JavaScriptCore:
2009-05-13 Darin Adler <Darin Adler>
Reviewed by Cameron Zwarich.
Bug 25674: syntax tree nodes should use arena allocation
https://bugs.webkit.org/show_bug.cgi?id=25674
Step 3: Add some actual arena allocation. About 1% SunSpider speedup.
- bytecompiler/BytecodeGenerator.cpp:
(JSC::BytecodeGenerator::BytecodeGenerator): Updated since VarStack
contains const Identifier* now.
(JSC::BytecodeGenerator::emitPushNewScope): Updated to take a const
Identifier&.
- bytecompiler/BytecodeGenerator.h: Ditto
- bytecompiler/SegmentedVector.h: Added isEmpty.
- debugger/Debugger.cpp:
(JSC::Debugger::recompileAllJSFunctions): Moved this function here from
WebCore so WebCore doesn't need the details of FunctionBodyNode.
- debugger/Debugger.h: Ditto.
- interpreter/Interpreter.cpp:
(JSC::Interpreter::execute): Updated since VarStack contains const
Identifier* now.
- jit/JITStubs.cpp:
(JSC::JITStubs::cti_vm_lazyLinkCall): Call isHostFunction on the body
rather than on the function object, since we can't easily have inlined
access to the FunctionBodyNode in JSFunction.h since WebCore needs
access to that header.
(JSC::JITStubs::cti_op_construct_JSConstruct): Ditto.
- profiler/Profiler.cpp:
(JSC::Profiler::createCallIdentifier): Ditto.
- parser/Grammar.y: Use JSGlobalData* to pass the global data pointer
around whenever possible instead of using void*. Changed
SET_EXCEPTION_LOCATION from a macro to an inline function. Marked
the structure-creating functions inline. Changed the VarStack to use
identifier pointers instead of actual identifiers. This takes
advantage of the fact that all identifier pointers come from the
arena and avoids referenc count churn. Changed Identifier* to
const Identifier* to make sure we don't modify any by accident.
Used identifiers for regular expression strings too, using the new
scanRegExp that has out parameters instead of the old one that relied
on side effects in the Lexer. Move the creation of numeric identifiers
out of this file and into the PropertyNode constructor.
- parser/Lexer.cpp:
(JSC::Lexer::setCode): Pass in ParserArena, used for identifiers.
(JSC::Lexer::makeIdentifier): Changed return type to const Identifier*
and changed to call ParserArena.
(JSC::Lexer::scanRegExp): Added out arguments that are const Identifier*
as well as a prefix character argument so we can handle the /= case
without a string append.
(JSC::Lexer::skipRegExp): Added. Skips a regular expression without
allocating Identifier objects.
(JSC::Lexer::clear): Removed the code to manage m_identifiers, m_pattern,
and m_flags, and added code to set m_arena to 0.
- parser/Lexer.h: Updated for changes above.
- parser/NodeConstructors.h:
(JSC::ParserArenaFreeable::operator new): Added. Calls allocateFreeable
on the arena.
(JSC::ParserArenaDeletable::operator new): Changed to call the
allocateDeletable function on the arena instead of deleteWithArena.
(JSC::RegExpNode::RegExpNode): Changed arguments to Identifier instead
of UString since these come from the parser which makes identifiers.
(JSC::PropertyNode::PropertyNode): Added new constructor that makes
numeric identifiers. Some day we might want to optimize this for
integers so it doesn't create a string for each one.
(JSC::ContinueNode::ContinueNode): Initialize m_ident to nullIdentifier
since it's now a const Identifier& so it can't be left uninitialized.
(JSC::BreakNode::BreakNode): Ditto.
(JSC::CaseClauseNode::CaseClauseNode): Updated to use SourceElements*
to keep track of the statements rather than a separate statement vector.
(JSC::BlockNode::BlockNode): Ditto.
(JSC::ForInNode::ForInNode): Initialize m_ident to nullIdentifier.
- parser/Nodes.cpp: Moved the comment explaining emitBytecode in here.
It seemed strangely out of place in the header.
(JSC::ThrowableExpressionData::emitThrowError): Added an overload for
UString as well as Identifier.
(JSC::SourceElements::singleStatement): Added.
(JSC::SourceElements::lastStatement): Added.
(JSC::RegExpNode::emitBytecode): Updated since the pattern and flags
are now Identifier instead of UString. Also changed the throwError code
to use the substitution mechanism instead of doing a string append.
(JSC::SourceElements::emitBytecode): Added. Replaces the old
statementListEmitCode function, since we now keep the SourceElements
objects around.
(JSC::BlockNode::lastStatement): Added.
(JSC::BlockNode::emitBytecode): Changed to use emitBytecode instead of
statementListEmitCode.
(JSC::CaseClauseNode::emitBytecode): Added.
(JSC::CaseBlockNode::emitBytecodeForBlock): Changed to use emitBytecode
instead of statementListEmitCode.
(JSC::ScopeNodeData::ScopeNodeData): Changed to store the
SourceElements* instead of using releaseContentsIntoVector.
(JSC::ScopeNode::emitStatementsBytecode): Added.
(JSC::ScopeNode::singleStatement): Added.
(JSC::ProgramNode::emitBytecode): Call emitStatementsBytecode instead
of statementListEmitCode.
(JSC::EvalNode::emitBytecode): Ditto.
(JSC::EvalNode::generateBytecode): Removed code to clear the children
vector. This optimization is no longer possible since everything is in
a single arena.
(JSC::FunctionBodyNode::emitBytecode): Call emitStatementsBytecode
insetad of statementListEmitCode and check for the return node using
the new functions.
- parser/Nodes.h: Changed VarStack to store const Identifier* instead
of Identifier and rely on the arena to control lifetime. Added a new
ParserArenaFreeable class. Made ParserArenaDeletable inherit from
FastAllocBase instead of having its own operator new. Base the Node
class on ParserArenaFreeable. Changed the various Node classes
to use const Identifier& instead of Identifier to avoid the need to
call their destructors and allow them to function as "freeable" in the
arena. Removed extraneous JSC_FAST_CALL on definitions of inline functions.
Changed ElementNode, PropertyNode, ArgumentsNode, ParameterNode,
CaseClauseNode, ClauseListNode, and CaseBlockNode to use ParserArenaFreeable
as a base class since they do not descend from Node. Eliminated the
StatementVector type and instead have various classes use SourceElements*
instead of StatementVector. This prevents those classes from having th
use ParserArenaDeletable to make sure the vector destructor is called.
- parser/Parser.cpp:
(JSC::Parser::parse): Pass the arena to the lexer.
- parser/Parser.h: Added an include of ParserArena.h, which is no longer
included by Nodes.h.
- parser/ParserArena.cpp:
(JSC::ParserArena::ParserArena): Added. Initializes the new members,
m_freeableMemory, m_freeablePoolEnd, and m_identifiers.
(JSC::ParserArena::freeablePool): Added. Computes the pool pointer,
since we store only the current pointer and the end of pool pointer.
(JSC::ParserArena::deallocateObjects): Added. Contains the common
memory-deallocation logic used by both the destructor and the
reset function.
(JSC::ParserArena::~ParserArena): Changed to call deallocateObjects.
(JSC::ParserArena::reset): Ditto. Also added code to zero out the
new structures, and switched to use clear() instead of shrink(0) since
we don't really reuse arenas.
(JSC::ParserArena::makeNumericIdentifier): Added.
(JSC::ParserArena::allocateFreeablePool): Added. Used when the pool
is empty.
(JSC::ParserArena::isEmpty): Added. No longer inline, which is fine
since this is used only for assertions at the moment.
- parser/ParserArena.h: Added an actual arena of "freeable" objects,
ones that don't need destructors to be called. Also added the segmented
vector of identifiers that used to be in the Lexer.
- runtime/FunctionConstructor.cpp:
(JSC::extractFunctionBody): Use singleStatement function rather than
getting at a StatementVector.
- runtime/FunctionPrototype.cpp:
(JSC::functionProtoFuncToString): Call isHostFunction on the body
rather than the function object.
- runtime/JSFunction.cpp:
(JSC::JSFunction::JSFunction): Moved the structure version of this in
here from the header. It's not hot enough that it needs to be inlined.
(JSC::JSFunction::isHostFunction): Moved this in here from the header.
It's now a helper to be used only within the class.
(JSC::JSFunction::setBody): Moved this in here. It's not hot enough that
it needs to be inlined, and we want to be able to compile the header
without the definition of FunctionBodyNode.
- runtime/JSFunction.h: Eliminated the include of "Nodes.h". This was
exposing too much JavaScriptCore dependency to WebCore. Because of this
change and some changes made to WebCore, we could now export a lot fewer
headers from JavaScriptCore, but I have not done that yet in this check-in.
Made a couple functions non-inline. Removes some isHostFunction() assertions.
- wtf/FastAllocBase.h: Added the conventional using statements we use in
WTF so we can use identifiers from the WTF namespace without explicit
namespace qualification or namespace directive. This is the usual WTF style,
although it's unconventional in the C++ world. We use the namespace primarily
for link-time disambiguation, not compile-time.
- wtf/FastMalloc.cpp: Fixed an incorrect comment.
WebCore:
2009-05-13 Darin Adler <Darin Adler>
Reviewed by Cameron Zwarich.
Bug 25674: syntax tree nodes should use arena allocation
https://bugs.webkit.org/show_bug.cgi?id=25674
- bindings/js/JSDOMBinding.h: Removed include of JSFunction.h.
We don't want the entire DOM binding to depend on that file.
- bindings/js/JSAudioConstructor.cpp: Added include of Error.h.
Before we inherited this automatically because JDDOMBinding.h
included JSFunction.h, but that was excessive.
- bindings/js/JSDOMWindowCustom.cpp: Ditto.
- bindings/js/JSHTMLInputElementCustom.cpp: Ditto.
- bindings/js/JSImageConstructor.cpp: Ditto.
- bindings/js/JSLazyEventListener.cpp: Ditto, but for JSFunction.h.
- bindings/js/JSMessageChannelConstructor.cpp: Ditto.
- bindings/js/JSOptionConstructor.cpp: Ditto.
- bindings/js/JSWorkerConstructor.cpp: Ditto.
- bindings/js/JSXMLHttpRequestConstructor.cpp: Ditto.
- bridge/jni/jni_jsobject.mm: Ditto, but for SourceCode.h.
- inspector/InspectorController.cpp: Ditto.
- inspector/JavaScriptDebugServer.cpp:
(WebCore::JavaScriptDebugServer::recompileAllJSFunctions):
Moved mose of this function into the base class in JavaScriptCore,
so the details of compilation don't have to be exposed.
WebKit/mac:
2009-05-13 Darin Adler <Darin Adler>
Reviewed by Cameron Zwarich.
Bug 25674: syntax tree nodes should use arena allocation
https://bugs.webkit.org/show_bug.cgi?id=25674
- Plugins/Hosted/NetscapePluginInstanceProxy.mm: Updated includes.
New ones needed due to reducing includes of JSDOMBinding.h.
- WebView/WebScriptDebugger.mm: Ditto.
