[Info-vax] gnulib test-fmal-ieee failure help needed.

[John E. Malmberg]

On 7/23/2017 6:54 PM, John E. Malmberg wrote:
> On 7/23/2017 6:38 PM, Craig A. Berry wrote:
>> On 7/23/17 4:51 PM, John E. Malmberg wrote:
>>> On 7/22/2017 11:01 PM, John E. Malmberg wrote:
>>>> On 7/22/2017 10:52 PM, John E. Malmberg wrote:
>>>>> On 7/22/2017 10:09 PM, Craig A. Berry wrote:
>>>>>> On 7/22/17 8:54 PM, John E. Malmberg wrote:
>>>>>>> The gnulib test-fmal-ieee is failing for the gnulib provided fmal 
>>>>>>> routine.
>>>>>>>
>>>>>>> Actually a lot of the IEEE tests are failing.
>>>>>>>
>>>>>>> My compiler command line on OpenVMS 8.4 is:
>>>>>>>
>>>>>>> CC/STANDARD=(RELAXED)/ACCEPT=(NOVAXC,RESTR,C99)-
>>>>>>>    /LIST/SHOW=(EXPAN,INCLU)/NAMES=(AS_IS,SHORT)-
>>>>>>>    /MAIN=POSIX_EXIT-
>>>>>>>    /FLOAT=IEEE/IEEE_MODE=DENORM_RESULTS-
>>>>>>>    /NESTED=NONE-
>>>>>>>    /PREFIX=EXCEPT=(STRTOIMAX,STRTOUMAX,ISWBLANK,ISBLANK)-
>>>>>>>    /WARN=(DISABLE=QUESTCOMPARE)-
>>>>>>>    /DEBUG/NOOPTIMIZE/DEB=ALL/NOOPT-
>>>>>>>    /DEF=-(HAVE_CONFIG_H,EXEEXT="",EXEEXT="",NO_XMALLOC,EXEEXT="",-
>>>>>>>           GNULIB_STRICT_CHECKING=1,_POSIX_EXIT,_USE_STD_STAT,-
>>>>>>>        _  _USE_STD_IOSTREAM)-
>>>>>>>     /INC=(./,../)/OBJECT=SYS$DISK:[]fmal.o SYS$DISK:[]fmal.c
>>>>>>>
>>>>>>> The floating variables involved are declared double.
>>>>>>>
>>>>>>> The input to the failed call is:
>>>>>>>
>>>>>>>   fmal(Infinity, 2.0L, -Infinity) and the result is expected to 
>>>>>>> be NaN.
>>>>>>>
>>>>>>> Is there some other compile settings that I should be using to be 
>>>>>>> more compatible with GNULIB's expectations for IEEE math?
>>>>>>
>>>>>> fmal is for long doubles, so passing doubles to it is unlikely to 
>>>>>> work.
>>>>>> For example, you probably want LDBL_INFINITY instead of Infinity.
>>>>>> Possibly related is that none of the IEEE test macros (isnan, isinf,
>>>>>> etc.) in math.h on VMS works for floats or long doubles -- they only
>>>>>> work for doubles. Which is a bug. We have some workarounds in Perl 
>>>>>> for
>>>>>> some of it.
>>>>>
>>>>> Your right, the gnulib fmal code is long doubles, not doubles.
>>>>>
>>>>> VMS has a number of the declarations that should be in math.h in fp.h.
>>>>>
>>>>> In this case INFINITY is defined as decc$gt_gbl_infinity by fp.h 
>>>>> and is reported as Infinity in the VMS debugger.
>>>>
>>>> The test-fmal1 fails in the same way as test-fmal-ieee.
>>>>
>>>> The test-malm2 fails when fma returns what the debugger says is 
>>>> Infinity and the test was expecting "-Infinity".
>>>>
>>>> DBG> eval result
>>>> Infinity
>>>> DBG> eval -result
>>>> -Infinity
>>>
>>> Latest failure:
>>>
>>> ASSERT (ILOGB (L_(0.0)) == FP_ILOGB0)
>>> ILOG_B is ilogb()
>>> FP_ILOGB0 is - 2147483647
>>> L_(0.0) is 0.0L
>>
>> So it's passing a long double to a function that expects a double. That
>> doesn't seem likely to work. It should be calling ilogbl not ilogb if
>> the argument is a long double. See:
>>
>> <http://pubs.opengroup.org/onlinepubs/9699919799/functions/ilogb.html>
> 
> I can not seem to copy things right.
> 
> The ILOGB in the last test was
> 
>   I2   1   42156   ASSERT (ILOGB (L_(0.0)) == FP_ILOGB0);
>         E           do { if ( ! ( ilogbl ( 0.0L ) == ( - 2147483647 ) ) )
>                          { rpl_fprintf ( ( decc$ga_stderr ) , 
>                              "%s:%d: assertion '%s' failed\n" ,
> 
> "GNULIB_ROOT:[gnulib.gltests]test-ilogb.h;1" ,
>                                          60 , "ILOGB (L_(0.0)) == 
> FP_ILOGB0" ) ;                          rpl_fflush ( ( decc$ga_stderr ) 
> ) ; abort ( ) ; }                       } while ( 0 )
> 
> The gnulib code should be correct with regards to doubles, long doubles, 
> etc as the code works on many other platforms.  If I post a mis-match 
> between doubles and long doubles etc, it is likely a transcription 
> error.  I realize that makes it harder to diagnose what is going wrong.
> 
> So the problem is likely one of:
> 
> 1. Wrong CC command qualifiers.
> 2. Bug in math.h/fp.h
> 3. Bug in VMS CRTL
> 4. Bug in VMS C compiler.
> 
>>>> These results are showing up while am running the failed tests in 
>>>> debug to get more information.  So as soon as I exit one failed 
>>>> test, the script moves on to the next.  I am not sure how many more 
>>>> tests are going to run in this pass.
>>>
>>> And to clarify, this is on HPE OpenVMS 8.4 on Itanium.

The test-log-ieee seems to be more clear on the failure.

   x = 0.6;
   y = log (x);
   ASSERT (y >= -0.5108256238 && y <= -0.5108256237);
Debugger reports y = -0.510825623765991 so assert passes.

Failed assert:

   minus_zerod = 0.0 in the debugger.

  I2   2   41962     DOUBLE z = LOG (MINUS_ZERO);
        E             double
        E                        log
        E                             minus_zerod
  I2   2   41963     ASSERT (z == - HUGEVAL);
        E             z == - decc$gt_dbl_infinity

  Debugger reports z is NaN

I am seeing a pattern where NaN is seen where infinity is expected in 
some of the failed asserts.

Regards,
   -John
    wb8tyw at qsl.net_work