Skip to content

Latest commit

 

History

History
256 lines (184 loc) · 9.47 KB

README.md

File metadata and controls

256 lines (184 loc) · 9.47 KB

Foundation

Note

Foundation 3 is currently in the testing phase and may undergo changes without notice. Foundation 3 aims to be semi-compatible with Foundation 2 projects, but due to the change in its backing library, full compatibility cannot be guaranteed.

Opus #7

Foundation is a tool for writing virtual machine code, a class file assembler, and a library for creating compilers that target JVM bytecode.

Foundation 3 is modularised into the following libraries:

  • foundation-assembler: for assembling class files in a linear format (like ASM).
  • foundation-factory: a helper layer for building method code in the Java paradigm.
  • foundation-legacy: the former Foundation 2, for preserving compatibility.

Motivations

Foundation is designed to simplify the process of building classes and methods with bytecode. Tools like ASM's 'visitor' are very effective but make simple tasks unnecessarily difficult and add cumbersome boilerplate.

This over-complication makes writing simple bytecode-generating utilities more difficult than it needs to be and makes the process unpleasant for beginners, leading people to rely on tools such as 'ASMifier' to create the bytecode-generating code.

While such tools are undoubtedly useful (and appropriate in many situations) users can develop an unnecessary dependency on them and find themselves incapable of designing the bytecode without a tool to generate it for them.

Foundation is designed to set a balance between doing too much and too little, while also matching the line format and structure of Java code.

Foundation 3: Assembler

Foundation 3 entirely removes the dependency on ObjectWeb's ASM library, and contains its own fully-supported class file assembler.

Design Goals

In order to achieve its speed and low memory footprint, ASM has some incredibly strict design limitations:

  • Elements have to be visited in a specific order.
  • Operation codes have to be entered with a particular method.
  • Once visited, an element cannot be re-visited, accessed or altered in any way.
  • Safety checking and verification are left to the user.
  • Errors (particularly in frame calculation) are almost impossible to trace.

Foundation versions 1 and 2 attempted to overcome some of these limitations by storing all the details and then building the class file in a single pass at the end. While this was effective for overcoming the ordering constraints, it could not address any of the other problems listed above.

Foundation 3 includes its own assembler (a replacement for ASM) that was designed with overcoming these limitations in mind.

Ordering of Access & Creation

A key requirement for Foundation 3's assembler was not to limit the order in which class and code members could be created, as much as was feasibly possible. Since the class file itself has a fixed order in which elements appear (meta, fields, methods, attributes), this meant that the builder could not write the class bytes directly to the file, since a user might create a method and then go back and make a field.

The solution was to create a data graph format that was flexible enough to be edited. This brought with it an additional problem: lots of class file elements use indices to reference each other (e.g. constant index 3, jump forward 4 spaces), and if elements can be re-ordered then these indices will be invalidated.

This is most relevant to the class file's constant pool, which stores everything from the class names to the literal values used in code. The easiest solution (which ASM uses) is to enter things as they appear and then save the index at which they were entered. However, this has a drawback of leaving the constant pool in a very untidy state with no proper ordering, and also creates a forward referencing issue (where a constant pool entry might reference something that was entered after it, e.g. 3: class named #4, 4: 'java/lang/String'). This in itself is not prohibited by the virtual machine specification, but it is a problem for anything that is searching for a pool entry in order.

Foundation's solution to all of these problems is the tabular reference; a two-byte dynamic pointer to an entry in a table, such as a constant pool entry or an instruction in a code vector. These references mean the constant pool can be sorted, re-ordered or can have items added or removed without breaking any of the elements that reference it. During a phase called 'finalisation', which occurs before the class code is built, the references can be baked into their actual number format since no more changes to the table will occur.

Operation Codes

Definition: a 'well-formed instruction' consists of an operation code (opcode) followed by the bytes required to complete the instruction.

Method code is written in a sequence of operation codes. Some codes are atomic (followed by nothing else) and well-formed instructions in their own right, (e.g. iload_0/26) whereas others are followed by a fixed number of data bytes (e.g. iload/21 <index> or goto/167 <index>), and need other data to be provided in order to constitute a well-formed instruction (i.e. writing goto on its own does nothing, it must be followed by two branch bytes).

As a result, the process to write a bytecode changes according to the opcode used. This gives the designer a choice: either the designer allows the user to write whatever they want (e.g. write(byte opcode, byte... bytes), in which case it becomes the user's responsibility to read and understand the virtual machine specification and what each opcode requires, or the designer has to create individual processes for opcodes based on the information they require (e.g. writeJump(byte opcode, short target)). ASM follows the latter option, having individual visit methods for different instruction types.

The approach used by Foundation's assembler is a combination of the two. A single form of the write(Element) method is provided, and the process to create a well-formed instruction is attached to the opcode constants in mx.kenzie.foundation.assembler.code.OpCode.

For atomic codes, where the opcode alone is a well-formed instruction, the opcode constant can be entered as-is:

import mx.kenzie.foundation.assembler.tool.CodeBuilder;

class Example {

    void example(CodeBuilder builder) {
        builder.write(POP);
        builder.write(ALOAD_0, ILOAD_1);
        builder.write(SWAP);
    }

}

In the other case, a well-formed instruction can be created from a non-atomic opcode constant:

import mx.kenzie.foundation.assembler.tool.CodeBuilder;

class Example {

    void example(CodeBuilder builder) {
        builder.write(ALOAD.var(6), CHECKCAST.type(String.class));
    }

}

The advantage of this approach is that a well-formed instruction can be stored (or provided by some other part of a program) and then be entered into the code vector as-is without having to know which entry method to use.

Safety and Verification

Foundation 2 & Legacy Library

Since Foundation 2, instructions for the class builder do not directly correspond to JVM 'opcodes'. This was changed to make it easier and more intuitive to write code.

Foundation 2 aims to reflect Java code line structure, so that users do not need to worry about managing the stack.

// java
Object var1 = "hello there"
// builder
STORE_VAR.object(1, CONSTANT.of("hello there"))
// java
return var1
// builder
RETURN.object(LOAD_VAR.object(1)))

Maven Information

<repository>
    <id>kenzie</id>
    <url>https://repo.kenzie.mx/releases</url>
</repository>
<dependency>
    <groupId>mx.kenzie</groupId>
    <artifactId>foundation</artifactId>
    <version>3.0.0</version>
    <scope>compile</scope>
</dependency>

Supported Functionality

Foundation supports almost all basic instructions from Java, including variable and field access, method calls, branches, arithmetic and instantiation.

Foundation 2 has dropped support for individual bytecode instructions to provide more safety for beginners handling stack operations.

Examples

Generate and load a very simple class.

class MyClass {

    Class<?> test() {
        final PreClass builder = new PreClass("org.example", "Thing");
        final PreMethod method = new PreMethod(PUBLIC, STATIC, VOID, "main", String[].class);
        method.line(RETURN.none());
        builder.add(method);
        return builder.load(Loader.DEFAULT); // built-in basic class loader
    }

}

This would generate the code:

public class Thing {

    public static void main(String[] args) {
        return;
    }

}

Generate a runnable class.

class MyClass {

    Class<?> test() {
        // references to System.out and out.println(..)
        final CallMethod.Stub target = METHOD.of(PrintStream.class, "println", String.class);
        final AccessField.Stub field = FIELD.of(System.class, "out", PrintStream.class);

        final PreClass builder = new PreClass("org.example", "Thing");
        builder.addInterfaces(Runnable.class);
        final PreMethod method = new PreMethod(PUBLIC, VOID, "run");
        method.line(target.call(field.get(), CONSTANT.of("hello there!")));
        method.line(RETURN.none());
        builder.add(method);
        return builder.load(Loader.DEFAULT); // built-in basic class loader
    }

}

This would be the equivalent of:

class Thing implements Runnable {

    @Override // Overrides are implicit.
    public void run() {
        System.out.println("hello there!");
    }

}