Hello,
I've committed a revision of src/float.lisp so that the branch cuts
of atanh(bfloat(x)) match those of bfloat(rectform(atanh(x))).
With this change and the changes to floating-point evaluation of
asin, acos, and atanh mentioned yesterday, run_testsuite reports
no unexpected errors with Maxima built from the 5.9.3 branch in CVS
on GCL, Clisp, and SBCL (all Linux).
COMPLEX-ATANH applies the identity arg(x - y i) = - arg(x + y i)
but that fails if y = 0 and x < 0. So I plugged in a special case for y = 0.
The patched code needs a little more work but I think we're
heading in the right direction.
Hope this helps,
Robert