Tuesday, June 18, 2013

JikesRVM Optimizing (JIT) Compiler - I

Jikes RVM includes two compilers. One is the baseline compiler, and the other is optimizing compiler, which is commonly referred to as the JIT compiler. BaseBase builds run JikesRVM with only the Baseline Compiler, where the other builds such as Production (FastAdaptive), Development (FullAdaptive), and the ExtremeAssertion builds have both the baseline and optimizing compilers.
Here, we will study the optimizing compiler in Jikes RVM. The package org.jikesrvm.compilers contains the source code of the compilers, where The package org.jikesrvm.compilers.opt contains the code of the optimizing compiler.

SharedBooleanOptions.dat found in the directory /rvm/src-generated/options defines the boolean options for the optimizing compiler. Similarly, SharedValueOptions.dat in the same directory defines the non-boolean options.

1. Transformations of Methods

Jikes RVM takes methods as the fundamental unit of optimization, and optimize them by transforming the intermediate representation as below.

Byte Code → [Optimizing Compiler] → Machine Code + Mapping Information

Mapping information consists of garbage collection maps, source code maps, and execution tables. Optimizing compiler has the following intermediate phase transitions, for the intermediate representation.

High-level Intermediate Representation (HIR) → Low-level Intermediate Representation → Machine Intermediate Representation.

The general structure of the master plan consists of elements,

  1. Converting, byte codes → HIR, HIR → LIR, LIR → MIR, and MIR → Machine Code.
  2. Performing optimization transformations on the HIR, LIR, and MIR.
  3. Performing optimization.

2. CompilationPlan

CompilationPlan is the major class, which instructs the optimizing compiler how to optimize a given method. Its constructor constructs a compilation plan, as the name indicates. This includes the instance of NormalMethod, the method to be optimized. An array of TypeReference is used in place of those defined in the method.

This also contains an object of InstrumentationPlan, defining how to instrument a method to gather the run time measurement information. The methods initInstrumentation() and finalizeInstrumentation() are called at the beginning of the compilation, and after the compilation just before the method is executed, respectively.

A compilation plan is executed by executing each element in the optimization plan, by calling the method execute(). The compiler queries the InlineOracle interface to decide whether to inline a call site. If there is some instrumentation to be performed, initialization takes place. After the instrumentation, finalization takes place to clean up. However, the finalization will not be executed, if this fails with an exception. 

3. OptimizationPlanElement
Class diagram of OptimizationPlanElement

An array of OptimizationPlanElement defines the compilation steps. OptimizationPlanElement thus represents an element in the optimization plan. Instances of the sub classes of the abstract class OptimizationPlanElement are held in OptimizationPlanner.masterPlan, and hence they represent the global state. It is incorrect to store any per-compilation state in the instance field of these objects. OptimizationPlanner specifies the order of execution during the HIR and LIR phases for a method. The method reportStats() generates a report on time spent performing the element.

The method shouldPerform() determines whether the optimization plan should be performed and be part of the compilation plan, by consulting the passed OptOptions object. When an element is included, all the aggregate elements that the element is a component of, are included too. The work represented by the element is done in the optimization plan, by perform(). The work done is assumed to modify the Intermediate Representation (IR) in some way. The perform() of the aggregate element will invoke all the elements' perform().

A client is a compiler driver that constructs a specific optimization plan by including all the OptimizationPlanElements from the master plan, that are appropriate for this compilation instance.

4. OptimizationPlanAtomicElement

The final class OptimizationPlanAtomicElement extends OptimizationPlanElement. This object consists of a single compiler phase in the compiler plan. The main work of the class is done by its phase, which is an instance of the class CompilerPhase. Each phase overrides this abstract class, and specifically the abstract methods perform(), which does the actual job, and getName(), which gets the name of the phase. A new instance of the phase is created when shouldPerform() is called, and is discarded as soon as the method is finished. Hence, the per-compilation state is contained in the instances of the sub classes of CompilerPhase, as the state is not stored in the element.

5. OptimizationPlanCompositeElement

Similarly, OptimizationPlanCompositeElement is the base class of all the elements in the compiler plan that aggregates together the other OptimizationPlan elements, as depicted in the Figure. Here an array of OptimizationPlanElement is kept to store the elements that compose this composite element. The constructor composes together the elements passed through as a method parameter, into the composite element. If the phase wants the IR to be dumped before or after it runs, it can be enabled using printingEnabled(). By default, this is disabled, where the sub classes are free to override this method, and make it true.

This post continues at JikesRVM Optimizing (JIT) Compiler - II.

No comments:

Post a Comment

You are welcome to provide your opinions in the comments. Spam comments and comments with random links will be deleted.