Skip to main content.

Web Based Programming Tutorials

Homepage | Forum - Join the forum to discuss anything related to programming! | Programming Resources

Java Unleashed Second Edition

Chapter 34 -- Java Under the Hood: Inside the Virtual Machine

Chapter 34

Java Under the Hood: Inside the Virtual Machine

by Stephen Ingram

CONTENTS

This chapter takes an in-depth look at the internals of the Java virtual machine (VM). Although an understanding of Java's internals is not required to be an effective Java programmer, comprehension of this chapter provides the basis for making the transition to expert-level Java coding. In any event, the VM internals shed light on the mindset of Java's original designers. Exploration at this level is a fascinating journey because of the elegance behind the Java architecture.

Looking at a virtual machine from the outside in is probably the best way to understand its workings. Incremental learning results when you move from the known to the unknown, so this chapter starts with the item you are most familiar with: the class file.

The Class File

The class file is similar to standard language object modules. When a C language file is compiled, the output is an object module. Multiple object modules are linked together to form an executable program. In Java, the class file replaces the object module and the Java virtual machine replaces the executable program.

You'll find all the information needed to execute a class contained within the class file; you'll also find extra information that aids debugging and source file tracking. Remember that Java has no "header" include files, so the class file format also has to fully convey class layout and members. Parsing a class file yields a wealth of class information, not just its runtime architecture.

Layout

The overall layout of the class file uses an outer structure and a series of substructures that contain an ever-increasing amount of detail. The outer layer is described by the following ClassFile structure:

ClassFile
{
    u4 magic;
    u2 minor_version;
    u2 major_version;
    u2 constant_pool_count;
    cp_info constant_pool[constant_pool_count - 1];
    u2 access_flags;
    u2 this_class;
    u2 super_class;
    u2 interfaces_count;
    u2 interfaces[interfaces_count];
    u2 fields_count;
    field_info fields[fields_count];
    u2 methods_count;
    method_info methods[methods_count];
    u2 attributes_count;
    attribute_info attributes[attribute_count];
}

In addition to the generic class information (this_class, super_class, version, and so on), there are three major substructures: cp_info, field_info, and method_info. The attribute_info structure is considered a minor substructure because attributes recur throughout the class file at various levels. Fields and methods contain their own set of attributes and some individual attributes also contain their own private attribute arrays.

The symbols u2 and u4 represent unsigned 2-byte and unsigned 4-byte quantities.

The Class Viewer

Simply regurgitating class file structures does not provide the best basis for actual learning. To better convey the overall class file structure, I wrote an interactive Java application that presents a class file in a tree format. Figure 34.1 shows the tool in action. With this tool, you can view any class and save it in ASCII format. Navigation is performed with the keyboard: The arrow keys provide movement and the spacebar expands and contracts the nodes.

Figure 34.1 : The class Viewer application in action.

The Viewer.class tool and source code are provided on the CD-ROM that accompanies this book. Because it's an application, run the tool from the command line with this statement: java Viewer

Now that you have a tool, you need a simple class to explore. Here's one, just waiting for you:

public class test
{
    public static int st_one;

    public test()
    {
        st_one = 100;
    }

    public test(int v)
    {
        st_one = v;
    }

    public native boolean getData(int data[] );

    public int do_countdown()
    {
        int x = st_one;

        System.out.println("Performing countdown:");
        while ( x-- != 0 )
            System.out.println(x);
        return st_one;
    }

    public int do_countdown(int x)
    {
        int save = x;

        System.out.println("Performing countdown:");
        while ( x-- != 0 )
            System.out.println(x);
        return save;
    }
}

This class doesn't actually do very much, but it does provide a basis for class file exploration. Once compiled, the outer layer of the resulting class file is as follows:

----file name: test.class
       |....magic: cafebabe
       |....version: 45.3
       |+---ConstantPool [42]
       |....access: PUBLIC
       |....this: Class: test
       |....parent: Class: java/lang/Object
       |....No Interfaces
       |+---Fields [1]
       |+---Methods [5]
       |+---Attributes [1]

The magic number of a class is 0xcafebabe. If a class file does not contain this number, the Java virtual machine will refuse to load the class. This number must appear in a class file or the file is assumed to be invalid. The current major version is 45; the current minor version is 3. Future Java compilers will increment these numbers when the underlying class file format changes. Version numbers allow future Java machines to recognize and disallow execution of older class files.

Access Flags

Access flags are used throughout the class file to convey the access characteristics of various items. The flag itself is a collection of 11 individual bits. Table 34.1 lays out the masks.

Table 34.1. Access flag bit values.

Flag ValueIndication
ACC_PUBLIC = 0x0001Global visibility
ACC_PRIVATE = 0x0002Local class visibility
ACC_PROTECTED = 0x0004Subclass visibility
ACC_STATIC = 0x0008One occurrence in system (not per class)
ACC_FINAL = 0x0010No changes allowed
ACC_SYNCHRONIZED = 0x0020Access with a monitor
ACC_VOLATILE = 0x0040No local caching
ACC_TRANSIENT = 0x0080Not a persistent value
ACC_NATIVE = 0x0100Native method implementation
ACC_INTERFACE = 0x0200Class is an interface
ACC_ABSTRACT = 0x0400Class or method is abstract

Access flags are present for a class and its fields and methods. Only a subset of values appears in any given item. Some bits apply only to fields (for example, VOLATILE and TRANSIENT); others apply only to methods (for example, SYNCHRONIZED and NATIVE).

Attributes

Attributes, like access flags, appear throughout a class file. They have the following form:

GenericAttribute_info
{
    u2 attribute_name;
    u4 attribute_length;
    u1 info[attribute_length];
}

A generic structure exists to enable loaders to skip over attributes they don't understand. The actual attribute has a unique structure that can be read if the loader understands the format. As an example, the following structure specifies the format of a source file attribute:

SourceFile_attribute
{
    u2 attribute_name;
    u4 attribute_length;
    u2 sourcefile_index;
}

The name of an attribute is an index into the constant pool. You learn about the constant pool in the next section. If a loader does not understand the source file attribute structure, it can skip the data by reading the number of bytes specified in the length parameter. For the source file attribute, the length is 2.

Constant Pool

The constant pool forms the basis for all numbers and strings within a class file. Nowhere else do you find strings or numbers. Any time you need to reference a string or number, you substitute an index into the constant pool. Consequently, the constant pool is the dominant feature of a class. The pool is even used directly within the virtual machine itself.

There are 12 different types of constants:

Each constant structure leads off with a tag identifying the structure type. Following the type is data specific to each individual structure. The layout of each constant structure is shown in Listing 34.1.


Listing 34.1. The layout of all 12 constant structures.
CONSTANT_Utf8_info
{
    u1 tag;
    u2 length;
    u1 bytes[length];
}

CONSTANT_Unicode_info
{
    u1 tag;
    u2 length;
    u2 words[length];
}

CONSTANT_Integer_info
{
    u1 tag;
    u4 bytes;
}

CONSTANT_Float_info
{
    u1 tag;
    u4 bytes;
}

CONSTANT_Long_info
{
    u1 tag;
    u4 high_bytes;
    u4 low_bytes;
}

CONSTANT_Double_info
{
    u1 tag;
    u4 high_bytes;
    u4 low_bytes;
}

CONSTANT_Class_info
{
    u1 tag;
    u2 name_index;
}

CONSTANT_String_info
{
    u1 tag;
    u2 string_index;
}

CONSTANT_Fieldref_info
{
    u1 tag;
    u2 class_index;
    u2 name_and_type_index;
}

CONSTANT_Methodref_info
{
    u1 tag;
    u2 class_index;
    u2 name_and_type_index;
}

CONSTANT_InterfaceMethodref_info
{
    u1 tag;
    u2 class_index;
    u2 name_and_type_index;
}

CONSTANT_NameAndType_info
{
    u1 tag;
    u2 name_index;
    u2 signature_index;
}

The CONSTANT_Utf8 structure contains standard ASCII text strings. These are not null-terminated because they use an explicit length parameter. Notice that most of the constants reference other constants for information. Methods, for example, specify a class and type by providing indexes to other constant pool members. Constant pool cross-references eliminate repetition of data.

The constant pool for the sample test class appears as follows:

|---file name: test.class
     |...magic: cafebabe
     |...version: 45.3
     |---ConstantPool [42]
     |    |...String #28 -> Performing countdown:
     |    |...Class: java/lang/System
     |    |...Class: java/lang/Object
     |    |...Class: java/io/PrintStream
     |    |...Class: test
     |    |...Method: java/lang/Object.<init>()V
     |    |...Field: test.st_one I
     |    |...Field: java/lang/System.out Ljava/io/PrintStream;
     |    |...Method: java/io/PrintStream.println(I)V
     |    |...Method: java/io/PrintStream.println(Ljava/lang/String;)V
     |    |...NameAndType: st_one I
     |    |...NameAndType: println (I)V
     |    |...NameAndType: out Ljava/io/PrintStream;
     |    |...NameAndType: println (Ljava/lang/String;)V
     |    |...NameAndType: <init> ()V
     |    |...Utf8 [7] println
     |    |...Utf8 [4] (I)V
     |    |...Utf8 [3] ()I
     |    |...Utf8 [13] ConstantValue
     |    |...Utf8 [4] (I)I
     |    |...Utf8 [19] java/io/PrintStream
     |    |...Utf8 [10] Exceptions
     |    |...Utf8 [15] LineNumberTable
     |    |...Utf8 [1] I
     |    |...Utf8 [10] SourceFile
     |    |...Utf8 [14] LocalVariables
     |    |...Utf8 [4] Code
     |    |...Utf8 [21] Performing countdown:
     |    |...Utf8 [3] out
     |    |...Utf8 [21] (Ljava/lang/String;)V
     |    |...Utf8 [16] java/lang/Object
     |    |...Utf8 [6] <init>
     |    |...Utf8 [21] Ljava/io/PrintStream;
     |    |...Utf8 [16] java/lang/System
     |    |...Utf8 [12] do_countdown
     |    |...Utf8 [5] ([I)Z
     |    |...Utf8 [6] st_one
     |    |...Utf8 [7] getData
     |    |...Utf8 [9] test.java
     |    |...Utf8 [3] ()V
     |    |...Utf8 [4] test
     |...access: PUBLIC
     |...this: Class: test
     |...parent: Class: java/lang/Object
     |...No Interfaces
     |+--Fields [1]
     |+--Methods [5]
     |+--Attributes [1]

NOTE
The Viewer tool substitutes pool indexes with actual pool data whenever possible.

Fields

Field structures contain the individual data members of a class. Any class item that is not a method is placed into the fields section of the class file. The field structure looks like this:

field_info
{
    u2 access_flags;
    u2 name_index;
    u2 signature_index;
    u2 attribute_count;
    attribute_info attributes[attribute_count];
}

The sample test class contains one field:

|---file name: test.class
     |...magic: cafebabe
     |...version: 45.3
     |+--ConstantPool [42]
     |...access: PUBLIC
     |...this: Class: test
     |...parent: Class: java/lang/Object
     |...No Interfaces
     |---Fields [1]
     |    |---st_one
     |         |...I
     |         |...PUBLIC STATIC
     |         |...No attributes
     |+--Methods [5]
     |+--Attributes [1]

Methods

The method section of the class file contains all the executable content of a class. In addition to the method name and signature, the structure contains a set of attributes. One of these attributes has the actual bytecodes that the virtual machine will execute. The method structure is shown here:

method_info
{
    u2 access_flags;
    u2 name_index;
    u2 signature_index;
    u2 attributes_count;
    attribute_info attributes[attribute_count];
}

The sample test class contains the following method section:

|---file name: test.class
     |...magic: cafebabe
     |...version: 45.3
     |+--ConstantPool [42]
     |...access: PUBLIC
     |...this: Class: test
     |...parent: Class: java/lang/Object
     |...No Interfaces
     |+--Fields [1]
     |---Methods [5]
     |    |---<init>
     |    |    |...()V
     |    |    |...PUBLIC
     |    |    |---Attributes [1]
     |    |         |---Code
     |    |              |...max_stack = 1
     |    |              |...max_locals = 1
     |    |              |+--Byte Codes [10]
     |    |              |...No exceptions
     |    |              |+--Attributes [1]
     |    |---<init>
     |    |    |...(I)V
     |    |    |...PUBLIC
     |    |    |---Attributes [1]
     |    |         |---Code
     |    |              |...max_stack = 1
     |    |              |...max_locals = 2
     |    |              |+--Byte Codes [9]
     |    |              |...No exceptions
     |    |              |+--Attributes [1]
     |    |---getData
     |    |    |...([I)Z
     |    |    |...PUBLIC NATIVE
     |    |    |...No attributes
     |    |---do_countdown
     |    |    |...()I
     |    |    |...PUBLIC
     |    |    |---Attributes [1]
     |    |         |---Code
     |    |              |...max_stack = 2
     |    |              |...max_locals = 2
     |    |              |+--Byte Codes [33]
     |    |              |...No exceptions
     |    |              |+--Attributes [1]
     |    |---do_countdown
     |         |...(I)I
     |         |...PUBLIC
     |         |---Attributes [1]
     |              |---Code
     |                   |...max_stack = 2
     |                   |...max_locals = 3
     |                   |+--Byte Codes [29]
     |                   |...No exceptions
     |                   |+--Attributes [1]
     |+--Attributes [1]

Each method has a name and signature. Signatures are used by Java to determine calling arguments and return types. The format of a signature is as follows:

"(args*)return_type"

Arguments can be any combination of the characters listed in Table 34.2. Class name arguments are written as follows:

Lclass_name;

The semicolon (;) signals the end of the class name, just as the right parenthesis signals the end of an argument list. Arrays are followed by the array type:

[B for an array of bytes
[Ljava/langString; for an array of objects (in this case, Strings)

Table 34.2. Method signature symbols.

Type
Signature Character
byte
B
char
C
class
L
end of class
;
float
F
double
D
function
(
end of function
)
int
I
long
J
short
S
void
V
boolean
Z

All the methods except getData() have a code attribute. This method is marked as NATIVE, so the Java virtual machine expects the code to be in a native library. Each non-native method contains a code attribute that has the following format:

Code_attribute
{
    u2 attribute_name;
    u4 attribute_length;
    u2 max_stack;
    u2 max_locals;
    u4 code_length;
    u1 code[code_length];
    u2 exception_table_length;
    ExceptionItem exceptions[exception_table_length];
    u2 attributes_count;
    attribute_info attributes[attribute_count];
}

Code attributes contain a private list of other attributes. Typically, these are debugging lists, such as line-number information.

Exception Handling

The pc register points to the next bytecode to execute. Whenever an exception is thrown, the method's exception table is searched for a handler. Each exception table entry has this format:

ExceptionItem
{
    u2 start_pc;
    u2 end_pc;
    u2 handler_pc;
    u2 catch_type;
}

If the pc register is within the proper range and the thrown exception is the proper type, the entry's handler code block is executed. If no handler is found, the exception propagates up to the calling method. The procedure repeats itself until either a valid handler is found or the program exits.

Now that you've hit the code attribute, it's time to jump into the virtual machine.

The Virtual Machine

The Java virtual machine interprets Java bytecodes that are contained in code attributes. The virtual machine is stack based. Most computer architectures perform their operations on a mixture of memory locations and registers. The Java virtual machine performs its operations exclusively on a stack. This was done primarily to support portability. No assumptions could be made about the size or number of registers in a given CPU. Intel microprocessors are especially limited in their register composition.

Registers

The virtual machine does contain some registers, but these are used for tracking the current state of the machine:

All these registers are 32 bits wide and point into separate storage blocks. The blocks, however, can be allocated all at once because the code attribute specifies the size of the operand stack, the number of local variables, and the length of the bytecodes.

Operand Stack

Most Java bytecodes work on the operand stack. For example, to add two integers together, each integer is pushed onto the operand stack. The addition operator removes the top two integers, adds them, and places the result in their place back on the stack:

..., 4, 5 -> ..., 9

NOTE
Operand stack notation is used throughout the remainder of this chapter. The stack reads from left to right, with the stack top on the extreme right. Ellipses indicate indeterminate data buried on the stack. The arrow indicates an operation; the data to the right of the arrow represents the stack after the operation is performed.

Each stack location is 32 bits wide. Longs and doubles are 64 bits wide, so they take up two stack locations.

Local Variables

Each code attribute specifies the size of the local variables. A local variable is 32 bits wide, so long and double primitives take up two variable slots. Unlike C, all method arguments appear as local variables. The operand stack is reserved exclusively for operations.

Before detailing the bytecodes, I think a more holistic view of the virtual machine's operations would be instructive.

The Virtual Machine in Action

Consider the following method from the sample test class:

public int do_countdown(int x)
    {
        int save = x;

        System.out.println("Performing countdown:");
        while ( x-- != 0 )
            System.out.println(x);
        return save;
    }

The class file specifies that this method has a maximum operand stack of two and a local variable block of three. Figure 34.2 shows the initial state of these two blocks.

Figure 34.2 : The initial state of the operand stack and local variable blocks.

Notice how the method's input argument is placed into the local variable block. Java contains an extensive set of bytecodes dedicated to moving data between the local variable block and the operand stack. All data must be moved to the stack before it can be used. This is an important distinction between Java machine code and register-based machine architectures. Register machines reference memory directly from the contents of a register. Java must first move the contents of a variable to its operand stack before it can reference the location.

The do_countdown(int x) method contains 29 bytecodes. These two instructions store the input argument x into the local variable save:

0 iload_1    Move argument x onto the operand stack
1 istore_2   Store top of stack (x) into variable save

The next three bytecodes first load a target object (System.out) and a String onto the operand stack. At this point, println(String) is invoked. Notice that before the invocation, both elements of the operand stack are occupied. After the invocation, the operand stack is completely empty because invokevirtual removes all the stack elements it uses:

2 getstatic #8 <Field: java/lang/System.out Ljava/io/PrintStream;>
5 ldc1 #1 <String #28 -> Performing countdown:>
7 invokevirtual #10 <Method: java/io/PrintStream.println(Ljava/lang/String;)V>

This instruction transfers control to the test portion of the method's while loop. Most modern compilers code a loop with the loop test at the bottom:

10 goto 20

The loop body calls println(int):

13 getstatic #8 <Field: java/lang/System.out Ljava/io/PrintStream;>
16 iload_1    Move argument x onto the operand stack
17 invokevirtual #9 <Method: java/io/PrintStream.println(I)V>

Okay, this next bit is a little tricky. First, the current value of x is placed onto the stack. Next, 1 is subtracted from variable x. Finally, the value on the stack is checked for zero. Notice that the value on the stack is x before it was decremented. If the stack value is not zero, the body of the loop will be executed:

20 iload_1    Move argument x onto the operand stack
21 iinc 1 -1    Decrement argument x by one
24 ifne 13    Branch if stack top (x before decrement) not zero

The final two bytecodes move the value of variable save onto the stack and then return this value to the calling method:

27 iload_2    Move variable save onto the operand stack
28 ireturn    Move integer on stack onto calling method's operand stack

The Verifier

When a class is loaded, it is passed through a bytecode verifier before it is executed. The verifier checks the internal consistency of the class and the validity of the code. Java uses a late binding scheme that puts the code at risk. In traditional languages, the object linker binds all the method calls and variable accesses to specific addresses. In Java, the virtual machine doesn't perform this service until the last possible moment. As a result, it is possible for a called class to have changed since the original class was compiled. Method names or their arguments may have been altered, or the access levels may have been changed. One of the verifier's jobs is to make sure that all external object references are correct and allowed.

No assumptions can be made about the origin of bytecodes. A hostile compiler could be used to create executable bytecodes that conform to the class file format, but specify illegal codes.

The verifier uses a conservative four-pass verification algorithm to check bytecodes.

Pass 1

The first pass through the verifier reads in the class file and ensures that it is valid. The magic number must be present and all the class data must be present with no truncation or extra data after the end of the class. Any recognized attributes must have the correct length and the constant pool must not have any unrecognized entries.

Pass 2

The second pass through the verifier involves validating class features other than the bytecodes. All methods and fields must have a valid name and signature, and every class must have a super class. Signatures are not actually checked, but they must appear valid. The next pass is more specific.

Pass 3

The third pass is the most complex because the bytecodes are validated. The bytecodes are analyzed to make sure that they have the correct type and number of arguments. In addition, a data-flow analysis is performed to determine each path through the method. Each path must arrive at a given point with the same stack size and types. Each path must call methods with the proper arguments, and fields must be modified with values of the appropriate type. Class accesses are not checked in this pass. Only the return type of external functions is verified.

Forcing all paths to arrive with the same stack and registers can lead the verifier to fail some otherwise legitimate bytecodes. This is a small price to pay for this high level of security.

Pass 4

The fourth pass loads externally referenced classes and checks that the method name and signatures match. This pass also validates that the current class has access rights to the external class. After complete validation, each instruction is replaced with a _quick alternative. These _quick bytecodes indicate that the class has been verified and need not be checked again.

Bytecodes

The bytecodes can be divided into 11 major categories:

Each bytecode has a unique tag and is followed by a fixed number of additional arguments. Notice that there is no way to work directly with class fields or local variables. They must be moved to the operand stack before any operations can be performed on the contents.

Generally, there are multiple formats for each individual operation. The addition operation provides a good example. There are actually four forms of addition: iadd, ladd, fadd, and dadd. Each type assumes the top two stack items are of the correct format: integers, longs, floats, or doubles.

Pushing Constants onto the Stack

Java uses the following instructions for moving object data and local variables to the operand stack.

Push One-Byte Signed Integer:

bipush=16 byte1      Stack: ... -> ..., byte1

Push Two-Byte Signed Integer:

sipush=17 byte1 byte2    Stack: ... -> ..., word1

Push Item from the Constant Pool (8-bit index):

ldc1=18 indexbyte1    Stack: ... -> ..., item

Push Item from the Constant Pool (16-bit index):

ldc2=19 indexbyte1 indexbyte2   Stack: ... -> ..., item

Push Long or Double from Constant Pool (16-bit index):

ldc2w=20 indexbyte1 indexbyte2   Stack: ... -> ..., word1, word2

Push Null Object:

aconst_null=1    Stack: ... -> ..., null

Push Integer Constant -1:

iconst_m1=2   Stack: ... -> ..., -1

Push Integer Constants:

iconst_0=3   Stack: ... -> ..., 0
iconst_1=4   Stack: ... -> ..., 1
iconst_2=5   Stack: ... -> ..., 2
iconst_3=6   Stack: ... -> ..., 3
iconst_4=7   Stack: ... -> ..., 4
iconst_5=8   Stack: ... -> ..., 5

Push Long Constants:

lconst_0=9   Stack: ... -> ..., 0, 0
lconst_1=10  Stack: ... -> ..., 0, 1

Push Float Constants:

fconst_0=11   Stack: ... -> ..., 0
fconst_1=12   Stack: ... -> ..., 1
fconst_2=13   Stack: ... -> ..., 2

Push Double Constants:

dconst_0=14   Stack: ... -> ..., 0, 0
dconst_1=15   Stack: ... -> ..., 0, 1

Accessing Local Variables

The most commonly referenced local variables are at the first four offsets from the vars register. Because of this, Java provides single-byte instructions to access these variables for both reading and writing. A two-byte instruction is needed to reference variables greater than four deep. The variable at location zero is the class pointer itself (the this pointer).

Load Integer from Local Variable:

iload=21 vindex  Stack: ... -> ..., contents of varaible at vars[vindex]
iload_o=26       Stack: ... -> ..., contents of variable at vars[0]
iload_1=27       Stack: ... -> ..., contents of variable at vars[1]
iload_2=28       Stack: ... -> ..., contents of variable at vars[2]
iload_3=29       Stack: ... -> ..., contents of variable at vars[3]

Load Long Integer from Local Variable:

lload=22 vindex  Stack: .. -> ..., word1, word2  from vars[vindex] & vars[vindex+1]
lload_0=30       Stack: .. -> ..., word1, word2  from vars[0] & vars[1]
lload_1=31       Stack: .. -> ..., word1, word2  from vars[1] & vars[2]
lload_2=32       Stack: .. -> ..., word1, word2  from vars[2] & vars[3]
lload_3=33       Stack: .. -> ..., word1, word2  from vars[3] & vars[4]

Load Float from Local Variable:

fload=23 vindex  Stack: ... -> ..., contents from vars[vindex]
fload_0=34       Stack: ... -> ..., contents from vars[0]
fload_1=35       Stack: ... -> ..., contents from vars[1]
fload_2=36       Stack: ... -> ..., contents from vars[2]
fload_3=37       Stack: ... -> ..., contents from vars[3]

Load Double from Local Variable:

dload=24 vindex  Stack: ... -> ..., word1, word2  from vars[vindex] & vars[vindex+1]
dload_0=38       Stack: ... -> ..., word1, word2  from vars[0] & vars[1]
dload_1=39       Stack: ... -> ..., word1, word2  from vars[1] & vars[2]
dload_2=40       Stack: ... -> ..., word1, word2  from vars[2] & vars[3]
dload_3=41       Stack: ... -> ..., word1, word2  from vars[3] & vars[4]

Load Object from Local Variable:

aload=25 vindex  Stack: ... -> ..., object  from vars[vindex]
aload_0=42       Stack: ... -> ..., object  from vars[0]
aload_1=43       Stack: ... -> ..., object  from vars[1]
aload_2=44       Stack: ... -> ..., object  from vars[2]
aload_3=45       Stack: ... -> ..., object  from vars[3]

Store Integer into Local Variable:

istore=54 vindex Stack: ..., INT -> ...  into vars[vindex]
istore_0=59      Stack: ..., INT -> ...  into vars[0]
istore_1=60      Stack: ..., INT -> ...  into vars[1]
istore_2=61      Stack: ..., INT -> ...  into vars[2]
istore_3=62      Stack: ..., INT -> ...  into vars[3]

Store Long Integer into Local Variable:

lstore=55 vindex Stack: ..., word1, word2 -> ...  into vars[vindex] & vars[vindex+1]
lstore_0=63      Stack: ..., word1, word2 -> ...  into vars[0] & vars[1]
lstore_1=64      Stack: ..., word1, word2 -> ...  into vars[1] & vars[2]
lstore_2=65      Stack: ..., word1, word2 -> ...  into vars[2] & vars[3]
lstore_3=66      Stack: ..., word1, word2 -> ...  into vars[3] & vars[4]

Store Float into Local Variable:

fstore=56 vindex Stack: ..., FLOAT -> ...  into vars[vindex]
fstore_0=67      Stack: ..., FLOAT -> ...  into vars[0]
fstore_1=68      Stack: ..., FLOAT -> ...  into vars[1]
fstore_2=69      Stack: ..., FLOAT -> ...  into vars[2]
fstore_3=70      Stack: ..., FLOAT -> ...  into vars[3]

Store Double into Local Variable:

dstore=57 vindex Stack: ..., word1, word2 -> ...  into vars[vindex] & vars[vindex+1]
dstore_0=71      Stack: ..., word1, word2 -> ...  into vars[0] & vars[1]
dstore_1=72      Stack: ..., word1, word2 -> ...  into vars[1] & vars[2]
dstore_2=73      Stack: ..., word1, word2 -> ...  into vars[2] & vars[3]
dstore_3=74      Stack: ..., word1, word2 -> ...  into vars[3] & vars[4]

Store Object into Local Variable:

astore=58 vindex Stack: ..., OBJ -> ...  into vars[vindex]
astore_0=75      Stack: ..., OBJ -> ...  into vars[0]
astore_1=76      Stack: ..., OBJ -> ...  into vars[1]
astore_2=77      Stack: ..., OBJ -> ...  into vars[2]
astore_3=78      Stack: ..., OBJ -> ...  into vars[3]

Increment Local Variable (incrementing applies only to integers):

iinc=132 vindex constant   Stack: ... -> ...  vars[vindex] += constant

Managing Arrays

Arrays are treated as objects, but they don't use a method table pointer. Because of this uniqueness, arrays have special bytecodes to create and access them.

Allocate a New Array:

newarray=188 type Stack: ..., size -> ..., OBJ

Allocate a New Array of Objects:

anewarray=189 classindex1 classindex2 Stack: ..., size -> ..., OBJ

Allocate a New Multidimensional Array:

multianewarray=197 indexbyte1 indexbyte1 indexbyte2  Stack: ..., size1, size2, etc. -> ..., OBJ

Get the Array Length:

arraylength=190 Stack: ..., OBJ -> ..., length

Load Primitives from the Array:

iaload=46 Stack: ..., OBJ, index -> ..., INT
laload=47 Stack: ..., OBJ, index -> ..., LONG1, LONG2
faload=48 Stack: ..., OBJ, index -> ..., FLOAT
daload=49 Stack: ..., OBJ, index -> ..., DOUBLE1, DOUBLE2
aaload=50 Stack: ..., OBJ, index -> ..., OBJ
baload=51 Stack: ..., OBJ, index -> ..., BYTE
caload=52 Stack: ..., OBJ, index -> ..., CHAR
saload=53 Stack: ..., OBJ, index -> ..., SHORT

Store Primitives into the Array:

iastore=79 Stack: ..., OBJ, index, INT -> ...
lastore=80 Stack: ..., OBJ, index, LONG1, LONG2 -> ...
fastore=81 Stack: ..., OBJ, index, FLOAT -> ...
dastore=82 Stack: ..., OBJ, index, DOUBLE1, DOUBLE2 -> ...
aastore=83 Stack: ..., OBJ, index, OBJ -> ...
bastore=84 Stack: ..., OBJ, index, BYTE -> ...
castore=85 Stack: ..., OBJ, index, CHAR -> ...
sastore=86 Stack: ..., OBJ, index, SHORT -> ...

Generic Stack Instructions

Following are the basic operations that alter the stack.

Do Nothing:

nop=0   Stack: ... -> ...

Pop Stack Values:

pop=87  Stack: ..., VAL -> ...
pop2=88  Stack: ..., VAL1, VAL2 -> ...

Duplicate Stack Values and Possibly Insert Below Stack Top:

dup=89     Stack: ..., V -> ..., V, V
dup2=92    Stack: ..., V1, V2 -> ..., V1, V2, V1, V2
dup_x1=90  Stack: ..., V1, V2 -> ..., V2, V1, V2
dup2_x1=93 Stack: ..., V1, V2, V3 -> ..., V2, V3, V1, V2, V3
dup_x2=91  Stack: ..., V1, V2, V3 -> ..., V3, V1, V2, V3
dup2_x2=94 Stack: ..., V1, V2, V3, V4 -> ..., V3, V4, V1, V2, V3, V4

Swap Two Stack Items:

swap=95    Stack: ..., V1, V2 -> ..., V2, V1

Arithmetic and Logical Instructions

All the arithmetic operations operate on four possible types: integer, long, float, or double. Logical instructions operate only on integer and long types.

Addition:

iadd=96  Stack: ..., INT1, INT2 -> ..., INT1+INT2
ladd=97  Stack: ..., L1_1, L1_2, L2_1, L2_2 -> ..., L1+L2 (high), L1+L2 (low)
fadd=98  Stack: ..., FLOAT1, FLOAT2 -> ..., FLOAT1+FLOAT2
dadd=99  Stack: ..., D1_1, D1_2, D2_1, D2_2 -> ..., D1+D2 (high), D1+D2 (low)

Subtraction:

isub=100 Stack: ..., INT1, INT2 -> ..., INT1-INT2
lsub=101 Stack: ..., L1_1, L1_2, L2_1, L2_2 -> ..., L1-L2 (high), L1-L2 (low)
fsub=102 Stack: ..., FLOAT1, FLOAT2 -> ..., FLOAT1-FLOAT2
dsub=103 Stack: ..., D1_1, D1_2, D2_1, D2_2 -> ..., D1-D2 (high), D1-D2 (low)

Multiplication:

imul=104  Stack: ..., INT1, INT2 -> ..., INT1*INT2
lmul=105  Stack: ..., L1_1, L1_2, L2_1, L2_2 -> ..., L1*L2 (high), L1*L2 (low)
fmul=106  Stack: ..., FLOAT1, FLOAT2 -> ..., FLOAT1*FLOAT2
dmul=107  Stack: ..., D1_1, D1_2, D2_1, D2_2 -> ..., D1*D2 (high), D1*D2 (low)

Division:

idiv=108  Stack: ..., INT1, INT2 -> ..., INT1/INT2
ldiv=109  Stack: ..., L1_1, L1_2, L2_1, L2_2 -> ..., L1/L2 (high), L1/L2 (low)
fdiv=110  Stack: ..., FLOAT1, FLOAT2 -> ..., FLOAT1/FLOAT2
ddiv=111  Stack: ..., D1_1, D1_2, D2_1, D2_2 -> ..., D1/D2 (high), D1/D2 (low)

Remainder:

irem=112  Stack: ..., INT1, INT2 -> ..., INT1%INT2
lrem=113  Stack: ..., L1_1, L1_2, L2_1, L2_2 -> ..., L1%L2 (high), L1%L2 (low)
frem=114  Stack: ..., FLOAT1, FLOAT2 -> ..., FLOAT1%FLOAT2
drem=115  Stack: ..., D1_1, D1_2, D2_1, D2_2 -> ..., D1%D2 (high), D1%D2 (low)

Negation:

ineg=116  Stack: ..., INT -> ..., -INT
lneg=117  Stack: ..., LONG1, LONG2 -> ..., -LONG1, -LONG2
fneg=118  Stack: ..., FLOAT -> ..., -FLOAT
dneg=119  Stack: ..., DOUBLE1, DOUBLE2 -> ..., -DOUBLE1, -DOUBLE2

Integer Logical Instructions:

>>> denotes an unsigned right shift

ishl=120  Stack: ..., INT1, INT2 -> INT1<<(INT2 & 0x1f)
ishr=122  Stack: ..., INT1, INT2 -> INT1>>(INT2 & 0x1f)
iushr=124 Stack: ..., INT1, INT2 -> INT1>>>(INT2 & 0x1f)

Long Integer Logical Instructions:

>>> denotes an unsigned right shift

lshl=121  Stack: ..., L1, L2, INT -> L1<<(INT & 0x3f), L2<<(INT & 0x3f)
lshr=123  Stack: ..., L1, L2, INT -> INT1>>(INT & 0x3f), L2>>(INT & 0x03)
lushr=125 Stack: ..., L1, L2, INT -> INT1>>>(INT & 0x3f), L2>>>(INT & 0x3f)

Integer Boolean Operations:

iand=126  Stack: ..., INT1, INT2 -> ..., INT1&INT2
ior=128   Stack: ..., INT1, INT2 -> ..., INT1|INT2
ixor=130  Stack: ..., INT1, INT2 -> ..., INT1^INT2

Long Integer Boolean Operations:

land=127  Stack: ..., L1_1, L1_2, L2_1, L2_2 -> ..., L1_1&L2_1, L1_2&L2_2
lor=129   Stack: ..., L1_1, L1_2, L2_1, L2_2 -> ..., L1_1|L2_1, L1_2|L2_2
lxor=131  Stack: ..., L1_1, L1_2, L2_1, L2_2 -> ..., L1_1^L2_1. L1_2^L2_2

Conversion Instructions

Because most of the previous bytecodes expect the stack to contain a homogenous set of operands, Java uses conversion functions. In code, you can add a float and an integer, but Java will first convert the integer to a float type before performing the addition.

Integer Conversions:

i2l=133       Stack: .., INT -> ..., LONG1, LONG2
i2f=134       Stack: .., INT -> ..., FLOAT
i2d=135       Stack: .., INT -> ..., DOUBLE1, DOUBLE2
int2byte=145  Stack: .., INT -> ..., BYTE
int2char=146  Stack: .., INT -> ..., CHAR
int2short=147 Stack: .., INT -> ..., SHORT

Long Integer Conversions:

l2i=136       Stack: .., LONG1, LONG2 -> ..., INT
l2f=137       Stack: .., LONG1, LONG2 -> ..., FLOAT
l2d=138       Stack: .., LONG1, LONG2 -> ..., DOUBLE1, DOUBLE2

Float Conversions:

f2i=139       Stack: .., FLOAT -> ..., INT
f2l=140       Stack: .., FLOAT -> ..., LONG1, LONG2
f2d=141       Stack: .., FLOAT -> ..., DOUBLE1, DOUBLE2

Double Conversions:

d2i=142       Stack: .., DOUBLE1, DOUBLE2 -> ..., INT
d2l=143       Stack: .., DOUBLE1, DOUBLE2 -> ..., LONG1, LONG2
d2f=144       Stack: .., DOUBLE1, DOUBLE2 -> ..., FLOAT

Control Transfer and Function Return

All branch indexes are signed 16-bit offsets from the current pc register.

Comparisons with Zero:

ifeq=153 branch1 branch2   Stack: ..., INT -> ...
ifne=154 branch1 branch2   Stack: ..., INT -> ...
iflt=155 branch1 branch2   Stack: ..., INT -> ...
ifge=156 branch1 branch2   Stack: ..., INT -> ...
ifgt=157 branch1 branch2   Stack: ..., INT -> ...
ifle=158 branch1 branch2   Stack: ..., INT -> ...

Comparisons with Null:

ifnull=198 branch1 branch2    Stack: ..., OBJ -> ...
ifnonnull=199 branch1 branch2 Stack: ..., OBJ -> ...

Compare Two Integers:

if_icmpeq=159 branch1 branch2  Stack: ..., INT1, INT2 -> ...
if_icmpne=160 branch1 branch2  Stack: ..., INT1, INT2 -> ...
if_icmplt=161 branch1 branch2  Stack: ..., INT1, INT2 -> ...
if_icmpge=162 branch1 branch2  Stack: ..., INT1, INT2 -> ...
if_icmpgt=163 branch1 branch2  Stack: ..., INT1, INT2 -> ...
if_icmple=164 branch1 branch2  Stack: ..., INT1, INT2 -> ...

Compare Two Long Integers:

lcmp=148  Stack: ..., L1_1, L1_2, L2_1, L2_2 -> ..., INT (One of [-1, 0, 1])

Compare Two Floats:

l->-1 on NaN, g->1 on NaN.
fcmpl=149  Stack: ..., FLOAT1, FLOAT2 -> ..., INT (One of [-1, 0, 1])
fcmpg=150  Stack: ..., FLOAT1, FLOAT2 -> ..., INT (One of [-1, 0, 1])

Compare Two Doubles:

l->-1 on NaN, g->1 on NaN.
dcmpl=151  Stack: ..., D1_1, D1_2, D2_1, D2_2 -> ..., INT (One of [-1, 0, 1])
dcmpg=152  Stack: ..., D1_1, D1_2, D2_1, D2_2 -> ..., INT (One of [-1, 0, 1])

Compare Two Objects:

if_acmpeq=165 branch1 branch2  Stack: ..., OBJ1, OBJ2 -> ...
if_acmpne=166 branch1 branch2  Stack: ..., OBJ1, OBJ2 -> ...

Unconditional Branching (16-bit and 32-bit branching):

goto=167 branch1 branch2                    Stack: ... -> ...
goto_w=200 branch1 branch2 branch3 branch4  Stack: ... -> ...

Jump Subroutine (16-bit and 32-bit jumps):

jsr=168 branch1 branch2                     Stack: ... -> ..., returnAddress
jsr_w=201 branch1 branch2 branch3 branch4   Stack: ... -> ..., returnAddress

Return from Subroutine:

The return address is retrieved from a local variable, not the stack.

ret=169 vindex   Stack: ... -> ...   (returnAddress <- vars[vindex])
ret_w=209 vindex1 vindex2   Stack: ... -> ...   (returnAddress <- vars[vindex])

Return Primitives:

The current stack frame is destroyed. The top primitive is pushed onto the caller's operand stack.

ireturn=172  Stack: ..., INT -> [destroyed]
lreturn=173  Stack: ..., LONG1, LONG2 -> [destroyed]
freturn=174  Stack: ..., FLOAT -> [destroyed]
dreturn=175  Stack: ..., DOUBLE1, DOUBLE2 -> [destroyed]
areturn=176  Stack: ..., OBJ -> [destroyed]
return=177  Stack: ... -> [destroyed]

Calling the Breakpoint Handler:

breakpoint=202  Stack: ..., -> ...

Manipulating Object Fields

A 16-bit index into the constant pool is used to retrieve the class and field name. These names are used to determine the field offset and width. The object reference on the stack is used as the source or target. Values are 32 or 64 bits, depending on the field information in the constant pool.

Getstatic=178  index1 index2   Stack: ..., -> ..., VAL
Putstatic=179  index1 index2   Stack: ..., VAL -> ...
Getfield=180   index1 index2   Stack: ..., OBJ -> ..., VAL
Putfield=181   index1 index2   Stack: ..., OBJ, VAL -> ...

Method Invocation

There are four types of method invocation:

virtual index1 index2              Stack: ..., OBJ, [arg1, [arg2, ...]] -> ...
nonvirtual index1 index2           Stack: ..., OBJ, [arg1, [arg2, ...]] -> ...
static index1 index2               Stack: ..., [arg1, [arg2, ...]] -> ...
interface index1 index2 nargs resv Stack: ..., OBJ, [arg1, [arg2, ...]] -> ...

Miscellaneous Operations

The following instructions don't fall under any other heading; they deal with generic object operations, such as creation and casting.

Throw Exception:

athrow=191  Stack: ..., OBJ -> [undefined]

Create a New Object:

new=187 index1 index2  Stack: ... -> ..., OBJ

Check a Cast Operation:

checkcast=192 index1 index2  Stack: ..., OBJ -> ..., OBJ

Instanceof:

instanceof=193 index1 index2  Stack: ..., OBJ -> ... INT (1 or 0)

Monitors

Monitor instructions are used for synchronization.

Enter a Monitored Region of Code:

monitorenter=194  Stack: ..., OBJ -> ...

Exit a Monitored Region of Code:

monitorexit=195  Stack: ..., OBJ -> ...

Summary

This chapter revealed the internal structure of the Java virtual machine through use of the Viewer tool. You learned to interpret signatures and read Java machine code. The elegance of the Java class file was also presented.

Keep this information in mind when you are working with Java. An appreciation for the internal structure will help you become an expert Java programmer. I hope this introduction has heightened your curiosity; feel free to use the Viewer program to explore some of your own Java classes.