Re: [boost] Optimized portable isnan, isfinite, fpclassify etc

From: John Maddock (john_at_[hidden])
Date: 2006-11-10 05:27:40

Johan Råde wrote:
> John,
>
> Thanks a lot for running these tests.
>
> Here is a summary of what you have done
> (could you double check that I've got the
> operating systems / compilers / processors right),
> and my interpretation of the results:
>
> ---- pass ----
>
> Win32 / Intel 8.? / ia32
> Win32 / Intel 9.1 / ia32
> Linux / Intel 9.1 / ia32
> HP-UX / GCC ? / ia64

Add to that HP aCC with -O4.

>
> ---- fail ----
>
> Win32 / Intel 7.? / ia32
>
>   where: basic_test<float> line 242
>   what: !isinf(signalling_NaN())
>   why: broken numeric_limits<T>::signalling_NaN

Not sure about that: the std lib is the same as that used by VC6, however 
these are both old compilers now, so maybe ignor this one for now (your code 
doesn't compile on VC6 anyway).

> HP-UX / aCC / ia64
>
>   where: basic_test<float> line 122
>   what: fpclassify(denorm_min()) == FP_SUBNORMAL
>   why: multiple definitions of FP_SUBNORMAL ?

I don't think that's the cause: it passes with -O4 but not -Ofaster.  The 
docs for -Ofaster state that it breaks code that relies on denorms, and it's 
noticable that your first denorm test passes, but the second (which negates 
denorm_min()) fails.  I suspect that the negation of the denorm causes it to 
round down to zero maybe?

> Linux / Intel 9.1 / ia64
>
>   where: basic_test<float> line 139
>   what: !isnormal(min()/2)
>   why: broken numeric_limits<T>::min ?

I don't believe that's the cause: the test passes when built in debug mode, 
or with -O1, but fails with -O2 or -O3, so it's an optimiser issue again.

My suspicion would be that:

        x = (std::numeric_limits<T>::min)() / 2;

either underflows to zero, or is evaluated by the compiler at compile time 
maybe?

> Could you run the inspect tool on the last two platforms as well?

Here's HP-UX:

maddock_at_td176> aCC -AA -I../../ -I ~ -Ofaster *.cpp
maddock_at_td176> ./a.out

This platform is big-endian.
BOOST_LITTTLE_ENDIAN is defined.

---------- float --------------------

000 00 00 00
sn. min00 00 00 01
n. min00 80 00 00
13f 80 00 00
4/33f aa aa ab
max7f 7f ff ff
inf7f 80 00 00
q. nan7f c0 00 00
s. nan7f a0 00 00
-1bf 80 00 00

---------- double -------------------

000 00 00 00 00 00 00 00
sn. min00 00 00 00 00 00 00 01
n. min00 10 00 00 00 00 00 00
13f f0 00 00 00 00 00 00
4/33f f5 55 55 55 55 55 55
max7f ef ff ff ff ff ff ff
inf7f f0 00 00 00 00 00 00
q. nan7f f8 00 00 00 00 00 00
s. nan7f f4 00 00 00 00 00 00
-1bf f0 00 00 00 00 00 00

---------- long double --------------

000 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00
sn. min00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 01
n. min00 01 00 00 00 00 00 00 00 00 00 00 00 00 00 00
13f ff 00 00 00 00 00 00 00 00 00 00 00 00 00 00
4/33f ff 55 55 55 55 55 55 55 55 55 55 55 55 55 55
max7f fe ff ff ff ff ff ff ff ff ff ff ff ff ff ff
inf7f ff 00 00 00 00 00 00 00 00 00 00 00 00 00 00
q. nan7f ff 80 00 00 00 00 00 00 00 00 00 00 00 00 00
s. nan7f ff 40 00 00 00 00 00 00 00 00 00 00 00 00 00
-1bf ff 00 00 00 00 00 00 00 00 00 00 00 00 00 00

And Intel on Linux/IA64

maddock_at_[hidden]> icc -O3 -I../../ -I ~ *.cpp -static
maddock_at_[hidden]> ./a.out

This platform is little-endian.
BOOST_LITTTLE_ENDIAN is defined.

---------- float --------------------

0         00 00 00 00
sn. min   00 00 00 01
n. min    00 80 00 00
1         3f 80 00 00
4/3       3f aa aa ab
max       7f 7f ff ff
inf       7f 80 00 00
q. nan    7f e0 00 00
s. nan    7f a0 00 00
-1        bf 80 00 00

---------- double -------------------

0         00 00 00 00 00 00 00 00
sn. min   00 00 00 00 00 00 00 01
n. min    00 10 00 00 00 00 00 00
1         3f f0 00 00 00 00 00 00
4/3       3f f5 55 55 55 55 55 55
max       7f ef ff ff ff ff ff ff
inf       7f f0 00 00 00 00 00 00
q. nan    7f fc 00 00 00 00 00 00
s. nan    7f f4 00 00 00 00 00 00
-1        bf f0 00 00 00 00 00 00

---------- long double --------------

0         00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00
sn. min   00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 01
n. min    00 00 00 00 00 00 00 00 80 00 00 00 00 00 00 00
1         00 00 00 00 00 00 3f ff 80 00 00 00 00 00 00 00
4/3       00 00 00 00 00 00 3f ff aa aa aa aa aa aa aa ab
max       00 00 00 00 00 00 7f fe ff ff ff ff ff ff ff ff
inf       00 00 00 00 00 00 7f ff 80 00 00 00 00 00 00 00
q. nan    00 00 00 00 00 00 7f ff e0 00 00 00 00 00 00 00
s. nan    00 00 00 00 00 00 7f ff a0 00 00 00 00 00 00 00
-1        00 00 00 00 00 00 bf ff 80 00 00 00 00 00 00 00

Can I make a couple of suggestions with respect to your test program?

1)  It would be useful if the asserts were changed to something that didn't 
abort so we could see all the possible failures at once.
2) It would be useful if failed tests printed out the binary representation 
of the failed value (merge the inspect tool into the test.
3) It would help to check whether numeric_limits::has_infinity/has_denorm 
etc is true before testing those features.
4) It might pay to sanity check that denorm test values aren't zero or >= 
numeric_limits<>::min() before using them.  That would catch out clever 
compilers right away: there's no point in testing for a denorm if the value 
isn't a denorm! :-)

HTH, John.