[Qt-interest] Replace Qt Allocator

[Eric Clark]

> -----Original Message-----
> From: qt-interest-bounces at trolltech.com [mailto:qt-interest-
> bounces at trolltech.com] On Behalf Of Philippe
> Sent: Wednesday, September 16, 2009 10:52 AM
> To: Qt Interest (qt-interest at trolltech.com)
> Subject: Re: [Qt-interest] Replace Qt Allocator
> 
> As a side note, I use another allocator than nedMalloc, and as you I
> can
> see a real performance boost of my Qt application (under Windows, less
> under Mac where the allocator is good). I have not your problem
> certainly because new/delete/malloc/free/realloc are all hooked in the
> MSVC run time DLL.
> 
> Qt makes quickly hundreds of thousands of memory allocations (I made
> statistics, this is incredible), and for a big applications, this hurts
> performances if the memory allocator is not state of the art.

We have done the statistics as well. And using nedmalloc is about 150% faster
than the standard malloc. This is an incredible increase in speed and we are
not willing to give it up, considering we have been doing it for over 10 years
now.

> 
> Nokia programmers fight to improve Qt performances, and it's good, and
> certainly it's not their tasks to provide a good memory allocator. I
> hope they could consider providing all necessary hooks to override all
> memory allocations (qMalloc/qFree, etc. is not enough).

I couldn't agree more. The basis of what I am saying here is that Qt does provide
and document the ability to replace the allocator; however, the provided functions
are just not enough. I have to also add an implementation of new and delete in the
qglobal header file. And as many of you know, this file is included by everything, 
which causes me to have to recompile EVERYTHING. And those of you that have made 
a global change like this know that that takes a very long time. All Qt needed to
do was override new and delete in qglobal and have them call the provided qFree()
and qMalloc(). Then I would just have to change the implementations of qFree and
qMalloc to call nedfree and nedmalloc and rebuild ONLY QtCore.


> 
> Philippe
> 
> 
> about 150% faster
> 
> On Wed, 16 Sep 2009 15:02:35 +0000
> Eric Clark <eclark at ara.com> wrote:
> 
> >
> >
> > > -----Original Message-----
> > > From: Dan White [mailto:ygor at comcast.net]
> > > Sent: Wednesday, September 16, 2009 9:47 AM
> > > To: Qt Interest (qt-interest at trolltech.com)
> > > Cc: Eric Clark
> > > Subject: Re: [Qt-interest] Replace Qt Allocator
> > >
> > > Short Answer: Don't use nedmalloc with C++
> > >
> > > According to what appears to be the nedMalloc homepage
> > > <http://www.nedprod.com/programs/portable/nedmalloc/>, it is a
> variant
> > > on the old, classic C function "malloc"
> >
> > This is not an option. Nedmalloc is extremely fast and extremely
> optimized for threaded applications. Our application is on the order of
> about 150% faster when nedmalloc is used. You should read more about
> nedmalloc before saying that it is bad because it is not bad. What is
> bad is the fact that Qt is violating the ISO standard for templated
> classes. Of course nedmalloc is a variant of malloc, all allocators are
> a variant of malloc since it was the one and only allocator in the
> beginning.
> >
> > >
> > > This is dangerously bad
> > > Reference: <http://www.devx.com/tips/Tip/12491>
> > >
> > > "Avoid Using malloc() and free() in C++
> > > The use of malloc() and free() functions in a C++ file is not
> > > recommended and is even dangerous:
> > > 1. malloc() requires the exact number of bytes as an argument
> whereas
> > > new calculates the size of the allocated object automatically. By
> using
> > > new, silly mistakes such as the following are avoided:
> > >
> > >
> > > long * p = malloc(sizeof(short)); //p originally pointed to a
> short;
> > >  				//changed later (but  malloc's arg was not)
> > > 2. malloc() does not handle allocation failures, so you have to
> test
> > > the return value of malloc() on each and every call. This tedious
> and
> > > dangerous function imposes performance penalty and bloats your .exe
> > > files. On the other hand, new throws an exception of type
> > > std::bad_alloc when it fails, so your code may contain only one
> > > catch(std::bad_alloc) clause to handle such exceptions.
> > >
> > > 3. As opposed to new, malloc() does not invoke the object's
> > > constructor. It only allocates uninitialized memory. The use of
> objects
> > > allocated this way is undefined and should never occur. Similarly,
> > > free() does not invoke its object's destructor."
> > >
> > > ... which pretty much describes your problems.
> >
> > This does not explain my problem at all. We never invoke malloc(),
> free(), or the ned versions of these directly. Everything is done by
> overriding the new and delete operators. We always use new and we
> always use delete. This is a problem in Qt, not in our application or
> in the implementation of nedmalloc. Qt gives the functions qFree(),
> qMalloc() and qRealloc() to allow the overriding of the allocators used
> within Qt, but reimplementing these functions is simply not enough.
> >
> > As I said before, nedmalloc is extremely fast. Our application needs
> to be as fast as possible because it is a real time application. Simply
> saying not to use nedmalloc is not a solution. We have been using Qt
> and nedmalloc since Qt 3 and prior to Qt 4 we never had an issue. The
> container classes were modified dramatically in Qt 4 and due to the
> changes, this bug was introduced. If Qt is going to provide a means for
> reimplementing the allocator, there should be a full implementation,
> not just a partial one. Why have the qFree, qMalloc, and qRealloc
> functions there to allow the replacement of the system allocator if
> those functions are not called all the time, like they should be?
> >
> > >
> > > "Sometimes I think the surest sign that intelligent life exists
> > > elsewhere in the universe is that none of it has tried to contact
> us."
> > > Bill Waterson (Calvin & Hobbes)
> > >
> > > ----- Original Message -----
> > > From: Eric Clark <eclark at ara.com>
> > > To: Qt Interest (qt-interest at trolltech.com) <qt-
> interest at trolltech.com>
> > > Sent: Wed, 16 Sep 2009 14:31:42 +0000 (UTC)
> > > Subject: [Qt-interest] Replace Qt Allocator
> > >
> > > Hello All,
> > >
> > > Our program uses a system allocator called nedmalloc. We override
> the
> > > new and delete operators to use this system allocator. However,
> after
> > > porting to Qt 4 we have ran into some issues with the QList class
> and
> > > it's destructor. When the list is destructed, it attempts to free
> the
> > > memory created by the list, but nedfree() crashes because the
> memory
> > > was instantiated using the system allocator, malloc(). Here is the
> > > problem:
> > >
> > >
> > > The problem is indeed a violation of the ODR requirement in the ISO
> C++
> > > spec. For example in this line:
> > >
> > >
> > >
> > >    QModelIndexList indexes = list->selectionModel()-
> >selectedRows();
> > >
> > >
> > >
> > > QModelIndexList exists in the local binary due to being a templated
> > > type based on QList, therefore its destructor is called from the
> local
> > > binary where ::operator delete has been defined to invoke
> nedmalloc.
> > > Unfortunately the initial constructor of the list (selectedRows())
> > > exists in one of the Qt DLL binaries where ::operator new is
> defined to
> > > invoke the system allocator. Therefore the list is being
> constructed
> > > with the system allocator and destructed with nedmalloc - this
> > > obviously enough provokes a segfault as ODR is violated.
> > >
> > >
> > >
> > > I was told by some guys on this interest board that I should build
> > > nedmalloc into Qt. However, after doing so, I still get the same
> error.
> > > What I did to build nedmalloc into Qt, was I modified the body of
> > > qFree(), qMalloc(), and qRealloc() to use nedfree(), nedmalloc(),
> and
> > > nedRealloc() respectively. This apparently was not enough to fix
> the
> > > problem. Could anyone tell me what else needs to be done? Do I need
> to
> > > override the new and delete operators in Qt as well? Are there any
> > > direct calls to malloc() in the code that I am unaware of?
> > >
> > >
> > >
> > > Any help with this would be greatly appreciated!
> > >
> > >
> > >
> > > Eric
> > >
> >
> >
> > _______________________________________________
> > Qt-interest mailing list
> > Qt-interest at trolltech.com
> > http://lists.trolltech.com/mailman/listinfo/qt-interest
> 
> 
> 
> _______________________________________________
> Qt-interest mailing list
> Qt-interest at trolltech.com
> http://lists.trolltech.com/mailman/listinfo/qt-interest