For learning purposes, I'm trying to implement a similar class as the Value object in micrograd. To be more precise, I'm trying to implement it in template metaprogramming.
So far, I'm able to do addition and multiplication with the following code:
#include <iostream>
#include <array>
template<double N>
struct Value {
static constexpr double data = N;
friend std::ostream& operator<<(std::ostream& os, const Value &v) {
os << "Value(data=" << v.data << ")";
return os;
}
};
template<double N, double R>
constexpr auto operator+(const Value<N> lhs, const Value<R> rhs) {
return Value<lhs.data + rhs.data>();
}
template<double N, double R>
constexpr auto operator*(const Value<N> lhs, const Value<R> rhs) {
return Value<lhs.data * rhs.data>();
}
int main() {
Value<3.5> v;
Value<2.0> w;
std::cout << v << std::endl;
std::cout << v + w << std::endl;
std::cout << v * w << std::endl;
return 0;
}
Now, the problem arises when I try to keep reference of the children for each Value object. For example, I would like to be able to do v + w, this would give me a new Value object with data equal to 5.5 and a list of children {v, w}. Furthermore, I would like to keep the children field generic and not assume any number of children. In other words, I'm trying to build a N-ary tree, not a binary tree. So right now, I have:
#include <iostream>
#include <array>
template<double N, auto... Children>
struct Value {
static constexpr double data = N;
static constexpr Value children[sizeof...(Children)]{Children...};
friend std::ostream& operator<<(std::ostream& os, const Value &v) {
os << "Value(data=" << v.data << ", children=[";
for (auto &&child : children) {
os << child << ",";
}
os << "])";
return os;
}
};
template<double N, auto... Ns, double R, auto... Rs>
constexpr auto operator+(const Value<N, Ns...> lhs, const Value<R, Rs...> rhs) {
return Value<lhs.data + rhs.data, lhs, rhs>();
}
template<double N, auto... Ns, double R, auto... Rs>
constexpr auto operator*(const Value<N, Ns...> lhs, const Value<R, Rs...> rhs) {
return Value<lhs.data * rhs.data, lhs, rhs>();
}
int main() {
Value<3.5> v;
Value<2.0> w;
std::cout << v << std::endl;
std::cout << v + w << std::endl;
std::cout << v * w << std::endl;
return 0;
}
and this gives me the following errors:
main.cc: In instantiation of 'constexpr const Value<5.5e+0, Value<3.5e+0>(), Value<2.0e+0>()> Value<5.5e+0, Value<3.5e+0>(), Value<2.0e+0>()>::children [2]':
main.cc:11:23: required from 'std::ostream& operator<<(std::ostream&, const Value<5.5e+0, Value<3.5e+0>(), Value<2.0e+0>()>&)'
main.cc:34:19: required from here
main.cc:7:32: error: initializer for 'const Value<5.5e+0, Value<3.5e+0>(), Value<2.0e+0>()>' must be brace-enclosed
7 | static constexpr Value children[sizeof...(Children)]{Children...};
| ^~~~~~~~
main.cc: In instantiation of 'constexpr const Value<7.0e+0, Value<3.5e+0>(), Value<2.0e+0>()> Value<7.0e+0, Value<3.5e+0>(), Value<2.0e+0>()>::children [2]':
main.cc:11:23: required from 'std::ostream& operator<<(std::ostream&, const Value<7.0e+0, Value<3.5e+0>(), Value<2.0e+0>()>&)'
main.cc:35:19: required from here
main.cc:7:32: error: initializer for 'const Value<7.0e+0, Value<3.5e+0>(), Value<2.0e+0>()>' must be brace-enclosed
Unfortunately, I'm not understanding the error. It seems to say that I need to put braces around Children... but I already have some. If someone could give me a feedback on the error and maybe on how to improve the code, that would be highly appreciated.
Edit
This is seems to be caused by the operator<< function when accessing the children property, not sure why, but this might give a lead to someone more experienced.
Value<3.5> and Value<2.0> are different types. You cannot have an array of elements of different types, i.e., you can't put Value<3.5> and Value<2.0> in the same array. You would either need to have an array of polymorphic base or a tuple. The latter comes with less (none) runtime overhead:
#include <iostream>
#include <tuple>
template<double N, auto... Children>
struct Value {
static constexpr double data = N;
static constexpr auto children = std::make_tuple(Children...);
friend std::ostream& operator<<(std::ostream& os, const Value &v) {
os << "Value(data=" << v.data << ", children=[";
std::apply([&os](auto const&... child) {
((os << child << ","), ...);
}, children);
os << "])";
return os;
}
};
template<double N, auto... Ns, double R, auto... Rs>
constexpr auto operator+(const Value<N, Ns...> lhs, const Value<R, Rs...> rhs) {
return Value<lhs.data + rhs.data, lhs, rhs>();
}
template<double N, auto... Ns, double R, auto... Rs>
constexpr auto operator*(const Value<N, Ns...> lhs, const Value<R, Rs...> rhs) {
return Value<lhs.data * rhs.data, lhs, rhs>();
}
int main() {
Value<3.5> v;
Value<2.0> w;
std::cout << v << std::endl;
std::cout << v + w << std::endl;
std::cout << v * w << std::endl;
return 0;
}
This prints:
Value(data=3.5, children=[])
Value(data=5.5, children=[Value(data=3.5, children=[]),Value(data=2, children=[]),])
Value(data=7, children=[Value(data=3.5, children=[]),Value(data=2, children=[]),])
Related
i am currently working on a c++ project and now i am stuck already for a while. It's about delayed evaluation with expression templates and (for me at least) a strange bad_alloc.
If you try the code below, you'll notice runtime error bad_alloc due to the very last addition b+c. So thats the point where the delayed evaluation is done. Furthermore the code below compiles and runs fine if you remove the references of the members of "Expression" (left,right). But i need references there, due to performance, etc. . However i also dont see, why i cant use references there.
I've already spent a lot of time with it. Please let me know if somebody can help me.
Best Regards.
#include <iostream>
#include <vector>
template<typename value_t, typename left_t, typename right_t, typename op_t>
class Expression
{
public:
typedef value_t value_type;
explicit Expression(const left_t &left,
const right_t &right,
const op_t &op) :
left(left),
right(right),
op(op)
{
}
value_t operator [](const size_t &i) const
{
return op(left[i],right[i]);
}
size_t size() const { return left.size();}
private:
const left_t &left;
const right_t &right;
//const left_t left;
//const right_t right;
const op_t &op;
};
template<class left_t,
class right_t,
class value_t = typename left_t::value_type,
class op_t = std::plus<value_t>>
const Expression<value_t, left_t, right_t, op_t> operator +(const left_t &left,
const right_t &right)
{
return Expression<value_t,left_t,right_t,op_t>(left, right, op_t());
}
template<typename value_t, typename data_t = std::vector<value_t>>
class Vector : public data_t
{
public:
typedef value_t value_type;
using data_t::size;
Vector(const std::initializer_list<value_t> &list) :
data_t(list)
{
}
Vector(const size_t &n) :
data_t(n)
{
}
Vector(const Vector &v) :
data_t(v)
{
}
template<typename left_t, typename right_t, typename op_t>
Vector(const Expression<value_t,left_t,right_t,op_t> &v) :
data_t(v.size())
{
operator =(v);
}
template<typename vec_t>
Vector(const vec_t &v) :
data_t(v.size())
{
operator =(v);
}
template<typename vec_t>
Vector &operator =(const vec_t &v)
{
for(size_t i = 0; i < data_t::size(); ++i)
data_t::operator [](i) = v[i];
return (*this);
}
friend std::ostream &operator <<(std::ostream &os, const Vector &v)
{
if(v.size())
os << v[0];
for(size_t i = 1; i < v.size(); ++i)
os << " " << v[i];
return os;
}
};
int main()
{
Vector<double> a{0,1,2};
auto b = a+a+a;
auto c = a;
std::cout << a+a+a+a << std::endl;
std::cout << b+c << std::endl; // gives bad_alloc
return 0;
}
"But i need references there, due to performance, etc."
Prove it.
In expression templates, all¹ the information should be compile-time.
You can see my example here for a simple expression template:
// we have lazy placeholder types:
template <int N> struct placeholder {};
placeholder<1> _1;
placeholder<2> _2;
placeholder<3> _3;
// note that every type here is stateless, and acts just like a more
// complicated placeholder.
// We can have expressions, like binary addition:
template <typename L, typename R> struct addition { };
template <typename L, typename R> struct multiplication { };
// here is the "factory" for our expression template:
template <typename L, typename R> addition<L,R> operator+(L const&, R const&) { return {}; }
template <typename L, typename R> multiplication<L,R> operator*(L const&, R const&) { return {}; }
///////////////////////////////////////////////
// To evaluate/interpret the expressions, we have to define "evaluation" for each type of placeholder:
template <typename Ctx, int N>
auto eval(Ctx& ctx, placeholder<N>) { return ctx.arg(N); }
template <typename Ctx, typename L, typename R>
auto eval(Ctx& ctx, addition<L, R>) { return eval(ctx, L{}) + eval(ctx, R{}); }
template <typename Ctx, typename L, typename R>
auto eval(Ctx& ctx, multiplication<L, R>) { return eval(ctx, L{}) * eval(ctx, R{}); }
///////////////////////////////////////////////
// A simple real-life context would contain the arguments:
#include <vector>
struct Context {
std::vector<double> _args;
// define the operation to get an argument from this context:
double arg(int i) const { return _args.at(i-1); }
};
#include <iostream>
int main() {
auto foo = _1 + _2 + _3;
Context ctx { { 3, 10, -4 } };
std::cout << "foo: " << eval(ctx, foo) << "\n";
std::cout << "_1 + _2 * _3: " << eval(ctx, _1 + _2 * _3) << "\n";
}
So what you need is a literal type that holds a reference to the associated value, and defer all other evaluation to evaluation time.
I might prefer to add the size() operation as a free function, so that you don't have to encumber all the expression types with it (Separation Of Concerns).
¹ nearly, nl. except when encoding literals
Proof Of Concept
Using the strategy outlined:
Live On Coliru
#include <iostream>
#include <tuple>
namespace ETL {
template <typename T>
struct Literal {
T value;
T get() const { return value; }
};
/*
*template <typename T>
* static inline std::ostream& operator<<(std::ostream& os, ETL::Literal<T> const& lit) {
* return os << __PRETTY_FUNCTION__ << "\n actual: lit.value = " << lit.value;
* }
*/
template <class L, class R, class Op>
struct BinaryExpr : std::tuple<L, R, Op> { // tuple optimizes for empty element types
BinaryExpr(L l, R r, Op op)
: std::tuple<L, R, Op> { l, r, op }
{}
L const& get_lhs() const { return std::get<0>(*this); }
R const& get_rhs() const { return std::get<1>(*this); }
Op const& get_op() const { return std::get<2>(*this); }
};
template <class L, class R, class Op> auto cured(BinaryExpr<L,R,Op> _) { return _; }
template <class T> auto cured(Literal<T> l) { return std::move(l); }
template <class T> Literal<T> cured(T&& v) { return {std::forward<T>(v)}; }
template <class Op, class L, class R>
BinaryExpr<L, R, Op> make_binexpr(L&& l, R&& r) { return { std::forward<L>(l), std::forward<R>(r), Op{} }; }
template <class L, class R> auto operator +(L&& l, R&& r)
{ return make_binexpr<std::plus<>>(cured(std::forward<L>(l)), cured(std::forward<R>(r))); }
template <class L, class R> auto operator -(L&& l, R&& r)
{ return make_binexpr<std::minus<>>(cured(std::forward<L>(l)), cured(std::forward<R>(r))); }
template <class L, class R> auto operator *(L&& l, R&& r)
{ return make_binexpr<std::multiplies<>>(cured(std::forward<L>(l)), cured(std::forward<R>(r))); }
template <class L, class R> auto operator /(L&& l, R&& r)
{ return make_binexpr<std::divides<>>(cured(std::forward<L>(l)), cured(std::forward<R>(r))); }
template <class L, class R> auto operator %(L&& l, R&& r)
{ return make_binexpr<std::modulus<>>(cured(std::forward<L>(l)), std::forward<R>(r)); }
template <typename T> auto val(T const& v)
{ return cured(v); }
namespace impl {
template <class T>
static constexpr auto is_indexable(T const&) -> decltype(std::declval<T const&>()[0], std::true_type{}) { return {}; }
static constexpr auto is_indexable(...) -> decltype(std::false_type{}) { return {}; }
struct {
template <class T> size_t operator()(T const& v) const { return (*this)(v, is_indexable(v)); }
template <class T> size_t operator()(T const& v, std::true_type) const { return v.size(); }
template <class T> size_t operator()(T const&, std::false_type) const { return 0; }
template <class T> size_t operator()(Literal<T> const& l) const { return (*this)(l.value); }
template <class L, class R, class Op>
size_t operator()(BinaryExpr<L,R,Op> const& be) const { return (*this)(be.get_lhs()); }
} size;
struct {
template <class T>
auto operator()(size_t i, T const& v) const { return (*this)(i, v, is_indexable(v)); }
template <class T>
auto operator()(size_t i, T const& v, std::true_type) const { return v[i]; }
template <class T>
auto operator()(size_t, T const& v, std::false_type) const { return v; }
template <class T> auto operator()(size_t i, Literal<T> const& l) const { return (*this)(i, l.value); }
template <class L, class R, class Op>
auto operator()(size_t i, BinaryExpr<L,R,Op> const& be) const {
return be.get_op()((*this)(i, be.get_lhs()), (*this)(i, be.get_rhs()));
}
} eval_at;
}
template <typename T> size_t size(T const& v) { return impl::size(v); }
template <typename T> auto eval_at(size_t i, T const& v) { return impl::eval_at(i, v); }
}
#include <vector>
template <class value_t>
struct Vector : std::vector<value_t> {
using data_t = std::vector<value_t>;
typedef value_t value_type;
using data_t::data_t;
template <typename Expr>
Vector(Expr const& expr) { *this = expr; }
template <typename Expr>
Vector& operator=(Expr const& expr) {
this->resize(size(expr));
for (size_t i = 0; i < this->size(); ++i)
this->at(i) = eval_at(i, expr);
return *this;
}
friend std::ostream &operator<<(std::ostream &os, const Vector &v) {
for (auto& el : v) os << " " << el;
return os;
}
};
int main() {
Vector<double> a { 1, 2, 3 };
using ETL::operator+;
using ETL::operator*;
//std::cout << typeid(a + a * 4 / 2).name() << "\n";
#define DD(x) std::cout << typeid(x).name() << " size: " << ETL::size(x) << " result:" << (x) << "\n"
DD(a * -100.0);
auto b = a + a + a;
auto c = a;
std::cout << size(b) << "\n";
std::cout << (a + a + a + a) << "\n";
std::cout << a * 4.0 << "\n";
std::cout << b + c << "\n";
std::cout << (a + a + a + a) - 4 * a << "\n";
}
Prints
ETL::BinaryExpr<ETL::Literal<Vector<double>&>, ETL::Literal<double>, std::multiplies<void> > size: 3 result: -100 -200 -300
3
4 8 12
4 8 12
4 8 12
0 0 0
I wanted to write my own Vector class template and also wanted to add some specializations, for example a 3D vector type where the components can be accessed through x/y/z.
The template and the specializations work fine so far, but the issue is, that the specialized templates require a lot of copy/pasting from the base template to work. I would like to reduce that.
This is what it looks like right now:
template<class T, unsigned int dim>
class Vector;
template<class T, unsigned int dim>
Vector<T, dim> add(Vector<T, dim> const& lhs, Vector<T, dim> const& rhs)
{
Vector<T, dim> tmp;
for (unsigned int i = 0; i < dim; ++i)
{
tmp[i] = lhs[i] + rhs[i];
}
return tmp;
}
template<class T, unsigned int dim, class S>
Vector<T, dim> add(Vector<T, dim> const& lhs, S const& rhs)
{
Vector<T, dim> tmp;
for (unsigned int i = 0; i < dim; ++i)
{
tmp[i] = lhs[i] + rhs;
}
return tmp;
}
template<class T, unsigned int dim>
Vector<T, dim> operator+(Vector<T, dim> const& lhs, Vector<T, dim> const& rhs)
{
return vectors::add(lhs, rhs);
}
template<class T, unsigned int dim, class S>
Vector<T, dim> operator+(Vector<T, dim> const& lhs, S const& rhs)
{
return vectors::add(lhs, rhs);
}
template<class T, unsigned int dim>
class Vector
{
//...
protected:
T values[dim] __attribute((aligned(16)));
public:
template<class R, unsigned int fdim>
friend Vector<R, fdim> operator+(Vector<R, fdim> const& lhs, Vector<R, fdim> const& rhs);
template<class R, unsigned int fdim, class S>
friend Vector<R, fdim> operator+(Vector<R, fdim> const& lhs, S const& rhs);
template<class R, unsigned int fdim, class S>
friend Vector<R, fdim> operator+(S const& lhs, Vector<R, fdim> const& rhs);
//...
//constructors, etc.
};
template<class T>
class Vector<T, 3>
{
//...
protected:
T values[3] __attribute((aligned(16)));
public:
T& x = values[0];
T& y = values[1];
T& z = values[2];
//lots of copy-pasta :(
template<class R, unsigned int fdim>
friend Vector<R, fdim> operator+(Vector<R, fdim> const& lhs, Vector<R, fdim> const& rhs);
template<class R, unsigned int fdim, class S>
friend Vector<R, fdim> operator+(Vector<R, fdim> const& lhs, S const& rhs);
template<class R, unsigned int fdim, class S>
friend Vector<R, fdim> operator+(S const& lhs, Vector<R, fdim> const& rhs);
//...
//constructors, etc.
};
Now I thought the easy solution would be to simply define Vector3D as a sub-class of the Vector template, like so:
template<class T>
class Vector3D: public Vector<T, 3>
{
//...
public:
T& x = values[0];
T& y = values[1];
T& z = values[2];
//no copy-pasta :)
//...
//constructors, etc.
};
That doesn't work at all, due to ambiguity:
ambiguous overload for ‘operator+’ (operand types are ‘const vec3f {aka const math::vectors::Vector3D<float>}’ and ‘math::vectors::vec3f {aka math::vectors::Vector3D<float>}’)
../main.cpp:84:16: note: candidates are:
In file included from ../main.cpp:10:0:
../include/vector.hpp:720:16: note: math::vectors::Vector<T, dim> math::vectors::operator+(const math::vectors::Vector<T, dim>&, const math::vectors::Vector<T, dim>&) [with T = float; unsigned int dim = 3u]
Vector<T, dim> operator+(Vector<T, dim> const& lhs, Vector<T, dim> const& rhs)
^
../include/vector.hpp:726:16: note: math::vectors::Vector<T, dim> math::vectors::operator+(const math::vectors::Vector<T, dim>&, const S&) [with T = float; unsigned int dim = 3u; S = math::vectors::Vector3D<float>]
Vector<T, dim> operator+(Vector<T, dim> const& lhs, S const& rhs)
^
../include/vector.hpp:732:16: note: math::vectors::Vector<T, dim> math::vectors::operator+(const S&, const math::vectors::Vector<T, dim>&) [with T = float; unsigned int dim = 3u; S = math::vectors::Vector3D<float>]
Vector<T, dim> operator+(S const& lhs, Vector<T, dim> const& rhs)
So it seems like the template substitution fails, because S can also be substituted with the new Vector3D class as well, while it's supposed to handle only scalars.
So I tried to get rid of that issue by writing a small wrapper class for scalars like so:
template<class T>
class ScalarType
{
public:
T value;
ScalarType() :
value(0)
{
}
ScalarType(T const& _v) :
value(_v)
{
}
ScalarType(ScalarType<T> const& rhs) :
value(rhs.value)
{
}
operator T&()
{
return value;
}
operator T() const
{
return value;
}
};
And replace all instances of S const& (l|r)hs with ScalarType<S> const& (l|r)hs.
That got the operators with Vectors on both sides to work again, but the operators that are supposed to handle Vector-Scalar operations fail still.
This time it's due to the fact, that the scalar value has to be explicitly of type ScalarType, since implicit conversions to that don't work with template substitution.
So, is there any way of getting this to work at all or do I have to stick with the copy-paste code?
Done here with partial template specialisation and CRTP.
maybe_has_z<Container, N> is a class which translates Container::z() into Container::operator[](2), but only if Container::size() >= 3
#include <array>
#include <iostream>
#include <algorithm>
//
// some boilerplate - note the different indecies
//
// define some concepts
template<class Container, std::size_t N, typename= void>
struct maybe_has_x{};
template<class Container, std::size_t N, typename = void>
struct maybe_has_y{};
template<class Container, std::size_t N, typename = void>
struct maybe_has_z{};
// specialise the concepts into (sometimes) concrete accessors
template<class Container, std::size_t N>
struct maybe_has_x<Container, N, std::enable_if_t<(N > 0)>>
{
auto& x() const { return static_cast<const Container&>(*this)[0]; }
auto& x() { return static_cast<Container&>(*this)[0]; }
};
template<class Container, std::size_t N>
struct maybe_has_y<Container, N, std::enable_if_t<(N > 1)>>
{
auto& y() const { return static_cast<const Container&>(*this)[1]; }
auto& y() { return static_cast<Container&>(*this)[1]; }
};
template<class Container, std::size_t N>
struct maybe_has_z<Container, N, std::enable_if_t<(N > 2)>>
{
auto& z() const { return static_cast<const Container&>(*this)[2]; }
auto& z() { return static_cast<Container&>(*this)[2]; }
};
// define our vector type
template<class T, std::size_t N>
struct Vector
: std::array<T, N>
, maybe_has_x<Vector<T, N>, N> // include the maybe_ concepts
, maybe_has_y<Vector<T, N>, N>
, maybe_has_z<Vector<T, N>, N>
{
private:
using inherited = std::array<T, N>;
public:
Vector() : inherited {} {};
Vector(std::initializer_list<T> il)
: inherited { }
{
std::copy_n(il.begin(), std::min(il.size(), this->size()), std::begin(*this));
}
Vector(const inherited& rhs) : inherited(rhs) {}
public:
using value_type = typename inherited::value_type;
// offer arithmetic unary functions in class (example +=)
// note that this allows us to add integers to a vector of doubles
template<class Other, std::enable_if_t<std::is_convertible<value_type, Other>::value> * = nullptr>
Vector& operator+=(const Vector<Other, N>&rhs) {
auto lfirst = std::begin(*this);
auto rfirst = std::begin(rhs);
auto lend = std::end(*this);
while (lfirst != lend) {
*lfirst += *rfirst;
++lfirst;
++rfirst;
}
return *this;
}
};
// offer binary arithmetic as free functions
template<class T, std::size_t N, class Other>
Vector<T, N> operator+(Vector<T, N> lhs, const Vector<Other, N>& rhs) {
lhs += rhs;
return lhs;
}
// offer some streaming capability
template<class T, std::size_t N>
std::ostream& operator<<(std::ostream& os, const Vector<T, N>& rhs) {
auto sep = "";
os << '[';
for (auto& x : rhs) {
os << sep << x;
sep = ", ";
}
return os << ']';
}
// test
int main()
{
auto a = Vector<double, 3> { 2.1, 1.2, 3.3 };
auto b = a + a + Vector<int, 3> { 1, 1, 1 };
std::cout << a << std::endl;
std::cout << b << std::endl;
std::cout << a.x() << ", " << a.y() << ", " << a.z() << std::endl;
auto c = Vector<double, 2> { 4.4, 5.5 };
std::cout << c << std::endl;
std::cout << c.x() << std::endl;
std::cout << c.y() << std::endl;
// won't compile
// std::cout << c.z() << std::endl;
}
expected output:
[2.1, 1.2, 3.3]
[5.2, 3.4, 7.6]
2.1, 1.2, 3.3
[4.4, 5.5]
4.4
5.5
I've got the following code, and I'm struggling to understand why packing the structure causes a warning (and then subsequently segfault when run.) Firstly the code:
#include <iostream>
using namespace std;
template <typename T, std::size_t D, std::size_t S>
struct __attribute__((packed)) C {
union __attribute__((packed)) ValueType {
char fill[S];
T value;
};
ValueType data_;
const T& get() const {
return data_.value;
}
friend
std::ostream& operator<<(std::ostream& str, const C& f) {
return str << f.data_.value;
}
};
template <typename T, std::size_t D>
struct __attribute__((packed)) C<T, D, D> {
T value;
const T& get() const {
return value;
}
friend
std::ostream& operator<<(std::ostream& str, const C& f) {
return str << f.value;
}
};
template <typename T, std::size_t S = 0>
struct __attribute__((packed)) D {
C<T, sizeof(T), S> c;
const T& get() const {
return c.get();
}
friend
std::ostream& operator<<(std::ostream& str, const D& f) {
return str << f.get();
}
};
template <typename T>
struct __attribute__((packed)) D<T, 0> {
T v;
const T& get() const {
return v;
}
friend
std::ostream& operator<<(std::ostream& str, const D& f) {
return str << f.get();
}
};
int main(void) {
D<int64_t> d1;
cout << d1 << endl;
D<int64_t, 8> d2;
cout << d2 << endl;
}
With the packing (it's GCC specific) I get the following (gcc 5.2.1/5.3.1, c++14 -Wall):
packed.cpp: In instantiation of ‘const T& D<T, 0ul>::get() const [with T = long int]’:
packed.cpp:58:16: required from ‘std::ostream& operator<<(std::ostream&, const D<long int>&)’
packed.cpp:64:13: required from here
packed.cpp:53:12: warning: returning reference to temporary [-Wreturn-local-addr]
return v;
^
packed.cpp: In instantiation of ‘const T& C<T, D, D>::get() const [with T = long int; long unsigned int D = 8ul]’:
packed.cpp:40:18: required from ‘const T& D<T, S>::get() const [with T = long int; long unsigned int S = 8ul]’
packed.cpp:45:16: required from ‘std::ostream& operator<<(std::ostream&, const D<long int, 8ul>&)’
packed.cpp:66:13: required from here
packed.cpp:27:12: warning: returning reference to temporary [-Wreturn-local-addr]
return value;
If the structures are not packed - the code compiles normally and then runs fine (prints rubbish value - but that's expected.)
So my question is - why does packing cause a return of a temporary?
What you're looking for is pragma pack. It has the same effect as what you want, but is more common and cross-compiler.
NOTE: Adding modified version of the code with pragma pack
#include <iostream>
using namespace std;
#pragma pack(1)
template <typename T, std::size_t D, std::size_t S>
struct C {
union ValueType {
char fill[S];
T value;
};
ValueType data_;
const T& get() const {
return data_.value;
}
};
#pragma pack()
template <typename T, std::size_t D>
struct C<T, D, D> {
T value;
const T& get() const {
return value;
}
};
template <typename T, std::size_t D>
struct C<T, D, 0> {
T value;
const T& get() const {
return value;
}
};
template <typename T, std::size_t S = 0>
struct D : C<T, sizeof(T), S> {
friend
std::ostream& operator<<(std::ostream& str, const D& f) {
return str << f.get();
}
};
#pragma pack(1)
struct E {
D<int16_t> a;
D<int64_t, 9> b;
};
#pragma pack()
int main(void) {
D<int16_t> d1;
cout << d1.get() << ' ' << sizeof(d1) << endl;
D<int64_t, 9> d2;
cout << d2.get() << ' ' << sizeof(d2) << endl;
cout << sizeof(E) << endl;
}
hello Nim I just find this problem may be a gcc compiler's BUG, you can see the reference at here enter link description here
I am using Boost-Operatators to construct a matrix class. (A toy project). However, I run into issues when I want to mix matrices of different element types.
Basically I have a template class Matrix<T>, where T is the element type of that matrix. I'm using Boost-Operators to define operators between instances of Matrix<T> (e.g. element-wise add), between Matrix<T> and T (e.g. scalar multiplication), and if possible also between Matrix<T> and Matrix<U> (e.g. real matrix plus complex matrix).
The boost operators support one, or two template arguments. One if you want operators between two objects of the same type, and two if you want mixed operators.
template<typename T>
class Matrix : boost::addable<Matrix<T>> // Add another matrix of same type.
boost::multiplyable2<Matrix<T>,T> // Scalar multiplication with a `T`.
However, I cannot give Matrix<U> as a second argument, because then my class would have two template arguments and the type would depend on which matrices I can operate with.
template<typename T, typename U>
class Matrix : boost::addable2<Matrix<T,U>,Matrix<U,?>> // Now I have two template arguments.
// That's certainly not what I want!
I also tried implementing my own version of boost::addable, but this didn't work either. The compiler complains about an uncomplete type.
template<class Derived>
class Addable {
template<class Other>
friend Derived operator+(Derived lhs, const Other &rhs) {
return lhs += rhs;
}
template<class Other>
friend Derived operator+(const Other &lhs, Derived rhs) {
return rhs += lhs;
}
};
Another approach was to define a cast constructor from Matrix<U> to Matrix<T>. However, now I have the issue, that those are two different types, and I don't get access to the private members. So, I either need to make more stuff public than I want to, or find a different way of doing this.
How would you implement such a thing?
The full Code
#include <cassert>
#include <utility>
#include <complex>
#include <vector>
#include <algorithm>
#include <iostream>
#include <boost/operators.hpp>
typedef double Real;
typedef std::complex<Real> Complex;
template<typename T>
class Matrix : boost::addable<Matrix<T>>
{
public:
Matrix() = default;
template<typename U>
Matrix(const Matrix<U> &other)
: m_(other.m()), n_(other.n()),
data_(other.data_.begin(), other.data_.end()) { }
Matrix(size_t m, size_t n) : m_(m), n_(n), data_(m*n) { }
Matrix(size_t m, size_t n, const T &initial)
: m_(m), n_(n), data_(m*n, initial) { }
size_t m() const { return m_; }
size_t n() const { return n_; }
size_t size() const {
assert(m_*n_ == data_.size());
return data_.size();
}
const T &operator()(size_t i, size_t j) const { return data_[i*m_ + j]; }
T &operator()(size_t i, size_t j) { return data_[i*m_ + j]; }
void fill(const T &value) {
std::fill(data_.begin(), data_.end(), value);
}
Matrix &operator+=(const Matrix &other) {
assert(dim_match(other));
for (int i = 0; i < size(); ++i) {
data_[i] += other.data_[i];
}
return *this;
}
friend std::ostream &operator<<(std::ostream &o, const Matrix &m) {
if (m.size() == 0) {
o << "()" << std::endl;
return o;
}
for (int i = 0; i < m.m(); ++i) {
o << "( ";
for (int j = 0; j < m.n() - 1; ++j) {
o << m(i,j) << ", ";
}
o << m(i, m.n() - 1) << " )" << std::endl;
}
return o;
}
private:
bool dim_match(const Matrix &other) {
return n_ == other.n_ && m_ == other.m_;
}
private:
int m_, n_;
typedef std::vector<T> Store;
Store data_;
};
int main() {
Matrix<Real> A(2,3, 1.);
Matrix<Complex> B(2,3, Complex(0,1));
auto C = Matrix<Complex>(A) + B;
std::cout << A << std::endl;
std::cout << B << std::endl;
std::cout << C << std::endl;
}
This is how I'd do it: use a friend template function (see Operator overloading: The Decision between Member and Non-member):
template<typename T>
class Matrix
{
public:
template<typename> friend class Matrix;
And then later
template <typename T1, typename T2>
Matrix<typename std::common_type<T1, T2>::type>
operator+(Matrix<T1> const& a, Matrix<T2> const& b)
{
Matrix<typename std::common_type<T1, T2>::type> result(a);
return (result += b);
}
Note the use of common_type to arrive at a sensible resulting type (you might want to introduce your own trait there to cater for your specific requirements)
See it Live On Coliru
I have the following C++ code:
#include <iostream>
template <class T>
void assign(T& t1, T& t2) {
std::cout << "First method" << std::endl;
t1 = t2;
}
template <class T>
void assign(T& t1, const T& t2) {
std::cout << "Second method" << std::endl;
t1 = t2;
}
class A {
public:
A(int a) : _a(a) {};
private:
int _a;
friend A operator+(const A& l, const A& r);
};
A operator+(const A& l, const A& r) {
return A(l._a + r._a);
}
int main() {
A a = 1;
const A b = 2;
assign(a, a);
assign(a, b);
assign(a, a+b);
}
The output is:
First method
Second method
Second method
The output stays the same even if I comment out the the first 2 assigns in the main function.
Can someone please explain to me why the operator+ returns a const A?
The output is the same in both Linux Debian 64bit and Windows 7 64 bit.
It doesn't return a const A at all. It returns a temporary A, which may only bind to a const reference.