Java Bytecode

Mixins operate on Java bytecode, so to understand them one needs a grasp on their fundamentals.

To find out how to view the bytecode of a class in your IDE, please consult the Viewing Bytecode section of the Tips and Tricks page.

Names and Symbols

Many things in bytecode, like classes, fields, and methods, are still identified by the name (and the descriptor for fields and methods, we'll get to that), just like they are in source code. However, the exact format of those names differs a little bit.

Class Names

Classes are generally referred to by their internal name, which is roughly equivalent to the fully qualified class name (full name including the package), where all dots . are replaced with slashes /. For example, the internal name of the java.lang.Object class is java/lang/Object.

Nested classes use $ symbols to separate their name from the enclosing classes. For example, given:

package pkg;
class Foo {
    class Bar {
    }
}

... the internal name of Bar would be pkg/Foo$Bar.

Anonymous classes use numbers instead of names. For example, if there were two anonymous classes in the Foo class from the code block above, their internal names would be pkg/Foo$1 and pkg/Foo$2 respectively.

Local classes (classes defined within a method) have a number followed by their name. For example, a local class name might look like pkg/Foo$1Local.

Type Descriptors

When bytecode needs to refer to primitive types or arrays, type descriptors are used. Here is a table of data types and their respective type descriptors:

Type	Descriptor
boolean	Z
byte	B
char	C
double	D
float	F
int	I
long	J
short	S
void	V
Arrays	[ + the element type: int[] -> [I
Objects	L + the internal name + ;: String -> Ljava/lang/String;

Field and Method Descriptors

In bytecode, fields and methods are identified by combining their name and descriptor. For fields, that's the descriptor of their data type.

Methods, on the other hand, get theirs by combining the parameter types and the return type. For example, the following method:

void drawText(int x, int y, String text, int color) {
    // ...
}

... has descriptor (IILjava/lang/String;I)V.

The descriptors for the parameter types are directly concatenated together, with no separator. In this case, there is I for int twice (both x and y), then Ljava/lang/String; for String (text), and one more I for the last int (color).

Constructors and Static Initializers

At the bytecode level, constructors are just another method: the detailed differences between the two fall beyond the scope of this overview.

A constructor's method name is <init> (with the <> angled brackets), and the return type in its descriptor is V (void). All non-static field initializations, after compilation, will be found inside the <init> methods.

On the other hand, static initializers (the static {} block in the source code, as well as static field initializers with some exceptions) also are just another method that is run when a class is loaded: its name is <clinit>, and its descriptor is ()V.

Local Variables

In source code, local variables are identified by their name. In bytecode, they are instead identified by a number, or index into the Local Variable Table (LVT). Method parameters are included in the LVT, as is the this object in non-static methods.

Consider the following method as an example:

Source Code

public int getX(int offset) {
    int result = this.x + offset;
    return result;
}

Bytecode

public getX (I)I
  aload 0  // this
  getfield x
  iload 1  // offset
  iadd
  istore 2  // result

  iload 2  // result
  ireturn

In the bytecode, this gets index 0, offset gets index 1, and result gets index 2.

Static methods do not have this in the LVT, so the first parameter of static methods gets index 0 directly.

Longs and doubles take up 2 indexes in the LVT. For example, in the following static method:

Source Code

public static double add(double x, double y, double z) {
    return x + y + z;
}

Bytecode

static add (DDD)D
  dload 0  // x
  dload 2  // y
  dadd
  dload 4  // z
  dadd
  dreturn

... the x parameter gets index 0, the y parameter gets index 2, and the z parameter gets index 4.

INFO

We've seen that bytecode doesn't need the names of local variables because it identifies them by their LVT index. Despite this, many libraries will retain debug information, including the names of local variables, to ease debugging and allow you to target local variables by name when developing mixins.

However, Minecraft 1.21.11 does not provide that by default and is therefore said to be obfuscated. Note that future versions of Minecraft will be deobfuscated.

The Operand Stack

Just like native assembly uses processor registers, Java bytecode uses the Operand Stack to store temporary values.

Like any stack, values are added ("pushed") to the top of the stack, and removed ("popped") from the top of the stack. Think of it like a stack of plates: when you add a plate onto the stack, you put it on top, and when you need one, you take the top one. Such a data structure is said to be Last-In, First-Out, because the last "plate" pushed onto the stack will be popped first.

Let's take a look at the previous getX example again:

Source Code

public int getX(int offset) {
    int result = this.x + offset;
    return result;
}

Bytecode

public getX (I)I
  aload 0  // this
  getfield x
  iload 1  // offset
  iadd
  istore 2  // result

  iload 2  // result
  ireturn

Let's imagine getX(5) is called when this.x has the value 42, and let's follow what happens, instruction by instruction:

Index	Local Variable Table	Operand Stack
2
1	offset: 5
0	this

Navigate through the instructions by clicking on the buttons above.

The diagram above will show the state of the Local Variable Table and the Operand Stack after the instruction.

Notice how LVT slot 0 contains this: that's because getX is not a static method.

Conditional Instructions

We've seen how the JVM follows instructions sequentially, one after the other. However, certain instructions tell the JVM to jump to a different point in the bytecode:

• goto: Always jumps to the referenced instruction
• ifeq: Pops the top value off the Operand Stack, and, if it is equal to 0, jumps to the referenced instruction
• ifne: Pops the top value off the Operand Stack, and, if it is not equal to 0, jumps to the referenced instruction
• if_icmpXX: Pops the top two values of the Operand Stack, and compares them. If the comparison is true, then the JVM jumps to the referenced instruction. For example:
  • if_icmpeq (==): Succeeds if the two values are equal
  • if_icmpgt (>): Succeeds if the first is greater than the second
  • if_icmple (<=): Succeeds if the first is less than or equal to the second

For example, consider the following method:

Source Code

static String makeFoobar(boolean cond) {
    String result;
    if (cond) {
        result = "foo";
    } else {
        result = "bar";
    }
    return result;
}

Bytecode

static makeFoobar (Z)Ljava/lang/String;
  iload 0  // cond
  ifeq L1
  ldc "foo"
  astore 1  // result
  goto L2
L1
  ldc "bar"
  astore 1  // result
L2
  aload 1  // result
  areturn

Notice that jump targets are marked with L*.

The ifeq instruction compares the value on the top of the Operand Stack (which is cond due to the previous iload instruction) to 0, and jumps to L1 if it is equal to 0 (false).

If it is not equal to 0, which means cond is true, it continues through the next instructions until it gets to the goto instruction, which then skips over to L2.

The if block essentially becomes the lines from ifeq L1 to L1, while the else block is L1-L2. The conditional "jump" instructions, reminiscent of goto-era programming, are how if statements, loops, ternaries, and so on, are compiled.

Compilation can end up generating complex logic that is not only hard to read, but also hard to target with mixins. Consider the following classic example:

Source Code

static void doSomething(boolean cond1, boolean cond2) {
    if (cond1) {
        if (cond2) {
            System.out.println("Something is being done");
        }
        // inject here?
    }
}

Bytecode

static doSomething (ZZ)V
  iload 0  // cond1
  ifeq L1
  iload 1  // cond2
  ifeq L1
  getstatic System.out
  invokevirtual println
L1
  return

Because the bytecode for both if conditions jumps to the exact same label, there is no place in the bytecode corresponding to the // inject here? comment, meaning workarounds must be used to target it with mixins.

Common Bytecode Patterns

Here's a reference for the most common bytecode instructions and patterns you'll encounter while developing mixins. For a full advanced list of instructions, check out the List of JVM bytecode instructions on Wikipedia.

Constants

Constant instructions push a constant value onto the Operand Stack.

• iconst_m1, iconst_0, iconst_1, ..., iconst_5: the literal integers -1 through 5
• lconst_0, dconst_1, fconst_2, and the like: the literal numbers, as long, double, and float respectively
• bipush, sipush: pushes a larger integer constant
• ldc: can push several different types of constants, including strings and even larger integers

Variables

Loading instructions read a value from the LVT, and push it onto the Operand Stack.

Storing instructions pop the top value off the Operand Stack, and write it to a local variable.

• iload, istore: loads or stores variables of type int, boolean, byte, char, and short
• lload, lstore: loads or stores variables of type long
• fload, fstore: loads or stores variables of type float
• dload, dstore: loads or stores variables of type double
• aload, astore: loads or stores variables of non-primitive types

Fields

• getfield: reads a non-static field
• putfield: writes to a non-static field
• getstatic: reads a static field
• putstatic: writes to a static field

Method Invocations

• invokestatic: invokes a static method
• invokevirtual: invokes a non-static method. Takes polymorphism and inheritance into account, calling the overridden version where applicable
• invokespecial: invokes a non-static method, exactly the one declared, without taking into account polymorphism/inheritance. Uses include calling constructors and superclass methods
• invokeinterface: invokes a non-static interface method

Conditionals

See Conditional Instructions.

Operators

Operator instructions generally pop two values off the Operand Stack, perform an operation, and push the result. Here is a list of some common operator instructions:

• iadd, ladd, fadd, dadd: addition
• isub, lsub, fsub, dsub: subtraction
• imul, lmul, fmul, dmul: multiplication
• idiv, ldiv, fdiv, ddiv: division
• irem, lrem, frem, drem: modulo
• ineg, lneg, fneg, dneg: negation. Pops only one value off the stack

The prefixes i, l, f, d, as seen with variable instructions, determine the type of data on which to apply the operator.

Returns

Return instructions close the method call, returning the value at the top of the Operand Stack (except for void methods).

When prefixed with i, l, f, d, and a, just like variable instructions, the method returns a value of that type. The instruction for void methods is simply return.

New Object Creation

In source code, writing new MyClass() creates a new instance of MyClass and calls its constructor. In bytecode, these two steps become distinct operations. Take the following code, for example:

Source Code

static Creeper createCreeper(Level level) {
    return new Creeper(level);
}

Bytecode

static createCreeper (Lnet/minecraft/world/level/Level;)Lnet/mineraft/world/entity/monster/Creeper;
  new net/minecraft/world/entity/monster/Creeper
  dup
  aload 0  // level
  invokespecial net/minecraft/world/entity/monster/Creeper.<init> (Lnet/minecraft/world/level/Level;)V
  areturn

Let's examine what happens on the Operand Stack.

Index	Local Variable Table	Operand Stack
2
1
0	level

LVT slot 0 contains level. It does not contain this because the method is static.

Lambdas

Lambda expressions are compiled to a separate method, which is then called by a lambda instance that got instantiated by an invokedynamic instruction.

The details of the invokedynamic instruction are beyond the scope of this overview, but it's useful to know what kind of code to expect. Some invokedynamic operands have been omitted in this section for simplicity.

Here's an example:

Source Code

static void hello() {
    Runnable r = () -> System.out.println("Hello, World!");
    r.run();
}

Bytecode

static hello ()V
  invokedynamic run ()Ljava/lang/Runnable; java/lang/invoke/LambdaMetafactory.metafactory ()V lambda$hello$1 ()V
  astore 0  // r

  aload 0  // r
  invokeinterface run

  return

static lambda$hello$1 ()V
  getstatic System.out
  ldc "Hello, World!"
  invokevirtual println

  return

You can see here that the contents of the lambda have been moved into a separate method, in this case lambda$hello$1.

If you want to target the contents of a lambda with mixins, this is the method you'll want to be targeting. The lambda instance is then created with the invokedynamic instruction and then stored into the variable r.

If the lambda captures any variables, these variables will end up as parameters to the lambda methods. For example:

Source Code

static void hello(String name) {
    Runnable r = () -> System.out.println("Hello, " + name + "!");
    r.run();
}

Bytecode

static hello (Ljava/lang/String;)V
  aload 0  // name
  invokedynamic run (Ljava/lang/String;)Ljava/lang/Runnable; java/lang/invoke/LambdaMetafactory.metafactory ()V lambda$hello$1 (Ljava/lang/String;)V ()V
  astore 1  // r

  aload 1  // r
  invokeinterface run

  return

static lambda$hello$1 (Ljava/lang/String;)V
  getstatic System.out
  aload 0  // name
  invokedynamic makeConcatWithConstants (Ljava/lang/String;)Ljava/lang/String; java/lang/invoke/StringConcatFactory.makeConcatWithConstants "Hello, \1!"
  invokevirtual println

  return

Here, the name parameter is being passed as a parameter into the lambda. Notice also how string concatenation is implemented with invokedynamic.