Chapter 3. Compiling for the Java Virtual Machine

This chapter consists mainly of examples of source code together with annotated listings of the Java Virtual Machine code that the javac compiler in Oracle’s JDK release 1.0.2 generates for the examples. The Java Virtual Machine code is written in the informal “virtual machine assembly language” output by Oracle's javap utility, distributed with the JDK release. You can use javap to generate additional examples of compiled methods.

The format of the examples should be familiar to anyone who has read assembly code. Each instruction takes the form:

<index> <opcode> [ <operand1> [ <operand2>... ]] [<comment>]

The <index> is the index of the opcode of the instruction in the array that contains the bytes of Java Virtual Machine code for this method. Alternatively, the <index> may be thought of as a byte offset from the beginning of the method. The <opcode> is the mnemonic for the instruction's opcode, and the zero or more <operandN> are the operands of the instruction. The optional <comment> is given in end-of-line comment syntax:

8   bipush 100     // Push int constant 100

Some of the material in the comments is emitted by javap; the rest is supplied by the authors. The <index> prefacing each instruction may be used as the target of a control transfer instruction. For instance, a goto 8 instruction transfers control to the instruction at index 8. Note that the actual operands of Java Virtual Machine control transfer instructions are offsets from the addresses of the opcodes of those instructions; these operands are displayed by javap (and are shown in this chapter) as more easily read offsets into their methods.

We preface an operand representing a run-time constant pool index with a hash sign and follow the instruction by a comment identifying the run-time constant pool item referenced, as in:

10  ldc #1         // Push float constant 100.0

or:

9   invokevirtual #4    // Method Example.addTwo(II)I

For the purposes of this chapter, we do not worry about specifying details such as operand sizes.

Java Virtual Machine code exhibits a set of general characteristics imposed by the Java Virtual Machine's design and use of types. In the first example we encounter many of these, and we consider them in some detail.

The spin method simply spins around an empty for loop 100 times:

void spin() {
    int i;
    for (i = 0; i < 100; i++) {
        ;    // Loop body is empty
    }
}

A compiler might compile spin to:

0   iconst_0       // Push int constant 0
1   istore_1       // Store into local variable 1 (i=0)
2   goto 8         // First time through don't increment
5   iinc 1 1       // Increment local variable 1 by 1 (i++)
8   iload_1        // Push local variable 1 (i)
9   bipush 100     // Push int constant 100
11  if_icmplt 5    // Compare and loop if less than (i < 100)
14  return         // Return void when done

The Java Virtual Machine is stack-oriented, with most operations taking one or more operands from the operand stack of the Java Virtual Machine's current frame or pushing results back onto the operand stack. A new frame is created each time a method is invoked, and with it is created a new operand stack and set of local variables for use by that method (§2.6). At any one point of the computation, there are thus likely to be many frames and equally many operand stacks per thread of control, corresponding to many nested method invocations. Only the operand stack in the current frame is active.

The instruction set of the Java Virtual Machine distinguishes operand types by using distinct bytecodes for operations on its various data types. The method spin operates only on values of type int. The instructions in its compiled code chosen to operate on typed data (iconst_0, istore_1, iinc, iload_1, if_icmplt) are all specialized for type int.

The two constants in spin, 0 and 100, are pushed onto the operand stack using two different instructions. The 0 is pushed using an iconst_0 instruction, one of the family of iconst_<i> instructions. The 100 is pushed using a bipush instruction, which fetches the value it pushes as an immediate operand.

The Java Virtual Machine frequently takes advantage of the likelihood of certain operands (int constants -1, 0, 1, 2, 3, 4 and 5 in the case of the iconst_<i> instructions) by making those operands implicit in the opcode. Because the iconst_0 instruction knows it is going to push an int 0, iconst_0 does not need to store an operand to tell it what value to push, nor does it need to fetch or decode an operand. Compiling the push of 0 as bipush 0 would have been correct, but would have made the compiled code for spin one byte longer. A simple virtual machine would have also spent additional time fetching and decoding the explicit operand each time around the loop. Use of implicit operands makes compiled code more compact and efficient.

The int i in spin is stored as Java Virtual Machine local variable 1. Because most Java Virtual Machine instructions operate on values popped from the operand stack rather than directly on local variables, instructions that transfer values between local variables and the operand stack are common in code compiled for the Java Virtual Machine. These operations also have special support in the instruction set. In spin, values are transferred to and from local variables using the istore_1 and iload_1 instructions, each of which implicitly operates on local variable 1. The istore_1 instruction pops an int from the operand stack and stores it in local variable 1. The iload_1 instruction pushes the value in local variable 1 on to the operand stack.

The use (and reuse) of local variables is the responsibility of the compiler writer. The specialized load and store instructions should encourage the compiler writer to reuse local variables as much as is feasible. The resulting code is faster, more compact, and uses less space in the frame.

Certain very frequent operations on local variables are catered to specially by the Java Virtual Machine. The iinc instruction increments the contents of a local variable by a one-byte signed value. The iinc instruction in spin increments the first local variable (its first operand) by 1 (its second operand). The iinc instruction is very handy when implementing looping constructs.

The for loop of spin is accomplished mainly by these instructions:

5   iinc 1 1       // Increment local variable 1 by 1 (i++)
8   iload_1        // Push local variable 1 (i)
9   bipush 100     // Push int constant 100
11  if_icmplt 5    // Compare and loop if less than (i < 100)

The bipush instruction pushes the value 100 onto the operand stack as an int, then the if_icmplt instruction pops that value off the operand stack and compares it against i. If the comparison succeeds (the variable i is less than 100), control is transferred to index 5 and the next iteration of the for loop begins. Otherwise, control passes to the instruction following the if_icmplt.

If the spin example had used a data type other than int for the loop counter, the compiled code would necessarily change to reflect the different data type. For instance, if instead of an int the spin example uses a double, as shown:

void dspin() {
    double i;
    for (i = 0.0; i < 100.0; i++) {
        ;    // Loop body is empty
    }
}

the compiled code is:

Method void dspin()
0   dconst_0       // Push double constant 0.0
1   dstore_1       // Store into local variables 1 and 2
2   goto 9         // First time through don't increment
5   dload_1        // Push local variables 1 and 2 
6   dconst_1       // Push double constant 1.0 
7   dadd           // Add; there is no dinc instruction
8   dstore_1       // Store result in local variables 1 and 2
9   dload_1        // Push local variables 1 and 2 
10  ldc2_w #4      // Push double constant 100.0 
13  dcmpg          // There is no if_dcmplt instruction
14  iflt 5         // Compare and loop if less than (i < 100.0)
17  return         // Return void when done

The instructions that operate on typed data are now specialized for type double. (The ldc2_w instruction will be discussed later in this chapter.)

Recall that double values occupy two local variables, although they are only accessed using the lesser index of the two local variables. This is also the case for values of type long. Again for example,

double doubleLocals(double d1, double d2) {
    return d1 + d2;
}

becomes

Method double doubleLocals(double,double)
0   dload_1       // First argument in local variables 1 and 2
1   dload_3       // Second argument in local variables 3 and 4
2   dadd
3   dreturn

Note that local variables of the local variable pairs used to store double values in doubleLocals must never be manipulated individually.

The Java Virtual Machine's opcode size of 1 byte results in its compiled code being very compact. However, 1-byte opcodes also mean that the Java Virtual Machine instruction set must stay small. As a compromise, the Java Virtual Machine does not provide equal support for all data types: it is not completely orthogonal (Table 2.11.1-A).

For example, the comparison of values of type int in the for statement of example spin can be implemented using a single if_icmplt instruction; however, there is no single instruction in the Java Virtual Machine instruction set that performs a conditional branch on values of type double. Thus, dspin must implement its comparison of values of type double using a dcmpg instruction followed by an iflt instruction.

The Java Virtual Machine provides the most direct support for data of type int. This is partly in anticipation of efficient implementations of the Java Virtual Machine's operand stacks and local variable arrays. It is also motivated by the frequency of int data in typical programs. Other integral types have less direct support. There are no byte, char, or short versions of the store, load, or add instructions, for instance. Here is the spin example written using a short:

void sspin() {
    short i;
    for (i = 0; i < 100; i++) {
        ;    // Loop body is empty
    }
}

It must be compiled for the Java Virtual Machine, as follows, using instructions operating on another type, most likely int, converting between short and int values as necessary to ensure that the results of operations on short data stay within the appropriate range:

Method void sspin()
0   iconst_0
1   istore_1
2   goto 10
5   iload_1        // The short is treated as though an int
6   iconst_1
7   iadd
8   i2s            // Truncate int to short
9   istore_1
10  iload_1
11  bipush 100
13  if_icmplt 5
16  return

The lack of direct support for byte, char, and short types in the Java Virtual Machine is not particularly painful, because values of those types are internally promoted to int (byte and short are sign-extended to int, char is zero-extended). Operations on byte, char, and short data can thus be done using int instructions. The only additional cost is that of truncating the values of int operations to valid ranges.

The long and floating-point types have an intermediate level of support in the Java Virtual Machine, lacking only the full complement of conditional control transfer instructions.

The Java Virtual Machine generally does arithmetic on its operand stack. (The exception is the iinc instruction, which directly increments the value of a local variable.) For instance, the align2grain method aligns an int value to a given power of 2:

int align2grain(int i, int grain) {
    return ((i + grain-1) & ~(grain-1));
}

Operands for arithmetic operations are popped from the operand stack, and the results of operations are pushed back onto the operand stack. Results of arithmetic subcomputations can thus be made available as operands of their nesting computation. For instance, the calculation of ~(grain-1) is handled by these instructions:

5   iload_2        // Push grain
6   iconst_1       // Push int constant 1
7   isub           // Subtract; push result
8   iconst_m1      // Push int constant -1
9   ixor           // Do XOR; push result

First grain-1 is calculated using the contents of local variable 2 and an immediate int value 1. These operands are popped from the operand stack and their difference pushed back onto the operand stack. The difference is thus immediately available for use as one operand of the ixor instruction. (Recall that ~x == -1^x.) Similarly, the result of the ixor instruction becomes an operand for the subsequent iand instruction.

The code for the entire method follows:

Method int align2grain(int,int)
0   iload_1
1   iload_2
2   iadd
3   iconst_1
4   isub
5   iload_2
6   iconst_1
7   isub
8   iconst_m1
9   ixor
10  iand
11  ireturn

Many numeric constants, as well as objects, fields, and methods, are accessed via the run-time constant pool of the current class. Object access is considered later (§3.8). Data of types int, long, float, and double, as well as references to instances of class String, are managed using the ldc, ldc_w, and ldc2_w instructions.

The ldc and ldc_w instructions are used to access values in the run-time constant pool (including instances of class String) of types other than double and long. The ldc_w instruction is used in place of ldc only when there is a large number of run-time constant pool items and a larger index is needed to access an item. The ldc2_w instruction is used to access all values of types double and long; there is no non-wide variant.

Integral constants of types byte, char, or short, as well as small int values, may be compiled using the bipush, sipush, or iconst_<i> instructions (§3.2). Certain small floating-point constants may be compiled using the fconst_<f> and dconst_<d> instructions.

In all of these cases, compilation is straightforward. For instance, the constants for:

void useManyNumeric() {
    int i = 100;
    int j = 1000000;
    long l1 = 1;
    long l2 = 0xffffffff;
    double d = 2.2;
    ...do some calculations...
}

are set up as follows:

Method void useManyNumeric()
0   bipush 100   // Push small int constant with bipush
2   istore_1
3   ldc #1       // Push large int constant (1000000) with ldc
5   istore_2
6   lconst_1     // A tiny long value uses small fast lconst_1
7   lstore_3
8   ldc2_w #6    // Push long 0xffffffff (that is, an int -1)
        // Any long constant value can be pushed with ldc2_w
11  lstore 5
13  ldc2_w #8    // Push double constant 2.200000
        // Uncommon double values are also pushed with ldc2_w
16  dstore 7
...do those calculations...

Compilation of for statements was shown in an earlier section (§3.2). Most of the Java programming language's other control constructs (if-then-else, do, while, break, and continue) are also compiled in the obvious ways. The compilation of switch statements is handled in a separate section (§3.10), as are the compilation of exceptions (§3.12) and the compilation of finally clauses (§3.13).

As a further example, a while loop is compiled in an obvious way, although the specific control transfer instructions made available by the Java Virtual Machine vary by data type. As usual, there is more support for data of type int, for example:

void whileInt() {
    int i = 0;
    while (i < 100) {
        i++;
    }
}

is compiled to:

Method void whileInt()
0   iconst_0
1   istore_1
2   goto 8
5   iinc 1 1
8   iload_1
9   bipush 100
11  if_icmplt 5
14  return

Note that the test of the while statement (implemented using the if_icmplt instruction) is at the bottom of the Java Virtual Machine code for the loop. (This was also the case in the spin examples earlier.) The test being at the bottom of the loop forces the use of a goto instruction to get to the test prior to the first iteration of the loop. If that test fails, and the loop body is never entered, this extra instruction is wasted. However, while loops are typically used when their body is expected to be run, often for many iterations. For subsequent iterations, putting the test at the bottom of the loop saves a Java Virtual Machine instruction each time around the loop: if the test were at the top of the loop, the loop body would need a trailing goto instruction to get back to the top.

Control constructs involving other data types are compiled in similar ways, but must use the instructions available for those data types. This leads to somewhat less efficient code because more Java Virtual Machine instructions are needed, for example:

void whileDouble() {
    double i = 0.0;
    while (i < 100.1) {
        i++;
    }
}

is compiled to:

Method void whileDouble()
0   dconst_0
1   dstore_1
2   goto 9
5   dload_1
6   dconst_1
7   dadd
8   dstore_1
9   dload_1
10  ldc2_w #4      // Push double constant 100.1
13  dcmpg          // To compare and branch we have to use...
14  iflt 5         // ...two instructions
17  return

Each floating-point type has two comparison instructions: fcmpl and fcmpg for type float, and dcmpl and dcmpg for type double. The variants differ only in their treatment of NaN. NaN is unordered (§2.3.2), so all floating-point comparisons fail if either of their operands is NaN. The compiler chooses the variant of the comparison instruction for the appropriate type that produces the same result whether the comparison fails on non-NaN values or encounters a NaN. For instance:

int lessThan100(double d) {
    if (d < 100.0) {
        return 1;				
    } else {
        return -1;				
    }
}

compiles to:

Method int lessThan100(double)
0   dload_1
1   ldc2_w #4      // Push double constant 100.0
4   dcmpg          // Push 1 if d is NaN or d > 100.0;
                   // push 0 if d == 100.0
5   ifge 10        // Branch on 0 or 1
8   iconst_1
9   ireturn
10  iconst_m1
11  ireturn

If d is not NaN and is less than 100.0, the dcmpg instruction pushes an int -1 onto the operand stack, and the ifge instruction does not branch. Whether d is greater than 100.0 or is NaN, the dcmpg instruction pushes an int 1 onto the operand stack, and the ifge branches. If d is equal to 100.0, the dcmpg instruction pushes an int 0 onto the operand stack, and the ifge branches.

The dcmpl instruction achieves the same effect if the comparison is reversed:

int greaterThan100(double d) {
    if (d > 100.0) {
        return 1;			
    } else {
        return -1;			
    }
}

becomes:

Method int greaterThan100(double)
0   dload_1
1   ldc2_w #4      // Push double constant 100.0
4   dcmpl          // Push -1 if d is NaN or d < 100.0;
                   // push 0 if d == 100.0
5   ifle 10        // Branch on 0 or -1
8   iconst_1
9   ireturn
10  iconst_m1
11  ireturn

Once again, whether the comparison fails on a non-NaN value or because it is passed a NaN, the dcmpl instruction pushes an int value onto the operand stack that causes the ifle to branch. If both of the dcmp instructions did not exist, one of the example methods would have had to do more work to detect NaN.

If n arguments are passed to an instance method, they are received, by convention, in the local variables numbered 1 through n of the frame created for the new method invocation. The arguments are received in the order they were passed. For example:

int addTwo(int i, int j) {
    return i + j;
}

compiles to:

Method int addTwo(int,int)
0   iload_1        // Push value of local variable 1 (i)
1   iload_2        // Push value of local variable 2 (j)
2   iadd           // Add; leave int result on operand stack
3   ireturn        // Return int result

By convention, an instance method is passed a reference to its instance in local variable 0. In the Java programming language the instance is accessible via the this keyword.

Class (static) methods do not have an instance, so for them this use of local variable 0 is unnecessary. A class method starts using local variables at index 0. If the addTwo method were a class method, its arguments would be passed in a similar way to the first version:

static int addTwoStatic(int i, int j) {
    return i + j;
}

compiles to:

Method int addTwoStatic(int,int)
0   iload_0
1   iload_1
2   iadd
3   ireturn

The only difference is that the method arguments appear starting in local variable 0 rather than 1.

The normal method invocation for a instance method dispatches on the run-time type of the object. (They are virtual, in C++ terms.) Such an invocation is implemented using the invokevirtual instruction, which takes as its argument an index to a run-time constant pool entry giving the internal form of the binary name of the class type of the object, the name of the method to invoke, and that method's descriptor (§4.3.3). To invoke the addTwo method, defined earlier as an instance method, we might write:

int add12and13() {
    return addTwo(12, 13);
}

This compiles to:

Method int add12and13()
0   aload_0             // Push local variable 0 (this)
1   bipush 12           // Push int constant 12
3   bipush 13           // Push int constant 13
5   invokevirtual #4    // Method Example.addtwo(II)I
8   ireturn             // Return int on top of operand stack;
                        // it is the int result of addTwo()

The invocation is set up by first pushing a reference to the current instance, this, on to the operand stack. The method invocation's arguments, int values 12 and 13, are then pushed. When the frame for the addTwo method is created, the arguments passed to the method become the initial values of the new frame's local variables. That is, the reference for this and the two arguments, pushed onto the operand stack by the invoker, will become the initial values of local variables 0, 1, and 2 of the invoked method.

Finally, addTwo is invoked. When it returns, its int return value is pushed onto the operand stack of the frame of the invoker, the add12and13 method. The return value is thus put in place to be immediately returned to the invoker of add12and13.

The return from add12and13 is handled by the ireturn instruction of add12and13. The ireturn instruction takes the int value returned by addTwo, on the operand stack of the current frame, and pushes it onto the operand stack of the frame of the invoker. It then returns control to the invoker, making the invoker's frame current. The Java Virtual Machine provides distinct return instructions for many of its numeric and reference data types, as well as a return instruction for methods with no return value. The same set of return instructions is used for all varieties of method invocations.

The operand of the invokevirtual instruction (in the example, the run-time constant pool index #4) is not the offset of the method in the class instance. The compiler does not know the internal layout of a class instance. Instead, it generates symbolic references to the methods of an instance, which are stored in the run-time constant pool. Those run-time constant pool items are resolved at run-time to determine the actual method location. The same is true for all other Java Virtual Machine instructions that access class instances.

Invoking addTwoStatic, a class (static) variant of addTwo, is similar, as shown:

int add12and13() {
    return addTwoStatic(12, 13);
}

although a different Java Virtual Machine method invocation instruction is used:

Method int add12and13()
0   bipush 12
2   bipush 13
4   invokestatic #3     // Method Example.addTwoStatic(II)I
7   ireturn

Compiling an invocation of a class (static) method is very much like compiling an invocation of an instance method, except this is not passed by the invoker. The method arguments will thus be received beginning with local variable 0 (§3.6). The invokestatic instruction is always used to invoke class methods.

The invokespecial instruction must be used to invoke instance initialization methods (§3.8). It is also used when invoking methods in the superclass (super). For instance, given classes Near and Far declared as:

class Near {
    int it;
    int getItNear() {
        return it;
    }
}
class Far extends Near {
    int getItFar() {
        return super.getItNear();
    }
}

The method Far.getItFar (which invokes a superclass method) becomes:

Method int getItFar()
0   aload_0
1   invokespecial #4    // Method Near.getItNear()I
4   ireturn

Note that methods called using the invokespecial instruction always pass this to the invoked method as its first argument. As usual, it is received in local variable 0.

To invoke the target of a method handle, a compiler must form a method descriptor that records the actual argument and return types. A compiler may not perform method invocation conversions on the arguments; instead, it must push them on the stack according to their own unconverted types. The compiler arranges for a reference to the method handle object to be pushed on the stack before the arguments, as usual. The compiler emits an invokevirtual instruction that references a descriptor which describes the argument and return types. By special arrangement with method resolution (§5.4.3.3), an invokevirtual instruction which invokes the invokeExact or invoke methods of java.lang.invoke.MethodHandle will always link, provided the method descriptor is syntactically well-formed and the types named in the descriptor can be resolved.

Java Virtual Machine class instances are created using the Java Virtual Machine's new instruction. Recall that at the level of the Java Virtual Machine, a constructor appears as a method with the compiler-supplied name <init>. This specially named method is known as the instance initialization method (§2.9). Multiple instance initialization methods, corresponding to multiple constructors, may exist for a given class. Once the class instance has been created and its instance variables, including those of the class and all of its superclasses, have been initialized to their default values, an instance initialization method of the new class instance is invoked. For example:

Object create() {
    return new Object();
}

compiles to:

Method java.lang.Object create()
0   new #1              // Class java.lang.Object
3   dup
4   invokespecial #4    // Method java.lang.Object.<init>()V
7   areturn

Class instances are passed and returned (as reference types) very much like numeric values, although type reference has its own complement of instructions, for example:

int i;                                  // An instance variable
MyObj example() {
    MyObj o = new MyObj();
    return silly(o);
}
MyObj silly(MyObj o) {
    if (o != null) {
        return o;
    } else {
        return o;
    }
}

becomes:

Method MyObj example()
0   new #2              // Class MyObj
3   dup
4   invokespecial #5    // Method MyObj.<init>()V
7   astore_1
8   aload_0
9   aload_1
10  invokevirtual #4    // Method Example.silly(LMyObj;)LMyObj;
13  areturn

Method MyObj silly(MyObj)
0   aload_1
1   ifnull 6
4   aload_1
5   areturn
6   aload_1
7   areturn

The fields of a class instance (instance variables) are accessed using the getfield and putfield instructions. If i is an instance variable of type int, the methods setIt and getIt, defined as:

void setIt(int value) {
    i = value;
}
int getIt() {
    return i;
}

become:

Method void setIt(int)
0   aload_0
1   iload_1
2   putfield #4    // Field Example.i I
5   return

Method int getIt()
0   aload_0
1   getfield #4    // Field Example.i I
4   ireturn

As with the operands of method invocation instructions, the operands of the putfield and getfield instructions (the run-time constant pool index #4) are not the offsets of the fields in the class instance. The compiler generates symbolic references to the fields of an instance, which are stored in the run-time constant pool. Those run-time constant pool items are resolved at run-time to determine the location of the field within the referenced object.

Java Virtual Machine arrays are also objects. Arrays are created and manipulated using a distinct set of instructions. The newarray instruction is used to create an array of a numeric type. The code:

void createBuffer() {
    int buffer[];
    int bufsz = 100;
    int value = 12;
    buffer = new int[bufsz];
    buffer[10] = value;
    value = buffer[11];
}

might be compiled to:

Method void createBuffer()
0   bipush 100     // Push int constant 100 (bufsz)
2   istore_2       // Store bufsz in local variable 2
3   bipush 12      // Push int constant 12 (value)
5   istore_3       // Store value in local variable 3
6   iload_2        // Push bufsz...
7   newarray int   // ...and create new int array of that length
9   astore_1       // Store new array in buffer
10  aload_1        // Push buffer
11  bipush 10      // Push int constant 10
13  iload_3        // Push value
14  iastore        // Store value at buffer[10]
15  aload_1        // Push buffer
16  bipush 11      // Push int constant 11
18  iaload         // Push value at buffer[11]...
19  istore_3       // ...and store it in value
20  return

The anewarray instruction is used to create a one-dimensional array of object references, for example:

void createThreadArray() {
    Thread threads[];
    int count = 10;
    threads = new Thread[count];
    threads[0] = new Thread();
}

becomes:

Method void createThreadArray()
0   bipush 10           // Push int constant 10
2   istore_2            // Initialize count to that
3   iload_2             // Push count, used by anewarray
4   anewarray class #1  // Create new array of class Thread
7   astore_1            // Store new array in threads
8   aload_1             // Push value of threads
9   iconst_0            // Push int constant 0
10  new #1              // Create instance of class Thread
13  dup                 // Make duplicate reference...
14  invokespecial #5    // ...for Thread's constructor
                        // Method java.lang.Thread.<init>()V
17  aastore             // Store new Thread in array at 0
18  return

The anewarray instruction can also be used to create the first dimension of a multidimensional array. Alternatively, the multianewarray instruction can be used to create several dimensions at once. For example, the three-dimensional array:

int[][][] create3DArray() {
    int grid[][][];
    grid = new int[10][5][];
    return grid;
}

is created by:

Method int create3DArray()[][][]
0   bipush 10                // Push int 10 (dimension one)
2   iconst_5                 // Push int 5 (dimension two)
3   multianewarray #1 dim #2 // Class [[[I, a three-dimensional
                             // int array; only create the
                             // first two dimensions
7   astore_1                 // Store new array...
8   aload_1                  // ...then prepare to return it
9   areturn

The first operand of the multianewarray instruction is the run-time constant pool index to the array class type to be created. The second is the number of dimensions of that array type to actually create. The multianewarray instruction can be used to create all the dimensions of the type, as the code for create3DArray shows. Note that the multidimensional array is just an object and so is loaded and returned by an aload_1 and areturn instruction, respectively. For information about array class names, see §4.4.1.

All arrays have associated lengths, which are accessed via the arraylength instruction.

Compilation of switch statements uses the tableswitch and lookupswitch instructions. The tableswitch instruction is used when the cases of the switch can be efficiently represented as indices into a table of target offsets. The default target of the switch is used if the value of the expression of the switch falls outside the range of valid indices. For instance:

int chooseNear(int i) {
    switch (i) {
        case 0:  return  0;
        case 1:  return  1;
        case 2:  return  2;
        default: return -1;
    }
}

compiles to:

Method int chooseNear(int)
0   iload_1             // Push local variable 1 (argument i)
1   tableswitch 0 to 2: // Valid indices are 0 through 2
      0: 28             // If i is 0, continue at 28
      1: 30             // If i is 1, continue at 30
      2: 32             // If i is 2, continue at 32
      default:34        // Otherwise, continue at 34
28  iconst_0            // i was 0; push int constant 0...
29  ireturn             // ...and return it
30  iconst_1            // i was 1; push int constant 1...
31  ireturn             // ...and return it
32  iconst_2            // i was 2; push int constant 2...
33  ireturn             // ...and return it
34  iconst_m1           // otherwise push int constant -1...
35  ireturn             // ...and return it

The Java Virtual Machine's tableswitch and lookupswitch instructions operate only on int data. Because operations on byte, char, or short values are internally promoted to int, a switch whose expression evaluates to one of those types is compiled as though it evaluated to type int. If the chooseNear method had been written using type short, the same Java Virtual Machine instructions would have been generated as when using type int. Other numeric types must be narrowed to type int for use in a switch.

Where the cases of the switch are sparse, the table representation of the tableswitch instruction becomes inefficient in terms of space. The lookupswitch instruction may be used instead. The lookupswitch instruction pairs int keys (the values of the case labels) with target offsets in a table. When a lookupswitch instruction is executed, the value of the expression of the switch is compared against the keys in the table. If one of the keys matches the value of the expression, execution continues at the associated target offset. If no key matches, execution continues at the default target. For instance, the compiled code for:

int chooseFar(int i) {
    switch (i) {
        case -100: return -1;
        case 0:    return  0;
        case 100:  return  1;
        default:   return -1;
    }
}

looks just like the code for chooseNear, except for the lookupswitch instruction:

Method int chooseFar(int)
0   iload_1
1   lookupswitch 3:
         -100: 36
            0: 38
          100: 40
      default: 42
36  iconst_m1
37  ireturn
38  iconst_0
39  ireturn
40  iconst_1
41  ireturn
42  iconst_m1
43  ireturn

The Java Virtual Machine specifies that the table of the lookupswitch instruction must be sorted by key so that implementations may use searches more efficient than a linear scan. Even so, the lookupswitch instruction must search its keys for a match rather than simply perform a bounds check and index into a table like tableswitch. Thus, a tableswitch instruction is probably more efficient than a lookupswitch where space considerations permit a choice.

The Java Virtual Machine has a large complement of instructions that manipulate the contents of the operand stack as untyped values. These are useful because of the Java Virtual Machine's reliance on deft manipulation of its operand stack. For instance:

public long nextIndex() { 
    return index++;
}

private long index = 0;

is compiled to:

Method long nextIndex()
0   aload_0        // Push this
1   dup            // Make a copy of it
2   getfield #4    // One of the copies of this is consumed
                   // pushing long field index,
                   // above the original this
5   dup2_x1        // The long on top of the operand stack is 
                   // inserted into the operand stack below the 
                   // original this
6   lconst_1       // Push long constant 1 
7   ladd           // The index value is incremented...
8   putfield #4    // ...and the result stored in the field
11  lreturn        // The original value of index is on top of
                   // the operand stack, ready to be returned

Note that the Java Virtual Machine never allows its operand stack manipulation instructions to modify or break up individual values on the operand stack.

Exceptions are thrown from programs using the throw keyword. Its compilation is simple:

void cantBeZero(int i) throws TestExc {
    if (i == 0) {
        throw new TestExc();
    }
}

becomes:

Method void cantBeZero(int)
0   iload_1             // Push argument 1 (i)
1   ifne 12             // If i==0, allocate instance and throw
4   new #1              // Create instance of TestExc
7   dup                 // One reference goes to its constructor
8   invokespecial #7    // Method TestExc.<init>()V
11  athrow              // Second reference is thrown
12  return              // Never get here if we threw TestExc

Compilation of try-catch constructs is straightforward. For example:

void catchOne() {
    try {
        tryItOut();
    } catch (TestExc e) {
        handleExc(e);
    }
}

is compiled as:

Method void catchOne()
0   aload_0             // Beginning of try block
1   invokevirtual #6    // Method Example.tryItOut()V
4   return              // End of try block; normal return
5   astore_1            // Store thrown value in local var 1
6   aload_0             // Push this
7   aload_1             // Push thrown value
8   invokevirtual #5    // Invoke handler method: 
                        // Example.handleExc(LTestExc;)V
11  return              // Return after handling TestExc
Exception table:
From    To      Target      Type
0       4       5           Class TestExc

Looking more closely, the try block is compiled just as it would be if the try were not present:

Method void catchOne()
0   aload_0             // Beginning of try block
1   invokevirtual #6    // Method Example.tryItOut()V
4   return              // End of try block; normal return

If no exception is thrown during the execution of the try block, it behaves as though the try were not there: tryItOut is invoked and catchOne returns.

Following the try block is the Java Virtual Machine code that implements the single catch clause:

5   astore_1            // Store thrown value in local var 1
6   aload_0             // Push this
7   aload_1             // Push thrown value
8   invokevirtual #5    // Invoke handler method: 
                        // Example.handleExc(LTestExc;)V
11  return              // Return after handling TestExc
Exception table:
From    To      Target      Type
0       4       5           Class TestExc

The invocation of handleExc, the contents of the catch clause, is also compiled like a normal method invocation. However, the presence of a catch clause causes the compiler to generate an exception table entry (§2.10, §4.7.3). The exception table for the catchOne method has one entry corresponding to the one argument (an instance of class TestExc) that the catch clause of catchOne can handle. If some value that is an instance of TestExc is thrown during execution of the instructions between indices 0 and 4 in catchOne, control is transferred to the Java Virtual Machine code at index 5, which implements the block of the catch clause. If the value that is thrown is not an instance of TestExc, the catch clause of catchOne cannot handle it. Instead, the value is rethrown to the invoker of catchOne.

A try may have multiple catch clauses:

void catchTwo() {
    try {
        tryItOut();
    } catch (TestExc1 e) {
        handleExc(e);
    } catch (TestExc2 e) {
        handleExc(e);
    }
}

Multiple catch clauses of a given try statement are compiled by simply appending the Java Virtual Machine code for each catch clause one after the other and adding entries to the exception table, as shown:

Method void catchTwo()
0   aload_0             // Begin try block
1   invokevirtual #5    // Method Example.tryItOut()V
4   return              // End of try block; normal return
5   astore_1            // Beginning of handler for TestExc1;
                        // Store thrown value in local var 1
6   aload_0             // Push this
7   aload_1             // Push thrown value
8   invokevirtual #7    // Invoke handler method:
                        // Example.handleExc(LTestExc1;)V
11  return              // Return after handling TestExc1
12  astore_1            // Beginning of handler for TestExc2;
                        // Store thrown value in local var 1
13  aload_0             // Push this
14  aload_1             // Push thrown value
15  invokevirtual #7    // Invoke handler method:
                        // Example.handleExc(LTestExc2;)V
18  return              // Return after handling TestExc2
Exception table:
From    To      Target      Type
0       4       5           Class TestExc1
0       4       12          Class TestExc2

If during the execution of the try clause (between indices 0 and 4) a value is thrown that matches the parameter of one or more of the catch clauses (the value is an instance of one or more of the parameters), the first (innermost) such catch clause is selected. Control is transferred to the Java Virtual Machine code for the block of that catch clause. If the value thrown does not match the parameter of any of the catch clauses of catchTwo, the Java Virtual Machine rethrows the value without invoking code in any catch clause of catchTwo.

Nested try-catch statements are compiled very much like a try statement with multiple catch clauses:

void nestedCatch() {
    try {
        try {
            tryItOut();
        } catch (TestExc1 e) {
            handleExc1(e);
        }
    } catch (TestExc2 e) {
        handleExc2(e);
    }
}

becomes:

Method void nestedCatch()
0   aload_0             // Begin try block
1   invokevirtual #8    // Method Example.tryItOut()V
4   return              // End of try block; normal return
5   astore_1            // Beginning of handler for TestExc1;
                        // Store thrown value in local var 1
6   aload_0             // Push this
7   aload_1             // Push thrown value
8   invokevirtual #7    // Invoke handler method: 
                        // Example.handleExc1(LTestExc1;)V
11  return              // Return after handling TestExc1
12  astore_1            // Beginning of handler for TestExc2;
                        // Store thrown value in local var 1
13  aload_0             // Push this
14  aload_1             // Push thrown value
15  invokevirtual #6    // Invoke handler method:
                        // Example.handleExc2(LTestExc2;)V
18  return              // Return after handling TestExc2
Exception table:
From    To      Target      Type
0       4       5           Class TestExc1
0       12      12          Class TestExc2

The nesting of catch clauses is represented only in the exception table. The Java Virtual Machine does not enforce nesting of or any ordering of the exception table entries (§2.10). However, because try-catch constructs are structured, a compiler can always order the entries of the exception handler table such that, for any thrown exception and any program counter value in that method, the first exception handler that matches the thrown exception corresponds to the innermost matching catch clause.

For instance, if the invocation of tryItOut (at index 1) threw an instance of TestExc1, it would be handled by the catch clause that invokes handleExc1. This is so even though the exception occurs within the bounds of the outer catch clause (catching TestExc2) and even though that outer catch clause might otherwise have been able to handle the thrown value.

As a subtle point, note that the range of a catch clause is inclusive on the "from" end and exclusive on the "to" end (§4.7.3). Thus, the exception table entry for the catch clause catching TestExc1 does not cover the return instruction at offset 4. However, the exception table entry for the catch clause catching TestExc2 does cover the return instruction at offset 11. Return instructions within nested catch clauses are included in the range of instructions covered by nesting catch clauses.

(This section assumes a compiler generates class files with version number 50.0 or below, so that the jsr instruction may be used. See also §4.10.2.5.)

Compilation of a try-finally statement is similar to that of try-catch. Prior to transferring control outside the try statement, whether that transfer is normal or abrupt, because an exception has been thrown, the finally clause must first be executed. For this simple example:

void tryFinally() {
    try {
        tryItOut();
    } finally {
        wrapItUp();
    }
}

the compiled code is:

Method void tryFinally()
0   aload_0             // Beginning of try block
1   invokevirtual #6    // Method Example.tryItOut()V
4   jsr 14              // Call finally block
7   return              // End of try block
8   astore_1            // Beginning of handler for any throw
9   jsr 14              // Call finally block
12  aload_1             // Push thrown value
13  athrow              // ...and rethrow value to the invoker
14  astore_2            // Beginning of finally block
15  aload_0             // Push this
16  invokevirtual #5    // Method Example.wrapItUp()V
19  ret 2               // Return from finally block
Exception table:
From    To      Target      Type
0       4       8           any

There are four ways for control to pass outside of the try statement: by falling through the bottom of that block, by returning, by executing a break or continue statement, or by raising an exception. If tryItOut returns without raising an exception, control is transferred to the finally block using a jsr instruction. The jsr 14 instruction at index 4 makes a "subroutine call" to the code for the finally block at index 14 (the finally block is compiled as an embedded subroutine). When the finally block completes, the ret 2 instruction returns control to the instruction following the jsr instruction at index 4.

In more detail, the subroutine call works as follows: The jsr instruction pushes the address of the following instruction (return at index 7) onto the operand stack before jumping. The astore_2 instruction that is the jump target stores the address on the operand stack into local variable 2. The code for the finally block (in this case the aload_0 and invokevirtual instructions) is run. Assuming execution of that code completes normally, the ret instruction retrieves the address from local variable 2 and resumes execution at that address. The return instruction is executed, and tryFinally returns normally.

A try statement with a finally clause is compiled to have a special exception handler, one that can handle any exception thrown within the try statement. If tryItOut throws an exception, the exception table for tryFinally is searched for an appropriate exception handler. The special handler is found, causing execution to continue at index 8. The astore_1 instruction at index 8 stores the thrown value into local variable 1. The following jsr instruction does a subroutine call to the code for the finally block. Assuming that code returns normally, the aload_1 instruction at index 12 pushes the thrown value back onto the operand stack, and the following athrow instruction rethrows the value.

Compiling a try statement with both a catch clause and a finally clause is more complex:

void tryCatchFinally() {
    try {
        tryItOut();
    } catch (TestExc e) {
        handleExc(e);
    } finally {
        wrapItUp();
    }
}

becomes:

Method void tryCatchFinally()
0   aload_0             // Beginning of try block
1   invokevirtual #4    // Method Example.tryItOut()V
4   goto 16             // Jump to finally block
7   astore_3            // Beginning of handler for TestExc;
                        // Store thrown value in local var 3
8   aload_0             // Push this
9   aload_3             // Push thrown value
10  invokevirtual #6    // Invoke handler method:
                        // Example.handleExc(LTestExc;)V
13  goto 16             // This goto is unnecessary, but was
                        // generated by javac in JDK 1.0.2
16  jsr 26              // Call finally block
19  return              // Return after handling TestExc
20  astore_1            // Beginning of handler for exceptions
                        // other than TestExc, or exceptions
                        // thrown while handling TestExc
21  jsr 26              // Call finally block
24  aload_1             // Push thrown value...
25  athrow              // ...and rethrow value to the invoker
26  astore_2            // Beginning of finally block
27  aload_0             // Push this
28  invokevirtual #5    // Method Example.wrapItUp()V
31  ret 2               // Return from finally block
Exception table:
From    To      Target      Type
0       4       7           Class TestExc
0       16      20          any

If the try statement completes normally, the goto instruction at index 4 jumps to the subroutine call for the finally block at index 16. The finally block at index 26 is executed, control returns to the return instruction at index 19, and tryCatchFinally returns normally.

If tryItOut throws an instance of TestExc, the first (innermost) applicable exception handler in the exception table is chosen to handle the exception. The code for that exception handler, beginning at index 7, passes the thrown value to handleExc and on its return makes the same subroutine call to the finally block at index 26 as in the normal case. If an exception is not thrown by handleExc, tryCatchFinally returns normally.

If tryItOut throws a value that is not an instance of TestExc or if handleExc itself throws an exception, the condition is handled by the second entry in the exception table, which handles any value thrown between indices 0 and 16. That exception handler transfers control to index 20, where the thrown value is first stored in local variable 1. The code for the finally block at index 26 is called as a subroutine. If it returns, the thrown value is retrieved from local variable 1 and rethrown using the athrow instruction. If a new value is thrown during execution of the finally clause, the finally clause aborts, and tryCatchFinally returns abruptly, throwing the new value to its invoker.

Synchronization in the Java Virtual Machine is implemented by monitor entry and exit, either explicitly (by use of the monitorenter and monitorexit instructions) or implicitly (by the method invocation and return instructions).

For code written in the Java programming language, perhaps the most common form of synchronization is the synchronized method. A synchronized method is not normally implemented using monitorenter and monitorexit. Rather, it is simply distinguished in the run-time constant pool by the ACC_SYNCHRONIZED flag, which is checked by the method invocation instructions (§2.11.10).

The monitorenter and monitorexit instructions enable the compilation of synchronized statements. For example:

void onlyMe(Foo f) {
    synchronized(f) {
        doSomething();
    }
}

is compiled to:

Method void onlyMe(Foo)
0   aload_1             // Push f
1   dup                 // Duplicate it on the stack
2   astore_2            // Store duplicate in local variable 2
3   monitorenter        // Enter the monitor associated with f
4   aload_0             // Holding the monitor, pass this and...
5   invokevirtual #5    // ...call Example.doSomething()V
8   aload_2             // Push local variable 2 (f)
9   monitorexit         // Exit the monitor associated with f
10  goto 18             // Complete the method normally
13  astore_3            // In case of any throw, end up here
14  aload_2             // Push local variable 2 (f)
15  monitorexit         // Be sure to exit the monitor!
16  aload_3             // Push thrown value...
17  athrow              // ...and rethrow value to the invoker
18  return              // Return in the normal case
Exception table:
From    To      Target      Type
4       10      13          any
13      16      13          any

The compiler ensures that at any method invocation completion, a monitorexit instruction will have been executed for each monitorenter instruction executed since the method invocation. This is the case whether the method invocation completes normally (§2.6.4) or abruptly (§2.6.5). To enforce proper pairing of monitorenter and monitorexit instructions on abrupt method invocation completion, the compiler generates exception handlers (§2.10) that will match any exception and whose associated code executes the necessary monitorexit instructions.

The representation of annotations in class files is described in §4.7.16-§4.7.22. These sections make it clear how to represent annotations on declarations of classes, interfaces, fields, methods, method parameters, and type parameters, as well as annotations on types used in those declarations. Annotations on package declarations require additional rules, given here.

When the compiler encounters an annotated package declaration that must be made available at run time, it emits a class file with the following properties:

A compilation unit that contains a module declaration (JLS §7.7) is compiled to a class file that contains a Module attribute.

By convention, the name of a compilation unit that contains a module declaration is module-info.java, echoing the package-info.java convention for a compilation unit that contains solely a package declaration. Consequently, by convention, the name for the compiled form of a module declaration is module-info.class.

A flag in the access_flags item of the ClassFile structure, ACC_MODULE (0x8000), indicates that this class file declares a module. ACC_MODULE plays a similar role to ACC_ANNOTATION (0x2000) and ACC_ENUM (0x4000) in flagging this class file as "not an ordinary class". ACC_MODULE does not describe accessibility of a class or interface.

The Module attribute is explicit about the module's dependences; there are no implicit requires directives at the ClassFile level. If the requires_count item is zero, then the Java SE Platform does not infer the existence of a requires table nor any particular entry therein. java.base is the only module in which a zero requires_count is legal, because it is the primordial module. For every other module, the Module attribute must have a requires table of at least length one, because every other module depends on java.base. If a compilation unit contains a module declaration (except java.base) that does not state its dependence on java.base explicitly, then a compiler must emit an entry for java.base in the requires table and flag it as ACC_MANDATED to denote that it was implicitly declared.

For encapsulation, the Module attribute is explicit about the packages exported and opened by a normal module; there are no implicit exports or opens directives at the ClassFile level for a normal module. If the exports_count item or opens_count item is zero, then the Java SE Platform does not infer the existence of an exports table or opens table, nor any particular entry therein. On the other hand, for an open module, the Module attribute is implicit about the packages opened by the module. All packages of an open module are opened to all other modules, even though the opens_count item is zero.

The Module attribute is explicit about the module's consumption and provision of services; there are no implicit uses or provides directives at the ClassFile level.