difficult to implement efficiently in XLA. The implementation imposes
a total ordering on floats. The ordering is consistent with the usual
partial order.  Positive NaNs are greater than positive
infinity. Negative NaNs are less than negative infinity. NaNs with
distinct payloads are treated as distinct. Subnormal numbers are
preserved (not flushed to zero). Positive infinity is greater than all

		for _, arg := range n.Args[1:] {
			x = op(x, s.expr(arg))
		}
		return x
	}

	typ := n.Type()

	if typ.IsFloat() || typ.IsString() {
		// min/max semantics for floats are tricky because of NaNs and
		// negative zero. Some architectures have instructions which
		// we can use to generate the right result. For others we must
		// call into the runtime instead.
		//

var s []string
_ = min(s...)               // invalid: slice arguments are not permitted
t := max("", "foo", "bar")  // t == "foo" (string kind)
</pre>

<p>
For numeric arguments, assuming all NaNs are equal, <code>min</code> and <code>max</code> are
commutative and associative:
</p>

<pre>
min(x, y)    == min(y, x)
min(x, y, z) == min(min(x, y), z) == min(x, min(y, z))
</pre>

<p>