ASM: visitLabel generates too many labels and nop instructions

Question

ASM documentation says a label represent a basic block, and it is a node in the control graph. So I test the visitLabel method on this simple example:

public static void main(String[] args) {
    int x = 3, y = 4;
    if (x < y) {
        x++;
    }
}

For the visitLabel method, I instrument it with a native API: setID(int id), where the id is incremental. In this example, a CFG should have 3 nodes: one at the beginning, and one for each branch of the if statement. So I expect setID would be called in 3 locations. However, it is called 5 times, and there are a lot of nop instructions. Could anybody explain for me why?

Here is the instrumented bytecode for the above program.

  public static void main(java.lang.String[]);
    Code:
       0: iconst_2
       1: invokestatic  #13                 // Method setId:(I)V
       4: iconst_3
       5: istore_1
       6: iconst_3
       7: invokestatic  #13                 // Method setId:(I)V
      10: iconst_4
      11: istore_2
      12: iconst_4
      13: invokestatic  #13                 // Method setId:(I)V
      16: iload_1
      17: iload_2
      18: if_icmpge     28
      21: iconst_5
      22: invokestatic  #13                 // Method setId:(I)V
      25: iinc          1, 1
      28: bipush        6
      30: invokestatic  #13                 // Method setId:(I)V
      33: return
      34: nop
      35: nop
      36: nop
      37: nop
      38: athrow

What I don't understand is why there is a label before each istore instruction. There is no branching to make it a new node in the CFG.

score 1 · Accepted Answer · answered Nov 12 '18 at 11:56

The primary purpose of a Label is to denote a position in the bytecode sequence. Since this is needed for branch targets, you can use them to identify basic blocks. But you have to be aware, that they are also used for reporting line numbers, when a LineNumberTable attribute is present and for reporting local variable scopes when a LocalVariableTable attribute is present as well as, for newer class files, their type annotations recorded in a RuntimeVisibleTypeAnnotations attribute. Further, labels may mark the protected area of an exception handler. For code generated from Java source code, this protected area matches the try block, so its a basic block, but that doesn’t need to hold for other bytecode.

See

Since the scope of local variables may span the last return instruction, it is possible to encounter labels after that last instruction, which is what happens in your case. You are injecting a bipush 7, invokestatic #13 after the return instruction, resulting in unreachable code.

Apparently, you are also using the COMPUTE_FRAMES options to let ASM recalculate stack map frames from scratch, but it is impossible to calculate frames for unreachable code, due to the unknown initial stack state. ASM solves this problem by replacing unreachable code with nop instructions, followed by a single athrow statement. For this sequence, it’s possible to specify a valid initial stack frame and it has no impact on the execution (as the code is unreachable).

As you can see, four nop instructions plus one athrow instruction span five bytes, which is the same size as the replaced bipush 7, invokestatic #13 sequence had.

You can get rid of most of these reported labels, by specifying ClassReader.SKIP_DEBUG to its accept method. Then, you get only one reported label for your example, the branch target associated with the if statement. But you have to handle the visitJumpInsn to identify the start of the conditional code.

So, to identify all basic blocks, you have to handle all branch instructions, i.e. via visitJumpInsn, visitLookupSwitchInsn, and visitTableSwitchInsn, as well as all end instructions, i.e. athrow and all variants of return. Further, you need to process all visitTryCatchBlock calls. If you need to identify potential targets of branch instructions in a single pass, I’d use visitFrame instead of labels, as frames are mandatory at all branch targets for class file version of 51 (Java 7) or higher.

By the way, when all you’re injecting are these sequences of loading a constant and calling a static method (at reachable locations), I’d use COMPUTE_MAXS instead of COMPUTE_FRAMES, as an expensive recalculation is not necessary when the general code structure does not change.

thanks for the awesome answer, I understand the problem now. Using `SKIP_DEBUG` does remove unnecessary labels. However, I want to instrument each basic block with an ID, and how do I instrument returns instructions and ATHROW. I tried to override `visitInsn`, and instrumented before the visiting the return instructions, but doing this just adds many `nop` instructions like before, and no invocation method calls are added. — sean, Nov 12 '18 at 19:22
That sounds like a new question. You should include the code doing the transformation in the new question then. Otherwise, it’s really hard to tell. — Holger, Nov 13 '18 at 07:27

ASM: visitLabel generates too many labels and nop instructions

1 Answers1