NS_NewAtom currently flattens a string to hash it. We need a hash function that takes an nsAString& so we don't need to allocate. (This seems to be responsible for about 2000 allocations on startup.)

FWIW, the allocation sites are: 2033 operator new(unsigned long, const std::nothrow_t&) operator new(unsigned long) NS_AllocateContiguousHandleWithData ... nsSharableString::assign(const nsAString&) nsPromiseFlatString::nsPromiseFlatString(const nsAString&) PromiseFlatString(const nsAString&) GetAtomHashEntry(const nsAString&) NS_NewAtom(const nsAString&) nsXMLContentSink::SplitXMLName(const nsAString&, nsIAtom**, ... nsXMLContentSink::AddAttributes(const nsIParserNode&, ... nsXMLContentSink::OpenContainer(const nsIParserNode&) 648 operator new(unsigned long, const std::nothrow_t&) operator new(unsigned long) NS_AllocateContiguousHandleWithData ... nsSharableString::assign(const nsAString&) nsPromiseFlatString::nsPromiseFlatString(const nsAString&) PromiseFlatString(const nsAString&) GetAtomHashEntry(const nsAString&) NS_NewAtom(const nsAString&) nsXMLContentSink::SplitXMLName(const nsAString&, nsIAtom**, ... nsXMLContentSink::OpenContainer(const nsIParserNode&) 276 operator new(unsigned long, const std::nothrow_t&) operator new(unsigned long) NS_AllocateContiguousHandleWithData ... nsSharableString::assign(const nsAString&) nsPromiseFlatString::nsPromiseFlatString(const nsAString&) PromiseFlatString(const nsAString&) GetAtomHashEntry(const nsAString&) NS_NewAtom(const nsAString&) nsXMLContentSink::SplitXMLName(const nsAString&, nsIAtom**, ... nsXMLContentSink::AddAttributes(const nsIParserNode&, ... nsXMLContentSink::OpenContainer(const nsIParserNode&) 194 operator new(unsigned long, const std::nothrow_t&) operator new(unsigned long) NS_AllocateContiguousHandleWithData ... nsSharableString::assign(const nsAString&) nsPromiseFlatString::nsPromiseFlatString(const nsAString&) PromiseFlatString(const nsAString&) GetAtomHashEntry(const nsAString&) NS_NewAtom(const nsAString&) nsXMLContentSink::SplitXMLName(const nsAString&, nsIAtom**, ... nsXMLContentSink::AddAttributes(const nsIParserNode&, ... nsXMLContentSink::OpenContainer(const nsIParserNode&)

The above is from http://www.mozilla.org/performance/mallocsites20011031.html

I never managed to think of a good approach to fixing this, given our hashtable interfaces.

hrmm, is this still a problem, with the work alecf has been doing with atoms? I have a hashkey that uses nsAString, but it still has to allocate a copy of the string buffer, because we're not using sharable strings.

well, I'll bet its still a problem - I think that everywhere that I changed to permanent atoms had flat strings already.. so those call sites are probably slightly different, but still applicable...

No longer relevant with our new string code.