Skip to main content
Jump to: navigation, search

JDT Core Programmer Guide/ECJ/Parse


In classical compiler construction, the input text is tokenized by a Scanner. The resulting token stream is then fed into the Parser which creates the abstract syntax tree (AST). ECJ essentially follows this design, but in reality things are more complex, because the Java syntax is no longer amenable to a strict implementation of this approach.


Responsibilities of the scanner:

  • interpret input unicode
  • recognize comments
    • recognize task tags within comments
  • record line ends (for translation between linear offsets and line based positions)
  • disambiguate tokens which depend on context. Some examples, why scanning needs more context than it should according to text books (cf. JLS §3.9):
    • token -> is ambiguous with respect to the two single tokens - and >
    • restricted keywords are opportunistically recognized when this enables the parser to recognize a legal module declaration (inside the modular compilation unit "").
    • restricted identifiers like var and yield must also be recognized by the scanner in certain contexts.

Two main concepts are used for token disambiguation:

  • A VanguardParser, is a stripped-down parser that runs the parse automaton at a location of ambiguity to check if a specific goal can be reached using the remaining input. In particular, the VanguardParser does not produce any AST, and it does not really consume any input.
  • The scanner keeps track of the current ScanContext in order to decide if a restricted keyword should be classified as a keyword or as an identifier.

To support the above, the scanner allows limited look ahead and the option to "unget" an erroneously consumed token (in order to allow re-classification).

The above implies a few restrictions on the use of a scanner:
  • many locations in JDT request partial scanning starting from a given context. If the input could possibly be a modular compilation unit, then the scanner needs to actually "parse" starting from the beginning of the text, in order to determine the ScanContext at the requested offset. This is done using ScanContextDetector which is a special VanguardParser.
  • ...
Matters are further complicated in variant CompletionScanner as used for code completion.

Scanner itself is 100% handcoded, but ScannerHelper contains several tables for unicode handling. Olivier Thomann is the expert for unicode handling.

Scanner has a public variant, PublicScanner which must be updated after each relevant change in Scanner. PublicScanner shields clients from generated constants in interface TerminalTokens and maps those to stable constants in ITerminalSymbols.


Constituents of the parser:

  • java.g is the grammar definition that is used by jikespg to create the following:
    • method Parser.consumeRule(int)
    • constants in interface TerminalTokens: every token detected by the scanner is represented using one of the constants in this interface. Constant names are composed of TokenName + the terminal name used in the grammar.
    • constants in interface ParserBasicInformation (used mostly by the parser automaton).
  • Most of the main Parser class, however, is hand-written.

Generating the parser (see also bug 562044):

  • install jikespg on your local machine
  • define a per-workspace String Substitution "JIKESPG" pointing to the location where jikespg is installed
  • run scripts/build-parser.launch

Main parsing mechanism:

  • the parser state is captured in various stacks, each of which is implementated by a pair of fields:
    • a field of an array type to contain the stack data
    • an int field representing the pointer to the current top element (initially -1 for: empty stack)
  • method Parser.parse() contains the main loop of the parse automaton
  • method Parser.consumeToken(int) generically records some information purely based on a token being consumed
  • method Parser.consumeRule(int) (generated) dispatches into the individual consumeXYZ methods.
  • each consume method may peek or pop elements from some stacks and push elements on one or more other stacks

Parsing modes

  • in diet mode, any blocks like method bodies are skipped, which can later be filled in using dedicated methods like Parser.parse(MethodDeclaration,CompilationUnitDeclaration).
  • fields Parser.methodRecoveryActivated and Parser.statementRecoveryActivated control whether or not attempts should be made to create AST despite syntax errors.

The Grammar

Grammar java.g is the input to the parser generator "jikespg". For lack of documentation concerning "jikespg" here is a set of documents from "LPG" (which is a variant derived from jikespg): -- use with caution, some details may not apply, most details are unnecessary for ecj

For normal maintenance work on ecj, however, the following hints might suffice for editing java.g:

  • Declarations at the top of the file are rarely modified, including options, defines (macros), terminal definitions, aliases.
    • Lines starting with Goal ::= are parser rules with the special intention to setup the parser for selectively parsing a subset of the grammar.
    E.g., the rule Goal ::= '&&' FieldDeclaration defines that assigning TokenNameAND_AND to Parser.firstToken will allow to parse an individual field declaration (see also Parser.goForFieldDeclaration() and friends).
    Goals are also used by the special VanguardParser when trying to disambiguate a token that could mean different things
  • The actual parser rules are written in a BNF dialect, below the directive $Rules
    • Two different symbols are used to map a non-terminal to a grammar expression: ::= and ->.
    Only rules using the former symbol carry actions to be performed by the parser (it is not clear if this is enforced or by convention only)

what’s the difference between ::= and ->, here is what we gathered by experimentation:

Both are equivalent in the sense they are both valid grammars without any difference in terms of meaning. However, when you use ->, then you are giving the parser generator the freedom to optimize this production out if there is a chance. If you use ::=, then you are mandating that this production needs to be there – don’t optimize this out even if there is a chance.

So why is this relevant to us? If you have a consume*() method which is dependent on the reduction of this production, and if that consume*() is *important* [it should be important or atleast it should be having a future use, otherwise you would not put that there in the first place], then you should use ::= so that we have that production and consequently that consume*() method in the Parser[.java]. Otherwise, you are free to use ->. You can see examples of both in java.g.

  • For lack of operators like |, alternatives are written by a set of rules with the same LHS non-terminal
  • The special word $empty matches the empty input
    • Non-terminals ending in opt by convention have one alternative producing $empty
  • Keywords are written in single quotes.
  • Parser actions are written below a rule
    • By convention each action looks like this: /. $putCase consumeXYZ(); $break ./, which will create one case statement for method Parser.consumeRule().
    • By convention, each consumeXYZ() method (which itself is written by hand) should contain as a code comment the grammar rules through which it is invoked
  • Minimal compliance level for the applicability of a rule is denoted like /:$compliance 1.8:/ (also on a separate line below its rule)
  • For display in syntax error messages each rule should define a readable name like /:$readableName QualifiedName:/

The tricky part about editing the grammar is the LALR(1) requirement. If this requirement is not fulfilled, jikespg will simply report this as an error and abort. Information will be available file java.l, but that's an overwhelming amount of it. It has been said that left-factoring can be a means to avoid conflicts, but the Java grammar is intrinsically not LALR(1), which gave rise to many tweaks and kludges like forced lookahead (VanguardParser and Scanner.lookBack) unusual Parser-Scanner communication etc. For ultimate confusion see JDT Core Programmer Guide/Completion#Lambda_Specifics

Peculiar parser rules
Grammar java.g contains a few unexpected rules:
  • Rules starting with Invalid or invoking a method consumeInvalid... explicitly expect some common error situations, in order to provide specific error messages.
  • Rules starting with Recovery are only invoked during syntax recovery, i.e., those may expect a valid Parser.currentElement (see below).
  • Rules of the shape ForceNoDiet ::= $empty are markers in the grammar, which do not correspond to any input text. Whenever the parser reaches such an empty non-terminal, a side effect is performed in the parser.
    • E.g.: EnumConstant ::= EnumConstantHeader ForceNoDiet ClassBody RestoreDiet ensures that the body of an enum constant is always parsed in non-diet mode.
    • Don't confuse such marker non-terminals with those rules terminating a recursion, like Argumentsopt ::= $empty

Parser Variants:

  • DiagnoseParser: after detecting a syntax error, this parser heuristically tries several "repairs", and selects the repair with the smallest "distance" to the actual input. During this process, the parser automaton is repeatedly run, without producing any output, other then accept or reject.
  • VanguardParser: as mentioned above, this parser is used for nested parse attempts in order to disambiguate certain tokens, that could be classified in different ways according to context.
  • AssistParser with subclasses CompletionParser and SelectionParser: these parsers focus on a specific text location and implement incomplete parsing optimized for the task at hand.
  • CodeSnippetParser: parses incomplete code for the sake of evaluation
  • CommentRecorderParser: used for created DOM AST
  • SourceElementParser: used for building structure of Java Elements (see Java Model)
  • IndexingParser: used for creating the JDT/Core index
  • DocumentElementParser: used for creating the obsolete JDOM (packages org.eclipse.jdt.core.jdom, org.eclipse.jdt.internal.core.jdom).
  • MatchLocatorParser: used during Java Search to provide input to the MatchLocator (which refines index matches into resolved java matches).

Stacks in detail

astStack is the target: in absence of syntax errors everything will eventually be aggregated into this stack and finally combined into a single ASTNode. astStack will only take some higher level ASTNodes, while specific stacks exist for "smaller" nodes.

astLengthStack is a companion to the above that provides a view as a stack of lists. Each element on this stack is the number of elements on the main astStack that should be seen as siblings at the same level. Examples are members of the same class, statements of the same block. If, e.g., the astStack contains 2 TypeDeclarations and 3 MethodDeclarions then an astLengthStack [ 2 3 ] will indicate that the three methods are siblings in the same list of methods, and likewise the two types are siblings in the same list of types.

  • method pushOnAstStack() always starts a new list (by advancing astLengthPtr) of initial length 1.
  • method concatNodeLists() combines the top two lists into a single list, typically the top list is a singleton (the most recently produced node) that is being appended to the current list, following a grammar rule like TypeDeclarations ::= TypeDeclarations TypeDeclaration

The same kind of companion stack (*LengthStack) exists for other stacks as well.

expressionStack as the name suggests, manages expressions while they are being assembled and before they are included into statements on the main astStack.

genericsStack is where type parameters and complex type references are assembled.

identifierStack is for identifiers that haven't even been integrated in any ASTNode.

intStack is a multipurpose stack mostly used for source positions. Other purposes include array dimensions, modifiers, and more. Due to the mix of ints with different meaning, bugs related to an unbalanced intStack are particularly difficult to analyze. So balancing pushs and pops for this stack for all cases should be done with great care.

Automaton stack

While all the above stacks hold work-in-progress elements of the target AST, parsing itself is controlled by the master stack stack. Each value on this stack is a "state" capturing information what tokens have been seen, and which tokens can legally follow. Sometimes these states are called "act" or "action". The central method tAction takes the previous act plus the current token to produce a new act (its implementation directly leverages the tables generated by jikespg). acts are grouped as shift, shift-reduce, reduce and ERROR states. During shift, a new token will be accepted from the scanner. When the parser has enough information for reduce, the current act is passed to consumeRule where it will select which of the consume* methods is invoked (consumeRule being generated by jikespg as well). This is when contents of the other stacks will be modified.

To see the automaton in action consider setting DEBUG_AUTOMATON to true (but don't commit this change! :) ).


Upon any syntax error, syntax recovery happens using the following concepts:

  • field Parser.currentElement holds a "recovered" element which wraps a possibly incomplete AST node.
    • each class below RecoveredElement models a node in the recovered tree, with generic functions for adding new child nodes etc.
    • currentElement acts like a cursor into the tree of recovered elements, with parent links to the rest of the tree.
  • field Parser.lastCheckpoint marks a position up-to which parsing is considered "done". When hitting a syntax error, parsing will restart at the last checkpoint.
  • method Parser.resumeOnSyntaxError() applies a number of heuristics to put the parser into a state where parsing can resume.
    • initially buildInitialRecoveryState() converts existing ASTNodes in Parser.astStack into RecoveredElements.
    • More elements will be added later using RecoveredElement.add(..). Each execution of add(..) will return either
      • this to signal that more elements should be added to the same node, or
      • parent to signal that the current element is complete and subsequent details should be added to the enclosing element.
    • Expressions are explicitly omitted from the structure of Recovered*. To enable recovering details from a lambda body, the following tweaks are applied:
      • The lambda block is inserted as a direct child of the current element (omitting the lambda itself)
      • If the current element is a local variable declaration (which cannot accept a block as its child), the local variable is considered as complete and the lambda block is inserted into the parent, i.e., as a sibling of the local variable declaration.
  • much of the incremental updating of recovered elements relates to source positions:
    • updateOnClosingBrace(), e.g., may set source end positions of a method being recovered.
    • source end positions, on the other hand, are used in add(..) to detect whether the current element is "complete".
  • recovery may proceed in several iterations starting at a slowly moving lastCheckpoint. Each iteration will start with empty stacks, but ast nodes contained in the recovered elements tree may survive this reset.
  • in the end updateParseTree() will collect the information from the recovered elements tree, and update the actual parse tree accordingly, such that Parser.compilationUnit will contain the fully recovered AST.

HOWTO: Grammar Changes

eg: Let us say you want to add a new statement, MyNewStatement. Grammar changes start with java.g - It should be a context free grammar since we use an LALR(1) parser You can add your MyNewStatement similar to that of EnhancedForStatement to the RHS of Statement, ie Statement -> MyNewStatement and then followed by the new rules (obeying the context free rule) with a few consumeMyNewStatement(), other consume* methods.

The parser generator jikespg is used to generate the parser from this grammar. You may want to take the latest bits of the jikespg from [1] since it contains a few bug fixes and then make. Additionally, the generation of parser has been made easier - ref [2]. If you want to look at the original documentation of how the parser is generated see [3].

Let us say, now you have the modified java.g and the jikespg parser ready. First run jikespg.exe java.g to check whether your grammar is LALR(1).

Assuming that it is, then you need to just run org.eclipse.jdt.core/scripts/build-parser.launch. In order to provide the path to jikespg you need to define a per-workspace String Substitution "JIKESPG", see also the usage notes inside build-parser.xml.

This would generate the files as described in [3], but what would interest you most would the - here you will see your consumeMyNewStatement() and other consume* stubs inserted - both the method call as per the generated automata (which you should not touch) and the declaration - you would need to add code to the declaration, to reduce and possibly create a new AST Node.

The AST Node created would be the internal compiler AST Node and not the DOM AST node - they have similar/identical definitions - so you need to create something similar to org.eclipse.jdt.internal.compiler.ast.ForeachStatement, which is the enhanced for statement node of compiler AST - look at org.eclipse.jdt.internal.compiler.parser.consumeEnhancedForStatement() - internally, it creates a ForeachStatement - [we are not talking of DOM AST node - the conversion to the DOM AST Node EnhancedForStatement comes later in the ASTConverter method and is not relevant here - mentioning to avoid confusion]

The most tricky part about constructing new AST nodes is to correctly manage the various stacks of the parser. As a general rule, each consume method should consume all ast nodes mentioned on its RHS and replace them by one node representing the LHS. The first distinction to be made is between astStack and expressionStack. Less obvious is the use of intStack which is also modified inside consumeToken, e.g.

Generating the parser separately with jikespg

(1) You can also run the jikespg separately and then bring the generated files. Once you copy all the generated files into the grammar folder [where java.g lives], set the env in Eclipse JIKESPG=JIKESPG_EXTERNAL to signal to the script to skip running the jikespg. Read below on how to use a gerrit job to run jikespg for your change

(2) To make things easier for generation of java.g, a job has been created - refer(Details verbatim from)

- Go to: - Build with Parameters -> here give the Egit refs (GERRIT_REFSPEC) - Goto to the patch and use copy option and click the Egit ref - something similar to refs/changes/31/189931/1

- Once the job completes, you'll see the output under build artifacts of the job run.

- Clicking in build artifacts will give you an option to download all files as zip or individual files.-





Back to the top