breadcrumbs: Blink (Rendering Engine) > chromium_link: Chromium chromium_os_link: Chromium OS home_link: Home page_name: webidl title: Web IDL in Blink title_hdr: |
Blink developers (non-bindings development): for general IDL use, see Web IDL interfaces; for configuring bindings, see Blink IDL Extended Attributes; for IDL dictionaries use, see IDL dictionaries in Blink.
Web IDL is a language that defines how Blink interfaces are bound to V8. You need to write IDL files (e.g. xml_http_request.idl, element.idl, etc) to expose Blink interfaces to those external languages. When Blink is built, the IDL files are parsed, and the code to bind Blink implementations to V8 interfaces automatically generated.
This document describes practical information about how the IDL bindings work and how you can write IDL files in Blink. The syntax of IDL files is fairly well documented in the Web IDL spec, but it is too formal to read :-) and there are several differences between the Web IDL spec and the Blink IDL due to implementation issues. For design docs on bindings generation, see IDL build and IDL compiler.
For Blink developers, the main details of IDLs are the extended attributes, which control implementation-specific behavior: see Blink IDL Extended Attributes for extensive details.
Our goal is to converge Blink's IDL and Web IDL. The grammar is almost identical; see below.
Here is an example of IDL files:
[CustomToV8] interface Node { const unsigned short ELEMENT_NODE = 1; attribute Node parentNode; [TreatReturnedNullStringAs=Null] attribute DOMString nodeName; [Custom] Node appendChild(Node newChild); void addEventListener(DOMString type, EventListener listener, optional boolean useCapture); };
Let us introduce some terms:
Node
interface.ELEMENT_NODE
is a constant of the Node
interface.parentNode
and nodeName
are attributes of the Node
interface.appendChild(...)
and addEventListener(...)
are operations of the Node
interface.type
, listener
and useCapture
are arguments of the addEventListener
operation.[CustomToV8]
, [TreatReturnedNullStringAs=Null]
and [Custom]
are extended attributes.The key points are as follows:
The valid extended attributes depend on where they attach: interfaces and methods have different extended attributes.
A simple IDL file template looks like:
interface INTERFACE_NAME { const unsigned long value = 12345; attribute Node node; void func(long argument, ...); };
With extended attributes, this looks like:
[ EXTATTR, EXTATTR, ..., ] interface INTERFACE_NAME { const unsigned long value = 12345; [EXTATTR, EXTATTR, ...] attribute Node node; [EXTATTR, EXTATTR, ...] void func([EXTATTR, EXTATTR, ...] optional [EXTATTR] long argument, ...); };
Blink IDL is a dialect of Web IDL. The lexical syntax is identical, but the phrase syntax is slightly different.
Implementation-wise, the lexer and parser are written in PLY (Python lex-yacc), an implementation of lex and yacc for Python. A standard-compliant lexer is used (Chromium tools/idl_parser/idl_lexer.py). The parser (Blink bindings/scripts/blink_idl_parser.py) derives from a standard-compliant parser (Chromium tools/idl_parser/idl_parser.py).
Blink deviations from the Web IDL standard can be seen as the BNF production rules in the derived parser.
Style
Style guidelines are to generally follow Blink style for C++, with a few points highlighted, addenda, and exceptions. These are not enforced by a pre-submit test, but do assist legibility:
// TODO(name or bug link):
comments[]
list if it's empty. Examples of Blink style:[ A, B /* No trailing commas on the last extended attribute */ ] interface Foo { ... }; interface Bar { ... };
getter DOMString (unsigned long index); // Not: DOMString(unsigned long index)
index
, and the (first) argument to named property operations should be called name
; the second argument in property setters should be called value
. For example:// Indexed property operations getter DOMString (unsigned long index); setter DOMString (unsigned long index, DOMString value); deleter boolean (unsigned long index); // Named property operations getter DOMString (DOMString name); setter DOMString (DOMString name, DOMString value); deleter boolean (DOMString name);
Web IDL exposes an interface to JavaScript, which is implemented in C++. Thus its semantics bridge these two languages, though it is not identical to either. Web IDL's semantics are much closer to C++ than to JavaScript – in practice, it is a relatively thin abstraction layer over C++. Thus C++ implementations are quite close to the IDL spec, though the resulting interface is somewhat unnatural from a JavaScript perspective: it behaves differently from normal JavaScript libraries.
See: Web IDL types.
Primitive types in Web IDL are very close to fundamental types in C++ (booleans, characters, integers, and floats), though note that there is no int
type in Web IDL (specs usually use long
instead).
JavaScript has two special values, undefined
and null
, which are often confusing and do not fit easily into C++. Indeed, precise behavior of undefined
in Web IDL has varied over time and is under discussion (see W3C Bug 23532 - Dealing with undefined).
Behavior on undefined
and null
MUST be tested in web tests, as these can be passed and are easy to get wrong. If these tests are omitted, there may be crashes (which will be caught by ClusterFuzz) or behavioral bugs (which will show up as web pages or JavaScript libraries breaking).
For the purposes of Blink, behavior can be summarized as follows:
undefined
and null
are valid values for basic types, and are automatically converted."undefined"
and "null"
for DOMString
.[Clamp]
and [EnforceRange]
.[Clamp]
changes the value so that it is valid.[EnforceRange]
throws a TypeError
on these invalid values.undefined
and null
are both treated as null
, which maps to nullptr
.Node?
, null
is a valid value, and thus nullptr
is passed to the C++ implementationNode
(no ?
), null
is not a valid value, and a TypeError
is thrown, as in JavaScript ToObject.[LegacyInterfaceTypeChecking]
is specified on the interface or member (see Bug 249598: Throw TypeError when null is specified to non-nullable interface parameter)[LegacyInterfaceTypeChecking]
is specified in the IDL, you do not need to have a null check in the Blink implementation, as the bindings code already does this, but if [LegacyInterfaceTypeChecking]
is not specified, you do need to have a null check in the Blink implementation.undefined
and null
both correspond to an empty dictionaryundefined
and null
are assigned to a type that can accept them, if possible: null, empty dictionary, or conversion to basic typeundefined
affects function resolution, both as an optional argument and for overloaded operations, basically being omitted if trailing (but some exceptions apply). This is complicated (see W3C Bug 23532 - Dealing with undefined) and not currently implemented. Further, note that in some cases one wants different behavior for f()
and f(undefined)
, which requires an explicit overload, not an optional argument; a good example is Window.alert()
, namely alert()
vs. alert(undefined)
(see W3C Bug 25686).null
affects function resolution for overloaded operations, due to preferring nullable types, but this is the only effect.Web IDL has required arguments and optional arguments. JavaScript does not: omitted arguments have undefined
value. In Web IDL, omitting optional arguments is not the same as explicitly passing undefined
: they call have different behavior (defined in the spec prose), and internally call different C++ functions implementing the operation.
Thus if you have the following Web IDL function declaration:
interface A { void foo(long x); };
...the JavaScript a = new A(); a.foo()
will throw a TypeError
. This is specified in Web IDL, and thus done by the binding code.
However, in JavaScript the corresponding function can be called without arguments:
function foo(x) { return x } foo() // undefined
Note that foo()
and foo(undefined)
are almost identical calls (and for this function have identical behavior): it only changes the value of arguments.length
.
To get similar behavior in Web IDL, the argument can be explicitly specified as optional
(or more precisely, optional
with undefined
as a default value). However, these do not need to have the same behavior, and do not generate the same code: the spec may define different behavior for these calls, and the bindings call the implementing C++ functions with a different number of arguments, which is resolved by C++ overloading, and these may be implemented by different functions.
For example, given an optional argument such as:
interface A { void foo(optional long x); };
This results in a = new A(); a.foo() being legal, and calling the underlying Blink C++ function implementing foo
with no arguments, while a.foo(undefined)
calls the underlying Blink function with one argument.
For overloaded operations, the situation is more complicated, and not currently implemented in Blink (Bug 293561). See the overload resolution algorithm in the spec for details.
Pragmatically, passing undefined
for an optional argument is necessary if you wish to specify a value for a later argument, but not earlier ones, but does not necessarily mean that you mean to pass in undefined
explicitly; these instead get the special value “missing”.
Passing undefined
to the last optional argument has unclear behavior for the value of the argument, but does mean that it resolves it to the operation with the optional argument, rather than others. (It then prioritizes nullable types and dictionary types, or unions thereof.) For example:
interface A { void foo(optional long x); void foo(Node x); };
This results in a = new A(); a.foo(undefined) resolving to the first foo
, it is not clear if the resulting call is a.foo()
, to a.foo
with “missing”, or (most likely) a.foo(undefined)
(here using the first overloaded function): it affect overload resolution, but perhaps not argument values. Note that undefined
is also a legitimate value for the argument of Node
type, so it would not be illegal, but the overload resolution algorithm first picks optional arguments in this case.
Note that Blink code implementing a function can also check arguments, and similarly, JavaScript functions can check arguments, and access the number of arguments via arguments.length
, but these are not specified by the language or checked by bindings.
Warning: undefined
is a valid value for required arguments, and many interfaces depend on this behavior particularly booleans, numbers, and dictionaries. Explicitly passing undefined
, as in a.foo(undefined)
, does not cause a type error (assuming foo
is unary). It is clearer if the parameter is marked as optional
(this changes semantics: the argument can now also be omitted, not just passed explicitly as undefined
), but this is not always done in the spec or in Blink's IDL files.
The Web IDL spec treats the Web API as a single API, spread across various IDL fragments. In practice these fragments are .idl
files, stored in the codebase alongside their implementation, with basename equal to the interface name. Thus for example the fragment defining the Node
interface is written in node.idl
, which is stored in the third_party/blink/renderer/core/dom
directory, and is accompanied by node.h
and node.cc
in the same directory. In some cases the implementation has a different name, in which case there must be an [ImplementedAs=...]
extended attribute in the IDL file, and the .h/.cc
files have basename equal to the value of the [ImplementedAs=...]
.
For simplicity, each IDL file contains a single interface or dictionary, and contains all information needed for that definition, except for dependencies (below), notably any enumerations, implements statements, typedefs, and callback functions.
In principle (as a matter of the Web IDL spec) any IDL file can depend on any other IDL file, and thus changing one file can require rebuilding all the dependents. In practice there are 4 kinds of dependencies (since other required definitions, like enumerations and typedefs, are contained in the IDL file for the interface):
partial interface
– a single interface can be spread across a single main interface
statement (in one file) and multiple other partial interface
statements, each in a separate file (each partial interface
statement is associated with a single main interface
statement). In this case the IDL file containing the partial interface has some other name, often the actual interface name plus some suffix, and is generally named after the implementing class for the members it contains. From the point of view of spec authors and compilation, the members are just treated as if they appeared in the main definition. From the point of view of the build, these are awkward to implement, since these are incoming dependencies, and cannot be determined from looking at the main interface IDL file itself, thus requiring a global dependency resolution step.implements
– this is essentially multiple inheritance: an interface can implement multiple other interfaces, and a given interface can be implemented by multiple other interfaces. This is specified by implements statements in the implementing file (these are outgoing dependencies), though from the perspective of the build the interface → .idl filename of that interface data is required, and is global information (because the .idl files are spread across the source tree).In practice, what happens is that, when compiling a given interfaces, its partial interfaces and the other interfaces it implements are merged into a single data structure, and that is compiled. There is a small amount of data recording where exactly a member came from (so the correct C++ class can be called), and a few other extended attributes for switching the partial/implemented interface on or off, but otherwise it is as if all members were specified in a single interface
statement. This is a deep dependency relationship: any change in the partial/implemented interface changes the bindings for the overall (merged) interface, since all the data is in fact used.
Bindings for interfaces in general do not depend on their ancestors, beyond the name of their immediate parent. This is because the bindings just generate a class, which refers to the parent class, but otherwise is subject to information hiding. However, in a few cases bindings depend on whether the interface inherits from some other interface (notably EventHandler or Node), and in a few cases bindings depend on the extended attributes of ancestors (these extended attributes are “inherited”; the list is compute_dependencies.INHERITED_EXTENDED_ATTRIBUTES, and consists of extended attributes that affect memory management). There is thus a shallow dependency on ancestors, specifically only on the ancestor chain and on inherited extended attributes, not on the other contents of ancestors.
On the other hand, the dependencies on used interfaces – so-called cross dependencies – are generally shallow dependency relationships: the using interface does not need to know much about the used interface (currently just the name of the implementing class, and whether the interface is a callback interface or not). Thus almost all changes in the used interface do not change the bindings for the using interface: the public information used by other bindings is minimal. There is one exception, namely the [PutForwards]
extended attribute (in standard Web IDL), where the using interface needs to know the type of an attribute in the used interface. This “generally shallow” relationship may change in future, however, as being able to inspect the used interface can simplify the code generator. This would require the using interface to depend on used interfaces, either rebuilding all using interfaces whenever a used interface is changed, or clearly specifying or computing the public information (used by code generator of other interfaces) and depending only on that.
To avoid bugs caused by typos in extended attributes in IDL files, the extended attribute validator was introduced to the Blink build flow to check if all the extended attributes used in IDL files are implemented in the code generator. If you use an extended attribute not implemented in code generators, the extended attribute validator fails, and the Blink build fails.
A list of IDL attributes implemented in code generators is described in IDLExtendedAttributes.txt. If you want to add a new IDL attribute, you need to
Note that the validator checks for known extended attributes and their arguments (if any), but does not enforce correct use of the the attributes. A warning will not be issued if, for example, [Clamp]
is specified on an interface.
third_party/blink/tools/run_bindings_tests.py tests the code generator, including its treatment of extended attributes. Specifically, run_bindings_tests.py compiles the IDL files in bindings/tests/idls, and then compares the results against reference files in bindings/tests/results. For example, run_bindings_tests.py reads test_object.idl, and then compares the generated results against v8_test_object.h and v8_test_object.cc, reporting any differences.
If you change the behavior of the code generator or add a new extended attribute, please add suitable test cases, preferably reusing existing IDL files (this is to minimize size of diffs when making changes to overall generated bindings). You can reset the run-bindings-tests results using the --reset-results option:
third_party/blink/tools/run_bindings_tests.py --reset-results
run_bindings_tests.py is run in a presubmit script for any changes to Source/bindings: this requires you to update test results when you change the behavior of the code generator, and thus if test results get out of date, the presubmit test will fail: you won't be able to upload your patch via git-cl, and the CQ will refuse to process the patch.
The objective of run-bindings-tests is to show you and reviewers how the code generation is changed by your patch. If you change the behavior of code generators, you need to update the results of run-bindings-tests.
Despite these checks, sometimes the test results can get out of date; this is primarily due to dcommitting or changes in real IDL files (not in Source/bindings) that are used in test cases. If the results are out of date prior to your CL, please rebaseline them separately, before committing your CL, as otherwise it will be difficult to distinguish which changes are due to your CL and which are due to rebaselining due to older CLs.
Note that using real interfaces in test IDL files means changes to real IDL files can break run-bindings-tests (e.g., Blink r174804/CL 292503006: Oilpan: add [WillBeGarbageCollected] for Node., since Node is inherited by test files). This is ok (we‘re not going to run run_bindings_tests.py on every IDL edit, and it’s easy to fix), but something to be aware of.
It is also possible for run_bindings_tests.py to break for other reasons, since it use the developer's local tree: it thus may pass locally but fail remotely, or conversely. For example, renaming Python files can result in outdated bytecode (.pyc files) being used locally and succeeding, even if run_bindings_tests.py is incompatible with current Python source (.py), as discussed and fixed in CL 301743008.
To test behavior, use web tests, most simply actual interfaces that use the behavior you‘re implementing. If adding new behavior, it’s preferable to make code generator changes and the first actual use case in the same CL, so that it is properly tested, and the changes appear in the context of an actual change. If this makes the CL too large, these can be split into a CG-only CL and an actual use CL, committed in sequence, but unused features should not be added to the CG.
For general behavior, like type conversions, there are some internal tests, like bindings/webidl-type-mapping.html, which uses testing/type_conversions.idl. There are also some other IDL files in testing, like testing/internals.idl.
By reading this document you can learn how extended attributes work. However, the best practice to understand extended attributes is to try to use some and watch what kind of bindings code is generated.
If you change an IDL file and rebuild (e.g., with ninja or Make), the bindings for that IDL file (and possibly others, if there are dependents) will be rebuilt. If the bindings have changed (in ninja), or even if they haven't (in other build systems), it will also recompile the bindings code. Regenerating bindings for a single IDL file is very fast, but regenerating all of them takes several minutes of CPU time.
In case of xxx.idl in the Release build, the bindings code is generated in the following files (“Release” becomes “Debug” in the Debug build).
out/Release/gen/third_party/blink/renderer/bindings/{core,modules}/v8_xxx.{h,cc}
A few parts of the Web IDL spec are not implemented; features are implemented on an as-needed basis. See component:Blink>Bindings for open bugs; please feel free to file bugs or contact bindings developers (members of blink-reviews-bindings, or bindings/OWNERS) if you have any questions, problems, or requests.
Bindings generation can be controlled in many ways, generally by adding an extended attribute to specify the behavior, sometimes by special-casing a specific type, interface, or member. If the existing extended attributes are not sufficient (or buggy), please file a bug and contact bindings developers!
Some commonly encountered limitations and suitable workarounds are listed below. Generally limitations can be worked around by using custom bindings, but these should be avoided if possible. If you need to work around a limitation, please put a TODO
with the bug number (as demonstrated below) in the IDL so that we can remove the hack when the feature is implemented.
Some syntax errors cause the IDL parser to enter an infinite loop (Bug 363830). Until this is fixed, if the compiler hangs, please terminate the compiler and check your syntax.
The bindings do not do full type checking (Bug 321518). They do some type checking, but not all. Notably nullability is not strictly enforced. See [TypeChecking]
under undefined and null above to see how to turn on more standard type checking behavior for interfaces and members.
Derived from: http://trac.webkit.org/wiki/WebKitIDL Licensed under BSD:
BSD License
Copyright (C) 2009 Apple Inc. All rights reserved.
Redistribution and use in source and binary forms, with or without modification, are permitted provided that the following conditions are met:
Redistributions of source code must retain the above copyright notice, this list of conditions and the following disclaimer.
Redistributions in binary form must reproduce the above copyright notice, this list of conditions and the following disclaimer in the documentation and/or other materials provided with the distribution.
THIS SOFTWARE IS PROVIDED BY APPLE INC. AND ITS CONTRIBUTORS “AS IS” AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL APPLE INC. OR ITS CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.