Reviewed by Darin.
Fix leaks during plugin tests (which actually excercise background JS), and potential
PAC brokenness that was not reported, but very likely.
The leaks shadowed a bigger problem with Identifier destruction. Identifier::remove involves
an IdentifierTable lookup, which is now a per-thread instance. Since garbage collection can
currently happen on a different thread than allocation, a wrong table was used.
No measurable change on SunSpider total, ~1% variation on individual tests.
* kjs/ustring.cpp:
(KJS::):
(KJS::UString::Rep::create):
(KJS::UString::Rep::destroy):
* kjs/ustring.h:
Replaced isIdentifier with a pointer to IdentifierTable, so that destruction can be done
correctly. Took one bit from reportedCost, to avoid making UString::Rep larger (performance
effect was measurable on SunSpider).
* kjs/identifier.cpp:
(KJS::IdentifierTable::IdentifierTable):
(KJS::IdentifierTable::~IdentifierTable):
(KJS::IdentifierTable::add):
(KJS::IdentifierTable::remove):
Make IdentifierTable a real class. Its destructor needs to zero out outstanding references,
because some identifiers may briefly outlive it during thread destruction, and we don't want
them to use their stale pointers.
(KJS::LiteralIdentifierTable):
(KJS::Identifier::add):
Now that LiteralIdentifierTable is per-thread and can be destroyed not just during application
shutdown, it is not appropriate to simply bump refcount for strings that get there; changed
the table to hold RefPtrs.
(KJS::CStringTranslator::translate):
(KJS::UCharBufferTranslator::translate):
(KJS::Identifier::addSlowCase):
(KJS::Identifier::remove):
* kjs/identifier.h:
(KJS::Identifier::add):
Use and update UString::Rep::identifierTable as appropriate. Updating it is now done in
IdentifierTable::add, not in translators.
git-svn-id: http://svn.webkit.org/repository/webkit/trunk@32222 268f45cc-cd09-0410-ab3c-d52691b4dbfc
5 files changed
