Approved. These changes are a reasonable solution and, as you say, we can improve it further in future releases.
Mike On Dec 22 2013, at 06:54 , Joe Darcy <joe.da...@oracle.com> wrote: > Hello, > > Testing (eventually) revealed that changing the streams floating point sum > and average algorithms to use compensated summation (JDK-8006572 > DoubleStream.sum() & DoubleSummaryStats implementations that reduce numerical > errors), changed the behavior of the code on streams with infinite values: > NaN was returned instead of infinity: > > 8030212: Several api.java.util.stream tests got "NaN" value instead of > "Infinity" or "-Infinity" > > The specification doesn't explicitly state how non-finite values in streams > should be handled (and it should be updated to do so in 9, JDK-8030942 > Explicitly state floating-point summation requirements on non-finite inputs), > but it isn't unreasonable to assume that a properly signed infinity should > result. > > Towards that end, I've prepared a webrev that uses an additional simple sum > to distinguish a spurious NaN sum in the result: > > http://cr.openjdk.java.net/~darcy/8030212.1/ > > This may not be the optimal way to track this situation, additional logic in > the compensated summation code is possible, but if needed the implementation > can be refined in JDK 9 and the 8 updates. > > Thanks, > > -Joe
