Re: [dpdk-stable] [PATCH] test/common: fix log2 check

[David Marchand <david.marchand@redhat.com>]

On Wed, Dec 4, 2019 at 10:20 PM Aaron Conole <aconole@redhat.com> wrote:
>
> David Marchand <david.marchand@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: stable@dpdk.org
> >
> > Signed-off-by: David Marchand <david.marchand@redhat.com>
> > ---
>
> Acked-by: Aaron Conole <aconole@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" ? :-)


>
> 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?


-- 
David Marchand