While look at a some assembly on sparc I noticed that C2 doesn't sometimes generate an extra useless instruction to produce constants of pointer type. In particular the effects the generation the initial mark word 0x1 and the card mark byte_map_base values. For example:

  0xf90b53c4: sethi %hi(0), %i1
  0xf90b53c8: or %i1, 1, %i1 ! 0x00000001
  0xf90b53cc: sethi %hi(0xfd62a000), %i0
  0xf90b53d0: mov %i0, %i0 ! word_map_base

64-bit seems to handle this slightly better for some reason. I think the problem is in sparc.ad in loadConP which forces the emission of the sethi/add pair for all pointers unlike the 64-bit version which checks for an oop first. Look at SetPtr64. This could probably be unified and remove the LP64 test and would end up producing slightly better code. Here's the trivial program I used to get that assembly.

public class p {
    public static void main(String[] args) {
        while (true) {
            for (int i = 0; i < args.length; i++) {
                args[i] = new String(args[i]);
            }
        }
    }
}

Often these values get commoned up as they do in this case but not always so it's worth correcting.
###@###.### 11/3/04 22:48 GMT