec;
(gdb) 
34        int nsd = ns - TIMESPEC_HZ;
(gdb) 
35        int rns = ns;
(gdb) 
37        if (0 <= nsd)
(gdb) print rs
$13 = -9223372036854775808
(gdb) print bs
$14 = -9223372036854775808
(gdb) print ns
$15 = 0
(gdb) print nsd
$16 = -1000000000
(gdb) next
49        if (INT_ADD_WRAPV (rs, bs, &rs))
(gdb) next
64        return make_timespec (rs, rns);

> that suggests a GCC bug. (Yes, I know, everybody at 
> first blames the compiler. :-)

I can't determine whether it's a GCC or a QEMU bug. The QEMU bug hypothesis
looks more probable to me, so I've opened this ticket:
https://gitlab.com/qemu-project/qemu/-/issues/616

Bruno

Re: timespec test failure on Linux/s390x

[Paul Eggert]

On 9/12/21 8:28 AM, Bruno Haible wrote:
> I can't determine whether it's a GCC or a QEMU bug.

Surely it is a QEMU bug, as QEMU isn't properly implementing the 
z/Architecture Principles of Operation (POP). Page 7-25 of the current 
edition (SA22-7832-12) says that the AGR instruction treats its operands 
as 64-bit signed binary integers, and sets condition code 3 (the 
overflow flag) if the result overflows. You can get a copy of the POP 
here (requires registration):

https://www.ibm.com/support/pages/zarchitecture-principles-operation

> The QEMU bug hypothesis
> looks more probable to me, so I've opened this ticket:
> https://gitlab.com/qemu-project/qemu/-/issues/616

The patch you submitted to QEMU looks reasonable. However, there are 
similar bugs in cc_calc_sub_32 and cc_calc_sub_64 that should also be 
fixed. I attempted to submit a comment to that effect on the gitlab.com 
website but couldn't easily figure out how to make commenting work there.

Re: timespec test failure on Linux/s390x

[Bruno Haible]

Paul Eggert wrote:
> > I can't determine whether it's a GCC or a QEMU bug.
> 
> Surely it is a QEMU bug, as QEMU isn't properly implementing the 
> z/Architecture Principles of Operation (POP).

Thanks; I didn't know about this document.

> Page 7-25 of the current 
> edition (SA22-7832-12) says that the AGR instruction treats its operands 
> as 64-bit signed binary integers, and sets condition code 3 (the 
> overflow flag) if the result overflows. You can get a copy of the POP 
> here (requires registration):
> 
> https://www.ibm.com/support/pages/zarchitecture-principles-operation

The previous edition (available without registration at
http://publibfi.boulder.ibm.com/epubs/pdf/dz9zr011.pdf )
says the same thing.

> The patch you submitted to QEMU looks reasonable. However, there are 
> similar bugs in cc_calc_sub_32 and cc_calc_sub_64 that should also be 
> fixed.

Oh, indeed! I had looked at these functions and did not see the bug.
Reported at https://gitlab.com/qemu-project/qemu/-/issues/618 .

Bruno