Currently, approxEqual doesn't take the maximum absolute difference into account when rhs is zero while lhs is nonzero. An example:

  double epsrel = 0.01;  // ...or whatever, it doesn't matter
  double epsabs = 1e-5; // This matters when rhs or lhs is zero!

  assert (approxEqual(0.0, 1e-10, epsrel, epsabs));  // OK
  assert (approxEqual(1e-10, 0.0, epsrel, epsabs));  // Fails!

This is very unintuitive -- I think the order of the "operands" shouldn't matter here. The offending piece of code is at line 3087 of std.math (rev. 1233):

    if (rhs == 0) {
        return (lhs == 0 ? 0 : 1) <= maxRelDiff;
    }

This could be changed to:

    if (rhs == 0) {
        return (lhs == 0 ? 0 : 1) <= maxRelDiff
            || (maxAbsDiff != 0 && fabs(rhs-lhs) <= maxAbsDiff);
    }

Another option, if abs(lhs-rhs)/rhs and abs(lhs-rhs)/lhs could be considered equally good definitions of the relative difference, would be this:

    if (rhs == 0) {
        if (lhs == 0) return true;
        // Switch lhs and rhs
        return approxEqual(rhs, lhs, maxRelDiff, maxAbsDiff);
    }

I actually prefer this one, because the name "approxEqual" doesn't in any way imply that the order of its arguments matter. It should simply return true if lhs and rhs are approximately equal, regardless of which is which.