Changeset 37297 in webkit for trunk/JavaScriptCore/profiler/ProfileGenerator.cpp

Timestamp:

Oct 4, 2008, 2:12:54 PM (17 years ago)

Author:

Darin Adler

Message:

JavaScriptCore:

2008-10-04 Darin Adler <Darin Adler>

Reviewed by Cameron Zwarich.

• ​https://bugs.webkit.org/show_bug.cgi?id=21295 Bug 21295: Replace ExecState with a call frame Register pointer

10% faster on Richards; other v8 benchmarks faster too.
A wash on SunSpider.

This does the minimum necessary to get the speedup. Next step in
cleaning this up is to replace ExecState with a CallFrame class,
and be more judicious about when to pass a call frame and when
to pass a global data pointer, global object pointer, or perhaps
something else entirely.

• VM/CTI.cpp: Remove the debug-only check of the exception in ctiVMThrowTrampoline -- already checked in the code the trampoline jumps to, so not all that useful. Removed the exec argument from ctiTrampoline. Removed emitDebugExceptionCheck -- no longer needed. (JSC::CTI::emitCall): Removed code to set ExecState::m_callFrame. (JSC::CTI::privateCompileMainPass): Removed code in catch to extract the exception from ExecState::m_exception; instead, the code that jumps into catch will make sure the exception is already in eax.
• VM/CTI.h: Removed exec from the ctiTrampoline. Also removed the non-helpful "volatile". Temporarily left ARG_exec in as a synonym for ARG_r; I'll change that on a future cleanup pass when introducing more use of the CallFrame type. (JSC::CTI::execute): Removed the ExecState* argument.

• VM/ExceptionHelpers.cpp: (JSC::InterruptedExecutionError::InterruptedExecutionError): Take JSGlobalData* instead of ExecState*. (JSC::createInterruptedExecutionException): Ditto.
• VM/ExceptionHelpers.h: Ditto. Also removed an unneeded include.

• VM/Machine.cpp: (JSC::slideRegisterWindowForCall): Removed the exec and exceptionValue arguments. Changed to return 0 when there's a stack overflow rather than using a separate exception argument to cut down on memory accesses in the calling convention. (JSC::Machine::unwindCallFrame): Removed the exec argument when constructing a DebuggerCallFrame. Also removed code to set ExecState::m_callFrame. (JSC::Machine::throwException): Removed the exec argument when construction a DebuggerCallFrame. (JSC::Machine::execute): Updated to use the register instead of ExecState and also removed various uses of ExecState. (JSC::Machine::debug): (JSC::Machine::privateExecute): Put globalData into a local variable so it can be used throughout the interpreter. Changed the VM_CHECK_EXCEPTION to get the exception in globalData instead of through ExecState. (JSC::Machine::retrieveLastCaller): Turn exec into a registers pointer by calling registers() instead of by getting m_callFrame. (JSC::Machine::callFrame): Ditto. Tweaked exception macros. Made new versions for when you know you have an exception. Get at global exception with ARG_globalData. Got rid of the need to pass in the return value type. (JSC::Machine::cti_op_add): Update to use new version of exception macros. (JSC::Machine::cti_op_pre_inc): Ditto. (JSC::Machine::cti_timeout_check): Ditto. (JSC::Machine::cti_op_instanceof): Ditto. (JSC::Machine::cti_op_new_func): Ditto. (JSC::Machine::cti_op_call_JSFunction): Optimized by using the ARG values directly instead of through local variables -- this gets rid of code that just shuffles things around in the stack frame. Also get rid of ExecState and update for the new way exceptions are handled in slideRegisterWindowForCall. (JSC::Machine::cti_vm_compile): Update to make exec out of r since they are both the same thing now. (JSC::Machine::cti_op_call_NotJSFunction): Ditto. (JSC::Machine::cti_op_init_arguments): Ditto. (JSC::Machine::cti_op_resolve): Ditto. (JSC::Machine::cti_op_construct_JSConstruct): Ditto. (JSC::Machine::cti_op_construct_NotJSConstruct): Ditto. (JSC::Machine::cti_op_resolve_func): Ditto. (JSC::Machine::cti_op_put_by_val): Ditto. (JSC::Machine::cti_op_put_by_val_array): Ditto. (JSC::Machine::cti_op_resolve_skip): Ditto. (JSC::Machine::cti_op_resolve_global): Ditto. (JSC::Machine::cti_op_post_inc): Ditto. (JSC::Machine::cti_op_resolve_with_base): Ditto. (JSC::Machine::cti_op_post_dec): Ditto. (JSC::Machine::cti_op_call_eval): Ditto. (JSC::Machine::cti_op_throw): Ditto. Also rearranged to return the exception value as the return value so it can be used by op_catch. (JSC::Machine::cti_op_push_scope): Ditto. (JSC::Machine::cti_op_in): Ditto. (JSC::Machine::cti_op_del_by_val): Ditto. (JSC::Machine::cti_vm_throw): Ditto. Also rearranged to return the exception value as the return value so it can be used by op_catch.

• kjs/DebuggerCallFrame.cpp: (JSC::DebuggerCallFrame::functionName): Pass globalData. (JSC::DebuggerCallFrame::evaluate): Eliminated code to make a new ExecState.
• kjs/DebuggerCallFrame.h: Removed ExecState argument from constructor.

• kjs/ExecState.h: Eliminated all data members and made ExecState inherit privately from Register instead. Also added a typedef to the future name for this class, which is CallFrame. It's just a Register* that knows it's a pointer at a call frame. The new class can't be constructed or copied. Changed all functions to use the this pointer instead of m_callFrame. Changed exception-related functions to access an exception in JSGlobalData. Removed functions used by CTI to pass the return address to the throw machinery -- this is now done directly with a global in the global data.

• kjs/FunctionPrototype.cpp: (JSC::functionProtoFuncToString): Pass globalData instead of exec.

• kjs/InternalFunction.cpp: (JSC::InternalFunction::name): Take globalData instead of exec.
• kjs/InternalFunction.h: Ditto.

• kjs/JSGlobalData.cpp: Initialize the new exception global to 0.
• kjs/JSGlobalData.h: Declare two new globals. One for the current exception and another for the return address used by CTI to implement the throw operation.

• kjs/JSGlobalObject.cpp: (JSC::JSGlobalObject::init): Removed code to set up globalExec, which is now the same thing as globalCallFrame. (JSC::JSGlobalObject::reset): Get globalExec from our globalExec function so we don't have to repeat the logic twice. (JSC::JSGlobalObject::mark): Removed code to mark the exception; the exception is now stored in JSGlobalData and marked there. (JSC::JSGlobalObject::globalExec): Return a pointer to the end of the global call frame.
• kjs/JSGlobalObject.h: Removed the globalExec data member.

• kjs/JSObject.cpp: (JSC::JSObject::putDirectFunction): Pass globalData instead of exec.

• kjs/collector.cpp: (JSC::Heap::collect): Mark the global exception.

• profiler/ProfileGenerator.cpp: (JSC::ProfileGenerator::addParentForConsoleStart): Pass globalData instead of exec to createCallIdentifier.

• profiler/Profiler.cpp: (JSC::Profiler::willExecute): Pass globalData instead of exec to createCallIdentifier. (JSC::Profiler::didExecute): Ditto. (JSC::Profiler::createCallIdentifier): Take globalData instead of exec. (JSC::createCallIdentifierFromFunctionImp): Ditto.
• profiler/Profiler.h: Change interface to take a JSGlobalData instead of an ExecState.

WebKit/mac:

2008-10-04 Darin Adler <Darin Adler>

Reviewed by Cameron Zwarich.

• ​https://bugs.webkit.org/show_bug.cgi?id=21295 Bug 21295: Replace ExecState with a call frame Register pointer

• WebView/WebScriptDebugger.mm: (WebScriptDebugger::WebScriptDebugger): Remove 0 passed for ExecState.

File:

• trunk/JavaScriptCore/profiler/ProfileGenerator.cpp (modified) (1 diff)

trunk/JavaScriptCore/profiler/ProfileGenerator.cpp

@@ -63 +63 @@
 
     exec->machine()->retrieveLastCaller(exec, lineNumber, sourceID, sourceURL, function);
-    m_currentNode = ProfileNode::create(Profiler::createCallIdentifier(exec, function ? function->toThisObject(exec) : 0, sourceURL, lineNumber), m_head.get(), m_head.get());
+    m_currentNode = ProfileNode::create(Profiler::createCallIdentifier(&exec->globalData(), function ? function->toThisObject(exec) : 0, sourceURL, lineNumber), m_head.get(), m_head.get());
     m_head->insertNode(m_currentNode.get());
 }