RE: Patches to java.util.zip by Christian Schlichtherle

[Christian Schlichtherle]

Hi everyone,

> Yes, if someone can make a little testcase where we fail now 
> and show how we are not properly converting the 
> signed/unsigned ints at the moment that would be appreciated. 
> Note that at the moment all public methods take and return 
> longs already, we only store it as int internally and try 
> treat it as unsigned already. It might be that we need to 
> internally store everything in an long and not just treat the 
> signed ints as unsigned ints as we do now in the code, but 
> the patch contained so many unrelated changes that it was 
> hard to make out which changes were explicitly for this 
> issue. So if someone could extract those parts of the patch 
> that would help a lot.

More specifically, the size and compressed size field in the ZipEntry class
are causing the problems as some comparisons are happening on these. The
issue is that once a big integer equal or greater than 2*1024^3 and smaller
than 4*1024^3 is stored into a Java int, it is hard to use this int as if it
were unsigned. You would have to do something like this on an int
(untested):

final static long NUM_FOUR_BYTE_INTS = 256L * 256L * 256L * 256L; // Number
of 4 byte ints

private static long int2long(int i) {
        return i >= 0 ? (long)i : (long)NUM_FOUR_BYTE_INTS - (long)i;
}

private static void doIt() {
        int ok = 2 * 1024 * 1024 * 1024 - 1; // max signed int
        int tooBig = 3 * 1024 * 1024 * 1024; // Creates a negative integer

        // Compare ok and tooBig
        assert int2long(ok) < int2long(tooBig) : "This is ok!";
        assert ok < tooBig : "This fails!";
}

This is way too much computational overhead and it is way too easy to forget
this on a comparison. Thus, the better way to deal with this issue is hold 4
byte signed ints in a Java long if you don't need to care for the memory
overhead.

With best regards,
Christian

smime.p7s
Description: S/MIME cryptographic signature