msvc and gcc different behavior - c++

#include <iostream>
template<typename T>
const T unit{ 1 };
template<typename T>
struct data
{
typename T::value_type val;
};
struct decimal_def
{
using value_type = int;
static const value_type unit_val() {
return 10;
}
};
template<typename T>
const data<T> unit<data<T>> { T::unit_val() };
auto decimal_data_unit = unit<data<decimal_def>>;
int main(){
std::cout << decimal_data_unit.val << std::endl;
return 0;
}
exoect: 10
gcc: 10
msvc: 0
I think the reason of the difference between gcc and msvc is that compilers optimize unit_val() differently.

Dynamic initialization of templated non-local variables is always unordered, even relative to other variables defined in the same translation unit (in part because templated variables might also be defined in other translation units). In this simple case, you can avoid the issue by using constant initialization instead: add constexpr to decimal_def::unit_val to enable that, and to the variable template definitions for clarity.

Related

constexpr differences between GCC and clang

The following compiles in GCC 9 but not in clang 10 and I'm wondering which of the two compilers is standard conforming:
template<typename T>
struct A {
static const T s;
static const T v;
};
template<typename T>
constexpr const T A<T>::s = T(1);
template<typename T>
constexpr const T A<T>::v = A<T>::s;
int main(int, char**) {
constexpr auto a = A<double>::v;
return 0;
}
This is intended to be a minimal example of a bigger issue which is why the fields s and v are explicitly declared as const but are defined as constexpr, this is intentional.
Is GCC correct to compile that code or is clang correct to reject it?
Compilers are only required to treat static const variables of integral and enum types as constexpr if they are initialize with a constant expression. This made it possible to use them as array lengths before constexpr was added to the language.

Specialize if value of a variable is known/unknown at compile time

How to specialize a template function for the case that the value of one of its argument is known/unknown during compile time (before actually compile and run the program)?
I can't figure out how yet.
idea 1:
#include <type_traits>
#include <iostream>
int main(void){
int a; //value of a is not known at compile time
bool b = (a == a); //value of b is known at compile time.
std::is_assignable< constexpr bool, bool >::value
}
//g++ magic.cpp -std=c++14
//error: wrong number of template arguments (1, should be 2)
// std::is_assignable< constexpr bool, bool >::value
Idea 2:
#include <type_traits>
#include <iostream>
int main(void){
const int a=1;
int b = (a == a);
std::cout << __builtin_constant_p (a) << std::endl;
std::cout << __builtin_constant_p (b) << std::endl;
}
//prints 0 and 0.
Well, I guess you mean the type of the argument, right? Values don't matter for partial template specializations...
Then: This can't be done.
Parameter types for templates must be known at compile time. How else should the compiler produce the correct code?
Also for partial template specializations, the types must be known at compile time for the same reason.
I'm a bit late to the party, and my partial answer may not be very satisfying, but here goes:
Situations where we can't tell from inspection whether a value is known at compile time:
Non-template input values to constexpr functions
Members of types provided by template arguments
I don't know what to do about problem 1, but for problem 2 we can use SFINAE: looking for a specific member with known name (in the below example it's X), and trying to send it as a template argument, like so:
// like std::void_t, but for value inputs:
template <auto>
using v_to_void_t = void;
//////////////////////////////////////////////////////////////
template <class, class = void>
struct has_constexpr_X
: std::false_type {};
template <class T>
struct has_constexpr_X <T, v_to_void_t<T().X>>
: std::true_type {};
template <class T>
constexpr bool has_constexpr_X_v
= has_constexpr_X<T>::value;
Example uses:
struct HasStaticConstexprX {
static constexpr int X = 2;
};
struct HasStaticConstX {
static const int X; // implied constexpr
};
const int HasStaticConstX::X = 3;
struct HasStaticX {
static int X;
};
int HasStaticX::X = 4;
struct HasConstX {
const int X;
};
int main () {
static_assert(has_constexpr_X_v<HasStaticConstexprX>);
static_assert(has_constexpr_X_v<HasStaticConstX>);
static_assert(! has_constexpr_X_v<HasStaticX>);
static_assert(! has_constexpr_X_v<HasConstX>);
}
Demos:
c++17
c++14

Type-dependent constant in template function

I want a static array in a templated function whose length depends on the type with which the function is specialized. My first attempt was:
Header:
template<typename T>
struct Length {
const static size_t len;
};
template<typename T>
void func(){
static T vars[Length<T>::len]; // len not const. according to compiler!
// ...
}
Source file:
template<> const size_t Length<double>::len = 2;
template<> const size_t Length<float>::len = 1;
// ...
However, g++ does not compile this and complains
error: storage size of ‘vars’ isn’t constant
So, what exactly is the problem here? I know that the size of a fixed-length array needs to be a constant and known on compile time, but that seems to be the case here.
When I write
const size_t len = 2;
void func(){
static double vars[len];
}
it compiles without problems.
Question:
What is wrong with the code and which alternatives are there for achieving the desired behavior? I would not like to allocate the memory during runtime...
For a const variable to be considered a compile-time constant (formally, a constant expression), its value must be available at point of use. Which means the specialised definitions would have to go to the header file.
If done as just specialisations of the member, as you did, I believe that would give you a multiple-definition error. You should be fine with specialising the entire class template, and keeping the static member definition inline:
template<typename T>
struct Length;
template <>
struct Length<double>
{
static const size_t len = 2;
};
As a side note, your program was originally invalid: an explicit specialisation must be declared before use. Which means you'd have to at least declare the specialisation of len in the header (or everywhere you intended to use it).
The following code compiles fine for me with g++ 4.6.3 and produces the output
2
1
array.cpp:
#include <cstddef>
#include <iostream>
template<typename T>
struct Length {
const static size_t len;
};
template<typename T>
void func(){
static T vars[Length<T>::len];
std::cout << (sizeof(vars) / sizeof(*vars)) << std::endl;
}
template<> const size_t Length<double>::len = 2;
template<> const size_t Length<float>::len = 1;
int main(){
func<double>();
func<float>();
}
$ make array
g++ array.cpp -o array
$ ./array
2
1

Static constants of templated class

I have been playing around with template classes and constants in gcc 4.8.2 and came across an interesting linker error:
#include <iostream>
using namespace std;
template <class T, int m>
class A {
public:
static const T param = m;
T value;
A(const T &value, const T &dummy = T()) : value(value) {}
};
// Works with this
// template <class T, int m>
// const T A<T, m>::param = m;
template <class T, int m>
A<T, m> operator +(const A<T, m> &a, const A<T, m> &b) {
if (a.param != b.param) exit(1);
// Works if I replace a.param with 0
return A<T, m>(a.value + b.value, a.param);
}
int main() {
A<int, 2> v = A<int, 2>(1) + A<int, 2>(2);
cout << v.value << endl;
return 0;
}
Compiling the code in the current state gives a linker error, telling me that A::param is not defined.
Trying this code in Visual Studio 2008 as well, it compiles and links without any problems. On gcc, it compiles when I either use the external declaration of the param constant, or if I replace a.param with 0 or nothing on the line indicated.
What I do not understand is why the line containing the if statement can use a.param without compilation error, while I cannot pass a.param to the constructor without declaring it externally.
So my question is: When do I need to declare param externally, and what is the difference between the access in the if statement and the constructor call? Which of the two compilers I tested does the "right" thing here, and which one has extended the standard?
Playing around a little more, I realized it also works when I specify the -O2 flag to g++.
template <class T, int m>
class A {
public:
static const T param = m;
T value;
A(const T &value, const T &dummy = T()) : value(value) {}
};
// Works with this
// template <class T, int m>
// const T A<T, m>::param = m;
//gotta define it here!
template <class T, int m>
const T A<T, m>::param;
template <class T, int m>
A<T, m> operator +(const A<T, m> &a, const A<T, m> &b) {
if (a.param != b.param) exit(1);
// Works if I replace a.param with 0
return A<T, m>(a.value + b.value, a.param);
}
int main() {
A<int, 2> v = A<int, 2>(1) + A<int, 2>(2);
std::cout << v.value << std::endl;
return 0;
}
from here
If a static data member is of const integral or const enumeration type, you may specify a constant initializer in the static data member's declaration. This constant initializer must be an integral constant expression. Note that the constant initializer is not a definition. You still need to define the static member in an enclosing namespace.
In the cases that is working, the compilers are being non-compliant. Microsoft is non-compliant cuz Microsoft is almost never compliant, the g++ with -O2 is kinda funny, but in either case you've got no definition!
Edit:
As a rule of thumb, I always define my static constant member variables outside of the class because then I never have to remember it. In this case, class A would be invalid if T was anything other than an integral type, such as a float. So rather than do any patterns around the exception, I tend to just stick with my patterns around the rule.

compile time loops

I would like to know if it is possible to have sort of compile time loops.
For example, I have the following templated class:
template<class C, int T=10, int B=10>
class CountSketch
{
public:
CountSketch()
{
hashfuncs[0] = &CountSketch<C>::hash<0>;
hashfuncs[1] = &CountSketch<C>::hash<1>;
// ... for all i until i==T which is known at compile time
};
private:
template<int offset>
size_t hash(C &c)
{
return (reinterpret_cast<int>(&c)+offset)%B;
}
size_t (CountSketch::*hashfuncs[T])(C &c);
};
I would thus like to know if I can do a loop to initialize the T hash functions using a loop. The bounds of the loops are known at compile time, so, in principle, I don't see any reason why it couldn't be done (especially since it works if I unroll the loop manually).
Of course, in this specific example, I could just have made a single hash function with 2 parameters (although it would be less efficient I guess). I am thus not interested in solving this specific problem, but rather knowing if "compile time loops" existed for similar cases.
Thanks!
Nope, it's not directly possible. Template metaprogramming is a pure functional language. Every value or type defined through it are immutable. A loop inherently requires mutable variables (Repeatedly test some condition until X happens, then exit the loop).
Instead, you would typically rely on recursion. (Instantiate this template with a different template parameter each time, until you reach some terminating condition).
However, that can solve all the same problems as a loop could.
Edit: Here's a quick example, computing the factorial of N using recursion at compile-time:
template <int N>
struct fac {
enum { value = N * fac<N-1>::value };
};
template <>
struct fac<0> {
enum { value = 1 };
};
int main() {
assert(fac<4>::value == 24);
}
Template metaprogramming in C++ is a Turing-complete language, so as long as you don't run into various internal compiler limits, you can solve basically any problem with it.
However, for practical purposes, it may be worth investigating libraries like Boost.MPL, which contains a large number of data structures and algorithms which simplify a lot of metaprogramming tasks.
Yes. Possible using compile time recursion.
I was trying with your code but since it was not compilable here is a modified and compiling exmaple:
template<class C, int T=10>
class CountSketch
{
template<int N>
void Init ()
{
Init<N-1>();
hashfuncs[N] = &CountSketch<C>::template hash<N>;
cout<<"Initializing "<<N<<"th element\n";
}
public:
CountSketch()
{
Init<T>();
}
private:
template<int offset>
size_t hash(C &c)
{
return 0;
}
size_t (CountSketch::*hashfuncs[T])(C &c);
};
template<>
template<>
void CountSketch<int,10>::Init<0> ()
{
hashfuncs[0] = &CountSketch<int,10>::hash<0>;
cout<<"Initializing "<<0<<"th element\n";
}
Demo. The only constraint of this solution is that you have to provide the final specialized version as, CountSketch<int,10>::Init<0> for whatever type and size.
You need a combination of boost::mpl::for_each and boost::mpl::range_c.
Note: This will result in run-time code and this is what you actually need. Because there is no way to know the result of operator& at compile time. At least none that I'm aware of.
The actual difficulty with this is to build a struct that is templated on an int parameter (mpl::int_ in our case) and that does the assignment when operator() is called and we also need a functor to actually capture the this pointer.
This is somewhat more complicated than I anticipated but it's fun.
#include <boost/mpl/range_c.hpp>
#include <boost/mpl/vector.hpp>
#include <boost/mpl/for_each.hpp>
#include <boost/mpl/transform.hpp>
#include <boost/mpl/copy.hpp>
// aforementioned struct
template<class C, class I>
struct assign_hash;
// this actually evaluates the functor and captures the this pointer
// T is the argument for the functor U
template<typename T>
struct my_apply {
T* t;
template<typename U>
void operator()(U u) {
u(t);
}
};
template<class C, int T=10, int B=10>
class CountSketch
{
public:
CountSketch()
{
using namespace boost::mpl;
// we need to do this because range_c is not an ExtensibleSequence
typedef typename copy< range_c<int, 0, T>,
back_inserter< vector<> > >::type r;
// fiddle together a vector of the correct types
typedef typename transform<r, typename lambda< assign_hash<C, _1 > >::type >
::type assignees;
// now we need to unfold the type list into a run-time construct
// capture this
my_apply< CountSketch<C, T, B> > apply = { this };
// this is a compile-time loop which actually does something at run-time
for_each<assignees>(apply);
};
// no way around
template<typename TT, typename I>
friend struct assign_hash;
private:
template<int offset>
size_t hash(C& c)
{
return c;
// return (reinterpret_cast<int>(&c)+offset)%B;
}
size_t (CountSketch::*hashfuncs[T])(C &c);
};
// mpl uses int_ so we don't use a non-type template parameter
// but get a compile time value through the value member
template<class C, class I>
struct assign_hash {
template<typename T>
void operator()(T* t) {
t->hashfuncs[I::value] = &CountSketch<C>::template hash<I::value>;
}
};
int main()
{
CountSketch<int> a;
}
with C++20 and consteval compile time loops became possible without doing template hell unless the value can have multiple types:
consteval int func() {
int out = 0;
for(int i = 10; i--;) out += i;
return out;
}
int main() {
std::cout << func(); // outputs 45
}
There are compilers that will see the loop and unroll it. But it's not part of the language specification that it must be done (and, in fact, the language specification throws all sorts of barriers in the way of doing it), and there's no guarantee that it will be done, in a particular case, even on a compiler that "knows how".
There are a few languages that explicitly do this, but they are highly specialized.
(BTW, there's no guarantee that the "unrolled" version of your initializations would be done "at compile time" in a reasonably efficient fashion. But most compilers will, when not compiling to a debug target.)
Here is, I think, a better version of the solution given above.
You can see that we use the compile-time recursive on the function params.
This enables putting all the logic inside your class, and the base case of Init(int_<0>) is very clear - just do nothing :)
Just so you won't fear performance penalty, know that the optimizer will throw away these unused parameters.
As a matter of fact, all these function calls will be inlined anyway. that's the whole point here.
#include <string.h>
#include <stdio.h>
#include <algorithm>
#include <iostream>
using namespace std;
template <class C, int N = 10, int B = 10>
class CountSketch {
public:
CountSketch() {
memset(&_hashFunctions, sizeof(_hashFunctions), 0); // for safety
Init(int_<N>());
}
size_t HashAll(C& c)
{
size_t v = 0;
for(const auto& h : _hashFunctions)
{
v += (this->*h)(c); // call through member pointer
}
return v;
}
private:
template<int offset>
size_t hash(C &c)
{
return (reinterpret_cast<size_t>(&c)+offset)%B;
}
size_t (CountSketch::*_hashFunctions[N])(C &c);
private: // implementation detail
// Notice: better approach.
// use parameters for compile-time recursive call.
// you can just override for the base case, as seen for N-1 below
template <int M>
struct int_ {};
template <int M>
void Init(int_<M>) {
Init(int_<M - 1>());
_hashFunctions[M - 1] = &CountSketch<C, N, B>::template hash<M>;
printf("Initializing %dth element\n", M - 1);
}
void Init(int_<0>) {}
};
int main() {
int c;
CountSketch<int, 10> cs;
int i;
cin >> i;
printf("HashAll: %d", cs.HashAll(c));
return 0;
}
Compiler Explorer