Date: Tue, 24 Jul 2018 05:54:37 +1000 (EST) From: Bruce Evans <brde@optusnet.com.au> To: enh <enh@google.com> Cc: freebsd-numerics@freebsd.org Subject: Re: fmod nan_mix usage Message-ID: <20180724050141.Q2280@besplex.bde.org> In-Reply-To: <CAJgzZopb_0fxM9jbVjUEZ0JPOfcrgeQo_Ki-afZ5aRNr38tKVg@mail.gmail.com> References: <CAJgzZopb_0fxM9jbVjUEZ0JPOfcrgeQo_Ki-afZ5aRNr38tKVg@mail.gmail.com>
next in thread | previous in thread | raw e-mail | index | archive | help
On Mon, 23 Jul 2018, enh via freebsd-numerics wrote: > the recent change from > > return (x*y)/(x*y); > > to > > return nan_mix(x, y)/nan_mix(x, y); > > in e_fmod.c broke some of our unit tests. for example, fmod(3.f, 0.f) in > one of the VM tests. This is a bug in my change. nan_mix(x, y) does essentially x+y, but here essentially x*y is needed so that y = 0 gives 0 unless x is NaN. In the example, adding gives 3 instead of 0, so the final result is 1 instead of 0.0/0.0 = NaN. The log message mentions avoiding this problem in s_ccosh[fl].c and s_sinh[fl].c. This list was supposed have all special cases. Unfortunately, this seems to prevent use of a single macro. I will try using a 2 macros with 1 using sums and the other products. The non-broken cases converted sums to sums. > bionic/tests/math_test.cpp:(784) Failure in test > math_h_force_long_double.fmod > Value of: isnan(fmod(3.0, 0.0)) > Actual: false > Expected: true > math_h_force_long_double.fmod exited with exitcode 1. > [ FAILED ] math_h_force_long_double.fmodf (13 ms) > bionic/tests/math_test.cpp:(798) Failure in test > math_h_force_long_double.fmodf > Value of: isnanf(fmodf(3.0f, 0.0f)) > Actual: false > Expected: true > math_h_force_long_double.fmodf exited with exitcode 1. > [ FAILED ] math_h_force_long_double.fmodl (12 ms) > bionic/tests/math_test.cpp:(812) Failure in test > math_h_force_long_double.fmodl > Value of: isnanl(fmodl(3.0L, 0.0L)) > Actual: false > Expected: true Do you have a lot of special tests like this? I mostly use generic tests that don't assert any particular result, but compare the results in different precisions. I apparently changed all precisions to be consistently wrong at the same time. > it looks like e_remainder.c might have the same issue, but Android's tests > didn't catch that :-( i'll improve the tests... Indeed. Also remquo* and ctanh* :-(. ctanh* should be more like csinh* and ccosh*, and it was. The only other complicated case seems to be hypot[fl](). This subtracts instead of adds, since it wants to convert Inf-Inf to NaN. Bruce
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?20180724050141.Q2280>