Consider the following code that implements a compile time counter.
#include <iostream>
template<int>
struct Flag { friend constexpr int flag(Flag); };
template<int N>
struct Writer
{
friend constexpr int flag(Flag<N>) { return 0; }
};
template<int N>
constexpr int reader(float, Flag<N>) { return N; }
template<int N, int = flag(Flag<N>{})>
constexpr int reader(int, Flag<N>, int value = reader(0, Flag<N + 1>{}))
{
return value;
}
template<int N = reader(0, Flag<0>{}), int = sizeof(Writer<N>) >
constexpr int next() { return N; }
int main() {
constexpr int a = next();
constexpr int b = next();
constexpr int c = next();
constexpr int d = next();
std::cout << a << b << c << d << '\n'; // 0123
}
For the second reader overload, if I put the default parameter inside the body of the function, like so:
template<int N, int = flag(Flag<N>{})>
constexpr int reader(int, Flag<N>)
{
return reader(0, Flag<N + 1>{});
}
Then the output will become:
0111
Why does this happen? What makes the second version not work anymore?
If it matters, I'm using Visual Studio 2015.2.
Without value being passed as a parameter, nothing stops the compiler from caching the call to reader(0, Flag<1>).
In both cases first next() call will work as expected since it will immediately result in SFINAEing to reader(float, Flag<0>).
The second next() will evaluate reader<0,0>(int, ...), which depends on reader<1>(float, ...) that can be cached if it does not depend on a value parameter.
Unfortunately (and ironically) the best source I found that confirms that constexpr calls can be cached is #MSalters comment to this question.
To check if your particular compiler caches/memoizes, consider calling
constexpr int next_c() { return next(); }
instead of next(). In my case (VS2017) the output turns into 0000.
next() is protected from caching by the fact that its default template arguments actually change with every instantiation, so it's a new separate function every time. next_c() is not a template at all, so it can be cached, and so is reader<1>(float, ...).
I do believe that this is not a bug and compiler can legitimately expect constexprs in compile-time context to be pure functions.
Instead it is this code that should be considered ill-formed - and it soon will be, as others noted.
The relevance of value is that it participates in overload resolution. Under SFINAE rules, template instantiation errors silently exclude candidates from overload resolution. But it does instantiate Flag<N+1>, which causes the overload resolution to become viable the next time (!). So in effect you're counting successful instantiations.
Why does your version behave differently? You still reference Flag<N+1>, but in the implementation of the function. This is important. With function templates, the declaration must be considered for SFINAE, but only the chosen overload is then instantiated. Your declaration is just template<int N, int = flag(Flag<N>{})> constexpr int reader(int, Flag<N>); and does not depend on Flag<N+1>.
As noted in the comments, don't count on this counter ;)
Related
I am working with a code example concerning stateful meta programming.
The original code example can be found from this link: http://b.atch.se/posts/non-constant-constant-expressions/#appendix-clang-workaround
In order to understand this technique, I try to modify the code example each time a bit. Here follows the latest version which still works:
constexpr int adl_flag(int);
template <class Tag> struct writer {
friend constexpr int adl_flag(int) {
return 0;
}
};
template <int = adl_flag(0)> constexpr bool is_flag_usable(int) {
return true;
}
constexpr bool is_flag_usable (...) {
return false;
}
template <
class Tag = int,
bool B = is_flag_usable(0),
int = sizeof (writer<Tag>) // replace Tag with int
>
constexpr int f() {
return B;
}
int main() {
constexpr int a = f();
constexpr int b = f();
static_assert(a != b, "fail");
return 0;
}
To me, it seems trivial to replace Tag with int. But this replacement in fact make the static_assert fail.
I guess this is because the compiler no longer implicitly instantiate f for the second call :
constexpr int b = f();
But the template type parameter Tag seems has nothing to do with this. Can somebody explain what's actually going on here, please.
The compiler I used is g++ 5.4.1. Compiling with -std=gnu++14.
sizeof(writer<int>) is non-dependent. That means that wrapper<int> is found in first phase lookup, when f is parsed (once).
sizeof(writer<Tag>), on the other hand, is dependent on Tag. Thus, its lookup is deferred until phase two, when f is instantiated (at each of its call sites).
Lets say that you have a function which generates some security token for your application, such as some hash salt, or maybe a symetric or asymetric key.
Now lets say that you have this function in your C++ as a constexpr and that you generate keys for your build based on some information (like, the build number, a timestamp, something else).
You being a diligent programmer make sure and call this in the appropriate ways to ensure it's only called at compile time, and thus the dead stripper removes the code from the final executable.
However, you can't ever be sure that someone else isn't going to call it in an unsafe way, or that maybe the compiler won't strip the function out, and then your security token algorithm will become public knowledge, making it more easy for would be attackers to guess future tokens.
Or, security aside, let's say the function takes a long time to execute and you want to make sure it never happens during runtime and causes a bad user experience for your end users.
Are there any ways to ensure that a constexpr function can never be called at runtime? Or alternately, throwing an assert or similar at runtime would be ok, but not as ideal obviously as a compile error would be.
I've heard that there is some way involving throwing an exception type that doesn't exist, so that if the constexpr function is not deadstripped out, you'll get a linker error, but have heard that this only works on some compilers.
Distantly related question: Force constexpr to be evaluated at compile time
In C++20 you can just replace constexpr by consteval to enforce a function to be always evaluated at compile time.
Example:
int rt_function(int v){ return v; }
constexpr int rt_ct_function(int v){ return v; }
consteval int ct_function(int v){ return v; }
int main(){
constexpr int ct_value = 1; // compile value
int rt_value = 2; // runtime value
int a = rt_function(ct_value);
int b = rt_ct_function(ct_value);
int c = ct_function(ct_value);
int d = rt_function(rt_value);
int e = rt_ct_function(rt_value);
int f = ct_function(rt_value); // ERROR: runtime value
constexpr int g = rt_function(ct_value); // ERROR: runtime function
constexpr int h = rt_ct_function(ct_value);
constexpr int i = ct_function(ct_value);
}
Pre C++20 workaround
You can enforce the use of it in a constant expression:
#include<utility>
template<typename T, T V>
constexpr auto ct() { return V; }
template<typename T>
constexpr auto func() {
return ct<decltype(std::declval<T>().value()), T{}.value()>();
}
template<typename T>
struct S {
constexpr S() {}
constexpr T value() { return T{}; }
};
template<typename T>
struct U {
U() {}
T value() { return T{}; }
};
int main() {
func<S<int>>();
// won't work
//func<U<int>>();
}
By using the result of the function as a template argument, you got an error if it can't be solved at compile-time.
A theoretical solution (as templates should be Turing complete) - don't use constexpr functions and fall back onto the good-old std=c++0x style of computing using exclusively struct template with values. For example, don't do
constexpr uintmax_t fact(uint n) {
return n>1 ? n*fact(n-1) : (n==1 ? 1 : 0);
}
but
template <uint N> struct fact {
uintmax_t value=N*fact<N-1>::value;
}
template <> struct fact<1>
uintmax_t value=1;
}
template <> struct fact<0>
uintmax_t value=0;
}
The struct approach is guaranteed to be evaluated exclusively at compile time.
The fact the guys at boost managed to do a compile time parser is a strong signal that, albeit tedious, this approach should be feasible - it's a one-off cost, maybe one can consider it an investment.
For example:
to power struct:
// ***Warning: note the unusual order of (power, base) for the parameters
// *** due to the default val for the base
template <unsigned long exponent, std::uintmax_t base=10>
struct pow_struct
{
private:
static constexpr uintmax_t at_half_pow=pow_struct<exponent / 2, base>::value;
public:
static constexpr uintmax_t value=
at_half_pow*at_half_pow*(exponent % 2 ? base : 1)
;
};
// not necessary, but will cut the recursion one step
template <std::uintmax_t base>
struct pow_struct<1, base>
{
static constexpr uintmax_t value=base;
};
template <std::uintmax_t base>
struct pow_struct<0,base>
{
static constexpr uintmax_t value=1;
};
The build token
template <uint vmajor, uint vminor, uint build>
struct build_token {
constexpr uintmax_t value=
vmajor*pow_struct<9>::value
+ vminor*pow_struct<6>::value
+ build_number
;
}
In the upcoming C++20 there will be consteval specifier.
consteval - specifies that a function is an immediate function, that is, every call to the function must produce a compile-time constant
Since now we have C++17, there is an easier solution:
template <auto V>
struct constant {
constexpr static decltype(V) value = V;
};
The key is that non-type arguments can be declared as auto. If you are using standards before C++17 you may have to use std::integral_constant. There is also a proposal about the constant helper class.
An example:
template <auto V>
struct constant {
constexpr static decltype(V) value = V;
};
constexpr uint64_t factorial(int n) {
if (n <= 0) {
return 1;
}
return n * factorial(n - 1);
}
int main() {
std::cout << "20! = " << constant<factorial(20)>::value << std::endl;
return 0;
}
Have your function take template parameters instead of arguments and implement your logic in a lambda.
#include <iostream>
template< uint64_t N >
constexpr uint64_t factorial() {
// note that we need to pass the lambda to itself to make the recursive call
auto f = []( uint64_t n, auto& f ) -> uint64_t {
if ( n < 2 ) return 1;
return n * f( n - 1, f );
};
return f( N, f );
}
using namespace std;
int main() {
cout << factorial<5>() << std::endl;
}
Lets say that you have a function which generates some security token for your application, such as some hash salt, or maybe a symetric or asymetric key.
Now lets say that you have this function in your C++ as a constexpr and that you generate keys for your build based on some information (like, the build number, a timestamp, something else).
You being a diligent programmer make sure and call this in the appropriate ways to ensure it's only called at compile time, and thus the dead stripper removes the code from the final executable.
However, you can't ever be sure that someone else isn't going to call it in an unsafe way, or that maybe the compiler won't strip the function out, and then your security token algorithm will become public knowledge, making it more easy for would be attackers to guess future tokens.
Or, security aside, let's say the function takes a long time to execute and you want to make sure it never happens during runtime and causes a bad user experience for your end users.
Are there any ways to ensure that a constexpr function can never be called at runtime? Or alternately, throwing an assert or similar at runtime would be ok, but not as ideal obviously as a compile error would be.
I've heard that there is some way involving throwing an exception type that doesn't exist, so that if the constexpr function is not deadstripped out, you'll get a linker error, but have heard that this only works on some compilers.
Distantly related question: Force constexpr to be evaluated at compile time
In C++20 you can just replace constexpr by consteval to enforce a function to be always evaluated at compile time.
Example:
int rt_function(int v){ return v; }
constexpr int rt_ct_function(int v){ return v; }
consteval int ct_function(int v){ return v; }
int main(){
constexpr int ct_value = 1; // compile value
int rt_value = 2; // runtime value
int a = rt_function(ct_value);
int b = rt_ct_function(ct_value);
int c = ct_function(ct_value);
int d = rt_function(rt_value);
int e = rt_ct_function(rt_value);
int f = ct_function(rt_value); // ERROR: runtime value
constexpr int g = rt_function(ct_value); // ERROR: runtime function
constexpr int h = rt_ct_function(ct_value);
constexpr int i = ct_function(ct_value);
}
Pre C++20 workaround
You can enforce the use of it in a constant expression:
#include<utility>
template<typename T, T V>
constexpr auto ct() { return V; }
template<typename T>
constexpr auto func() {
return ct<decltype(std::declval<T>().value()), T{}.value()>();
}
template<typename T>
struct S {
constexpr S() {}
constexpr T value() { return T{}; }
};
template<typename T>
struct U {
U() {}
T value() { return T{}; }
};
int main() {
func<S<int>>();
// won't work
//func<U<int>>();
}
By using the result of the function as a template argument, you got an error if it can't be solved at compile-time.
A theoretical solution (as templates should be Turing complete) - don't use constexpr functions and fall back onto the good-old std=c++0x style of computing using exclusively struct template with values. For example, don't do
constexpr uintmax_t fact(uint n) {
return n>1 ? n*fact(n-1) : (n==1 ? 1 : 0);
}
but
template <uint N> struct fact {
uintmax_t value=N*fact<N-1>::value;
}
template <> struct fact<1>
uintmax_t value=1;
}
template <> struct fact<0>
uintmax_t value=0;
}
The struct approach is guaranteed to be evaluated exclusively at compile time.
The fact the guys at boost managed to do a compile time parser is a strong signal that, albeit tedious, this approach should be feasible - it's a one-off cost, maybe one can consider it an investment.
For example:
to power struct:
// ***Warning: note the unusual order of (power, base) for the parameters
// *** due to the default val for the base
template <unsigned long exponent, std::uintmax_t base=10>
struct pow_struct
{
private:
static constexpr uintmax_t at_half_pow=pow_struct<exponent / 2, base>::value;
public:
static constexpr uintmax_t value=
at_half_pow*at_half_pow*(exponent % 2 ? base : 1)
;
};
// not necessary, but will cut the recursion one step
template <std::uintmax_t base>
struct pow_struct<1, base>
{
static constexpr uintmax_t value=base;
};
template <std::uintmax_t base>
struct pow_struct<0,base>
{
static constexpr uintmax_t value=1;
};
The build token
template <uint vmajor, uint vminor, uint build>
struct build_token {
constexpr uintmax_t value=
vmajor*pow_struct<9>::value
+ vminor*pow_struct<6>::value
+ build_number
;
}
In the upcoming C++20 there will be consteval specifier.
consteval - specifies that a function is an immediate function, that is, every call to the function must produce a compile-time constant
Since now we have C++17, there is an easier solution:
template <auto V>
struct constant {
constexpr static decltype(V) value = V;
};
The key is that non-type arguments can be declared as auto. If you are using standards before C++17 you may have to use std::integral_constant. There is also a proposal about the constant helper class.
An example:
template <auto V>
struct constant {
constexpr static decltype(V) value = V;
};
constexpr uint64_t factorial(int n) {
if (n <= 0) {
return 1;
}
return n * factorial(n - 1);
}
int main() {
std::cout << "20! = " << constant<factorial(20)>::value << std::endl;
return 0;
}
Have your function take template parameters instead of arguments and implement your logic in a lambda.
#include <iostream>
template< uint64_t N >
constexpr uint64_t factorial() {
// note that we need to pass the lambda to itself to make the recursive call
auto f = []( uint64_t n, auto& f ) -> uint64_t {
if ( n < 2 ) return 1;
return n * f( n - 1, f );
};
return f( N, f );
}
using namespace std;
int main() {
cout << factorial<5>() << std::endl;
}
#include <algorithm>
struct S
{
static constexpr int X = 10;
};
int main()
{
return std::min(S::X, 0);
};
If std::min expects a const int&, the compiler very likely would like to have the S::X also defined somewhere, i.e. the storage of S::X must exists.
See here or here.
Is there a way to force the compiler to evaluate my constexpr at compile time?
The reason is:
Initially, we had a problem in early initialization of static variables in the init priority. There was some struct Type<int> { static int max; };, and some global static int x = Type<int>::max;, and some other early code other_init used that x. When we updated GCC, suddenly we had x == 0 in other_init.
We thought that we could avoid the problem by using constexpr, so that it would always evaluate it at compile time.
The only other way would be to use struct Type<int> { static constexpr int max(); }; instead, i.e. letting it be a function.
For types that are allowed to exist as template value parameters, you can introduce a data structure like this:
template <typename T, T K>
struct force
{
static constexpr T value = K;
};
Usage:
force<int, std::min(S::X, 0)>::value
The constexpr is evaluated at compile time. Your problem is
due to the fact that std::min is not a constexpr, so
regardless of its input, the results are not a const expression
(and in particular, if you initialize a variable with static
lifetime using std::min, it is dynamic initialization).
The simplest solution is probably to define your own min,
something along the lines of:
template <typename T>
constexpr T staticMin( T a, T b )
{
return a > b ? b : a;
}
This should result in full evaluation at compile time, and
static initialization.
This is a simplified example from code meant to generate sequences of arbitrary values (in the sense of std::iota) and iterators of varying categories over them:
struct delta
{
template<typename I>
void inc(I& i) { ++i; }
template<typename I>
I next(I i) { inc(i); return i; }
};
There are many classes like delta, each defining inc differently, e.g. --i, i += step, i -= step, i *= step, f(i) etc. Function next remains the same and is actually shared in a base class.
We are generating the value-based operation of next from the mutating operation of inc. Doing the opposite would be equivalent, however we choose this design for performance, because next is only expected to be called at some initialization, while inc may be called a million times.
The problem is that in case some parameters are compile-time constant, I would like to call next at compile-time given a constexpr argument i.
This is not possible in C++11 because of the call to non-constexpr function inc. Simply changing to
template<typename I>
constexpr I next(I i) { inc(i); return i; }
or
template<typename I>
constexpr I next(I i) { return inc(i), i; }
will not work. Of course, we could provide another special function like
template<typename I>
constexpr I next(I i) { return i + 1; }
but this is too much code duplication, given that there are many classes like delta and many other operations like inc/next.
I have seen that constexpr function restrictions are to be relaxed in C++14, but I cannot achieve this in practice yet.
So:
will this work in C++14 eventually?
what is the status of standardization?
what is the status of compilers?
any possible workaround?
EDIT
It seems inc should be constexpr as well (though void). This works on clang 3.4:
struct delta
{
template<typename I>
constexpr void inc(I& i) { ++i; }
template<typename I>
constexpr I next(I i) { inc(i); return i; }
};
...but not on gcc 4.8.2. So is the code above correct C++14? Is it only a matter of time for gcc?
It is not surprising that this example does not work on gcc, according to this page C++14s generalized constexpr functions are not yet supported.
I believe the source code in your edit is valid C++14, the Draft Standard available here contains, on page 126(§5.19) an example that is very similiar to yours(without template, non member functions and they use a temporary):
constexpr int incr(int &n) {
return ++n;
}
constexpr int g(int k) {
constexpr int x = incr(k);// error: incr(k) is not a core constant
// expression because lifetime of k
// began outside the expression incr(k)
return x;
}
constexpr int h(int k) {
int x = incr(k);
// OK: incr(k) is not required to be a core
// constant expression
return x;
}
constexpr int y = h(1); // OK: initializes y with the value 2
// h(1) is a core constant expression because
// the lifetime of k begins inside h(1)
If my reading of the Standard is correct, the fact that yours are member functions should not matter, because "this" is not evaluated and it seems like the code does not violate any of the other rules outlaid in §5.19(2).