Bug #516

GModels container not updated appropriately

Added by Knödlseder Jürgen over 12 years ago. Updated about 12 years ago.

Status:ClosedStart date:09/20/2012
Priority:UrgentDue date:
Assigned To:Knödlseder Jürgen% Done:

100%

Category:-
Target version:00-07-00
Duration:

Description

In the following example, a point source model was appended to the GModels container, and later the point source model was replaced by an extended model. However, the container still returns a pointer to a point source model, although the content corresponds clearly to the extended source model.

>>> from gammalib import *
>>> m=GModels()
>>> p=GModelPointSource()
>>> p.name("Point source")
>>> e=GModelExtendedSource()
>>> e.name("Extended source")
>>> m.append(p)
>>> print m
=== GModels ===
 Number of models ..........: 1
 Number of parameters ......: 0
=== GModelPointSource ===
 Name ......................: Point source
 Instruments ...............: all
 Model type ................: 
 Number of parameters ......: 0
 Number of spatial par's ...: 0
 Number of spectral par's ..: 0
 Number of temporal par's ..: 0
>>> m[0]=e
>>> print m
=== GModels ===
 Number of models ..........: 1
 Number of parameters ......: 0
=== GModelPointSource ===
 Name ......................: Extended source
 Instruments ...............: all
 Model type ................: 
 Number of parameters ......: 0
 Number of spatial par's ...: 0
 Number of spectral par's ..: 0
 Number of temporal par's ..: 0
>>> print type(m[0])
<class 'gammalib.model.GModelPointSource'>
>>> print m[0].type()
PointSource

This seems not to be a problem of type casting at the moment of reading, as the object returns a type of “PointSource”.

Interestingly, the pointer has however be changed:

>>> print m[0].__repr__()
<gammalib.model.GModelPointSource; proxy of <Swig Object of type 'GModelPointSource *' at 0x101a58300> >
>>> m[0]=e
>>> print m[0].__repr__()
<gammalib.model.GModelPointSource; proxy of <Swig Object of type 'GModelPointSource *' at 0x101a583f0> >

Possible, this indicates a problem in the setitem method. Some debugging is needed here.


Recurrence

No recurrence.

History

#1 Updated by Knödlseder Jürgen over 12 years ago

The problem is related to the design of the GModel class and it’s derived classes. This is illustrated by the following code example:

#include <string>
#include <iostream>

class GModel {
public:
    GModel(void) {}
    virtual ~GModel(void) {}
    virtual std::string type(void) const = 0;
    virtual std::string name(void) const { return m_name; }
    virtual GModel& operator=(const GModel& model)
    {
        if (this != &model) {
            std::cout << "GModel& operator=(const GModel& model)" << std::endl;
            m_name = model.m_name;
        }
    }
    std::string m_name;
};

class GSkyModel : public GModel {
public:
    GSkyModel(void) { m_name="sky"; }
    virtual ~GSkyModel(void) {}
    virtual std::string type(void) const { return "sky"; }
};

class GDataModel : public GModel {
public:
    GDataModel(void) { m_name="data"; }
    virtual ~GDataModel(void) {}
    virtual std::string type(void) const { return "data"; }
};

int main(void)
{
    GModel& sky  = *(new GSkyModel);
    GModel& data = *(new GDataModel);
    std::cout << sky.type() << ":" << sky.name() << ": " << &sky << std::endl;
    if (dynamic_cast<GSkyModel*>(&sky) != NULL)
        std::cout << "is GSkyModel" << std::endl;
    else
        std::cout << "is GDataModel" << std::endl;
    sky = data;
    std::cout << sky.type() << ":" << sky.name() << ": " << &sky << std::endl;
    if (dynamic_cast<GSkyModel*>(&sky) != NULL)
        std::cout << "is GSkyModel" << std::endl;
    else                                     
        std::cout << "is GDataModel" << std::endl;
    return 0;
}

This code outputs the following:

sky:sky: 0x100100080
is GSkyModel
GModel& operator=(const GModel& model)
sky:data: 0x100100080
is GSkyModel

The assignment in line 43 just copy the members of GDataModel to GSkyModel, but the sky object remains of course of type GSkyModel.

That's exactly what happens in the GammaLib classes. When the GModels::operator[] is called, a reference to GModel is returned. If then operator= is used to assign one model to the reference, only the members are copied, but the target reference remains the same. In the above example (that captures the essential elements of the GammaLib implementation), the sky reference remains always a reference to a GSkyModel object.

#2 Updated by Knödlseder Jürgen over 12 years ago

A implementation of the main routine that has the expected behavior would be:

int main(void)
{
    GModel* sky  = new GSkyModel;
    GModel* data = new GDataModel;
    std::cout << sky->type() << ":" << sky->name() << ": " << sky << std::endl;
    if (dynamic_cast<GSkyModel*>(sky) != NULL)
        std::cout << "is GSkyModel" << std::endl;
    else
        std::cout << "is GDataModel" << std::endl;
    delete sky;
    sky = data;
    std::cout << sky->type() << ":" << sky->name() << ": " << sky << std::endl;
    if (dynamic_cast<GSkyModel*>(sky) != NULL)
        std::cout << "is GSkyModel" << std::endl;
    else
        std::cout << "is GDataModel" << std::endl;
    return 0;
}

which results in
sky:sky: 0x100100080
is GSkyModel
data:data: 0x1001000b0
is GDataModel

The essential difference here is that it handles pointers, not references. The assignment is from one pointer to another. But it requires a delete operation.

I guess it comes down to the question of whether assignments should be allowed between derived classes. Here a discussion on this topic on the web:
http://stackoverflow.com/questions/10868234/how-do-i-prevent-cross-assignment-between-2-classes-derived-from-string

Possible solutions seem to be:
  • use explicit to prevent conversion
  • declare private constructors and assignment operators for all the the types that one wants to prohibit direct copying from

#3 Updated by Knödlseder Jürgen over 12 years ago

  • % Done changed from 0 to 20

#4 Updated by Knödlseder Jürgen over 12 years ago

It is tempting to replace the reference operator by a pointer operator, yet only the reference operator allows assignment.

On the other hand, it’s the assignment that poses problems, so can we really use a reference operator for save assignment here? I would love to code an assignement such as

GModels m;
GModelPointSource p;
...
m[0] = p;

but can this really be done?

The question is: how can we implement model assignment to the model container class? In Python this is easy as there is the __setitem__ function. But in C++ we have to use a reference operator. The essential method is then the GModel::operator= method that executes the assignment. But this method gives no handle about the manipulation of derived classes. So we're stuck with our assignment operator.

Note that an assignment from a derived class to a base class is not safe, the extra bit from the derived class just get sliced off:

The slicing problem is serious because it can result in memory corruption, and it is very difficult to guarantee a program does not suffer from it. To design it out of the language, classes that support inheritance should be accessible by reference only (not by value).

#5 Updated by Knödlseder Jürgen over 12 years ago

Got the solution. We need a reference to pointer operator. See the code below:

#include <string>
#include <vector>
#include <iostream>

class GModel {
public:
    GModel(void) {}
    virtual ~GModel(void) {}
    virtual std::string type(void) const = 0;
    virtual GModel* clone(void) const = 0;
    virtual std::string name(void) const { return m_name; }
    virtual GModel& operator=(const GModel& model)
    {
        if (this != &model) {
            std::cout << "GModel& operator=(const GModel& model)" << std::endl;
            m_name = model.m_name;
        }
    }
    std::string m_name;
};

class GModels {
public:
    GModels(void) {}
    virtual ~GModels(void) {}
    virtual void append(const GModel& model) { m_models.push_back(model.clone()); }
    GModel*& operator[](int index) { return (m_models[index]); }
    std::vector<GModel*> m_models;
};

class GSkyModel : public GModel {
public:
    GSkyModel(void) { m_name="sky"; }
    virtual ~GSkyModel(void) {}
    virtual std::string type(void) const { return "sky"; }
    virtual GSkyModel* clone(void) const { return new GSkyModel(*this); }
};

class GDataModel : public GModel {
public:
    GDataModel(void) { m_name="data"; }
    virtual ~GDataModel(void) {}
    virtual std::string type(void) const { return "data"; }
    virtual GDataModel* clone(void) const { return new GDataModel(*this); }
};

int main(void)
{
    GModels    models;
    GSkyModel  sky;
    GDataModel data;
    models.append(sky);
    std::cout << models[0]->type() << ":" << models[0]->name() << std::endl;
    models[0] = &data;
    std::cout << models[0]->type() << ":" << models[0]->name() << std::endl;

    return 0;
}

In line 27 there is the reference to pointer operator. Instead of returning a reference to a value it provides a reference to a pointer. So we assign pointers, no copy of data is made.

In line 54 the data object is now assigned. It's what I want: being able to assign a model. I just have to specify the reference here instead of the value. This makes clear that nothing is really copied, and if we want indeed copy the data object, we would have to use the clone method, i.e.

models[0] = data.clone();

For the record, the output of this program is

sky:sky
data:data

#6 Updated by Knödlseder Jürgen over 12 years ago

  • Status changed from New to Feedback
  • % Done changed from 20 to 100

I now implemented the reference to pointer logic in the GModels::operator[] access operator - with success. This can be seen in the following Python test dump:

>>> from gammalib import *
>>> m=GModels()
>>> p=GModelPointSource()
>>> p.name("Point source")
>>> e=GModelExtendedSource()
>>> e.name("Extended source")
>>> m.append(p)
>>> print m
=== GModels ===
 Number of models ..........: 1
 Number of parameters ......: 0
=== GModelPointSource ===
 Name ......................: Point source
 Instruments ...............: all
 Model type ................: 
 Number of parameters ......: 0
 Number of spatial par's ...: 0
 Number of spectral par's ..: 0
 Number of temporal par's ..: 0
>>> m[0]=e
>>> print m
=== GModels ===
 Number of models ..........: 1
 Number of parameters ......: 0
=== GModelExtendedSource ===
 Name ......................: Extended source
 Instruments ...............: all
 Model type ................: 
 Number of parameters ......: 0
 Number of spatial par's ...: 0
 Number of spectral par's ..: 0
 Number of temporal par's ..: 0
>>> print type(m[0])
<class 'gammalib.model.GModelExtendedSource'>
>>> print m[0].type()
ExtendedSource

#7 Updated by Knödlseder Jürgen over 12 years ago

Just for the record:

I implemented the Python assignment operator as a deep copy of the class. This is to make the assignment analogous to the append() method. Otherwise we would always rely on the fact that the original object does still exist. This is also analogous to the way how Python lists work;

>>> a=[]
>>> a.append(78)
>>> print a
[78]
>>> b=13
>>> a[0]=b
>>> print a
[13]
>>> del b
>>> print a
[13]

To show that the same behavior is implemented in GModels, continuing the above example we get

>>> del e
>>> print m
=== GModels ===
 Number of models ..........: 1
 Number of parameters ......: 0
=== GModelExtendedSource ===
 Name ......................: Extended source
 Instruments ...............: all
 Model type ................: 
 Number of parameters ......: 0
 Number of spatial par's ...: 0
 Number of spectral par's ..: 0
 Number of temporal par's ..: 0

#8 Updated by Knödlseder Jürgen over 12 years ago

Although the above solution is possible and close to the desiderate, it is subject to memory leaks. Assigning a pointer to a preexisting memory cell will overwrite the pointer that lived there before, and the link to the respective memory will be lost.

In conclusion, it is not possible to write a save method for setting values of a container class holding pointers through an access operator.

In summary, here the problems that occur:
  • accessing the pointers through a reference leads to slicing
  • accessing the pointers through a reference to a pointer leads to memory leaks

Thus, a specific set method needs to be implemented for setting of the pointers. This should solve by the way Bug #515.

A section will be added to the coding & design standards document, explaining how this issue should be solved.

#9 Updated by Knödlseder Jürgen about 12 years ago

  • Status changed from Feedback to Closed

Also available in: Atom PDF