[webkit-dev] WebKit memory management?

[Mike Hommey]

On Mon, Sep 15, 2008 at 11:21:31AM +0200, Ariya Hidayat <ariya.hidayat at trolltech.com> wrote:
> 
> > our goal is to change standard malloc to tcmalloc under linux-qt port but
> > we couldn't solve operator new and delete problems.  As we see Paul's (
> > https://bugs.webkit.org/show_bug.cgi?id=20422 ) solution can
> > solve this, therefore if the community accepts this change then we would
> > gladly help to replace all new and delete operators to the new ones.
> 
> Because typically we build QtWebKit as a shared library, this raises the 
> problem that the overloaded new/delete will be global, i.e. this would also 
> affects the application that links QtWebKit, so not only at the library level 
> (QtWebKit). 
> 
> The mentioned patch seems to solve this problem, but at the cost of reduced 
> syntax readability because we can't use the familiar C++ new/delete 
> constructs anymore (AFAICS we are not even allowed to use it anymore). I 
> wonder whether this is a good compromise, i.e. what justifies it more than 
> just providing the new/delete override at the application (I wasn't involved 
> in previous memory management discussions before, so please point out what I 
> miss here).
> 
> Side note: we found out that the Linux's standard malloc is good enough, 
> enabling the JSC's (tcmalloc-based) FastMalloc does not add any significant 
> speed boost. This is however different on QtWebKit/Win32, where FastMalloc 
> improves SunSpider by 30%.
> 
> (Of course I might be wrong here, so feel free to provide corrections).

It might be beneficial on long term use, where memory allocation/deallocation
will have created fragmentation, making the process need a lot of virtual
memory because it can't fill the holes, making the whole thing be a memory
hog. Tcmalloc has usually less fragmentation.

Mike