Unexpected results with doubles

[Joseph Malle]

I am learning D.  I was working on Project Euler #199 (possible 
spoilers) and got some unexpected results.  It's probably a 
stupid mistake but I can't see it.

Here is an edited function from my program:

auto radius(const double r1, const double r2, const double r3) {
   auto const k1 = 1/r1;
   auto const k2 = 1/r2;
   auto const k3 = 1/r3;
   writeln();
   writeln("1   ", [k1, k2, k3]);
   writeln("2   ", [k1 * k2, k2 * k3, k3 * k1]);
   writeln("3   ", [k1 * k2 + k2 * k3 + k3 * k1]);
   assert(!isNaN(k1 * k2 + k2 * k3 + k3 * k1));
   writeln("4   ", [sqrt(k1 * k2 + k2 * k3 + k3 * k1)]);
   assert(!isNaN(sqrt(k1 * k2 + k2 * k3 + k3 * k1)));
   auto rv = 1 / (k1 + k2 + k3 + 2.0 * sqrt(k1 * k2 + k2 * k3 + k3 
* k1));
   assert(!isNaN(rv));
   writeln("radius ", [r1, r2, r3], " => ", rv);
   writeln();
   return rv;
}

Here is some output:

1   [41.7846, 6.4641, 6.4641]
2   [270.1, 41.7846, 270.1]
3   [581.985]
4   [24.1244]
radius [0.0239323, 0.154701, 0.154701] => 0.00971237


1   [41.7846, 6.4641, 6.4641]
2   [270.1, 41.7846, 270.1]
3   [581.985]
4   [nan]
radius [0.0239323, 0.154701, 0.154701] => 0.00971237


1   [41.7846, 6.4641, 6.4641]
2   [270.1, 41.7846, 270.1]
3   [581.985]
4   [nan]
radius [0.0239323, 0.154701, 0.154701] => 0.00971237

The "4   [nan]" is unexpected.  Each time they have the same 
input/same output.  But sometimes the 4th line is nan and 
sometimes it's not.  The asserts never fail.  I've seen this 
unexpected nan a few times with other inputs for this function.


If I change it to:

auto x = sqrt(k1 * k2 + k2 * k3 + k3 * k1);
writeln("4   ", [x]);
assert(!isNaN(x));

Then the assert fails.  I checked if the assert fails before the 
writeln too (as a sanity check) and yes, x is always NaN it seems.

I am doing $dmd -run on the command line.  Working on a 
reasonably up to date Mac.

$ dmd --version
DMD64 D Compiler v2.083.1
Copyright (C) 1999-2018 by The D Language Foundation, All Rights 
Reserved written by Walter Bright