A DESCRIPTION OF THE FIX :
  Bug Description : javac does not use iinc opcode for postfix decrement of local variable.
Generally, javac produces code that evaluates postfix decrement correctly. But looking to the sources I found some asymmetry and I think it was not made intentionally.
Subset of Releases affected : 1.4.2, 1.5.0, Mustang b65.
Platforms affected : all
Test case : See Test case section.
Diff baseline : Mustang b65
Diff :
--- j2se/src/share/classes/com/sun/tools/javac/jvm/Gen.java.prev Tue Dec 27 20:43:22 2005
+++ j2se/src/share/classes/com/sun/tools/javac/jvm/Gen.java Fri Jan 06 21:42:42 2006
@@ -1783,41 +1783,42 @@
  case JCTree.PREINC: case JCTree.PREDEC:
  od.duplicate();
  if (od instanceof LocalItem &&
  (operator.opcode == iadd || operator.opcode == isub)) {
  ((LocalItem)od).incr(tree.tag == JCTree.PREINC ? 1 : -1);
  result = od;
  } else {
  od.load();
  code.emitop0(one(od.typecode));
  code.emitop0(operator.opcode);
  // Perform narrowing primitive conversion if byte,
  // char, or short. Fix for 4304655.
  if (od.typecode != INTcode &&
  Code.truncate(od.typecode) == INTcode)
  code.emitop0(int2byte + od.typecode - BYTEcode);
  result = items.makeAssignItem(od);
  }
  break;
  case JCTree.POSTINC: case JCTree.POSTDEC:
  od.duplicate();
- if (od instanceof LocalItem && operator.opcode == iadd) {
+ if (od instanceof LocalItem &&
+ (operator.opcode == iadd || operator.opcode == isub)) {
  Item res = od.load();
  ((LocalItem)od).incr(tree.tag == JCTree.POSTINC ? 1 : -1);
  result = res;
  } else {
  Item res = od.load();
  od.stash(od.typecode);
  code.emitop0(one(od.typecode));
  code.emitop0(operator.opcode);
  // Perform narrowing primitive conversion if byte,
  // char, or short. Fix for 4304655.
  if (od.typecode != INTcode &&
  Code.truncate(od.typecode) == INTcode)
  code.emitop0(int2byte + od.typecode - BYTEcode);
  od.store();
  result = res;
  }
  break;
  case JCTree.NULLCHK:
  result = od.load();
  code.emitop0(dup);


JUnit TESTCASE :
This is not JUnit style test. This is just a proof that patch focuses on the problem.
Here is sample class:
// Main.java
public class Main {
private static boolean test() {
int x = 10;
int y = x--;
// check whether generated code still works correctly
return (x == 9) && (y == 10);
}

public static void main(String[] args) {
boolean ok = test();
System.out.println("ok = " + ok);
System.exit(ok? 0: 1);
}
}
And generated code for method "test()" differs in such a way:

 private static boolean test();
   Code:
- Stack=3, Locals=2, Args_size=0
+ Stack=2, Locals=2, Args_size=0
         bipush 10
         istore_0
         iload_0
- dup
- iconst_1
- isub
- istore_0
+ iinc 0, -1
         istore_1
         iload_0
         bipush 9
         if_icmpne @else
         iload_1
         bipush 10
         if_icmpne @else
         iconst_1
         goto @end
    @else:
         iconst_0
    @end:
         ireturn