Re: Sorting records using sort()

From: Elijah Bailey (geomrock_at_hotmail.com)
Date: 01/04/04 

Date: Sun, 4 Jan 2004 04:15:33 +0000 (UTC)

m@remove.this.part.rtij.nl (Martijn Lievaart) wrote in message news:<pan.2004.01.02.12.56.21.625773@remove.this.part.rtij.nl>...
> On Fri, 02 Jan 2004 03:13:26 -0800, Elijah Bailey wrote:
>
> [ Please don't top post, thank you. M4 ]
>
> [ Crossposted to csc++, is this standard complient? ]
>
> [ Short description of the problem ]
>
> We have an array of n*m bytes. It holds n objects of size m. Both are only
> known at runtime. For some strange reason we want to sort this using
> std::sort. We know there are other solutions, the question is, can it be
> done using std::sort.
>

Thanks Martijn. This is a much better description than i came up with.

> Constraints: Not clear, but memory seems to be an issue, creating an array
> of pointers and sort that is out.

Constraints: O(n) space is not allowed. So you can not create a
pointer
to each object and sort them. O(log(n)) space is perfectly ok.
O(log(n)*m) is also allowed if that simplifies things.

My data has n >> m, i.e. n is much greater than m.

>
> > "Ron Natalie" <ron@sensor.com> wrote in message
> > news:<3ff4297b$0$31857$9a6e19ea@news.newshosting.com>...
> >> "Jeff Schwab" <jeffplus@comcast.net> wrote in message
> >> news:WoidnaAaas0i8mmiRVn-hA@comcast.com...
> >>
> >> > Writing your own iterators would not be trivial, but probably would
> >> > be a educational. Having such iterators available would make your
> >> > data structure compatible with most of the standard library, and
> >> > might avoid countless future headaches.
> >>
> >> The trouble with writing iterators for this problem is that he needs it
> >> to be variable size at run time and that the iterators have to be
> >> derferenceable and the derefenced types assignable. Add his
> >> requirement to not allocate even enough memory to hold pointers to each
> >> object, I can't figure out how you could do it.
> >
> > Thanks Ron, I think that is definitely a problem. But I still "think" it
> > can be done, dont know how to thou...:)
>
> I also think it can be done. However, it requires so much trickery that it
> is not interesting to do so. Any of the other options quickly become very
> appealing. My approach uses some dynamic memory, but probably less than an
> array of pointers would take. The actual amount depends on the number of
> copies std::sort makes internally, there is no way around this. I guess
> for a typical implementation this would be around log(n) copies, unless
> the implementation takes special care to dispose of temporaries. You
> cannot count on any of this though.
>

Hopefully, by the end of this thread, we'll have something simple
enough
to implement...:)

> OK, how can we do it? Obviously every iterator must know the size of the
> object it is supposed to handle. The main problems we are facing:
>
> - We cannot create real objects (directly), the size of the real object is
> only known at runtime.
> - std::sort is allowed to (and most frequently does) make extra copies of
> objects.
>
> Asusmptions:
> - There are functions to copy these objects and to compare these objects.
> - The objects to be sorted don't need special construction or destruction
> and can be copied by the previously mentioned routine. They can be
> constructed by copying into a new memory area.
> - std::sort never uses pointers to objects, only iterators. I think the
> standard mandates this, but couldn't find it.
>

Would be interesting to know, what std::sort() is required to have
from
the standard. I think you are perfectly right that sort doesnt use
pointers
to the objects, only iterators and derefernces(*) the iterators...

> We create a proxy class and let the iterator return proxy objects. These
> proxy objects store a pointer to some master descriptor that stores the
> location and size of the array and the size of the individual objects. We
> could store this in the proxy object itself, but that seems wasteful.
> Also, the proxy object stores a void* to store the data for the real
> object.
>
> This proxy class is what our iterators value_type is.
>
> Now the proxy object implements some intelligent constructors. I doubt we
> need a default constructor, the value type needs to be assignable only
> (this follows from the iterator requirements). So we create a constructor
> that is used only inside the iterator, it sets the data pointer to point
> inside the array at the correct location. If the copy constructor is
> called (which would be by std::sort) we dynamically allocate memory to
> hold the object and use the copying routine to fill it.
>
> The assignment operator must use the copying routine if the destination is
> inside the array. If not, it can either use copy-construct-and-swap or the
> copying routine.
>
> The destructor would check if the pointer points inside the array and if
> not deletes the object.
>
> Obviously, we need to define an operator< for the proxy objects, this
> calls the comparison function above. If we template the proxy objects, we
> can use a functor for the comparison, potentially inlining the comaparison
> (this would be the only reason to use std::sort anyhow, so it makes
> sense).
>
> Looking at the standard, this would fullfill all requirements for
> std::sort I could find. Somehow I think I did overlook something here, so
> scrutiny by others is appreciated. I also left obvious details out,
> obviously :-)
>
> A note on exceptions. This would give the basic guarentee, if the copying
> and comparison routines give at least the basic guarentee.
>
> A note on memory usage. Assuming that std::sort holds log(n) copies of
> objects internally, this aproach uses less memory than the
> sort-array-of-pointers aproach if log(n)*m < n*sizeof(void*). So this
> greatly depends on these variables, but quickly holds for large n and gets
> less interesting for large m. Assume sizeof(void*)==4, we get m <
> 4n/log(n). For n=100, m should be less than 60. For n=1000, m should be
> less than 400. These are gross approximations but should give you a feel.
>

Usually this is certainly the case for my data.

> I hope anyone sees that you must have good reasons to do this. It is ugly,
> errorprone, prone to assumptions your implementation of std::sort makes
> (in this implementation &*it does not return a pointer to the real object,
> any pointer arithmetic will fail), difficult to maintain. There is no real
> reason why you should not call qsort in the first place, except as an
> accademic exercise.
>

First problem to qsort: It's slower than using sort()
Second problem to qsort: Later I might want to use other STL
algorithms on this data!
                                       like for_each()? find()? ...?

Thanks a lot for your comments,
--Elijah 

---
[ comp.std.c++ is moderated.  To submit articles, try just posting with ]
[ your news-reader.  If that fails, use mailto:std-c++@ncar.ucar.edu    ]
[              --- Please see the FAQ before posting. ---               ]
[ FAQ: http://www.jamesd.demon.co.uk/csc/faq.html                       ]

Relevant Pages

  • Re: Sorting records using sort()
    ... We have an array of n*m bytes. ... >> The trouble with writing iterators for this problem is that he needs it ... We create a proxy class and let the iterator return proxy objects. ... that is used only inside the iterator, it sets the data pointer to point ...
    (comp.lang.cpp)
  • Re: Question about arrays in objects
    ... >>I want to create a 2d array as an object member. ... > handled in the constructor, you just can't use a plain old array. ... > that is where using a pointer would come in. ...
    (alt.comp.lang.learn.c-cpp)
  • Re: Function lookup tables?
    ... Ignore any typos the source is good except howto initialize the function ... > I want to initialize the function lookup table array to my member ... > function calls of the class inside of the constructor for the class. ... > What is the correct code to initialize the function pointer array ...
    (comp.lang.cpp)
  • Re: Question about arrays in objects
    ... >then put the array in the constructor and use the pointer? ... There's at least one other way to do this, using a static const int ... handled in the constructor, you just can't use a plain old array. ...
    (alt.comp.lang.learn.c-cpp)
  • Re: minor confusion
    ... The compiler generated copy constructor makes gh from t. ... to int, data, will have the same value in both objects. ... They share that array. ... When the object passed by value is a pointer, ...
    (alt.comp.lang.learn.c-cpp)