[PATCH v2 4/9] math: Use an improved algorithm for hypot (dbl-64)

[Wilco Dijkstra <Wilco.Dijkstra@arm.com>]

Hi Adhemerval,

On AArch64 the new version is ~25% slower even after removing wrappers.
Using the faster checks for special cases helps here too of course. However
there is another issue: issignaling is not inlined, forcing a frame to be created
plus several callee-saves. This fixes it:

  if (!isfinite(x) || !isfinite(y))
  {
    if ((isinf (x) || isinf (y))
        && !issignaling_inline (x) && !issignaling_inline (y))
      return INFINITY;
    return x + y;
  }

This gave a 20% speedup, so it is now only a little slower than before.
We could rename the inline function to __issignaling I suppose, and then
it should be used as long as you include math_config.h.

+  h = math_narrow_eval (h * scale);
+  math_check_force_underflow_nonneg (h);
+  return h;

I don't think you need math_check_force_underflow_nonneg at all given
h * scale should already set the underflow flag correctly. Removing this
gives performance within ~2% of the original.

Wilco