2002-12-03 Maciej Stachowiak <[email protected]>

Reviewed by: Darin Adler

- fixed Deployment build.

* kjs/dtoa.cpp: Work around warnings.

2002-12-03 Maciej Stachowiak <[email protected]>

- fixed 3114790 - Gamespot reviews pages badly mis-rendering

because floating point numbers format wide

Reviewed by: David Hyatt

* kjs/dtoa.cpp: Imported float <--> string conversion routines

from David M. Gay. I changed this to fix warnings and avoid

colliding with names of standard library functions.

* kjs/dtoa.h: Added a header I made up for dtoa.cpp

* kjs/ustring.cpp:

(UString::from): Use new double to string routine (kjs_strtod).

(UString::toDouble): Use new string to double routine (kjs_dtoa).

* JavaScriptCore.pbproj/project.pbxproj: Added new files

2002-11-27 John Sullivan <[email protected]>

* kjs/collector.cpp:

removed puts("COLLECT") leftover debugging spam that was

buggin' gramps

=== Alexander-34 ===

2002-11-26 Maciej Stachowiak <[email protected]>

Change ActivationImp to be allocated via the garbage collector

again instead of on the stack. This fixes the following four

regressions but sadly it causes a 6% performance hit. It's

probably possibly to reduce the hit a bit by being smarter about

inlining and the way the marking list variant is implemented, but

I'll look into that later.

- fixed 3111500 - REGRESSION: crash in "KJS::ScopeChain::mark()" on www.posci.com

- fixed 3111145 - REGRESSION: reproducible crash in KJS hashtable lookup at time.com

- fixed 3110897 - REGRESSION: javascript crasher on http://bmwgallery.tripod.com/

- fixed 3109987 - REGRESSION: Reproducible crash in KJS ObjectImp at live365.com

Also:

- improved DEBUG_COLLECTOR mode a bit by never giving memory back

to the system.

* kjs/collector.cpp:

* kjs/context.h:

* kjs/function.cpp:

(ActivationImp::ActivationImp):

(ActivationImp::mark):

(ActivationImp::createArgumentsObject):

* kjs/function.h:

* kjs/internal.cpp:

(ContextImp::ContextImp):

(ContextImp::mark):

* kjs/list.cpp:

* kjs/list.h:

* kjs/value.cpp:

(Value::Value):

2002-11-26 Darin Adler <[email protected]>

* kjs/property_map.cpp:

(PropertyMap::save): Look at the attributes the same way in the single hash entry

case as in the actual hash table case. Change the rule for which attributes to save

to "attributes that don't have the ReadOnly, DontEnum, or Function bit set".

Also fix bug where saving an empty property map would leave the count set to the old value.

2002-11-26 Richard Williamson <[email protected]>

Remove debugging code. Could be cause of performance regresssion.

* kjs/nodes.cpp:

(FunctionCallNode::evaluate):

Restire attributes correctly.

* kjs/property_map.cpp:

2002-11-25 Richard Williamson <[email protected]>

Use delete[] (not delete) operator to delete array.

* kjs/property_map.cpp:

2002-11-25 Richard Williamson <[email protected]>

Added debugging info. Fixed property map save function.

* kjs/nodes.cpp:

(FunctionCallNode::evaluate):

* kjs/property_map.cpp:

2002-11-25 Richard Williamson <[email protected]>

Changes for back/forward. Currently disabled.

* kjs/property_map.cpp:

* kjs/property_map.h:

2002-11-25 Darin Adler <[email protected]>

* kjs/property_map.cpp: Rearrange code a little bit and tweak indentation.

This might provide a tiny speedup because we don't look at the single entry

any more in cases where the _table pointer is non-0.

2002-11-24 Darin Adler <[email protected]>

- changed ScopeChain to not ref each item in the chain, and use

marking instead; gains 1% on JavaScript iBench

* kjs/context.h: Return chain by reference.

* kjs/internal.cpp: (ContextImp::mark): Mark the scope chain.

* kjs/interpreter.cpp: (Context::scopeChain): Return chain by reference.

* kjs/interpreter.h: Make some Context methods inline.

* kjs/nodes.cpp:

(ThisNode::evaluate): Get at ContextImp directly.

(ResolveNode::evaluateReference): Ditto.

(VarDeclNode::evaluate): Ditto.

(VarDeclNode::processVarDecls): Ditto.

(FuncDeclNode::processFuncDecl): Pass ScopeChain directly to avoid copying.

(FuncExprNode::evaluate): Ditto.

* kjs/object.cpp: Make scope and setScope inline.

* kjs/object.h: Make scope return a chain by reference. Make scope and

setScope both be inline. Use a normal ScopeChain instead of NoRefScopeChain

since they are now one and the same.

* kjs/scope_chain.cpp: Remove all the code to ref and deref objects.

Merge NoRefScopeChain in with ScopeChain since they both work this way now.

* kjs/scope_chain.h: Remove NoRefScopeChain and simplify the ref counts.

Make more functions inline.

2002-11-24 Maciej Stachowiak <[email protected]>

- fixed 3098356 - Hard hang on movie search at www.movietickets.com

* kjs/string_object.cpp:

(StringProtoFuncImp::call): When doing a regexp replacement that

results in an empty match, always move on to the next character

after doing the replacement. The previous code would hit an

infinite loop if an initial empty match was replaced with the

empty string.

2002-11-24 Maciej Stachowiak <[email protected]>

- fixed 3095446 - Crash on AppleScript page due to very long argument list

* kjs/grammar.y: Don't try to construct the argument list in the

right order, since that blows out the parser stack.

* kjs/nodes.cpp:

(ArgumentsNode::ArgumentsNode): Instead reverse the argument list

here.

* kjs/nodes.h: Make ArgumentsNode a friend of ArgumentListNode.

* kjs/grammar.cpp: Updated from grammar.y.

2002-11-23 Maciej Stachowiak <[email protected]>

- completed Darin's mostly-fix for 3037795 - Resource use

increases when accessing very high index value in array

The two missing pieces were handling sparse properties when

shrinking the array, and when sorting. Thse are now both taken

care of.

* kjs/array_instance.h:

* kjs/array_object.cpp:

(ArrayInstanceImp::put):

(ArrayInstanceImp::deleteProperty):

(ArrayInstanceImp::resizeStorage):

(ArrayInstanceImp::setLength):

(ArrayInstanceImp::sort):

(ArrayInstanceImp::pushUndefinedObjectsToEnd):

* kjs/identifier.h:

* kjs/object.h:

* kjs/property_map.cpp:

* kjs/property_map.h:

* kjs/reference_list.cpp:

(ReferenceList::append):

(ReferenceList::length):

* kjs/reference_list.h:

* kjs/ustring.cpp:

(UString::toUInt32):

* kjs/ustring.h:

2002-11-23 Maciej Stachowiak <[email protected]>

Numerous collector changes for a net gain of 3% on JS ibench:

- Replaced per-block bitmap with free list.

- Increased number of empty blocks kept around to 2.

- Doubled block size.

- When scanning heap in collector, skip scanning the rest of a

block as soon as we see as many live cells as the the number of

used cells it had originally.

Also the following collector changes unrelated to performance:

- Made constants `const int' instead of `static const int'.

- Miscellaneous code cleanup.

* kjs/collector.cpp:

- Added debugging mode enabled by defining DEBUG_GC which asserts

when a destroyed ValueImp

* kjs/internal.cpp:

(ContextImp::mark):

* kjs/value.cpp:

(Value::Value):

* kjs/value.h:

* kjs/config.h:

2002-11-22 Darin Adler <[email protected]>

- replaced List class with a vector rather than a linked list, changed it

to use a pool of instances instead of all the nodes allocated off of the

heap; gives 10% gain on iBench

* kjs/list.h: Complete rewrite.

* kjs/list.cpp: Ditto.

* kjs/array_object.cpp: (compareWithCompareFunctionForQSort): Go back to

doing a clear and two appends here. Fast with the new list implementation.

* kjs/collector.h: Remove _COLLECTOR hack and just make rootObjectClasses

return a const void *.

* kjs/collector.cpp: Remove _COLLECTOR hack, and various other minor tweaks.

2002-11-22 Darin Adler <[email protected]>

- prepare to reimplement KJS::List; move to its own file, add statistics

* kjs/function_object.cpp: (FunctionProtoFuncImp::call): Use new copyTail()

function rather than copy() and removeFirst().

* kjs/identifier.cpp: Add statistics, off by default.

* kjs/property_map.cpp: Add statistics, off by default.

* kjs/list.cpp: Added. Moved code here. To be rewritten.

* kjs/list.h: Added. Moved interface here. To be rewritten.

* kjs/types.cpp: Removed.

* kjs/types.h: Now just an empty header that includes other headers.

* JavaScriptCore.pbproj/project.pbxproj: Add new files, rearrange.

2002-11-22 Maciej Stachowiak <[email protected]>

- reduce cell size to 56 bytes from 64, now that nearly all

objects fit in that size. .5% speed gain and probably some

footprint gain.

* kjs/collector.cpp: Change CELL_SIZE from 64 to 56.

2002-11-22 Darin Adler <[email protected]>

- change ScopeChain to be a singly linked list shares tails, gives 11% gain on iBench

* kjs/context.h:

(ContextImp::pushScope): Make inline, use push instead of prepend, and pass imp pointer.

(ContextImp::popScope): Make inline, use pop instead of removeFirst.

* kjs/function.cpp: (DeclaredFunctionImp::DeclaredFunctionImp): No need to copy.

* kjs/function_object.cpp: (FunctionObjectImp::construct): Use push instead of

prepend, and pass imp pointer.

* kjs/internal.cpp: (ContextImp::ContextImp): Use clear, push instead of prepend,

and pass imp pointers.

* kjs/nodes.cpp: (ResolveNode::evaluateReference): Use isEmpty, pop, and top instead

of ScopeChainIterator.

* kjs/object.h: Change _scope to be a NoRefScopeChain.

* kjs/object.cpp: No need to initialize _scope any more, since it's not a NoRefScopeChain.

* kjs/scope_chain.h: Rewrite, different implementation and interface.

* kjs/scope_chain.cpp: More of the same.

2002-11-22 Maciej Stachowiak <[email protected]>

- a simple change for .4% gain on ibench - instead of unmarking

all objects at the start of collection, instead unmark as part of

the sweep phase

* kjs/collector.cpp:

(Collector::collect): Remove separate unmarking pass and instead

unmark the objects that don't get collected during the sweep

phase.

2002-11-21 Darin Adler <[email protected]>

- stop garbage collecting the ActivationImp objects, gets 3% on iBench

- pave the way to separate the argument lists from scope chains

* kjs/context.h: Added. Moved ContextImp here so it can use things defined

in function.h

* kjs/scope_chain.h: Added. Starting as a copy of List, to be improved.

* kjs/scope_chain.cpp: Added. Starting as a copy of List, to be improved.

* JavaScriptCore.pbproj/project.pbxproj: Rearranged things, added context.h.

* kjs/function.cpp:

(FunctionImp::call): Pass InterpreterImp, not ExecState, to ContextImp.

(DeclaredFunctionImp::DeclaredFunctionImp): List -> ScopeChain.

(ActivationImp::createArgumentsObject): ArgumentList -> List.

(GlobalFuncImp::call): Pass InterpreterImp, not an ExecState, to ContextImp.

* kjs/function.h: List -> ScopeChain.

* kjs/function_object.cpp: (FunctionObjectImp::construct): List -> ScopeChain.

* kjs/internal.cpp:

(ContextImp::ContextImp): Set the context in the interpreter.

(ContextImp::~ContextImp): Set the context in the interpreter to the caller.

(ContextImp::mark): Mark all the activation objects.

(InterpreterImp::InterpreterImp): Initialize context to 0.

(InterpreterImp::mark): Mark the top context.

(InterpreterImp::evaluate): Pass InterpreterImp to ContextImp.

* kjs/internal.h: Move ContextImp to its own header. Add setContext to InterpreterImp.

* kjs/interpreter.cpp: (Context::scopeChain): List -> ScopeChain.

* kjs/interpreter.h: List -> ScopeChain.

* kjs/nodes.cpp:

(ResolveNode::evaluateReference): List -> ScopeChain.

(FuncDeclNode::processFuncDecl): List -> ScopeChain.

(FuncExprNode::evaluate): List -> ScopeChain.

* kjs/object.cpp: List -> ScopeChain.

* kjs/object.h: List -> ScopeChain.

* kjs/types.h: Remove needsMarking features from List.

* kjs/types.cpp: Ditto.

2002-11-21 Maciej Stachowiak <[email protected]>

- reduced the size of PropertyMap by storing sizes and such in the

dynamically allocated part of the object to reduce the size of

ObjectImp - .5% speed improvement on JS iBench.

* kjs/property_map.cpp:

* kjs/property_map.h:

2002-11-21 Maciej Stachowiak <[email protected]>

* Makefile.am: Pass symroots for this tree to pbxbuild.

=== Alexander-33 ===

2002-11-21 Darin Adler <[email protected]>

* kjs/property_map.cpp: More assertions.

2002-11-21 Darin Adler <[email protected]>

* kjs/property_map.cpp: Turn that consistency check back off.

2002-11-21 Darin Adler <[email protected]>

- someone somewhere must be defining a macro named check, causing a compile failure in WebCore

Rename check() to checkConsistency().

* kjs/property_map.h: Rename.

* kjs/property_map.cpp: Yes, rename.

2002-11-21 Darin Adler <[email protected]>

- add self-check to property map in hopes of finding the cnet.com bug

* kjs/property_map.h: Add check() function.

* kjs/property_map.cpp: Add the checking, controlled by DO_CONSISTENCY_CHECK.

- fixed UChar interface so it's not so slow in debug builds

* kjs/ustring.h: Nothing in UChar needs to be private.

* kjs/function.cpp: (GlobalFuncImp::call):

* kjs/function_object.cpp: (FunctionObjectImp::construct):

* kjs/identifier.cpp:

* kjs/lexer.cpp: (Lexer::setCode), (Lexer::shift):

* kjs/lookup.cpp: (keysMatch):

* kjs/ustring.cpp: (UString::Rep::computeHash), (KJS::compare):

Use the "uc" field instead of the "unicode()" inline function.

2002-11-21 Darin Adler <[email protected]>

- fixed a null-dereference I ran into while trying to reproduce bug 3107351

* kjs/function.h: Change ActivationImp constructor to take context parameter.

* kjs/function.cpp: (ActivationImp::ActivationImp): Take context parameter,

not execution state parameter.

* kjs/internal.cpp: (ContextImp::ContextImp): Initialize activation object

from context, not execution state, because the new context is not yet in the

execution state.

2002-11-20 Darin Adler <[email protected]>

- added a feature for Richard to use in his back/forward cache

* kjs/object.h: Added save/restoreProperties.

* kjs/property_map.h: Here too.

* kjs/property_map.cpp: Here too.

2002-11-20 Darin Adler <[email protected]>

- created argument list objects only on demand for a 7.5% speedup

* kjs/function.h: Change ActivationImp around.

* kjs/function.cpp:

(FunctionImp::call): Pass a pointer to the arguments list to avoid ref/unref.

(FunctionImp::get): Get the function pointer from the context directly,

not the activation object.

(ArgumentsImp::ArgumentsImp): Add an overload that takes no arguments.

(ActivationImp::ActivationImp): Store a context pointer and an arguments object pointer.

(ActivationImp::get): Special case for arguments, create it and return it.

(ActivationImp::put): Special case for arguments, can't be set.

(ActivationImp::hasProperty): Special case for arguments, return true.

(ActivationImp::deleteProperty): Special case for arguments, refuse to delete.

(ActivationImp::mark): Mark the arguments object.

(ActivationImp::createArgumentsObject): Do the work of actually creating it.

(GlobalFuncImp::call): Use stack-based objects for the ContextImp and ExecState.

* kjs/internal.h: Keep function and arguments pointer in the context.

* kjs/internal.cpp:

(ContextImp::ContextImp): Don't pass in the func and args when making an ActivationImp.

(InterpreterImp::evaluate): Use stack-based objects here.

* kjs/types.h: Add ArgumentList as a synonym for List, soon to be separate.

2002-11-20 Maciej Stachowiak <[email protected]>

Reduced the size of ValueImp by 8 bytes for a .5% speedup.

* kjs/value.h: Removed destructed flag. Made refcount and flag 16

bits each.

* kjs/value.cpp:

(ValueImp::~ValueImp): Don't set destructed flag.

2002-11-20 Darin Adler <[email protected]>

* kjs/types.cpp: Keep ref count for the whole lists of nodes.

Doesn't speed things up much, less than 1%.

2002-11-20 Maciej Stachowiak <[email protected]>

* kjs/collector.cpp:

(Collector::allocate): Clear the flags on newly allocated objects.

2002-11-20 Darin Adler <[email protected]>

- oops, checked in big regression instead of 5% speedup

* kjs/function.cpp: (ActivationImp::ActivationImp): Make a marking

list, not a refing list.

- a cut at the sparse array implementation

* kjs/array_instance.h: Keep storageLength separate from length.

* kjs/array_object.cpp:

(ArrayInstanceImp::ArrayInstanceImp): Start with storageLength == length.

(ArrayInstanceImp::get): Check against storage length.

(ArrayInstanceImp::put): Ditto.

(ArrayInstanceImp::hasProperty): Ditto.

(ArrayInstanceImp::deleteProperty): Ditto.

(ArrayInstanceImp::setLength): Only enlarge storage length up to a cutoff.

(ArrayInstanceImp::mark): Use storageLength.

(ArrayInstanceImp::pushUndefinedObjectsToEnd): Added FIXME.

2002-11-20 Darin Adler <[email protected]>

- decrease ref/deref -- 5% speedup in iBench

* JavaScriptCore.pbproj/project.pbxproj: Added array_instance.h

* kjs/array_instance.h: Added so it can be shared by function.h.

* kjs/array_object.cpp:

* kjs/array_object.h:

* kjs/bool_object.cpp:

* kjs/bool_object.h:

* kjs/collector.cpp:

* kjs/date_object.cpp:

* kjs/date_object.h:

* kjs/error_object.cpp:

* kjs/function.cpp:

* kjs/function.h:

* kjs/function_object.cpp:

* kjs/internal.cpp:

* kjs/internal.h:

* kjs/math_object.cpp:

* kjs/nodes.cpp:

* kjs/number_object.cpp:

* kjs/object.cpp:

* kjs/object.h:

* kjs/object_object.cpp:

* kjs/property_map.cpp:

* kjs/reference.cpp:

* kjs/reference.h:

* kjs/regexp_object.cpp:

* kjs/string_object.cpp:

* kjs/string_object.h:

* kjs/value.cpp:

* kjs/value.h:

Switched lots of interfaces so they don't require ref/deref.

2002-11-20 Maciej Stachowiak <[email protected]>

Fixed the two most obvious problems with the new GC for another 6%

improvement.

* kjs/collector.cpp:

(Collector::allocate): Don't bother doing the bit tests on a bitmap word if

all it's bits are on.

(Collector::collect): Track memoryFull boolean.

* kjs/collector.h: Inlined outOfMemory since it was showing up on profiles.

2002-11-20 Maciej Stachowiak <[email protected]>

Rewrote garbage collector to make blocks of actual memory instead

of blocks of pointers. 7% improvement on JavaScript

iBench. There's still lots of room to tune the new GC, this is

just my first cut.

* kjs/collector.cpp:

(Collector::allocate):

(Collector::collect):

(Collector::size):

(Collector::outOfMemory):

(Collector::finalCheck):

(Collector::numGCNotAllowedObjects):

(Collector::numReferencedObjects):

(Collector::liveObjectClasses):

* kjs/collector.h:

* kjs/function.cpp:

(ActivationImp::ActivationImp):

* kjs/function.h:

2002-11-20 Darin Adler <[email protected]>

- on the road to killing ActivationImp

* kjs/function.h: Add get/put to FunctionImp. Remove argumentsObject() from

ActivationImp. Add function() to ActivationImp.

* kjs/function.cpp:

(FunctionImp::FunctionImp): No arguments property.

(FunctionImp::call): No need to set up the arguments property.

(FunctionImp::parameterString): Remove ** strangeness.

(FunctionImp::processParameters): Ditto.

(FunctionImp::get): Added, handles arguments and length properties.

(FunctionImp::put): Ditto.

(FunctionImp::hasProperty): Ditto.

(FunctionImp::deleteProperty): Ditto.

(ActivationImp::ActivationImp): Store a function pointer so we can find it

in the context.

* kjs/function_object.cpp: (FunctionObjectImp::construct): No need to set up

arguments property.

* kjs/nodes.cpp: (FuncExprNode::evaluate): No need to set up length property.

* kjs/internal.h: Return ObjectImp * for activation object.

* kjs/interpreter.h: Remove stray declaration of ExecStateImp.

2002-11-20 Darin Adler <[email protected]>

- add a couple of list operations to avoid clearing lists so much during sorting; gives 1.5% iBench

* kjs/types.h: Added replaceFirst/replaceLast.

* kjs/types.cpp: (List::replaceFirst), (List::replaceLast): Added.

* kjs/array_object.cpp: (compareWithCompareFunctionForQSort): Use replaceFirst/replaceLast.

* kjs/property_map.cpp: Put in an ifdef so I can re-add/remove the single entry to see if

it has outlived its usefulness. (It hasn't yet.)

2002-11-20 Darin Adler <[email protected]>

- atomic identifiers; gives another 6.5% in the iBench suite

* kjs/identifier.h: Did the real thing.

* kjs/identifier.cpp: Ditto.

* kjs/property_map.h: _tableSizeHashMask -> _tableSizeMask

* kjs/property_map.cpp: The above, plus take advantage of comparing

by pointer instead of by comparing bytes.

2002-11-19 Darin Adler <[email protected]>

- a few more globals for often-used property names

- conversion to Identifier from UString must now be explicit

* kjs/error_object.cpp:

* kjs/function.cpp:

* kjs/function_object.cpp:

* kjs/identifier.cpp:

* kjs/identifier.h:

* kjs/lexer.cpp:

* kjs/nodes.cpp:

* kjs/number_object.cpp:

* kjs/object.cpp:

* kjs/object.h:

* kjs/string_object.cpp:

* kjs/testkjs.cpp:

* kjs/ustring.cpp:

* kjs/ustring.h:

2002-11-19 Darin Adler <[email protected]>

- another step towards atomic identifiers; storing hash in the string rep. gives about

a 1.5% speedup in the JavaScript iBench

* kjs/ustring.h: Add a hash field to UString::Rep.

* kjs/ustring.cpp:

(UString::Rep::create): Set hash to uninitialized value.

(UString::Rep::destroy): Do the deleting in her, and call Identifier if needed.

(UString::Rep::computeHash): Added.

(UString::append): Set hash to 0 when modifying the string in place.

(UString::operator=): Ditto.

* kjs/property_map.cpp: Use the hash from UString.

* kjs/identifier.h: Added aboutToDestroyUStringRep.

* kjs/identifier.cpp: (Identifier::aboutToDestroyUStringRep): Added.

2002-11-19 Darin Adler <[email protected]>

- next step towards atomic identifiers; Identifier is no longer derived from UString

* kjs/identifier.h: Remove base class and add _ustring member.

* kjs/identifier.cpp: Add null and an == that works with const char *.

* kjs/property_map.cpp: Get rep through _ustring.

* kjs/function.cpp: (FunctionImp::parameterString): Call ustring().

* kjs/function_object.cpp: (FunctionProtoFuncImp::call): Ditto.

* kjs/nodes.cpp:

(PropertyNode::evaluate): Ditto.

(VarDeclNode::evaluate): Ditto.

(ForInNode::execute): Ditto.

* kjs/nodes2string.cpp: (SourceStream::operator<<): Add overload for Identifier.

* kjs/reference.cpp: (Reference::getValue): Call ustring().

* kjs/regexp_object.cpp: (RegExpObjectImp::get): Call ustring().

2002-11-19 Darin Adler <[email protected]>

- fixed memory trasher

* kjs/ustring.cpp: (UString::from): Fix "end of buffer" computation.

2002-11-19 Darin Adler <[email protected]>

- a first step towards atomic identifiers in JavaScript

Most places that work with identifiers now use Identifier

instead of UString.

* kjs/identifier.cpp: Added.

* kjs/identifier.h: Added.

* JavaScriptCore.pbproj/project.pbxproj: Added files.

* kjs/array_object.cpp:

* kjs/array_object.h:

* kjs/completion.cpp:

* kjs/completion.h:

* kjs/date_object.cpp:

* kjs/date_object.h:

* kjs/function.cpp:

* kjs/function.h:

* kjs/function_object.cpp:

* kjs/grammar.cpp:

* kjs/grammar.cpp.h:

* kjs/grammar.h:

* kjs/grammar.y:

* kjs/internal.cpp:

* kjs/internal.h:

* kjs/lexer.cpp:

* kjs/lookup.cpp:

* kjs/lookup.h:

* kjs/math_object.cpp:

* kjs/math_object.h:

* kjs/nodes.cpp:

* kjs/nodes.h:

* kjs/number_object.cpp:

* kjs/number_object.h:

* kjs/object.cpp:

* kjs/object.h:

* kjs/property_map.cpp:

* kjs/property_map.h:

* kjs/reference.cpp:

* kjs/reference.h:

* kjs/regexp_object.cpp:

* kjs/regexp_object.h:

* kjs/string_object.cpp:

* kjs/string_object.h:

2002-11-19 Darin Adler <[email protected]>

- fix hash function and key comparison for the other kind of hash table; yields 3%

* kjs/lookup.cpp:

(keysMatch): Added.

(Lookup::findEntry): Don't allocate and convert to ASCII just to search.

2002-11-19 Darin Adler <[email protected]>

- another hash table fix; yields a 2% improvement on iBench JavaScript

* kjs/property_map.cpp: A few more places where we use & instead of %.

- some List changes that don't affect speed yet

* kjs/types.cpp:

(List::prependList): Tighten up a tiny bit.

(List::copy): Use prependList.

* kjs/types.h: Remove appendList and globalClear.

* kjs/interpreter.cpp: (Interpreter::finalCheck): Remove List::globalClear().

2002-11-19 Darin Adler <[email protected]>

- fixed 3105026 -- REGRESSION: DHTML menus are broken all over the place

* kjs/types.cpp: (List::prepend): Fix backwards links in new node.

2002-11-19 Darin Adler <[email protected]>

- a fix that gives another 1.5% on the iBench JavaScript test

* kjs/ustring.cpp: (UString::from): Stop using sprintf to format integers.

2002-11-18 Darin Adler <[email protected]>

- reduced the creation of Value objects and hoisted the property map

into Object for another gain of about 6%

* JavaScriptCore.pbproj/project.pbxproj: Made property_map.h public.

* kjs/array_object.cpp:

(compareWithCompareFunctionForQSort): Don't wrap the ValueImp * in a Value

just to add it to a list.

(ArrayProtoFuncImp::call): Pass the globalObject directly so we don't have

to ref/deref.

* kjs/function.cpp:

(FunctionImp::call): Use a reference for the global object to avoid ref/deref.

(GlobalFuncImp::call): Ditto.

* kjs/internal.cpp:

(BooleanImp::toObject): Put the object directly into the list, don't create a Value.

(StringImp::toObject): Ditto.

(NumberImp::toObject): Ditto.

(InterpreterImp::evaluate): Use a reference for the global object.

* kjs/internal.h: Return a reference for the global object.

* kjs/interpreter.cpp: (Interpreter::globalObject): Ditto.

* kjs/interpreter.h: Ditto.

* kjs/object.cpp: Use _prop directly in the object, not a separate pointer.

* kjs/object.h: Ditto.

* kjs/types.cpp: Added List methods that work directly with ValueImp.

(List::append): Added a ValueImp version.

(List::prepend): Ditto.

(List::appendList): Work directly with the ValueImp's.

(List::prependList): Ditto.

(List::copy): Use appendList.

(List::empty): Use a shared global List.

* kjs/types.h: Update for above changes.

2002-11-18 Darin Adler <[email protected]>

* kjs/property_map.cpp: Oops, copyright goes to Apple, not me.

* kjs/property_map.h: Ditto.

2002-11-18 Darin Adler <[email protected]>

- property and string improvements giving a 7% or so improvement in JavaScript iBench

* kjs/property_map.h: Rewrite to use a hash table.

* kjs/property_map.cpp: Ditto.

* kjs/string_object.h:

* kjs/string_object.cpp:

(StringInstanceImp::StringInstanceImp): Construct a string with the right value

instead of putting the string in later.

(StringInstanceImp::get): Get the length from the string, not a separate property.

(StringInstanceImp::put): Ignore attempts to set length, since we don't put it in

the property map.

(StringInstanceImp::hasProperty): Return true for length.

(StringInstanceImp::deleteProperty): Return false for length.

(StringObjectImp::construct): Call new StringInstanceImp constructor. Don't try

to set a length property.

* kjs/ustring.h: Make the rep deref know how to deallocate the rep.

* kjs/ustring.cpp:

(UString::release): Move the real work to the rep's deref, since the hash table

now uses the rep directly.

* kjs/object.h: Remove unused field.

2002-11-18 Maciej Stachowiak <[email protected]>

Change List to completely avoid going through the GC

allocator. 3.6% performance improvement on JavaScript iBench.

* kjs/internal.cpp:

(InterpreterImp::mark): Don't mark the empty list.

For all the methods below I basically lifted the ListImp version

up to the List method with minor tweaks.

* kjs/types.cpp:

(ListIterator::ListIterator):

(List::List):

(List::operator=):

(List::~List):

(List::mark):

(List::append):

(List::prepend):

(List::appendList):

(List::prependList):

(List::removeFirst):

(List::removeLast):

(List::remove):

(List::clear):

(List::clearInternal):

(List::copy):

(List::begin):

(List::end):

(List::isEmpty):

(List::size):

(List::at):

(List::operator[]):

(List::empty):

(List::erase):

(List::refAll):

(List::derefAll):

(List::swap):

(List::globalClear):

* kjs/types.h:

2002-11-18 Maciej Stachowiak <[email protected]>

Fixed a horrible leak introduced with my last change that

somehow did not show up on my machine.

* kjs/types.cpp:

(List::List): Mark ListImp as GC allowed.

2002-11-18 Maciej Stachowiak <[email protected]>

Another step towards the List conversion: stop inheriting from Value.

* kjs/types.cpp:

(ListIterator::ListIterator):

(List::List):

(List::operator=):

(List::~List):

(List::mark):

(List::append):

(List::prepend):

(List::appendList):

(List::prependList):

(List::removeFirst):

(List::removeLast):

(List::remove):

(List::clear):

(List::copy):

(List::begin):

(List::end):

(List::isEmpty):

(List::size):

(List::at):

(List::operator[]):

* kjs/types.h:

2002-11-18 Maciej Stachowiak <[email protected]>

Partway to removing Value from List. Created a marking List

variant, used it in place of ListImp.

* kjs/internal.h: Removed List stuff.

* kjs/internal.cpp:

(InterpreterImp::mark): Call appropriate List method to do marking of

empty ListImp.

* kjs/object.h:

* kjs/object.cpp: Use marking List instead of ListImp *.

* kjs/types.h:

* kjs/types.cpp:

(List::List): New boolean needsMarking parameter.

(List::operator=): Perform trickery related to needsMarking.

(List::~List): Likewise.

(List::mark): Mark the ListImp.

(List::markEmptyList):

(ListImp::*): Moved here fron internal.cpp, they will be

integrated into the relevant List methods soon.

2002-11-18 Darin Adler <[email protected]>

- another string constant discovered that can be optimized

* kjs/object.h: Add a property name constant for "__proto__".

* kjs/object.cpp: Define it.

(ObjectImp::get): Use it.

(ObjectImp::hasProperty): Use it.

- prepare to turn PropertyMap into a hash table

* kjs/object.cpp:

(ObjectImp::mark): Use the new PropertyMap::mark().

(ObjectImp::put): Use the new overload of PropertyMap::get().

(ObjectImp::deleteProperty): Use the new overload of PropertyMap::get().

(ObjectImp::propList): Use PropertyMap::addEnumerablesToReferenceList().

* kjs/property_map.h: Remove PropertyMapNode and make all node-related methods private.

Add mark(), a new overload of get() that returns attributes, a clear() that takes no attributes,

and addEnumerablesToReferenceList().

* kjs/property_map.cpp:

(PropertyMap::get): Added new overload.

(PropertyMap::clear): Added new overload.

(PropertyMap::mark): Added.

(PropertyMap::addEnumerablesToReferenceList): Added.

* kjs/ustring.h: Added a hash function.

* kjs/ustring.cpp: (KJS::hash): Added.

2002-11-18 Darin Adler <[email protected]>

- simplified the ExecState class, which was showing up in profiles

Sped up JavaScript iBench by 6%.

* kjs/interpreter.h: Removed the level of indirection, and made it all inline.

* kjs/interpreter.cpp: Removed ExecState implementation from here altogether.

- fixed an oversight in my sort speedup

* kjs/array_object.h: Add pushUndefinedObjectsToEnd.

* kjs/array_object.cpp:

(ArrayInstanceImp::sort): Call pushUndefinedObjectsToEnd.

(ArrayInstanceImp::pushUndefinedObjectsToEnd): Added.

Pushes all undefined to the end of the array.

2002-11-18 Darin Adler <[email protected]>

- fix worst speed problems on the sort page of the iBench JavaScript test

Sped up JavaScript iBench by 70%, the sort page by 88%.

* kjs/array_object.h: Add array-specific sort functions.

* kjs/array_object.cpp:

(compareByStringForQSort): Added.

(ArrayInstanceImp::sort): Added.

(compareWithCompareFunctionForQSort): Added.

(ArrayProtoFuncImp::call): Use ArrayInstanceImp::sort if the object being

sorted is actually an array.

* kjs/object.h: Add argumentsPropertyName.

* kjs/object.cpp: Add argumentsPropertyName.

* kjs/function.cpp:

(FunctionImp::FunctionImp): Use argumentsPropertyName to avoid making a UString.

(FunctionImp::call): Ditto.

(ActivationImp::ActivationImp): Ditto.

* kjs/function_object.cpp: (FunctionObjectImp::construct): Ditto.

* kjs/ustring.h: Added compare function for -1/0/+1 comparison.

* kjs/ustring.cpp: (KJS::compare): Added.

2002-11-18 Maciej Stachowiak <[email protected]>

Change ArgumentListNode operations to be iterative instead of

recursive. This probably fixes 3095446 (Crash in

KJS::ArgumentListNode::ref()) but I can't reproduce it myself so

I'm not 100% sure. I think the original bug was a stack overflow

and this change would remove that possibility.

* kjs/nodes.cpp:

(ArgumentListNode::ref): Make iterative.

(ArgumentListNode::deref): Make iterative.

(ArgumentListNode::evaluateList): Make iterative.

=== Alexander-32 ===

2002-11-14 Darin Adler <[email protected]>

- fixed 3101243 -- excite passes date that can't be parsed, results in bogus date at top right corner

* kjs/date_object.cpp: (KJS::KRFCDate_parseDate): Handle errors from strtol

by checking errno. Check the "string in a haystack" to be sure it's a multiple

of 3. Add case that allows year to be after time.

2002-11-14 Darin Adler <[email protected]>

- fixed 3101191 -- REGRESSION: Hang loading excite.com

* kjs/date_object.cpp:

(mktimeUsingCF): Pick an arbitrary cutoff of 3000, and return -1 if the

year passed in is that big so we don't infinite loop. Also validate the

rest of the date with CFGregorianDateIsValid.

(DateProtoFuncImp::call): Handle a -1 result from mktime.

(DateObjectImp::construct): Check for NaN before calling mktime, and also

handle a -1 result from mktime.

(DateObjectFuncImp::call): Check for NaN before calling mktime, and also

handle a -1 result from mktime.

2002-11-13 Darin Adler <[email protected]>

- fixed 3099930 -- dates/times without time zones are parsed as UTC by kjs,

local time by other browsers

* kjs/date_object.cpp:

(DateProtoFuncImp::call): Handle the NaN case better, like Mozilla and OmniWeb.