David Marchand <david.march...@redhat.com> writes:
> On Wed, Dec 4, 2019 at 10:20 PM Aaron Conole <acon...@redhat.com> wrote:
>>
>> David Marchand <david.march...@redhat.com> writes:
>>
>> > We recently started to get random failures on the common_autotest ut with
>> > clang on Ubuntu 16.04.6.
>> >
>> > Example: https://travis-ci.com/DPDK/dpdk/jobs/263177424
>> >
>> > Wrong rte_log2_u64(0) val 0, expected ffffffff
>> > Test Failed
>> >
>> > The ut passes 0 to log2() to get an expected value.
>> >
>> > Quoting log2 / log(3) manual:
>> > If x is zero, then a pole error occurs, and the functions return
>> > -HUGE_VAL, -HUGE_VALF, or -HUGE_VALL, respectively.
>> >
>> > rte_log2_uXX helpers handle 0 as a special value and return 0.
>> > Let's have dedicated tests for this case.
>> >
>> > Fixes: 05c4345ef5c2 ("test: add unit test for integer log2 function")
>> > Cc: sta...@dpdk.org
>> >
>> > Signed-off-by: David Marchand <david.march...@redhat.com>
>> > ---
>>
>> Acked-by: Aaron Conole <acon...@redhat.com>
>>
>> Somethings that concern me:
>>
>> 1. A log2(0) should probably be an undetermined value, but this
>> effectively makes log2(0) == log2(1) so that if anyone uses these
>> for some mathematical work, it will have an exceptional behavior. I
>> know it's documented from a programmer perspective, but I am all for
>> documenting the mathematical side effect as well.
>
> What do you have in mind, adding a big warning in the doxygen "THIS IS
> NOT REAL MATH STUFF" ? :-)
Is such a warning not reasonable? :-)
>>
>> 2. Why hasn't this been complaining for so long? Or has it and we just
>> haven't noticed? Were some compiler flags changed recently? (maybe
>> -funsafe-math was always set or something?)
>
> Yes, I wondered too how we did not see this earlier.
> Even now, it happens randomly.
>
> libc log2(0) is not undefined itself, it returns -infinity.
> Looking at the test code, ceilf (I would expect ceil) returns
> -infinity when getting it passed.
> So I'd say the problem lies in the cast to uint32_t.
>
> Both gcc and clang do not seem to bother with standard compilation
> flags, and the cast gives 0 on my RHEL.
>
> #include <stdio.h>
> #include <inttypes.h>
> #include <math.h>
>
> int main(int argc, char *argv[])
> {
> printf("%lf %f %"PRIu32"\n", log2(0), (float)log2(0), (uint32_t)log2(0));
> return 0;
> }
>
> $ ./log2
> -inf -inf 0
>
>
> Now, if I use UBSAN with clang, I get a complaint at runtime:
> $ ./log2
> (/home/dmarchan/log2+0x41dfa5): runtime error: value -inf is outside
> the range of representable values of type 'unsigned int'
> -inf -inf 0
>
> Not sure if it explains the random failures, but won't undefined
> behavior eat your babies?
Possibly. I would still expect it to be consistent when it eats babies,
but maybe it doesn't have to be.
