JDK 9 RFR of JDK-4851642: Add fused mac to Java math library

Dmitry Nadezhin dmitry.nadezhin at gmail.com
Wed Apr 13 12:41:58 UTC 2016


Here are my comments.

1) Probably in lines StrictMath:1171, StrictMath:1220, Math:1487, Math:1593
you meant
 {@code -0.0 * +0.0}
 {@code -0.0f * +0.0f}

2) Lines Math:1508-1525 could be simpler:
    double result = a * b + c;
    return Double.isNaN(result) && Double.isFinite(a) && Double.isFinite(b)
? c : result;
    JVM may use either double or double-extended-exponent value set for a *
    Values stored in the result variable might be different, but the value
returned by fusedMac(double,double.double) method will be the same.
    For example a = 0x1p1023, b = 0x1p1023, c = Double.NEGATIVE_INFINITY;
    a * b  is either Double.POSITIVE_INFINITY or 0x1p2046
    a * b + c is either Double.NaN or Double.NEGATIVE_INFINITY
    result is either Double.NaN or Double.NEGATIVE_INFINITY
    return value is Double.NEGATIVE_INFINITY in both cases.

3) As you say, this implementation favors code simplicity over speed.
    There are five special cases in lines Math:1526-1530 .
    Three of them (a==0.0, b==0.0, c==0.0) are essential because BigDecimal
can't handle signed zeroes.
    The other two (a==1.0, b==1.0) are optimizations. BigDecimal handles
them correctly.
    I think that the code would be simpler without the last two cases.
    Otherwise, one may ask why similar special cases (a==-1.0, b==-1.0) are
not considered as well.

Best Regards.

More information about the core-libs-dev mailing list