On Tue, 2002-05-21 at 17:35, Raymond Toy wrote:
> >>>>> "Valerij" == Valerij Pipin <pip@iszf.irk.ru> writes:
>
> Valerij> Currently, the output of cvs maxima tests (after "make
> Valerij> check") is defined to go to some log file and it causes
>
> When I run "make check" it just hangs. Don't know why. I can the
> command by hand and it works.
"make check" hangs becuase it is waiting for an answer to "is
<something> positive, negative or zero". This is due to a new bug in the
bessel function routines. Some arguments are being excessively
evaluated.
Previously, I could do this:
(C1) %j[1](sqrt(x));
(D1) %J (SQRT(x))
1
Now, I get this:
(C1) %j[1](sqrt(x));
Is x positive, negative, or zero?
This bug has to be fixed. I see you updated to tests to work around the
bug. Please roll back those changes. I haven't found the source of the
bug yet, but the above information should be a good lead.
> Valerij> the hang because maxima stops and needs the sign of some
> Valerij> parameters to continue. I think it's better to make
> Valerij> output for maxima tests to the screen instead, to get a
> Valerij> control on tests procedure.
>
> I think we should put in the appropriate assume's so that the user
> doesn't need to know how to answer the questions.
Right. That's the way the tests are set up. It is unfortunate, however,
that a bug such as the one above causes the tests to "hang." Obviously,
the bug has to be fixed, but the testing procedure leaves something to
be desired. My best idea so far is to make
make check VERBOSE=true
run the tests without redirecting the output. What does everyone think
of that?
> Valerij> Error(s) found in rtest6a.mac: (8)
> Valerij> Error(s) found in rtest14.mac: (21 17 16 15 3)
> Valerij> Error(s) found in rtest15.mac: (4 3)
>
> With cmucl and clisp, I get 2 errors in rtest14.mac (3 16) and 2 in
> rtest15.mac (3 4). For rtest15.mac, I think those errors are real
> errors because we incorrectly compute the integrals.
Right. We have a fix for one, but I haven't had time to commit it yet
because I was distracted by the talk for the conference in France and
the GCL issues. Now that those are gone, I should be able to get back to
getting the one fix in cvs.
> For rtest14.mac, test 3 has the wrong value for bessel(2,3), which is
> purely real, not complex.
Right. The only reason I hadn't updated that number is that I wanted to
get verify the number through a third party. It is obvious that some of
the tests in the special function section are checking against values
that are incorrect past single-precision. I would love to have a
volunteer audit all the special function numerical tests. Comparison
with a third party is really necessary.
> Don't know about test 16. I suspect this
> might be caused by my change to make %j[1/2](x) expand to elementary
> functions.
Is it related to the bug I described above?
--Jim