Bug #112964 for Statistics-Shannon: t/01base.t may fail (-Duselongdouble && hash randomization)

Sat Mar 12 16:56:37 2016 SREZIC [...] cpan.org - Ticket created

Subject:

t/01base.t may fail (-Duselongdouble && hash randomization)

If perl is compiled with -Duselongdouble, then t/01base.t may fail randomly, probably due to hash randomization. At least if I run for i in $(seq 100); do echo -n "$i "; env PERL_HASH_SEED=$i PERL_PERTURB_KEYS=NO /opt/perl-5.20.1D/bin/prove -b t/01base.t >&/dev/null && echo pass || echo fail; done twice, then the same fail/pass pattern happens.

Sun Mar 13 11:07:23 2016 JHI [...] cpan.org - Correspondence added

On Sat Mar 12 16:56:37 2016, SREZIC wrote: Show quoted text

> If perl is compiled with -Duselongdouble, then t/01base.t may fail > randomly, probably due to hash randomization. At least if I run > > for i in $(seq 100); do echo -n "$i "; env PERL_HASH_SEED=$i > PERL_PERTURB_KEYS=NO /opt/perl-5.20.1D/bin/prove -b t/01base.t

> >&/dev/null && echo pass || echo fail; done

> > twice, then the same fail/pass pattern happens.

Thanks for your report, but I cannot repro this in OS X with a -Duselongdouble Perl. Looking at the code I do not see how hash ordering could affect the result. The key is the index() sub, and while it does do 'keys' (without sorting) and 'values', I off-hand cannot see that this should matter, because what it does is sums. And furthermore, I don't see how longdouble or not should make a difference for (just) the base01.t failing. Could you debug a little bit more?

Sun Mar 13 11:07:23 2016 The RT System itself - Status changed from 'new' to 'open'

Sun Mar 13 11:10:24 2016 JHI [...] cpan.org - Correspondence added

Correction: now I was able to repro it. Stay tuned.

Sun Mar 13 11:23:42 2016 JHI [...] cpan.org - Correspondence added

On Sun Mar 13 11:10:24 2016, JHI wrote: Show quoted text

> Correction: now I was able to repro it. Stay tuned.

The test was expecting floating point equivalence, which in itself is a bad expectation, but what broke it was a bit more interesting: because of long doubles the summation order mattered, and the hash randomization affected the summation order.