JavaScriptCore: Simplified handling of 'arguments' -- 1.2% SunSpider speedup
authorggaren@apple.com <ggaren@apple.com@268f45cc-cd09-0410-ab3c-d52691b4dbfc>
Wed, 19 May 2010 05:10:11 +0000 (05:10 +0000)
committerggaren@apple.com <ggaren@apple.com@268f45cc-cd09-0410-ab3c-d52691b4dbfc>
Wed, 19 May 2010 05:10:11 +0000 (05:10 +0000)
commit7a249a5961b27f0fd53fd91a72308050b993dc71
tree13edc7deddf48fbe1c954ebd0159a8401d4d1ca8
parenteab104884048ef689afa0a0816cf1cee7c96edd3
JavaScriptCore: Simplified handling of 'arguments' -- 1.2% SunSpider speedup
https://bugs.webkit.org/show_bug.cgi?id=39200

Reviewed by Darin Adler.

Removed the reserved OptionalCalleeArguments slot from the CallFrame.
Now, slots for 'arguments' are allocated and initialized only by
functions that might need them.

* bytecode/CodeBlock.cpp:
(JSC::CodeBlock::dump): Updated for new bytecode operands.

(JSC::CodeBlock::CodeBlock):
* bytecode/CodeBlock.h:
(JSC::unmodifiedArgumentsRegister): Added a helper function for mapping
from the arguments register to its unmodified counterpart.

(JSC::CodeBlock::setArgumentsRegister):
(JSC::CodeBlock::argumentsRegister):
(JSC::CodeBlock::usesArguments):  Changed from a "usesArguments" bool to
an optional int index representing the arguments register.

* bytecode/Opcode.h: Updated for new bytecode operands.

* bytecompiler/BytecodeGenerator.cpp:
(JSC::BytecodeGenerator::addVar): Factored out a helper function for
allocating an anonymous var.

(JSC::BytecodeGenerator::BytecodeGenerator): Merged / simplified some
arguments vs activation logic, and added code to allocate the arguments
registers when needed.

(JSC::BytecodeGenerator::createArgumentsIfNecessary): Updated for new bytecode operands.

(JSC::BytecodeGenerator::emitCallEval): No need to create the arguments
object before calling eval; the activation object will lazily create the
arguments object if eval resolves it.

(JSC::BytecodeGenerator::emitReturn): Updated for new bytecode operands.

(JSC::BytecodeGenerator::emitPushScope):
(JSC::BytecodeGenerator::emitPushNewScope): Ditto emitCallEval.

* bytecompiler/BytecodeGenerator.h:
(JSC::BytecodeGenerator::addVar): Factored out a helper function for
allocating an anonymous var.

(JSC::BytecodeGenerator::registerFor): No more need for special handling
of the arguments registers; they're allocated just like normal registers
now.

* interpreter/CallFrame.h:
(JSC::ExecState::callerFrame):
(JSC::ExecState::init):
* interpreter/CallFrameClosure.h:
(JSC::CallFrameClosure::resetCallFrame): Nixed optionalCalleeArguments.

* interpreter/Interpreter.cpp:
(JSC::Interpreter::dumpRegisters):
(JSC::Interpreter::unwindCallFrame):
(JSC::Interpreter::privateExecute):
(JSC::Interpreter::retrieveArguments): Opcodes accessing 'arguments' now
take operands specifying registers, just like all other opcodes.
JSActivation::copyRegisters is no longer responsible for tearing off the
arguments object; instead, the VM is responsible for both.

Also, a behavior change: Each access to f.arguments creates a new object,
unless f itself uses 'arguments'. This matches Chrome, and is necessary
for the optimization. f.arguments is a nonstandard, deprecated feature,
so high fidelity to a given implementation is not necessarily a goal.
Also, as illustrated by the new test case, the identity of f.arguments
has been broken since 2008, except in the case where f itself accesses
f.arguments -- but nobody seemed to notice. So, hopefully this change won't
break the web.

* interpreter/Register.h: Nixed the special arguments accessor. It's no
longer needed.

* interpreter/RegisterFile.h:
(JSC::RegisterFile::):
* jit/JITCall.cpp:
(JSC::JIT::compileOpCallInitializeCallFrame):
(JSC::JIT::compileOpCall):
* jit/JITOpcodes.cpp:
(JSC::JIT::emit_op_tear_off_activation):
(JSC::JIT::emit_op_tear_off_arguments):
(JSC::JIT::emit_op_create_arguments):
(JSC::JIT::emit_op_init_arguments):
* jit/JITOpcodes32_64.cpp:
(JSC::JIT::emit_op_tear_off_activation):
(JSC::JIT::emit_op_tear_off_arguments):
(JSC::JIT::emit_op_create_arguments):
(JSC::JIT::emit_op_init_arguments): The actual optimization: Removed
OptionalCalleeArguments from the callframe slot. Now, it doesn't need
to be initialized for most calls.

* jit/JITStubs.cpp:
(JSC::DEFINE_STUB_FUNCTION):
* jit/JITStubs.h:
(JSC::): Updated stubs to support arbitrary 'arguments' registers,
instead of hard-coding something in the call frame.

* runtime/Arguments.h:
(JSC::JSActivation::copyRegisters): Removed some obfuscatory abstraction.

* runtime/Executable.h:
(JSC::FunctionExecutable::generatedByteCode): Added a helper for accessing
the 'arguments' register. In a future patch, that kind of data should
probably move out of CodeBlock and into Executable.

* runtime/JSActivation.cpp:
(JSC::JSActivation::getOwnPropertySlot):
(JSC::JSActivation::argumentsGetter):
* runtime/JSActivation.h: Simplified / fixed access to 'arguments' via
the activation object. It now implements the same behavior implemented
by optimized variable access in the VM. This simplifies some other
things, too -- like eval code generation.

LayoutTests: Simplified handling of 'arguments' -- 1.2% SunSpider speedup
https://bugs.webkit.org/show_bug.cgi?id=39200

Reviewed by Darin Adler.

* fast/js/function-dot-arguments-expected.txt:
* fast/js/script-tests/function-dot-arguments.js:
(argumentsIdentity): Updated to match new behavior.

* fast/js/function-dot-arguments2-expected.txt:
* fast/js/function-dot-arguments2.html: New tests for some things that
weren't covered before.

* fast/js/global-recursion-on-full-stack.html: Rejiggered the stack
usage in this test. Since stack usage is more efficient now, you
need a slightly different usage pattern to hit the exact thing this
test wanted to test.

* fast/js/kde/script-tests/function_arguments.js:
(f): Updated to more specifically test what this was trying to test,
to avoid just testing the identity of f.arguments.

git-svn-id: https://svn.webkit.org/repository/webkit/trunk@59742 268f45cc-cd09-0410-ab3c-d52691b4dbfc
27 files changed:
JavaScriptCore/ChangeLog
JavaScriptCore/bytecode/CodeBlock.cpp
JavaScriptCore/bytecode/CodeBlock.h
JavaScriptCore/bytecode/Opcode.h
JavaScriptCore/bytecompiler/BytecodeGenerator.cpp
JavaScriptCore/bytecompiler/BytecodeGenerator.h
JavaScriptCore/interpreter/CallFrame.h
JavaScriptCore/interpreter/CallFrameClosure.h
JavaScriptCore/interpreter/Interpreter.cpp
JavaScriptCore/interpreter/Register.h
JavaScriptCore/interpreter/RegisterFile.h
JavaScriptCore/jit/JITCall.cpp
JavaScriptCore/jit/JITOpcodes.cpp
JavaScriptCore/jit/JITOpcodes32_64.cpp
JavaScriptCore/jit/JITStubs.cpp
JavaScriptCore/jit/JITStubs.h
JavaScriptCore/runtime/Arguments.h
JavaScriptCore/runtime/Executable.h
JavaScriptCore/runtime/JSActivation.cpp
JavaScriptCore/runtime/JSActivation.h
LayoutTests/ChangeLog
LayoutTests/fast/js/function-dot-arguments-expected.txt
LayoutTests/fast/js/function-dot-arguments2-expected.txt [new file with mode: 0644]
LayoutTests/fast/js/function-dot-arguments2.html [new file with mode: 0644]
LayoutTests/fast/js/global-recursion-on-full-stack.html
LayoutTests/fast/js/kde/script-tests/function_arguments.js
LayoutTests/fast/js/script-tests/function-dot-arguments.js