Re: [Gcl-devel] some GCL bugs revealed by ansi test suite

[Vadim V. Zhytnikov]

Camm Maguire пишет:
> Greetings!
>
> "Vadim V. Zhytnikov" <address@hidden> writes:
>
>
> > Hi!
> >
> > I'd like to report some GCL bugs revealed during
> > clocc ansi compatibility tests.  The list is
> > very far from being complete and problem ranges
> > from very simple bugs to some quite general
> > problems which require modification in many many
> > places.
> >
> > First, a couple of very simple bugs:
> >
> > 1)  (make-array 5 :fill-pointer -1)
> > must signal an error for any negative value
> > of :fill-pointer.
>
>
> Done.  Please verify.
>
>
> > 2)  The built-in constant
> >
> >     array-total-size-limit
> >
> > returns symbol (another constant)
> >
> >     array-dimension-limit
> >
> > as its value but should return
> > numerical value (probably the same as
> > value of array-dimension-limit).
>
>
> Done.  Please verify.
>
>
> > 3) The path designator #P"aaa" (capital P)
> > result in error but should not be different
> > from #p"aaa".
>
>
> Done.  Please verify.
>
>
> > Now some more serious problems. ANSI tests
> > clocc suite report really huge number of
> > errors with floating point operations.
> > Here is some short extract from error log:

Fine!

> Umm -- clearly the AKCL numbers below are right, and the 'should be'
> are wrong, no?  I can even see this in my head :-).  Are the tests
> using 'legacy' broken values as the expected return?
>
>
> > Bugid: :NUMBER-LEGACY-2938 interpreted Form: (+ 0.52019S0 0.98203S0)
> > Should be: 1.50223S0
> > AKCL: 1.50222S0
> >
> > Bugid: :NUMBER-LEGACY-2942 interpreted Form: (+ 0.026268S0 0.6137S0)
> > Should be: 0.63997S0
> > AKCL: 0.639968S0
> >
> > Bugid: :NUMBER-LEGACY-2946 interpreted Form: (+ -0.338943S0 0.450523S0)
> > Should be: 0.11158S0
> > AKCL: 0.11158S0
> >
> > Bugid: :NUMBER-LEGACY-2950 interpreted Form: (+ -0.032799S0 0.995186S0)
> > Should be: 0.96239S0
> > AKCL: 0.962387S0
> >
> > Frankly I don't quite understand how to treat these
> > results. Are they really bugs? Clisp and cmucl
> > seems to be OK...
>
>
> Could you please give me some lisp showing this?
>
>
> > Another similar result which is reported as bug:
> > 1e-37     ->  1.0000000000000001E-37
> > 10.0e-38  ->  9.9999999999999986E-38

One thing confised me a bit with all these floating
point results.  Target values (Should be:) seems to
be exactly like in clisp.  I also checked few examples
with cmucl and once again got these "Should be" values.
Originally I thought that such deltas in least significant
digits are not important for floating point operations
but with such observation started wondering what maybe
I'm wrong (I'm not a floating point expert).
But after some more extensive check I see that all lisps
(clisp, cmucl, sbcl, ecls) give slightly different values.
So probably we have nothing to worry about.

Best wishes,

--
     Vadim V. Zhytnikov

      <address@hidden>
     <address@hidden>
     <address@hidden>
    <address@hidden>