Re: [eigen] FFT update

[Benoit Jacob <jacob.benoit.1@xxxxxxxxx>]

2010/1/24 Rohit Garg <rpg.314@xxxxxxxxx>:
> On Sun, Jan 24, 2010 at 12:33 PM, Benoit Jacob <jacob.benoit.1@xxxxxxxxx> wrote:
>> 2010/1/24 Benoit Jacob <jacob.benoit.1@xxxxxxxxx>:
>>> 2010/1/23 Benoit Jacob <jacob.benoit.1@xxxxxxxxx>:
>>>> By the way, this "ReturnByValue::coeff() crashes by infinite
>>>> recursion" thing is really ugly, I wonder how to fix it!!
>>>>
>>>> C++0x at last brings the ability to remove a base class function from
>>>> a derived class... I wonder how to implement that idea in C++98. Is
>>>> that enable_if?
>>>
>>> Forget enable_if, forget C++0x, forget boost, i'm your new idol.
>
> Alrighty, :)
>>>
>>> If we want to make sure that using ReturnByValue::coeff() generates an
>>> error at compile time without generating an error until it's actually
>>> used, we can define a class:
>>>
>>> class Unusable
>>> {
>>>  // implement copy ctor/operator as private to disable them
>>>  Unusable(const Unusable&) {}
>>>  void operator=(const Unusable&) {}
>>> };
>>>
>>> We then let ReturnByValue::coeff() return that:
>>>
>>> class ReturnByValue
>>> {
>>> ...
>>>  Unusable coeff(int, int) { return Unusable(); }
>>> ...
>>> };
>>>
>>> The result is that when the user does e.g.
>>>  x = retval.coeff(i);
>>> or
>>>  x = retval.coeff(i) + 123;
>>> he gets compilation errors.
>>>
>>> Just doing
>>>  retval.coeff(i);
>>> doesn't give an error, but that should be optimizable as just a NOP anyway.
>>>
>>> With this, i go to bed full^Wproud of myself.
>>
>> Ah, this needed a little adjustment. Doing
>>  Unusable coeff(int, int) { return Unusable(); }
>> was already copying a Unusable object, the return value of that
>> method. While the trick still worked with GCC, it was probably
>> fragile, and the error wasn't very explicit.
>>
>> Here is a much better variant:
>>    const Unusable& coeff(int) const { return *reinterpret_cast<const
>> Unusable*>(this); }
>>
>> Now this method is really not using any Unusable constructor/operator.
>>
>> The error message when I do:
>>     float x = rotate(v).coeff(0);
>> is now just:
>>    returnbyvalue.cpp: In function ‘int main()’:
>>    returnbyvalue.cpp:122: error: cannot convert ‘const
>> Eigen::Unusable’ to ‘float’ in initialization
>>
>> ok to commit?
>
> I'd suggest that since you are *forcing* a compiler error here, it
> would be better to add a long descriptive error to go with it. Some
> thing like this
>
> http://eigen.tuxfamily.org/index.php?title=FAQ#I_need_help_with_compiler_errors.21

Yes, good idea!

Benoit

>
>>
>> Benoit
>>
>>>
>>> Benoit
>>>
>>>>
>>>> Benoit
>>>>
>>>> 2010/1/23 Benoit Jacob <jacob.benoit.1@xxxxxxxxx>:
>>>>> Hi,
>>>>>
>>>>> Here's an example: see attached file.
>>>>> Enjoy!!
>>>>> Benoit
>>>>>
>>>>> 2010/1/22 Benoit Jacob <jacob.benoit.1@xxxxxxxxx>:
>>>>>> ok let me write an example...
>>>>>>
>>>>>> 2010/1/22 Jitse Niesen <jitse@xxxxxxxxxxxxxxxxx>:
>>>>>>> On Fri, 22 Jan 2010, Benoit Jacob wrote:
>>>>>>>
>>>>>>>> Our "policy" is now:
>>>>>>>> - try to return by value, optimizing with the ReturnByValue class,
>>>>>>>> wherever possible;
>>>>>>>> - otherwise you can just use references.
>>>>>>>
>>>>>>> Can somebody please explain how to use the ReturnByValue class?
>>>>>>> Or should I just extrapolate from the source code?
>>>>>>> I could not find any documentation.
>>>>>>>
>>>>>>> Jitse
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>
>>>>>
>>>>
>>>
>>
>>
>>
>
>
>
> --
> Rohit Garg
>
> http://rpg-314.blogspot.com/
>
> Senior Undergraduate
> Department of Physics
> Indian Institute of Technology
> Bombay
>
>
>