Date: Tue, 30 Jul 2024 18:16:21 +0400
Tiago,I seriously appreciate your contributions very much.But please try to read the thread from beginning and accumulate the idea's details then try to understand it as whole picture.Also I'm not discussing the current tools available with the current standard, I'm just using them to illustrate a point, and invent other tools to explain the idea.So, when i write, (int)<double>, i know it is not correct C++, but it is just merely a way to express a variable template of type int instantiated for type double. That's all.Sent from my Galaxy
-------- Original message --------From: Tiago Freire <tmiguelf_at_[hidden]> Date: 7/30/24 5:55 PM (GMT+04:00) To: std-proposals_at_[hidden], Jonathan Wakely <cxx_at_[hidden]> Cc: organicoman <organicoman_at_[hidden]> Subject: RE: [std-proposals] Function overload set type information loss
> template<typename T>
> int globalVar = 123;
> In theory the entity: glabalVar<int>
> Is not the same as : globalVar<double>
> Otherwise why we specialize based on template type at all.!!?? We can just write
globalVar
Nope. They are all “int”. The type of “globalVar<double>” is an int, the name is “globalVar<double>”.
There’s no “globalVar” whose type is “int<double>”.
From: Std-Proposals <std-proposals-bounces_at_[hidden]pp.org>
On Behalf Of organicoman via Std-Proposals
Sent: Tuesday, July 30, 2024 15:49
To: Jonathan Wakely <cxx_at_kayari.org>
Cc: organicoman <organicoman_at_[hidden]>; std-proposals_at_[hidden]
Subject: Re: [std-proposals] Function overload set type information loss
Thank you Jonathan for your discussion,
I just want to draw your attention that what I'm talking about is not available within the current state of the standard.
To understand this idea you must untie your mind from the rules set by the current state of C++. Be abstract.
Let me explain more:
In the current C++ standard we have 3 objects that lose their type information by design.
Namely:
1- functions
2- arrays
3- variable template.
At the compilation stage, we preserve the whole type signature of the above objects and we manipulate it, but as soon as we change the block scope we lose part of that information.
e.g
template<typename T>
int globalVar = 123;
In theory the entity: glabalVar<int>
Is not the same as : globalVar<double>
Otherwise why we specialize based on template type at all.!!?? We can just write
globalVar
Value wise they are both equal
globalVar<double> == globalVar<int> // true
But type wise they should not.
Yet the current standard says that:
std::is_same_v<decltype(globalVar<int>), decltype(globalVar<double>)>
// true.
This is incorrect in type's theory point of view.
The rule says that, if you specialize any type based on a template, the two types are different, even if their values are equal.
Given the above observations, and the current state of C++,
If we try to change the standard to implement the rule above for the 3 categories above, we will break a lot of code.
So what is the solution?
My solution is to divide the type of an object from the 3 categories above, into 2 parts:
1- an apparent type, which is implemented currently in the standard
2- an effective type, that reflects the real type at the time of its instantiation.
To implement the 2nd I suggested an operator called:
effective_decltype ();
How it works:
When you pass one object from the 3 listed above, it will return its full signature.
e.g.
decltype(globalVar<double>)
// is int , but
effective_decltype(globalVar<double>) // should return
(int)<double>, to say that
globalVar is a variable template of apparent type
int instantiated for
double
Take this concept and generalize it to arrays and functions.
So one can ask, how to implement this?
Remember that when you call, for example a function let say:
// declaration of function
void useGlobalVar(int v);
// usage of function
useGlobalVar(globalVar<double>);
The type is right there at the call site, you can capture it inside the function body using the suggested operator:
effective_decltype()
Even if you do:
auto var = globalVar<double>;
The auto key word behavior should be changed to capture the
effective type of right hand side of the assignment. So when we call
useGlobalVar(var);
The effective type is preserved.
One exception though is when you write:
int var = globalVar<double>;
Here you are explicitly casting the effective type of
var to be a plain int.
In conclusion:
I know it is a liitle bit twisted abstract math, but if you squint just a little bit, you can catch it.
Also the implications of this approach is very very beneficial.
Sent from my Galaxy
-------- Original message --------
From: Jonathan Wakely <cxx_at_[hidden]rg>
Date: 7/30/24 4:04 PM (GMT+04:00)
To: organicoman <organicoman_at_[hidden]>
Cc:
std-proposals_at_[hidden]ocpp.org
Subject: Re: [std-proposals] Function overload set type information loss
On Tue, 30 Jul 2024 at 11:09, organicoman <organicoman_at_[hidden]> wrote:
This technique of apparent and effective type, could be generalized to C style arrays and variable templates.
For example:
size_t foo(const char* arr)
{
//inside the function arr type is just a pointer
// but it's real type has an array extent N.
using eff_type = effective_decltype(arr);
return extract_array_size<eff_type>::value;
How would this be possible if 'foo' is separately compiled? How can the value returned by 'foo' depend on information only available at the call site, not in the function body? How would you avoid ODR violations, and how would you even
implement this?
The same way you implement
void foo(auto arg); // c++20
OK, that's creates a function template, equivalent to:
template<typename T> void foo(T arg);
So the equivalent for your example would be a function template too, which you can already write in C++ today (and all the way back to C++98):
template<std::size_t N> size_t foo(const char (&)[N]) { return N; }
And this already exists as std::size.
There's no need for your effective_decltype and extract_array_size magic, which are not implementable. For a non-template function, what you want is not possible. For a template function, you can do it already.
The reason you don't generally want to do that for more complicated functions is that you get a different foo<N> specialization for every different string length:
foo("1");
foo("12");
foo("123");
Each of these creates a separate specialization, so generates more code and increases your binary size.
}
template<typename T, typename V >
int gInteger = 1234;
int main()
{
char arr[]="cool_idea";
foo(arr); // prints 9
Why not 10? It's a const char[10] because it's a null terminated string.
auto i = gInteger<double,char>;
// type of i is integer
std::is_same_v<
effective_decltype(i),
(int)<double, char>)>; //<- this is a new notation to express a variable template
}
The whole idea is revolving around keeping the real type of each object, even if it decays to pointer ( ex. Functions and Arrays), or it loses the full type signature ( like variable templates, and function templates).
This sounds like magic.
No,
When you are writing a program you are explicitly passing variable names of fixed types between functions and objects, but as soon as you read data from IO, you will rely on a chosen type(picked by yourself)
to interpret that data.
So, when the compiler is crunching the source code, it keeps record of all declared types,
We can do that today with templates. You're trying to make non-templates behave like templates, and what you want is impossible in separately compiled code.
Just the current standard chose to omit and drop the effective type of arrays and functions to be of type pointer (decaying). I guess its a C legacy!
But now we have the tools to rectify this loss of information.
No you don't. If your 'foo' function is defined in a separate source file there is no way for it to get the information you want. It's impossible.
After all, a C++ source code, is just a series of inspections and manipulation of types. So the more we keep information about types the more it is useful.
What you probably want is reflection, which would not allow you to write your size_t foo(const char*) function, but if you write different functions, reflection would allow you to inspect and manipulate more information about your program.
The proposed idea will help in implementing reflection.
It's not needed.
Take this example:
template<typename T>
size_t
type_hash_value = typeid(T).hash_code();
void bar()
{
/* we will use the apparent type of a variable to instantiate a vector. And its effect type to record the type*/
/* remember the new notation of a variable template mentioned in example above:
(type_id)<Ts...> */
std::vector< (size_t)<T> > vec;
vec.push_back(type_hash_value<int>);
vec.push_back(type_hash_value<double>);
vec.push_back(type_hash_value<someT>);
size_t val;
std::cin >> val;
// IO input
vec.push_back(val);
/*apparent type of val is the same as its effective type => val is not a variable template yet it is still a valid entry to the vector*/
for(const auto& elem: vec)
if ( std::is_same_v<effective_decltype(elem),
(size_t)<double>>
)
return elem;
}
The elements of the vector above have two types of information: compile time and runtime.
The size and alignment of the elements is the same as std::size_t, that's for runtime.
The effective type of the elements is available at compile time only.
With this facility you can implement a program wide type_id hash table, and you can go back and forth between type<->value.
It is worth mentioning that, if the effective type and the apparent type are the same, this means that the inspected object is not templated, nor decayed.
I'm not 100% sure how this would be implemented, i need more research, but i guess compiler wizards out there have the necessary knowledge.
For now it is just a floating idea.
See "Reflection for C++26"
https://wg21.link/p2996 for the active work in this space.
Yes, I'm already following its progress closely.
Hello Gents,
Let:
O(f) be the Set of 'function overloads' of the function named
"f".
Let:
"Ret" be the return type of function
"f".
The elements of O(f) can have 2 forms:
1-> Ret f ( Vargs )
2-> Ret f <Ts...> ( Fargs )
Where:
1-> "Vargs"
is a variadic list of arguments types, participating or not in template type arguments deduction if
"f" is a template. e.g:
template<typename T, typename V>
void f(T, V, double, int)
Vargs = {T, V, double, int}
2-> "Fargs" is a fixed list of arguments types, not related to any template type parameter. e.g:
template <typename T, typename V>
void f(double, int)
Fargs= {double, int} // T,V are not in the list.
"Ts"
is just a variadic list of template parameters types.
The type of any function of the 1st form is:
decltype (f) = Ret(*)(Vargs),
which keeps information about the function's type template parameters participating in the function argument's list.
But the type of the 2nd form is:
decltype (f) = Ret(*)(Fargs);
No mater what the template parameters are, the type of the 2nd form always decays to:
Ret(*)(Fargs)
And always lose any type information about the function's type template parameters
Yet, when we want to get the address of such function, we are obligated to use the template types in the function name. e.g:
auto select_f = &f<Ts...>;
Otherwise we get overload ambiguity,
This is a proof that the template arguments participates in the function type.
In my opinion, the compiler should keep the template arguments type information.
I know that changing the type of "select_f" in the example above will break a lot of code.
But i have a suggestion.
If the compiler can keep record of :
* an apparent function type ; (the usual one)
decltype (&f<Ts...>) = Ret(*)(Fargs)
* and an effective function type which is:
decltype (&f<Ts...>) = Ret(*)<Ts...>(Fargs)
This would fix the problem without breaking any neck.
Why is this useful?
Take this example:
struct Erased{
std::any (*m_fun) (void);
template<auto Func>
constexpr Erased()
: m_fun(Func)
{ }
auto operator ()()
{
using f_type = effective_decltype (m_fun);
// imagin we have a type traits that
// extracts template types.
using T = extract_1st_template_type<f_type>;
return std::any_cast<T>( m_fun(void) );
}
};
template <typename Ret>
std::any foo()
{ return Ret{}; }
int main()
{
std::vector<Erased> vec;
vec.push_back(Erased<&foo<int>>{});
vec.push_back(Erased<&foo<double>>{});
vec.push_back(Erased<&foo<some_type>>{});
for(const auto& elem: vec)
DoSomethingBasedOnReturnType(elem());
}
Using this technique we can store the template type, then recall it back.
I guess it will make type erasure more effecient.
Any thoughts?
Regards
Nadir
Sent from my Galaxy
-------- Original message --------From: Tiago Freire <tmiguelf_at_[hidden]> Date: 7/30/24 5:55 PM (GMT+04:00) To: std-proposals_at_[hidden], Jonathan Wakely <cxx_at_[hidden]> Cc: organicoman <organicoman_at_[hidden]> Subject: RE: [std-proposals] Function overload set type information loss
> template<typename T>
> int globalVar = 123;
> In theory the entity: glabalVar<int>
> Is not the same as : globalVar<double>
> Otherwise why we specialize based on template type at all.!!?? We can just write
globalVar
Nope. They are all “int”. The type of “globalVar<double>” is an int, the name is “globalVar<double>”.
There’s no “globalVar” whose type is “int<double>”.
From: Std-Proposals <std-proposals-bounces_at_[hidden]pp.org>
On Behalf Of organicoman via Std-Proposals
Sent: Tuesday, July 30, 2024 15:49
To: Jonathan Wakely <cxx_at_kayari.org>
Cc: organicoman <organicoman_at_[hidden]>; std-proposals_at_[hidden]
Subject: Re: [std-proposals] Function overload set type information loss
Thank you Jonathan for your discussion,
I just want to draw your attention that what I'm talking about is not available within the current state of the standard.
To understand this idea you must untie your mind from the rules set by the current state of C++. Be abstract.
Let me explain more:
In the current C++ standard we have 3 objects that lose their type information by design.
Namely:
1- functions
2- arrays
3- variable template.
At the compilation stage, we preserve the whole type signature of the above objects and we manipulate it, but as soon as we change the block scope we lose part of that information.
e.g
template<typename T>
int globalVar = 123;
In theory the entity: glabalVar<int>
Is not the same as : globalVar<double>
Otherwise why we specialize based on template type at all.!!?? We can just write
globalVar
Value wise they are both equal
globalVar<double> == globalVar<int> // true
But type wise they should not.
Yet the current standard says that:
std::is_same_v<decltype(globalVar<int>), decltype(globalVar<double>)>
// true.
This is incorrect in type's theory point of view.
The rule says that, if you specialize any type based on a template, the two types are different, even if their values are equal.
Given the above observations, and the current state of C++,
If we try to change the standard to implement the rule above for the 3 categories above, we will break a lot of code.
So what is the solution?
My solution is to divide the type of an object from the 3 categories above, into 2 parts:
1- an apparent type, which is implemented currently in the standard
2- an effective type, that reflects the real type at the time of its instantiation.
To implement the 2nd I suggested an operator called:
effective_decltype ();
How it works:
When you pass one object from the 3 listed above, it will return its full signature.
e.g.
decltype(globalVar<double>)
// is int , but
effective_decltype(globalVar<double>) // should return
(int)<double>, to say that
globalVar is a variable template of apparent type
int instantiated for
double
Take this concept and generalize it to arrays and functions.
So one can ask, how to implement this?
Remember that when you call, for example a function let say:
// declaration of function
void useGlobalVar(int v);
// usage of function
useGlobalVar(globalVar<double>);
The type is right there at the call site, you can capture it inside the function body using the suggested operator:
effective_decltype()
Even if you do:
auto var = globalVar<double>;
The auto key word behavior should be changed to capture the
effective type of right hand side of the assignment. So when we call
useGlobalVar(var);
The effective type is preserved.
One exception though is when you write:
int var = globalVar<double>;
Here you are explicitly casting the effective type of
var to be a plain int.
In conclusion:
I know it is a liitle bit twisted abstract math, but if you squint just a little bit, you can catch it.
Also the implications of this approach is very very beneficial.
Sent from my Galaxy
-------- Original message --------
From: Jonathan Wakely <cxx_at_[hidden]rg>
Date: 7/30/24 4:04 PM (GMT+04:00)
To: organicoman <organicoman_at_[hidden]>
Cc:
std-proposals_at_[hidden]ocpp.org
Subject: Re: [std-proposals] Function overload set type information loss
On Tue, 30 Jul 2024 at 11:09, organicoman <organicoman_at_[hidden]> wrote:
This technique of apparent and effective type, could be generalized to C style arrays and variable templates.
For example:
size_t foo(const char* arr)
{
//inside the function arr type is just a pointer
// but it's real type has an array extent N.
using eff_type = effective_decltype(arr);
return extract_array_size<eff_type>::value;
How would this be possible if 'foo' is separately compiled? How can the value returned by 'foo' depend on information only available at the call site, not in the function body? How would you avoid ODR violations, and how would you even
implement this?
The same way you implement
void foo(auto arg); // c++20
OK, that's creates a function template, equivalent to:
template<typename T> void foo(T arg);
So the equivalent for your example would be a function template too, which you can already write in C++ today (and all the way back to C++98):
template<std::size_t N> size_t foo(const char (&)[N]) { return N; }
And this already exists as std::size.
There's no need for your effective_decltype and extract_array_size magic, which are not implementable. For a non-template function, what you want is not possible. For a template function, you can do it already.
The reason you don't generally want to do that for more complicated functions is that you get a different foo<N> specialization for every different string length:
foo("1");
foo("12");
foo("123");
Each of these creates a separate specialization, so generates more code and increases your binary size.
}
template<typename T, typename V >
int gInteger = 1234;
int main()
{
char arr[]="cool_idea";
foo(arr); // prints 9
Why not 10? It's a const char[10] because it's a null terminated string.
auto i = gInteger<double,char>;
// type of i is integer
std::is_same_v<
effective_decltype(i),
(int)<double, char>)>; //<- this is a new notation to express a variable template
}
The whole idea is revolving around keeping the real type of each object, even if it decays to pointer ( ex. Functions and Arrays), or it loses the full type signature ( like variable templates, and function templates).
This sounds like magic.
No,
When you are writing a program you are explicitly passing variable names of fixed types between functions and objects, but as soon as you read data from IO, you will rely on a chosen type(picked by yourself)
to interpret that data.
So, when the compiler is crunching the source code, it keeps record of all declared types,
We can do that today with templates. You're trying to make non-templates behave like templates, and what you want is impossible in separately compiled code.
Just the current standard chose to omit and drop the effective type of arrays and functions to be of type pointer (decaying). I guess its a C legacy!
But now we have the tools to rectify this loss of information.
No you don't. If your 'foo' function is defined in a separate source file there is no way for it to get the information you want. It's impossible.
After all, a C++ source code, is just a series of inspections and manipulation of types. So the more we keep information about types the more it is useful.
What you probably want is reflection, which would not allow you to write your size_t foo(const char*) function, but if you write different functions, reflection would allow you to inspect and manipulate more information about your program.
The proposed idea will help in implementing reflection.
It's not needed.
Take this example:
template<typename T>
size_t
type_hash_value = typeid(T).hash_code();
void bar()
{
/* we will use the apparent type of a variable to instantiate a vector. And its effect type to record the type*/
/* remember the new notation of a variable template mentioned in example above:
(type_id)<Ts...> */
std::vector< (size_t)<T> > vec;
vec.push_back(type_hash_value<int>);
vec.push_back(type_hash_value<double>);
vec.push_back(type_hash_value<someT>);
size_t val;
std::cin >> val;
// IO input
vec.push_back(val);
/*apparent type of val is the same as its effective type => val is not a variable template yet it is still a valid entry to the vector*/
for(const auto& elem: vec)
if ( std::is_same_v<effective_decltype(elem),
(size_t)<double>>
)
return elem;
}
The elements of the vector above have two types of information: compile time and runtime.
The size and alignment of the elements is the same as std::size_t, that's for runtime.
The effective type of the elements is available at compile time only.
With this facility you can implement a program wide type_id hash table, and you can go back and forth between type<->value.
It is worth mentioning that, if the effective type and the apparent type are the same, this means that the inspected object is not templated, nor decayed.
I'm not 100% sure how this would be implemented, i need more research, but i guess compiler wizards out there have the necessary knowledge.
For now it is just a floating idea.
See "Reflection for C++26"
https://wg21.link/p2996 for the active work in this space.
Yes, I'm already following its progress closely.
Hello Gents,
Let:
O(f) be the Set of 'function overloads' of the function named
"f".
Let:
"Ret" be the return type of function
"f".
The elements of O(f) can have 2 forms:
1-> Ret f ( Vargs )
2-> Ret f <Ts...> ( Fargs )
Where:
1-> "Vargs"
is a variadic list of arguments types, participating or not in template type arguments deduction if
"f" is a template. e.g:
template<typename T, typename V>
void f(T, V, double, int)
Vargs = {T, V, double, int}
2-> "Fargs" is a fixed list of arguments types, not related to any template type parameter. e.g:
template <typename T, typename V>
void f(double, int)
Fargs= {double, int} // T,V are not in the list.
"Ts"
is just a variadic list of template parameters types.
The type of any function of the 1st form is:
decltype (f) = Ret(*)(Vargs),
which keeps information about the function's type template parameters participating in the function argument's list.
But the type of the 2nd form is:
decltype (f) = Ret(*)(Fargs);
No mater what the template parameters are, the type of the 2nd form always decays to:
Ret(*)(Fargs)
And always lose any type information about the function's type template parameters
Yet, when we want to get the address of such function, we are obligated to use the template types in the function name. e.g:
auto select_f = &f<Ts...>;
Otherwise we get overload ambiguity,
This is a proof that the template arguments participates in the function type.
In my opinion, the compiler should keep the template arguments type information.
I know that changing the type of "select_f" in the example above will break a lot of code.
But i have a suggestion.
If the compiler can keep record of :
* an apparent function type ; (the usual one)
decltype (&f<Ts...>) = Ret(*)(Fargs)
* and an effective function type which is:
decltype (&f<Ts...>) = Ret(*)<Ts...>(Fargs)
This would fix the problem without breaking any neck.
Why is this useful?
Take this example:
struct Erased{
std::any (*m_fun) (void);
template<auto Func>
constexpr Erased()
: m_fun(Func)
{ }
auto operator ()()
{
using f_type = effective_decltype (m_fun);
// imagin we have a type traits that
// extracts template types.
using T = extract_1st_template_type<f_type>;
return std::any_cast<T>( m_fun(void) );
}
};
template <typename Ret>
std::any foo()
{ return Ret{}; }
int main()
{
std::vector<Erased> vec;
vec.push_back(Erased<&foo<int>>{});
vec.push_back(Erased<&foo<double>>{});
vec.push_back(Erased<&foo<some_type>>{});
for(const auto& elem: vec)
DoSomethingBasedOnReturnType(elem());
}
Using this technique we can store the template type, then recall it back.
I guess it will make type erasure more effecient.
Any thoughts?
Regards
Nadir
Sent from my Galaxy
-- Std-Proposals mailing list Std-Proposals_at_[hidden] https://lists.isocpp.org/mailman/listinfo.cgi/std-proposals
Received on 2024-07-30 14:16:33