A Hitchhiker's Guide to ECJ

What IS the Compiler / ECJ?

Strange enough this question does not have a single true answer.

Project organization before 4.27

Project refactoring
As of https://github.com/eclipse-jdt/eclipse.jdt.core/issues/181 the project structure has been refactored. This section describes the state before this refactoring.

The following locations contribute to the compiler:

org.eclipse.jdt.core/compiler
org.eclipse.jdt.core/batch
org.eclipse.jdt.compiler.tool
org.eclipse.jdt.compiler.apt

Since the compiler does not directly correspond to any project / plug-in the following measures are relevant:

Classes in source folders compiler and batch are not allowed to access classes in other source folders of org.eclipse.jdt.core. To avoid any violations, a secondary project has been created: org.eclipse.jdt.core.ecj.validation. This project should be imported into the workspace before working on the compiler. It contains only links to the two mentioned source folders and will signal errors, if any class outside this scope is used. The project is not intended for editing.
During production builds class files from different projects need to be merged into the single ecj.jar (this jar file is created as org.eclipse.jdt.core-*-SNAPSHOT-batch-compiler.jar and renamed to ecj.jar afterwards). Search for "batch-compiler" in pom files of the projects mentioned above, to see how the compiler is assembled.
Additionally, an ant script exists, org.eclipse.jdt.core/scripts/export-ecj.xml, that should allow manually creating ecj.jar from within Eclipse. This script is also executed when building org.eclipse.jdt.core using PDE/Build, probably happening also when interactively exporting org.eclipse.jdt.core as a deployable plug-in using the export wizard.

Project organization since 4.27 M1

Removed projects:
With the ecj project refactoring the following projects have been removed:

org.eclipse.jdt.compiler.tool - fragment merged into org.eclipse.jdt.core.compiler.batch
org.eclipse.jdt.compiler.apt - fragment merged into org.eclipse.jdt.core.compiler.batch
org.eclipse.jdt.core.ecj.validation - no longer relevant

The compiler resides in plug-in org.eclipse.jdt.core.compiler.batch. This plug-in has no dependencies at runtime and a (hidden) optional compile time dependency to ant.jar (see jars.extra.classpath entry in build.properties) and can thus be used in standalone mode. When deployed as a jar file, this plug-in serves as the "batch compiler" and is identical to the file ecj.jar.

For use in the IDE (or other OSGi-applications), clients may still refer to the compiler via org.eclipse.jdt.core which re-exports all relevant packages from org.eclipse.jdt.core.compiler.batch.

The ant adapter

For using ecj with ant, jdtCompilerAdapter.jar is created from selected classes inside org.eclipse.jdt.core.compiler.batch:

org.eclipse.jdt.internal.antadapter.*
org.eclipse.jdt.core.JDTCompilerAdapter

The same class files are also added to ecj.jar, but JDTCompilerAdapter is not exported in MANIFEST.MF and thus not visible in OSGI to other bundles.

Documentation about peculiarities of this jar file are described in some more detail in https://github.com/eclipse-jdt/eclipse.jdt.core/blob/master/org.eclipse.jdt.core.compiler.batch/src/org/eclipse/jdt/core/README.md

Creating jdtCompilerAdapter.jar

jdtCompilerAdapter.jar is packaged in project org.eclipse.jdt.core (!). The reason for doing this "remotely" (i.e., not in ..compiler.batch) is in the fact that org.eclipse.jdt.core needs a dedicated library containing only ant related classes for registration with extension point org.eclipse.ant.core.extraClasspathEntries - and this extension point has to be inside org.eclipse.jdt.core bundle, because JDTCompilerAdapter class from batch compiler should be "not visible" to OSGI to avoid conflicts between Ant and OSGI loaded classes at (OSGI) runtime (see the readme).

When interactively exporting org.eclipse.jdt.core from the IDE, creation of jdtCompilerAdapter.jar is managed by PDE/Build in a cooperation between customBuildCallbacks.xml and scripts/export-ecj.jar.

During maven builds, jdtCompilerAdapter.jar is created by copying required classes compiled in org.eclipse.jdt.core.compiler.batch module to the org.eclipse.jdt.core and bundling them together as jar.

Using jdtCompilerAdapter.jar

In all PDE/Build scenarios (e.g., when interactively exporting arbitrary plug-ins from the IDE) jdtCompilerAdapter.jar is used by ANT/PDE IDE code as an entry point into the compiler (see org.eclipse.ant.internal.core.AntClassLoader).

Interfacing with other components

Name Environments: To interface with its environment, the compiler needs an instance of org.eclipse.jdt.internal.compiler.env.INameEnvironment. During batch compilation, class org.eclipse.jdt.internal.compiler.batch.FileSystem is used. But using different implementations of this interface other components like the builder can provide required classes into the compiler.
- In terms of JLS the name environment implements the "host system", see in particular JLS §7.2.
IBinaryType: Different use cases use different implementations to represent existing .class files to which Java sources being compiled can refer.
ITypeRequestor: Whenever a new type is found from the name environment, it is first passed to methods of the type requestor. Normally, the Compiler itself acts as the type requestor, which will add a representation of the discovered type to the internal data structures of the compiler, but code assist, type hierarchy, search and indexing each have their own implementation of these hooks into the compiler.

Variants of the compiler

The central class is org.eclipse.jdt.internal.compiler.Compiler, which is used as-is in some use cases, but also a few subclasses exist, which are variants of the compiler, with purposes different from generating .class files. In other use cases, not the Compiler class, but org.eclipse.jdt.internal.compiler.parser.Parser is subclassed to achieve different functionality. The latter strategy is used notably for code select and code complete functionality.

Phases

As is standard in compiler technology, ecj operates on Java files in several phases, which are roughly outlined as:

Scan and parse, i.e., transform a character stream first into a stream of tokens, then into the abstract syntax tree (AST)
Build and connect type bindings, i.e., overlay the syntactic tree structure (AST) with a semantic graph of bindings.
Verify methods: analyse inheritance, overriding and overloading of methods
Resolve: interpret identifiers and link them to the bindings which they represent using various lookups.
Analyse: perform flow analysis in order to detect errors like variables read before assigned, final variables re-assigned, and also analysis of (potential) null pointers and resource leaks. This phase may also detect a few more errors that need the AST to be fully resolved.
Generate: Allocate positions to variables (as used in load and store operations of the byte code), then generate the byte code, in the steps shown below. Note that still during code generation some errors may be detected and reported.
- Generate the general class file structure with relevant byte code attributes
- Generate the Code attributes containing the actual byte code instructions for methods, constructors and initializers.

Looking at class Compiler the phases are written slightly differently:

beginToCompiler/internalBeginToCompile:
- parse or dietParse
- buildTypeBindings
- completeTypeBindings, here bindings are linked / connected with each other, which requires all bindings to already exist.
optionally: processAnnotations
processCompilationUnits / process -- at this point a separate compilation thread may be spawned: ProcessTaskManager
- getMethodBodies: the initial parse may have skipped method bodies, parse them now, perhaps only selectively
- faultInTypes: ensure that all bindings are properly created and initialized
- verifyMethods
- resolve
- analyseCode
- generateCode
- finalizeProblems: before errors and warnings are actually reported to the user, they are filtered by any @SuppressWarnings annotations found in the source.

Processing order

Sequentiel phases vs. demand-driven computations

In addition to the sequential process outlined by the phases above, some computations will be triggered on demand.

Caveat
JDT had several bugs that were caused by demand-driven computations happening at an unexpected times. The question, which computation can safely be invoked at which point during compilation, doesn't have a simple answer. Some assumptions have never been made explicit.

As an example, invoking ReferenceBinding.getAnnotationTagBits() can cause subtle errors when the receiver is a SourceTypeBinding. Then a typical effect is re-entrance of a compilation step that is not prepared for re-entrance.

Other candidates that have caused misunderstandings in the past are methods unResolvedMethods and unResolvedFields (declared in ReferenceBinding), which may be doing more than what their names suggest.

Hint
Throughout the compiler implementation, many fields are public and are directly accessed all over the place. Obviously, direct field access has no effect on compilation order, hence in terms of processing order, this is OK. If, however, a field only has an accessor method, or if the field has documentation saying it should not be accessed directly, this is a start of a hint that more stuff may be happening than just reading a field, which is the first step towards influencing the order of compilation steps.

Existing tricks to fine-tune order of processing steps:

ClassScope.deferredBoundChecks:

Inside ParameterizedQualifiedTypeReference.internalResolveLeafType we normally perform a boundCheck(). However, in some situations, notably during Scope.connectTypeVariables() we may not be ready yet to perform that check. If argument checkBounds is false, the check is deferred, adding an element to the list deferredBoundChecks, which will be processed via ClassScope.checkParameterizedTypeBounds()

Similarly MemberValuePair.resolveTypeExpecting(..) may add runnables to the same list deferredBoundChecks. This accounts for the fact that resolving annotations may happen at particularly unexpected points in time.

LookupEnvironment.deferredEnumMethods:

During scanMethodForNullAnnotation() we want to mark the generated enum methods "valueOf" and "values" as returning a nonnull type. To do so we want to add the configured nonnull annotation, but null annotations may not yet be initialized, in particular because "@NonNullByDefault" depends on an enum, whose "valueOf" and "values" methods should be marked as returning nonnull. To cut this circular dependency, we check if null annotations have been initialized, and if not we add the enum method to deferredEnumMethods for processing from usesNullTypeAnnotations().

LocalDeclaration.duplicateCheckObligation:

This paragraph is outdated!

To handle a specific scoping issue of instanceof pattern variables, the current implementation admits possibly-duplicate variables during LocalDeclaration.resolve(), because at that point we don't yet have the necessary flow information. Only later during analyseCode we process that deferred duplicateCheckObligation.

Note, that to be true to the spec, more situations need flow information right during resolving, so the entire design of separating phases resolve and analyseCode is at stake, see bug 562824

.

Subpages

JDT Core Programmer Guide/ECJ/AST	JDT Core Programmer Guide/ECJ/Analyse	JDT Core Programmer Guide/ECJ/Bindings
JDT Core Programmer Guide/ECJ/Generate	JDT Core Programmer Guide/ECJ/Investigating	JDT Core Programmer Guide/ECJ/Lambda
JDT Core Programmer Guide/ECJ/Lookups	JDT Core Programmer Guide/ECJ/Parse	JDT Core Programmer Guide/ECJ/Testing

Compilation phases

Data structures

Java concepts

Some Java concepts pose specific challenges for the compiler

/Lambda

Strategies for working on the compiler

Breadcrumbs

JDT Core Programmer Guide/ECJ

Contents