I'm writing a Foo class for a Metal compute shader and want users to be able to add Foos regardless of the address space in which they reside. If I take the values by reference, this means I have to declare many permutations of which address spaces things live in:
class Foo {
private:
uint8_t bunchOfData[160];
public:
operator+(const thread Foo&) const thread;
operator+(const constant Foo&) cosnt thread;
operator+(const thread Foo&) const constant;
operator+(const constant Foo&) const constant;
// ... More overloads for local. For device, users should manually unpack to maintain coalescing.
}
Is there a way to genericize the address space qualifiers using templates?
Should I just use pass by value and dodge this whole problem entirely?
The size of Foo is non-trivial, like 160 bytes each.
When I last wrote compute kernels (i.e. 2009 era OpenCL), GPUs couldn't make function calls so pass-by-value vs pass-by-reference didn't really make a difference since everything got inlined anyways. Is this still the case?
If pass-by-reference is still useful, is it conventional to simply have the caller copy the value into thread space and only provide the thread-thread variant?
I've tried writing:
class Foo {
private:
uint8_t bunchOfData[160];
public:
template <typename T> operator+(const T& rhs) const;
}
but this gives me the error
error: reference type must have explicit address space qualifier
template <typename T> operator+(const T& rhs);
Not that I'm even specifying the address space of this correctly anyways...
Related
Ok, muddling though Stack on the particulars about void*, books like The C Programming Language (K&R) and The C++ Programming Language (Stroustrup). What have I learned? That void* is a generic pointer with no type inferred. It requires a cast to any defined type and printing void* just yields the address.
What else do I know? void* can't be dereferenced and thus far remains the one item in C/C++ from which I have discovered much written about but little understanding imparted.
I understand that it must be cast such as *(char*)void* but what makes no sense to me for a generic pointer is that I must somehow already know what type I need in order to grab a value. I'm a Java programmer; I understand generic types but this is something I struggle with.
So I wrote some code
typedef struct node
{
void* data;
node* link;
}Node;
typedef struct list
{
Node* head;
}List;
Node* add_new(void* data, Node* link);
void show(Node* head);
Node* add_new(void* data, Node* link)
{
Node* newNode = new Node();
newNode->data = data;
newNode->link = link;
return newNode;
}
void show(Node* head)
{
while (head != nullptr)
{
std::cout << head->data;
head = head->link;
}
}
int main()
{
List list;
list.head = nullptr;
list.head = add_new("My Name", list.head);
list.head = add_new("Your Name", list.head);
list.head = add_new("Our Name", list.head);
show(list.head);
fgetc(stdin);
return 0;
}
I'll handle the memory deallocation later. Assuming I have no understanding of the type stored in void*, how do I get the value out? This implies I already need to know the type, and this reveals nothing about the generic nature of void* while I follow what is here although still no understanding.
Why am I expecting void* to cooperate and the compiler to automatically cast out the type that is hidden internally in some register on the heap or stack?
I'll handle the memory deallocation later. Assuming I have no understanding of the type stored in void*, how do I get the value out?
You can't. You must know the valid types that the pointer can be cast to before you can dereference it.
Here are couple of options for using a generic type:
If you are able to use a C++17 compiler, you may use std::any.
If you are able to use the boost libraries, you may use boost::any.
Unlike Java, you are working with memory pointers in C/C++. There is no encapsulation whatsoever. The void * type means the variable is an address in memory. Anything can be stored there. With a type like int * you tell the compiler what you are referring to. Besides the compiler knows the size of the type (say 4 bytes for int) and the address will be a multiple of 4 in that case (granularity/memory alignment). On top, if you give the compiler the type it will perform consistency checks at compilation time. Not after. This is not happening with void *.
In a nutshell, you are working bare metal. The types are compiler directives and do not hold runtime information. Nor does it track the objects you are dynamically creating. It is merely a segment in memory that is allocated where you can eventually store anything.
The main reason to use void* is that different things may be pointed at. Thus, I may pass in an int* or Node* or anything else. But unless you know either the type or the length, you can't do anything with it.
But if you know the length, you can handle the memory pointed at without knowing the type. Casting it as a char* is used because it is a single byte, so if I have a void* and a number of bytes, I can copy the memory somewhere else, or zero it out.
Additionally, if it is a pointer to a class, but you don't know if it is a parent or inherited class, you may be able to assume one and find out a flag inside the data which tells you which one. But no matter what, when you want to do much beyond passing it to another function, you need to cast it as something. char* is just the easiest single byte value to use.
Your confusion derived from habit to deal with Java programs. Java code is set of instruction for a virtual machine, where function of RAM is given to a sort of database, which stores name, type, size and data of each object. Programming language you're learning now is meant to be compiled into instruction for CPU, with same organization of memory as underlying OS have. Existing model used by C and C++ languages is some abstract built on top of most of popular OSes in way that code would work effectively after being compiled for that platform and OS. Naturally that organization doesn't involve string data about type, except for famous RTTI in C++.
For your case RTTI cannot be used directly, unless you would create a wrapper around your naked pointer, which would store the data.
In fact C++ library contains a vast collection of container class templates that are useable and portable, if they are defined by ISO standard. 3/4 of standard is just description of library often referred as STL. Use of them is preferable over working with naked pointers, unless you mean to create own container for some reason. For particular task only C++17 standard offered std::any class, previously present in boost library. Naturally, it is possible to reimplement it, or, in some cases, to replace by std::variant.
Assuming I have no understanding of the type stored in void*, how do I get the value out
You don't.
What you can do is record the type stored in the void*.
In c, void* is used to pass around a binary chunk of data that points at something through one layer of abstraction, and recieve it at the other end, casting it back to the type that the code knows it will be passed.
void do_callback( void(*pfun)(void*), void* pdata ) {
pfun(pdata);
}
void print_int( void* pint ) {
printf( "%d", *(int*)pint );
}
int main() {
int x = 7;
do_callback( print_int, &x );
}
here, we forget thet ype of &x, pass it through do_callback.
It is later passed to code inside do_callback or elsewhere that knows that the void* is actually an int*. So it casts it back and uses it as an int.
The void* and the consumer void(*)(void*) are coupled. The above code is "provably correct", but the proof does not lie in the type system; instead, it depends on the fact we only use that void* in a context that knows it is an int*.
In C++ you can use void* similarly. But you can also get fancy.
Suppose you want a pointer to anything printable. Something is printable if it can be << to a std::ostream.
struct printable {
void const* ptr = 0;
void(*print_f)(std::ostream&, void const*) = 0;
printable() {}
printable(printable&&)=default;
printable(printable const&)=default;
printable& operator=(printable&&)=default;
printable& operator=(printable const&)=default;
template<class T,std::size_t N>
printable( T(&t)[N] ):
ptr( t ),
print_f( []( std::ostream& os, void const* pt) {
T* ptr = (T*)pt;
for (std::size_t i = 0; i < N; ++i)
os << ptr[i];
})
{}
template<std::size_t N>
printable( char(&t)[N] ):
ptr( t ),
print_f( []( std::ostream& os, void const* pt) {
os << (char const*)pt;
})
{}
template<class T,
std::enable_if_t<!std::is_same<std::decay_t<T>, printable>{}, int> =0
>
printable( T&& t ):
ptr( std::addressof(t) ),
print_f( []( std::ostream& os, void const* pt) {
os << *(std::remove_reference_t<T>*)pt;
})
{}
friend
std::ostream& operator<<( std::ostream& os, printable self ) {
self.print_f( os, self.ptr );
return os;
}
explicit operator bool()const{ return print_f; }
};
what I just did is a technique called "type erasure" in C++ (vaguely similar to Java type erasure).
void send_to_log( printable p ) {
std::cerr << p;
}
Live example.
Here we created an ad-hoc "virtual" interface to the concept of printing on a type.
The type need not support any actual interface (no binary layout requirements), it just has to support a certain syntax.
We create our own virtual dispatch table system for an arbitrary type.
This is used in the C++ standard library. In c++11 there is std::function<Signature>, and in c++17 there is std::any.
std::any is void* that knows how to destroy and copy its contents, and if you know the type you can cast it back to the original type. You can also query it and ask it if it a specific type.
Mixing std::any with the above type-erasure techinque lets you create regular types (that behave like values, not references) with arbitrary duck-typed interfaces.
The following code will not compile on gcc 4.8.2.
The problem is that this code will attempt to copy construct an std::pair<int, A> which can't happen due to struct A missing copy and move constructors.
Is gcc failing here or am I missing something?
#include <map>
struct A
{
int bla;
A(int blub):bla(blub){}
A(A&&) = delete;
A(const A&) = delete;
A& operator=(A&&) = delete;
A& operator=(const A&) = delete;
};
int main()
{
std::map<int, A> map;
map.emplace(1, 2); // doesn't work
map.emplace(std::piecewise_construct,
std::forward_as_tuple(1),
std::forward_as_tuple(2)
); // works like a charm
return 0;
}
As far as I can tell, the issue isn't caused by map::emplace, but by pair's constructors:
#include <map>
struct A
{
A(int) {}
A(A&&) = delete;
A(A const&) = delete;
};
int main()
{
std::pair<int, A> x(1, 4); // error
}
This code example doesn't compile, neither with coliru's g++4.8.1 nor with clang++3.5, which are both using libstdc++, as far as I can tell.
The issue is rooted in the fact that although we can construct
A t(4);
that is, std::is_constructible<A, int>::value == true, we cannot implicitly convert an int to an A [conv]/3
An expression e can be implicitly converted to a type T if and only if the declaration T t=e; is well-formed,
for some invented temporary variable t.
Note the copy-initialization (the =). This creates a temporary A and initializes t from this temporary, [dcl.init]/17. This initialization from a temporary tries to call the deleted move ctor of A, which makes the conversion ill-formed.
As we cannot convert from an int to an A, the constructor of pair that one would expect to be called is rejected by SFINAE. This behaviour is surprising, N4387 - Improving pair and tuple analyses and tries to improve the situation, by making the constructor explicit instead of rejecting it. N4387 has been voted into C++1z at the Lenexa meeting.
The following describes the C++11 rules.
The constructor I had expected to be called is described in [pairs.pair]/7-9
template<class U, class V> constexpr pair(U&& x, V&& y);
7 Requires: is_constructible<first_type, U&&>::value is true and
is_constructible<second_type, V&&>::value is true.
8 Effects: The
constructor initializes first with std::forward<U>(x) and second with
std::forward<V>(y).
9 Remarks: If U is not implicitly convertible to
first_type or V is not implicitly convertible to second_type this
constructor shall not participate in overload resolution.
Note the difference between is_constructible in the Requires section, and "is not implicitly convertible" in the Remarks section. The requirements are fulfilled to call this constructor, but it may not participate in overload resolution (= has to be rejected via SFINAE).
Therefore, overload resolution needs to select a "worse match", namely one whose second parameter is a A const&. A temporary is created from the int argument and bound to this reference, and the reference is used to initialize the pair data member (.second). The initialization tries to call the deleted copy ctor of A, and the construction of the pair is ill-formed.
libstdc++ has (as an extension) some nonstandard ctors. In the latest doxygen (and in 4.8.2), the constructor of pair that I had expected to be called (being surprised by the rules required by the Standard) is:
template<class _U1, class _U2,
class = typename enable_if<__and_<is_convertible<_U1, _T1>,
is_convertible<_U2, _T2>
>::value
>::type>
constexpr pair(_U1&& __x, _U2&& __y)
: first(std::forward<_U1>(__x)), second(std::forward<_U2>(__y)) { }
and the one that is actually called is the non-standard:
// DR 811.
template<class _U1,
class = typename enable_if<is_convertible<_U1, _T1>::value>::type>
constexpr pair(_U1&& __x, const _T2& __y)
: first(std::forward<_U1>(__x)), second(__y) { }
The program is ill-formed according to the Standard, it is not merely rejected by this non-standard ctor.
As a final remark, here's the specification of is_constructible and is_convertible.
is_constructible [meta.rel]/4
Given the following function prototype:
template <class T>
typename add_rvalue_reference<T>::type create();
the predicate condition for a template specialization is_constructible<T, Args...> shall be satisfied if and only if the following variable definition would be well-formed for some invented variable t:
T t(create<Args>()...);
[Note: These tokens are never interpreted as a function declaration. — end note] Access checking is performed as if in a context unrelated to T and any of the Args. Only the validity of the immediate context of the variable initialization is considered.
is_convertible [meta.unary.prop]/6:
Given the following function prototype:
template <class T>
typename add_rvalue_reference<T>::type create();
the predicate condition for a template specialization is_convertible<From, To> shall be satisfied if and
only if the return expression in the following code would be well-formed, including any implicit conversions
to the return type of the function:
To test() {
return create<From>();
}
[Note: This requirement gives well defined results for reference types, void types, array types, and function types. — end note] Access checking is performed as if in a context unrelated to To and From. Only
the validity of the immediate context of the expression of the return-statement (including conversions to
the return type) is considered.
For your type A,
A t(create<int>());
is well-formed; however
A test() {
return create<int>();
}
creates a temporary of type A and tries to move that into the return-value (copy-initialization). That selects the deleted ctor A(A&&) and is therefore ill-formed.
I've always seen std::forward being utilized as below, utilized inside a template function
template<class T>
void foo(T&& arg): bar(std::forward<T>(arg)){}
Suppose I want to do this.
class A
{
private:
std::function<void()> bar;
public:
template<class T>
A(T&& arg):
bar(std::forward<T>(arg))
{}
};
Since bar already has its type defined. I can also directly specify T as std::function<void()> >.
class A
{
private:
std::function<void()> bar;
public:
A(std::function<void()>&& arg):
bar(std::forward<std::function<void()>>(arg))
{}
};
Both would be ok to compile. However, the second realization only support A(const std::function<void()>). While the first realization support A(const std::function<void()>&) and A(std::function<void()>&&) etc.
Forward is a conditional move of its argument. It is almost equivalent to std::move if and only if the type passed to it is a value type or rvalue reference type.
A move is a cast to an rvalue reference.
If you pass a different type to std::forward than its argument type, it will do horrible things. If convertible between, this would often involve creating a temporary within a function then returning a reference to it.
The proper thing to pass to std::forward(x) is X, where the type of x is X&&. Anything else is going to be extremely quirky and advanced use, and will probably cause unexpected behavior...
In your case, the second works fine, but is pointless. As std::forward is a conditional move, and we are passing it a fixed type, we know it is a std::move.
So we should replace std::forward<std::function<void()>>(arg) with std::move(arg), which is both clearer and more conventional. Also, equivalent in this case.
Generally std::forward should only be used in cases where you are using forwarding references.
Working on learning how to use smart pointers and C++ in general... Assume that I have the following class:
template<typename T>
class MyClass {
public:
MyClass(const T& def_val)
private:
std::unique_ptr<T> default_val;
};
What is the idiomatic way of implementing the constructor if I would only like to store a pointer to an object of type T with the value given in the default_val class member? My understanding is also that I don't have to define a destructor at all, since the unique_ptr will automatically taking care of cleaning up itself?
The way you have written your code, MyClass can only store a unique pointer to a copy of the constructor parameter:
MyClass::MyClass(const T& def_val)
: default_val(new T(def_val))
{
}
This means that T must be copy constructible.
My understanding is also that I don't have to define a destructor at all, since the unique_ptr will automatically taking care of cleaning up itself?
Correct. That is 1 of 2 main purposes for unique_ptr, the 2nd being the guarantee that it has only one owner.
If you're using C++11 you could add also a constructor that accepts an rvalue ref
template<typename T>
class MyClass {
public:
MyClass(T&& def_val) : default_val(new T(std::move(def_val))) {}
MyClass::MyClass(const T& def_val) : default_val(new T(def_val)) {}
private:
std::unique_ptr<T> default_val;
};
now you accept both const ref, generating a copy, or temporaries
I wanted to test a simple thing like the following:
#include <iostream>
#include <boost/variant.hpp>
template<typename T1,typename T2>
std::ostream& operator<<(std::ostream& os, const std::pair<T1,T2>& dt){
os << dt.first << dt.second;
return os;
}
int main(){
boost::variant<int, std::pair<int,int>, bool> v;
v = std::pair<int,int>(3,3);
std::cout << v << std::endl;
}
This should actually work, because for normal types, like int, double and so on, it compiles.
boost::variant has a printer vistor which it uses internally to output the content to the stream.
Actually this fails to compile, but I do not really know the problem:
The codes fails here: in variant_io.hpp
template <typename OStream>
class printer
: public boost::static_visitor<>
{
private: // representation
OStream& out_;
public: // structors
explicit printer(OStream& out)
: out_( out )
{
}
public: // visitor interface
template <typename T>
void operator()(const T& operand) const
{
out_ << operand; // HEEEEEEERRRRREE!!!!!!!!!!!!
}
private:
printer& operator=(const printer&);
};
With the message:
/usr/local/include/boost/variant/detail/variant_io.hpp|64|error: cannot bind 'std::basic_ostream<char>' lvalue to 'std::basic_ostream<char>&&'
Does someone know what I did wrong, and why?
Thanks a lot!
Most likely it's not finding your overload of operator <<, and then gets confused trying to match some other overload, leading to whatever message you're getting.
What you did wrong: You overloaded the stream operator in the global namespace instead of the namespace the right-hand-side class is defined in, so it's not found by ADL.
Trying to overload the stream operator for a standard class is a doomed exercise in the first place, unfortunately. You can't actually do that. I'm not sure if there is an explicit rule against it. However, if you place the operator in namespace std as you have to in order to make it properly findable by ADL, you violate the rule that you can't add your own stuff to namespace std except in very specific cases, this not being one of them.
The bottom line is that std::pair doesn't have a stream operator, and it's not possible to legally add a generic one that is useful. You can add one for a specific instantiation, if one of the parameters is a class you defined yourself; in this case the operator needs to be placed next to your own class.
Overloaded operator<< must be findable by argument dependent lookup. That means you have to put it in associated namespace of one of the arguments.
The first argument has only one associated namespace, std. The second also has only one associated namespace, std. However it is only permitted to overload symbols in std for user-defined types. Since std::pair<int, int> is not user-defined type, this is not allowed. However it is allowed for a structure or class you define yourself. Obviously in that case it is easier to place the overload to your namespace, not std.
That said if you put that overload in namespace std, it will actually work.
Also note, that boost::tuple does have operator<< (in separate header that you have to include, but it does), so you can use that instead.