Yesterday I encountered the first OpenJDK "bug". Sun uses IBM's ICU library for Unicode and Globalization support and it does a couple of DataInputStream.read(byte) calls without checking the return value. On IKVM the InputStream that ClassLoader.getSystemResourceAsStream() returns is an ikvm.io.InputStreamWrapper and that always returns 0 from available() when it is wrapping a non-seekable stream and the LZInputStream that decompresses compressed resources is a non-seekable stream. The reason this is significant is that ICU wraps a BufferedInputStream around the InputStreamWrapper and BufferedInputStream uses available() to figure out if it should read more data from the underlying stream if there's still room in the byte array passed in to read(byte) after the data from the buffer has been copied. So as long as the underlying stream has an implementation of available() that never returns 0, ICU works fine. I filed a bug report.
Issues like this demonstrate why cloning a large platform is so incredibly difficult. Real world code is bound to have dependencies on implementation details like this. So simply reimplementing a spec isn't sufficient.
For the time being I've added a workaround to map.xml (using the new <replace-method-call /> support) to modify the ICU method to call readFully(byte) instead of read(byte), but I'm considering rewriting the resource decompression to have available() return the number of bytes still available to read from the stream. There may be other code out there that depends on it.