The IDL compiler or bindings generator transcompiles Web IDL to C++ code, specifically bindings between V8 (the JavaScript engine) and Blink. That is, when an attribute or method in a Web IDL interface is used from JavaScript, V8 calls the bindings code, which calls Blink code.
As of early 2014, there are almost 700 IDL files, over 2,000 attributes, and almost 4,000 methods. This bindings code is highly repetitive, primarily passing values (arguments and return values) between V8 and Blink and doing various conversions and checks (e.g., type checking), hence it is mostly machine-generated.
This is complicated by numerous parameters and special cases, due to either wanting to automatically generate repetitive code (so Blink implementation is simpler), supporting quirks of Web interfaces, or Blink implementation details. The compiler is under active development, due to adding new features (bug 345506), adjusting to changes in V8 or Blink, and reducing technical debt, particularly custom bindings (manually written bindings, bug 345519). By far the majority of development occurs on the back end (code generator); the front end and overall structure is quite stable.
Users of the IDL compiler consist of:
IdlDefinitions).The compiler is in Source/bindings/scripts, and bindings are generated as part of the build (see IDL build). The top-level file idl_compiler.py, can be invoked directly (if you want to compile a single file), or via tools/run_bindings_tests.py. The rest of this document describes design, implementation, and use.
The compiler is factored into a pipeline, and the steps further divided into separate Python modules. While there are a number of components, each is simple and well-defined, and generally only one or two must be considered at a time.
There are complicated issues with global information such as dependencies (see below), but initially it is sufficient to consider compiling a single file.
The overall structure is standard for compilers (see compiler construction): a front end reads in input (IDL), yielding an intermediate representation (IR), and the back end takes the IR and writes out output (C++).
IDL (Foo.idl) --front end--> IR --back end--> C++ (V8Foo.h, V8Foo.cpp)
The flow is a (per file) pipeline: information flows one way, and, other than dependencies and global information, files are processed independently. In detail:
Front end: IDL --lexer--> tokens --parser--> AST --constructor--> IR Back end: IR --logic--> context --template processor--> C++
In terms of data types, objects, and modules, this is:
IdlCompiler: idl_filename --IdlReader--> IdlDefinitions --CodeGeneratorV8--> header_text, cpp_text
IdlReader: idl_filename --BlinkIDLLexer--> (internal) --BlinkIDLParser--> IDLNode
--idl_definitions_builder--> IdlDefinitions
CodeGeneratorV8: IdlDefinitions --(extract member)--> IdlInterface --v8_*.py--> dict
--jinja2--> header_text, cpp_text
One can trace the processing of each element of IDL input text through the pipeline, essentially 1-to-1:
p_Interface rule in idl_parser.IDLParser (base parser), producing an IDLNode (AST) named 'Interface'.IdlInterface is constructed from this tree in idl_definitions_builderv8_interface.interface_context function.cpp/.h output text by Jinja, from the templates interface.cpp (derived from interface_base.cpp) and interface.hThe boundary between the stages of the pipeline can move, primarily in the CG between the context and the template (depending on logic complexity: can it fit in the template, or should it be a Python function?). Changing the boundary between the FE and the CG is possible (e.g., moving “find indexed property getter” into the IdlInterface constructor, instead of the CG v8_interface module), but rare. In principle the constructor could be entirely integrated into the parser (have the yacc actions construct the object directly, instead of an AST), but this is unlikely, as it would require a significant amount of work for a relatively minor simplification.
Note that filenames, not IDL text, are what is input to the reader. This is fundamentally because the compiler and the parser assume they are reading in files (hence take a filename parameter, e.g. for reporting errors). Further, due to dependencies, processing one IDL file may require reading others (hence going out to the file system). This could be abstracted to taking in IDL text and passing interface names rather than filenames internally, but this isn't necessary at present.
The front end is familiar (regex lexer, LALR parser, and object constructor). The lexer-parser uses PLY (Python Lex-Yacc), a Python implementation of lex and yacc; the constructor is lengthy but straightforward.
However, the back end (code generator) differs from compilers that target machine code: since it is a source-to-source compiler, the back end outputs C++ (source code, not machine code). Thus we use a template processor to produce output; recall that templating (writing formatted output) is complementary to parsing (reading formatted input). The template processor is Jinja, which is used in several places in Chromium. Given an interface, containing various attributes and methods, we fill in a template for the corresponding C++ class and each attribute and method.
Further, the input language is declarative (the ‘D’ in IDL), so no optimizations of input code are necessary (there is no middle end): it's just filling in a template. The back end is itself divided into the templates themselves, and the Python code that fills in the templates (produces the context). There is also no separate semantic analysis step, except for validation of extended attributes (see below): the code generator assumes types are valid, and errors show up when the resulting C++ code fails to compile. This avoids the complexity and time cost of either a separate validation step, or of type-checking in the code generator, at the cost of typos showing up in compile failures instead of at IDL compile time. Notably, there is no type checking or name binding, since identifiers are assumed to refer to Web IDL interfaces (C++ classes) and the Web IDL namespace is global, and there is no assignment, since Web IDL is declarative.
Code-wise, the top-level file idl_compiler.py imports two modules: idl_reader for the front end, code_generator_v8 for the back end. Each of these is used to create an object (IdlReader and CodeGeneratorV8), which handles the library initialization (PLY or Jinja, which are slow, and thus reused if running multiple times during one run) and global information. The objects are then called one time each, for the IDL --> IR and IR --> C++ steps, respectively. By contrast, run_bindings_tests.py creates these objects as well, then calls them multiple times, once for each test file. Note that in the actual build, compilation is parallelized, which is why only one file is compiled per process, and there is a pre-caching step which significantly speeds up library initialization.
The code is mostly functional, except for a few module-level variables in the code generator (discussed below) for simplicity, and some objects used for initialization.
The main classes are as follows, with the module in which they are defined. Note that the relations are primarily composition, with inheritance significantly used for the Blink-specific lexer and parser. IdlCompiler, IdlReader, and CodeGeneratorV8 are expected to be used as singletons, and are just classes for initialization (avoids expensive re-initialization, and interface-wise separates initialization from execution). IdlDefinitions contains objects of a few other classes (for more minor data), and there is a bit of complex OO code in idl_definitions to simplify typedef resolution.
IdlCompiler :: idl_compiler
IdlReader :: idl_reader
BlinkIDLParser < IDLParser
BlinkIDLLexer < IDLLexer
IDLExtendedAttributeValidator
CodeGeneratorV8 :: code_generator_v8
IdlDefinitions :: idl_definitions
IdlInterface
IdlAttribute
IdlConstant
IdlOperation
IdlArgument
The front end is structured as a lexer --> parser --> object constructor pipeline:
Front end: IDL --lexer--> tokens --parser--> AST --constructor--> IR
There are two other steps:
The top-level module for the front end is idl_reader. This implements a class, IdlReader, whose constructor also constructs a lexer, parser, validator, and resolver. IdlReader implements a method to construct an IR object (IdlDefinitions) from an IDL filename. Thus to convert IDL to IR one instantiates a reader, then call its method to read an IDL file.
The lexer-parser uses PLY (Python Lex-Yacc). In fact, the lexer and parser for the Blink IDL dialect of Web IDL derive from a base lexer and base parser for standard Web IDL (in tools/idl_parser), and thus only need to include deviations from the standard. The lexical syntax of Blink IDL is standard (though the base lexer is slightly non-standard), so the Blink IDL lexer is very small and slated for removal (Bug 345137). The phrase syntax is slightly non-standard (primarily in extended attributes) and expected to stay this way, as extended attributes are implementation-dependent and the deviations are useful (see Blink IDL: syntax). We thus say that the base lexer/parser + Blink lexer/parser form 1 lexer and 1.1 parser (base + derived).
The base lexer class, defined in idl_lexer, is straightforward and short: it is just a list of regular expressions and keywords, wrapped in a class, IDLLexer; a lexer (object) itself is an instance of this class. There is some minor complexity with error handling (correct line count) and methods to assist with derived classes, but it‘s quite simple. The derived lexer class is very simple: it’s a class, BlinkIDLLexer, which derives from the base class. The only complexity is adding a method to remove a token class from the base class, and to remove comments (base lexer is non-standard here: standard lexer does not produce COMMENT tokens).
The base parser class, defined in idl_parser, is considerably longer, but consists almost entirely of production rules in (a form of) BNF, together with yacc actions that build an abstract syntax tree (AST, syntax tree). Recall that yacc traverses the concrete syntax tree (CST, parse tree) and can take whatever actions it chooses; in this case it generates an AST, though it could also generate the IR directly. See blink_idl_parser for a detailed explanation of the PLY yacc syntax and IDLParser methods used to construct the AST. The grammar is directly from the Web IDL grammar, which is an LL(1) grammar; the Blink IDL rules are exactly the deviations from the standard grammar, or overrides to irregularities in the base parser. A parser object is an instance of a class, IDLParser, from which BlinkIDLParser is derived. The parser constructor takes a lexer as an argument, and one passes the corresponding lexer.
The definitions classes, defined in idl_definitions, primarily consist of constructors, which take the AST and generate an intermediate representation (IR), in this case an object of type IdlDefinitions and contained objects for individual definitions and definition's members.
The classes are as follows (mostly composition, one case of inheritance); a few internal-use only classes are not shown:
IdlDefinitions
IdlInterface
IdlAttribute
IdlConstant
IdlOperation
IdlArgument
IdlException < IdlInterface
IdlCallbackFunction
IdlEnum
After reading an IDL file (producing an IR), the reader has two optional additional steps (these are run during the build, but aren't necessary for simpler use, such as compiling a test file without dependencies): validation and dependency resolution.
The front end largely does not do semantic analysis, as semantic errors (primarily name errors) are largely caught by the build visibly failing, either during the IDL compile or during the C++ compile (name lookup errors). Notably, there is no symbol table or name binding in the front end: each name is simply looked up on use (e.g., type names), or passed through to the output (e.g., attribute names), and catch errors by the lookup failing or the generated code failing to compile, respectively.
However, extended attributes are validated, both the keys and values, based on the list in IDLExtendedAttributes.txt. This is done because invalid extended attributes are ignored by the compiler, specifically the code generator, and thus errors are easy to miss. The Python code checks for specific extended attributes, so errors just result in the attribute not being found; for example, writing [EnforecRange] for [EnforceRange] would otherwise silently result in the range not being enforced. This is not perfect: extended attributes may be valid but used incorrectly (e.g., specified on an attribute when they only apply to methods), which is not caught; this is however a minor problem.
This is done in the front end since that is the proper place for semantic analysis, and simplifies the code generator.
IDL interface definitions have two types of dependencies:
implements (interface inheritance). We call these dependency interface definitions, dependency interfaces, or just dependencies.interface Foo { attribute Bar bar; }; We call these referenced interfaces.The dependency resolution phase consists of merging the dependencies into the main interface definition, and adding referenced interfaces to the list of all interfaces (available to the code generator). These dependencies are computed during the Global information stage; see below.
Members defined in partial interfaces and in implemented interfaces are simply appended to the list of members of the main interface, with a few caveats (below). Formally the difference between partial interfaces and implemented interfaces is that a partial interface is (external) type extension (it is part of and modifies the type, but is just defined separately), and a partial interface is associated with a single main interface that it extends; while implemented interfaces are interface inheritance (providing multiple inheritance in IDL) and several interfaces can implement a given (implemented) interface. This difference is only for IDL: in JavaScript these are both just exposed as properties on the object implementing the interface; and thus we mostly don't distinguish them in the code generator either, just merging them into one list.
The subtleties are:
[Conditional], [PerContextEnabled], and [RuntimeEnabled].[TypeChecking] extended attributes are also recognized on partial interfaces as a convenience. Its value is also transferred to its members, appending it to any [TypeChecking] values that the member already declares.Also, note that the compiler currently does not do significant type introspection of referenced interfaces: it mostly just uses certain global information about the interface (is it a callback interface?, [ImplementedAs], etc.). The only significant use of type introspection is in [PutForwards], where generating the code for an attribute with [PutForwards] requires looking up the referenced attribute in another interface. Type introspection may increase in future, to simplify complex cases and reduce hard-coding, but for now is in very limited use.
The back end (code generator, CG) is where the bulk of the complexity is. The structure is simple: the Python V8 CG modules (v8_*) generate a dictionary called the context, which is passed to Jinja, which uses this dict to fill in the appropriate templates (bindings/templates/*.{cpp,h}). The modules correspond to the class structure in IR, with the following exceptions:
v8_interface, because very simple; andv8_methods, because they are closely related to methods.Recall the overall flow of the CG:
CodeGeneratorV8: IdlDefinitions --(extract member)--> IdlInterface --v8_*.py--> dict
--jinja2--> header_text, cpp_text
The class structure of the IR, and corresponding Python module that processes it (in italics if included in a previously mentioned module), are:
IdlInterface :: v8_interface
IdlAttribute :: v8_attributes
IdlConstant :: _v8_interface_
IdlOperation :: v8_methods
IdlArgument :: _v8_methods_
The CG modules each have a top-level *_context function_,_ which takes an IR object (Idl*) as a parameter and returns a context (dict), looping through contained IR objects as necessary.
Note that the context is nested – the context for an interfaces contains a list of contexts for the members: attributes, constants, and methods. The context for a given member is used throughout the corresponding template, generally to generate several methods; e.g., in attributes.cpp there are macros for getter, getter callback, setter, and setter callback, which are called from a loop in interface_base.cpp. The context for members is also used at the overall interface level, notably for DOM configuration.
v8_attributes.py (Python) and attributes.cpp (Jinja) are a good guide to style: the getter and setter methods themselves are quite complex, while the callbacks are simple.
See Jinja: Style for general Jinja style tips; below is bindings generator specific guidance.
If nesting macros (so macros that call macros), use top-down order: top-level macro, followed by macros it calls. This is most notable in methods.cpp, where generating a method also requires generating code for the arguments, which comes after.
Assertions should not be used for validating input, and this includes IDL files: if an IDL file is invalid (e.g., it uses an extended attribute improperly), raise an exception explicitly. Assertions can be used for validating internal logic, but in practice this is rare in the code generator; see Using Assertions Effectively.
Deviations from the “One dictionary display per context” rule:
dict.update().No extended attributes in templates: A key point for the CG is that the raw extended attributes are not passed to the Jinja templates, and literal extended attribute names accordingly do not appear in templates. Instead, extended attributes are all passed via some variable in the context, either a content variable or a (boolean) control variable (e.g., deprecate_as for the [DeprecateAs] extended attribute and is_reflect for the [Reflect] extended attribute (this is only boolean since the value is used elsewhere, in the function for the getter/setter name)). This is primarily for consistency: in some cases variables must be used, notably if the extended attribute requires a side effect (includes), and thus putting all literal references to extended attributes in the Python logic keeps them together.
Includes: The CG code is primarily functional, with one significant exception: includes (for the .cpp file; the .h includes are simple). Includes are added to the .cpp file throughout the CG (as needed by various features), and account for a significant amount of the code bulk. This is handled by considering includes as a side effect, and storing them in a global variable (or rather a module level variable, as Python modules are singletons), specifically in v8_global. Alternatives include: being purely functional, and returning the includes from the functions, which makes the calls and returns bulky (due to all the passing backwards and forwards); or having all CG functions be methods of some object (CodeGeneratorV8), which adds a large amount of OO boilerplate. The only complexity of using module variables is that one needs to import it into each module, and must clear the includes if compiling multiple files in a single Python run (and can't compile multiple files concurrently in a single Python processes). This latter is ok because compilation is parallelized by distributing across separate processes, not by compiling multiple files in a single process.
Note that the context being computed is not a module-wide variable: for attributes, modules, and arguments, a context is computed for each of these (functionally), though for an interface only a single context is computed. In rare cases the context is also passed into functions as an argument (when later values depend on earlier computed values), which is primarily to flatten the module: it's clearer to pass explicitly, instead of having a nested function.
There are a few other module-level variables for the IDL global context, which are write-once or update-only: v8_globals.interfaces, for referenced interfaces, and a few in v8_types, for global and local type information (e.g., [ImplementedAs], enumerations). These are set at CodeGeneratorV8 construction (from interfaces_info), which is write-once, or CodeGeneratorV8.generate_code invocation (from definitions), which is update-only (can add some new enumerations, for instance, but shouldn't need or refer to any from other files).
Template rendering: the context is then passed into a Jinja template, via the Template.render method, which returns the generated code.
Global information is computed by a pre-processing step, dependency resolution and merging is done by the front end, and global information is used by the code generator in limited places.
FIXME
FIXME
Key goals of the compiler are as follows (in descending order of importance); these primarily affect the code generator:
Note that legibility of generated C++ source code is secondary to legibility of the code generator, as the CG is what is developed and reviewed, though the generated code should be legible if this does not conflict with other goals. As a rule, simple (one-line) changes to CG that simplify the generated source code are ok, but more complex changes that are only for source code legibility (that have no effect on object code) should be avoided. If you rely on the compiler (e.g., for dead code elimination), please write a comment to that effect.
Since the CG is complex enough as it is, a key part of the design is moving complexity out of the code generator, notably to the front end (building an easy-to-use IR) and to the global information step.
Fundamentally, why do we have our own compiler at all? Can't we reuse an existing program?
The answer is: we're reusing as much code as possible (lexer/parser and template library, and indeed a lexer/parser for the Web IDL standard), but we need custom code for 3 purposes:
Of these, the V8 CG is the main reason; Blink IDL dialect and the IR constructor exist to reduce complexity of the CG.
The compiler is written in Python, rather than C++, because hackability (ease of understanding and modification) is more important than build speed, and build speed is good enough.
A high-level language like Python is the usual choice for such tasks, and Python is the standard high-level language in Chromium: other languages (except JavaScript) are not widely used in Chromium, and JavaScript offers no advantages over Python.
Also, C++ code that generates C++ code risks being very confusing. In principle one could imagine writing the CG in template metaprogramming, but that would likely be hard to understand.
In principle C/C++ code may be used in the front end (lexer, parser, and constructor), particularly as the lexer and parser can be generated by lex-yacc, but the code generator is expected to be in Python.
We can‘t use SWIG because there’s in fact very little overlap. Essentially we use Jinja templates+Python logic for the backend, rather than SWIG templates.
While SWIG is a general “high-level language to C++ bindings generator”, and thus seems a natural fit, it doesn‘t fill our needs. We need a front end because we’re taking IDL files as input, and we need a code generator that supports all the specific features that Web IDL and Blink require: Web IDL has different semantics from JavaScript generally, and Blink has many IDL extended attributes and other implementation details. Further, there is a prototype V8 backend for SWIG, but it is incomplete (the JavaScript backend primarily targets JavaScriptCore). In principle we could have a compiler whose back end output SWIG templates, complete the general V8 backend for SWIG, and then modify this or fork it to make a Web IDL to V8 backend, customized for Blink. As this description makes clear, that involves significant complexity and few advantages over generating the C++ code directly, using Jinja instead of SWIG.
PLY (built on lex-yacc) and Jinja are standard and widely-used libraries for these tasks, and are used elsewhere in Chromium. We're open to other libraries if they are significantly superior (and we switch throughout Chromium), but these libraries are fit for purpose.
For overall IDL build performance, see: IDL build: Performance
The bindings are generated by running a separate compiler process for each interface (main IDL file), for parallelization and simplicity of build. This section discusses optimization of individual IDL compiler runs.
For compilation of an individual IDL file, the key factor in startup time, both Python startup time, and initialization time of libraries. This is because most IDL files are quite short and simple, so processing time (variable cost) is short relative to startup time (fixed cost). Currently (March 2014) the average sequential compilation time of a single IDL file on a fast Linux workstation is ~80 milliseconds, compared to Python startup time of ~4 milliseconds (skipping import of site with -S).
Naively, the key limiting factor is initialization time of the libraries (PLY and Jinja). This is expensive, as it requires compiling the lexer & parser or templates. Initialization is sped up by pre-caching these in a separate build step, which significantly improves build time. Note that run_bindings_tests.py will compile multiple test files in a single Python process_,_ which is quite fast, since initialization is done only once.
The compile has not been profiled (except by using time(1) on a build), since the low-hanging fruit were obvious (cache lexer and parser tables and templates), performance is now good enough, and significant additional improvements are unlikely without major work or reengineering.<
Processing time itself is presumably dominated by the main libraries (parsing via PLY and template rendering via Jinja), and possibly object construction (IdlDefinitions); the actual CG is likely quite cheap, as it's a simple iteration over the IR with a few conditionals and simple computations.
To our knowledge, there are no O(n2) or slower algorithms in the compiler – everything should be O(n) or faster – but some may be lurking somewhere.
Potential future improvements (without substantial reengineering) would primarily be to improve startup of some components, or rewrite some components in C.
.pyc or .pyo in separate step and execute that): v. small improvement (about .5 ms/process), would require other build step-O or -OO): likely little improvement (just shrinks size of bytecode), and currently causes serious regressions (likely due to breaking caching in a library)