The following two gvn.transform calls seem to be redundant.
It would be better to remove them.

```
  case Bytecodes::_l2f:
    if (Matcher::convL2FSupported()) {
      a = pop_pair();
      b = _gvn.transform( new ConvL2FNode(a));
      // For x86_32.ad, FILD doesn't restrict precision to 24 or 53 bits.
      // Rather than storing the result into an FP register then pushing
      // out to memory to round, the machine instruction that implements
      // ConvL2D is responsible for rounding.
      // c = precision_rounding(b);
      c = _gvn.transform(b); <-------- redundant
      push(c);
    } else {
      l2f();
    }
    break;

  case Bytecodes::_l2d:
    a = pop_pair();
    b = _gvn.transform( new ConvL2DNode(a));
    // For x86_32.ad, rounding is always necessary (see _l2f above).
    // c = dprecision_rounding(b);
    c = _gvn.transform(b); <------- redundant
    push_pair(c);
    break;
```