RVM-371 and r13964 improved the boot image size but regressed performance across the board. This degradation is most marked in the single iteration fop performance where a number of longs in the classes cause a notable delay when finding or creating their JTOC literals. One solution is to reintroduce a number of hopefully smaller hash tables. Another idea is that we could handle longs in a different manner. In particular we could take advantage of the empty constant pool slot following all double/long constant pool entries. We could use this to have a long/double stored in two halves in the JTOC.