acos (may apply to other trigonometric functions) have inconsistent results across different build options

When ECL is build with complex float support, then

CL-USER> (acos 2) ;; #C(0.0 -1.316958)

Otherwise, when it is build without the complex float support then a function ported from cmucl is used (see numlib.lsp):

(defun complex-acos (z)
  (declare (number z)
           (si::c-local))
  (let ((sqrt-1+z (sqrt (+ 1 z)))
        (sqrt-1-z (sqrt (- 1 z))))
    (complex (* 2 (atan (realpart sqrt-1-z) (realpart sqrt-1+z)))
             (asinh (imagpart (* (conjugate sqrt-1+z)
                                 sqrt-1-z))))))

CL-USER> (complex-acos 2) ;; #C(0.0 1.316958)

I'm not sure whether having the positive imaginary part is required by the spec, but the the discrepancy doesn't look good.

• check whether the result using inlined cacos(2 + I*0.0) is conforming
• think about how to make results consistent