Unrolling more when SLP auto-vectorization failed

For below Java code, which can't be auto-vectorized:
  public static void accessArrayConstants(int[] array) {
      for (int j = 0; j < 1024; j++) {
          array[0]++;
          array[1]++;
      }
  }
Current C2 unrolls the loop to 8 times on X86 and 4 times on AArch64, which is limited by architecture's vector size. Get better performance, we should unroll more until reaching the unroll limitation.