Rob Nugent on the javagaming.org forums has posted a small test case illustrating a change in behavior with the intrinsified version of Math.log() on x86 CPUs. Test case:

public class Test {
  public static void main(String[] args) {
    int w = 64;
    int levels = (int)(Math.log(w) / Math.log(2.0)) + 1;
    System.out.println("Width " + w + " gives " +
      levels + " mipmap levels");
  }
}

This produces an answer of 7 on JDK releases 1.5 and earlier and an answer of 6 on the current x86 builds of 1.6 with most configurations (-client -Xcomp seems to work, but -client -Xint and -server -Xint do not). It is not clear whether this change in behavior should be considered a bug due to the specification guarantees of Math.log(), but it should be investigated.

Relevant thread:
http://www.javagaming.org/cgi-bin/JGNetForums/YaBB.cgi?board=jogl;action=display;num=1103796449;start=30#30

  xxxxx@xxxxx   2005-1-04 22:35:29 GMT

N/A

The first log call was setting the fpu stack in 80bit mode, the second log call was being done with extended range which gave incorrect results.  Added a push and pop to revert the fpu stack back to 64bit mode

  xxxxx@xxxxx   2005-1-06 18:23:04 GMT