On Tue, 7 Mar 2017, Stefan Liebler wrote:
FAIL: math/test-ildouble-catan
FAIL: math/test-ldouble-catan
FAIL: math/test-ldouble-finite-catan
=> See attached test-ldouble-catan.out
As far as I've seen in generated file
<build-dir>/math/libm-test-catan.c, those cases are all evaluated
only with #if (TEST_COND_binary128).
In <src>/math/s_catan_template.c the imaginary part of the result
is calculated by M_LOG / M_LOG1P functions. I've added some of the
input values to auto-libm-test machinery for log / log1p functions and
verified that the results of log / log1p are correct.
Does anybody see those fails, too?
Thanks for pointing this out, it seems I missed these failures when
regenerating ulps for various platforms after moving catan/catanh tests to
the auto-libm-test machinery. I've reported this as an MPC bug
<https://lists.gforge.inria.fr/pipermail/mpc-discuss/2017-March/001404.html>.
Once fixed (i.e. test expectations regenerated with fixed MPC), recently
regenerated catan/catanh ulps should be deleted before regeneration so
that they can be reduced in the course of regeneration.