BUG: Test aginst +/-TANH_HUGE

ewmoore · ewmoore · commit 5362b35db178 · 2014-02-26T18:11:27.000-05:00
This is a bug introduced when adapting the FreeBSD implementation. Since
they play some games extracting parts of the double and comparing them
as unsigned ints and we don't. We don't since that doesn't generalize to
float, and long double as we do in npy_math_complex.c.src
diff --git a/numpy/core/src/npymath/npy_math_complex.c.src b/numpy/core/src/npymath/npy_math_complex.c.src
@@ -556,7 +556,7 @@ static NPY_INLINE @ctype@ cmuli@c@(@ctype@ a)
      * approximation sinh^2(huge) ~= exp(2*huge) / 4.
      * We use a modified formula to avoid spurious overflow.
      */
-    if (x >= TANH@C@_HUGE) {
+    if (npy_fabs@c@(x) >= TANH@C@_HUGE) {
         @type@ exp_mx = npy_exp@c@(-npy_fabs@c@(x));
         return (npy_cpack@c@(npy_copysign@c@(1, x),
             4 * npy_sin@c@(y) * npy_cos@c@(y) * exp_mx * exp_mx));
