template argument deduction for function pointer (g++ & ICC vs Clang++ & VC++ ) - c++

Consider following program:
#include <iostream>
template <typename T>
void foo(const T* x) {
x();
}
void bar() { std::cout<<"bar() is called\n"; }
int main() {
foo(bar);
}
It compiles fine on clang++ & VC++ but g++ gives following compiler error (See live demo here )
main.cpp: In function 'int main()':
main.cpp:10:9: error: no matching function for call to 'foo(void (&)())'
foo(bar);
^
main.cpp:3:6: note: candidate: template<class T> void foo(const T*)
void foo(const T* x) {
^~~
main.cpp:3:6: note: template argument deduction/substitution failed:
main.cpp:10:9: note: types 'const T' and 'void()' have incompatible cv-qualifiers
foo(bar);
^
I've used -pedantic-errors when using g++ & clang++ and I've used /W4 & /Zaoption when using VC++ compiler. See live demo here & here. So, I want to know how template type parameter T will be deduced here ? If I remove const from the program then it compiles fine on g++ also. If I use const T& then it compiles fine on all 3 compilers. So, how exactly type will be deduced here in these cases ?
Update:
This program fails in compilation on Intel C++ compiler also. See live demo here . So, is this bug in g++ & Intel C++ or bug in Clang++ & VC++ ?

This is essentially CWG issue 1584:
It is not clear whether the following is well-formed or not:
void foo(){}
template<class T> void deduce(const T*) { }
int main() {
deduce(foo);
}
Implementations vary in their treatment of this example.
Which is currently still active. It's not really possible to say which compiler is right. Though as the note from 2015 indicates, the consensus in the CWG is currently that this should be rejected.
To give a little more context, we must remember that a function type with a cv-qualifier-seq has special meaning (think member functions), and is not simply a type the designates something which may not be modified. Moreover, you can't even add the cv qualification in some sneaky manner, as [dcl.fct]/7 illustrates:
The effect of a cv-qualifier-seq in a function declarator is not the
same as adding cv-qualification on top of the function type. In the
latter case, the cv-qualifiers are ignored. [ Note: A function type
that has a cv-qualifier-seq is not a cv-qualified type; there are no
cv-qualified function types. — end note ][ Example:
typedef void F();
struct S {
const F f; // OK: equivalent to: void f();
};
— end example ]
There is no way in the language to form a const qualified function type. And yet, the deduction we need is to have const T deduced as void(). The former is a const qualified type, and it must also be a function type. But that's a type that cannot exist! So how can it be deduced?!
On the other hand there is machinery in the standard to deduce it if you were using a reference instead of a pointer.
So it isn't that clear how this should be resolved. On the one hand the wording today doesn't allow it per-se, but on the other hand machinery for it is already in place for references. So some implementations go ahead and do the same for pointers.

Related

In a C++ function template, why can't I use a lambda to specify the array size of a parameter?

I stumbled on the following while trying to implement some SFINAE trickery (what I was actually trying to achieve is irrelevant; I wan't to understand this behavior):
I define a constexpr function that takes a reference to an array of size 1, but I specify the array size through a lambda call:
constexpr bool f(const char(&)[+[](){return 1;}()]) {
return true;
}
(The + before the lambda is because the compiler complains about two consecutive left brackets.)
I add a caller function:
constexpr bool g() {
char x[1] = {};
return f(x);
}
This compiles fine.
Now I templatize and instantiate:
template<typename T>
constexpr bool f(const char(&)[+[](){return 1;}()]) {
return true;
}
constexpr bool g() {
char x[1] = {};
return f<int>(x);
}
This time I get a strange compiler error:
ERROR: maps/suggest/indexer/nhr/nhr_flume_flags.cc:134:45 no matching function for call to 'f'
constexpr bool g() { char x[1] = {}; return f<int>(x); }
^~~~~~~
maps/suggest/indexer/nhr/nhr_flume_flags.cc:130:16 candidate function [with T = void] not viable: no known conversion from 'char[1]' to 'const char[+[]() {
return 1;
}()]' for 1st argument
constexpr bool f(const char(&)[+[](){return 1;}()]) { return true; }
^
1 error generated.
Why am I getting this error?
The command I'm using is: /usr/lib/llvm-11/bin/clang++ -stdlib=libstdc++ -std=c++17 myprog.cc
The version info from the compiler is:
Debian clang version 11.1.0-4+build3
Target: x86_64-pc-linux-gnu
Thread model: posix
InstalledDir: /usr/lib/llvm-11/bin
Why am I getting this error?
/usr/lib/llvm-11/bin/clang++ -stdlib=libstdc++ -std=c++17 myprog.cc
Using lambdas in function signature isn't allowed in C++17:
[expr.prim.lambda]
A lambda-expression is a prvalue whose result object is called the closure object. A lambda-expression shall not appear in an unevaluated operand, in a template-argument, in an alias-declaration, in a typedef declaration, or in the declaration of a function or function template outside its function body and default arguments. [ Note: The intention is to prevent lambdas from appearing in a signature.  — end note ] [ Note: A closure object behaves like a function object. — end note ]
The program is ill-formed. The diagnostic message has room for improvement. Not diagnosing the non-template is a compiler bug.
It's easy to work around using a constant. Much easier to read too:
constexpr inline auto s = [](){return 1;}();
template<typename T>
constexpr bool f(const char(&)[s])
Since proposal P0315, it should be allowed in C++20 because the highlighted part of the rule is removed. Clang however still fails to compile it in C++20 which is a bug as far as I can tell. At the moment, Clang's support for P0315 is listed as "partial".

universal references and packed fields

Consider the following code:
#include <cstdint>
struct S {
uint32_t f1;
} __attribute__((packed)); // removing this attribute makes things work.
template <class T> void func(const T &x) {}
template <class T> void func(T &&x) {} // commenting out this line makes it "work"
int main() {
S s;
func(s.f1); // Error here in GCC, but not clang
}
GCC gives the following error:
<source>: In function 'int main()':
<source>:16:12: error: cannot bind packed field 's.S::f1' to 'unsigned int&'
16 | func(s.f1);
| ~~^~
It appears that GCC is choosing to not allow universal references to members of a packed struct, presumably due to alignment concerns. However, clang compiles the code just fine.
Adding to my confusion is that if I remove the (T &&x) overload, it works "just fine" if only the (const T &) overload exists. I would expect that if it can't use the universal-ref overload, then it would just fall back on const-ref version... but it doesn't.
Is clang incorrect here? Is GCC? Is it just undefined behavior and therefore they are both "right"?
The func(const T &x) is allowed because GCC will create a temporary to the packed member.
When adding a forwarding reference overload, the function call will resolve to a function that looks like func(uint32_t&). Since it's a mutable lvalue reference, no temporary can be created and the overload resolution fails, since there is no better match.
You can make it work by forcing const allowing the compiler to create a temporary again:
func(std::as_const(s).f1);

Why can't lambda, when cast to function pointer, be used in constexpr context?

Consider an example:
template <void (*Foo)()>
struct S {
};
int main() {
struct A {
static void x() { }
};
S<&A::x> s;
}
The code compiles in clang, gcc argue that x doesn't have a linkage...
For quite similar example just when using lambda expression:
template <void (*Foo)()>
struct S {
};
int main() {
auto lambda = []{};
S<+lambda> s;
}
Both gcc and clang agree not to compile the code: according to gcc the function returned by unary + doesn't have linkage, clang states in contrast that cast operator to the function isn't declared as constexpr. Are there any reasons to disallow lambda cast to function pointer to be used in constexpr context?
Find below errors produced by compilers and the live demos:
gcc:
prog.cc:7:14: error: 'main()::::_FUN' is not a valid template argument for type 'void (*)()' because 'static constexpr void main()::::_FUN()' has no linkage
clang:
prog.cc:7:8: note: non-constexpr function 'operator void (*)()' cannot be used in a constant expression
Clang hasn't implemented constexpr lambdas yet.
GCC is behind in other ways. [temp.arg.nontype]/2's only interesting constraint is that the argument shall be a constant expression. But [expr.const]/(5.2) makes it one, so that's perfectly valid. Perhaps GCC didn't implement N4198 yet, which eliminated the linkage requirement.
Note that both constexpr lambdas and no-linkage function pointer template arguments are post C++14 features.

rvalue reference in class template vs function template

#include <iostream>
template <typename T>
class test
{
public:
test(T&& t)
{
}
};
template <typename T>
void succeed(T&& t)
{
}
int main()
{
int i = 1;
test<int> t(i); // failed to compile
succeed(i); // OK
return 0;
}
Error from GCC 5.2:
main.cpp: In function 'int main()':
main.cpp:20:18: error: cannot bind 'int' lvalue to 'int&&'
test t(i);
^
main.cpp:7:5: note: initializing argument 1 of 'test::test(T&&) [with T = int]'
test(T&& t)
^~~~
Could someone explain why the class template cannot compile but function template is OK?
Thanks.
In succeed, T&& t is a forwarding reference, not an rvalue reference. But in test, it is an rvalue reference.
A forwarding reference happens only when the parameter is T&&, and T is a template parameter of that function. In your code T is a template parameter of the enclosing class, so it doesn't count as a forwarding reference.
A forwarding reference may bind to both lvalues and rvalues.
During the drafting of C++11 it was suggested to use different syntax for forwarding references than rvalue references (instead of using T&& t for both); however the committee eventually settled on the current behaviour.
For a more detailed description of template parameter deduction, including a more precise specification of when T&& becomes a forwarding reference, see here -- search for the term "forwarding reference" to find the special rules for forwarding references.
Your confusion is probably rooted in your assumption that in both cases T is int. This is why you presume that these two cases as similar. In reality they are not.
In the class version you are manually specifying what T is. You explicitly tell the compiler that T is int. Constructor parameter type T && in this case becomes int &&, which cannot bind to a regular lvalue. Hence the error.
In the function version you don't tell the compiler what T is, but instead you expect the compiler to deduce it. In situations like yours the language is deliberately designed to deduce T as int & (note: not as int, but rather as int &). Once T is deduced as int &, the so called "reference collapsing" rules lead to function parameter type T && becoming int & - an ordinary lvalue reference. This parameter can successfully bind to lvalue argument i.
That explains the difference you observe.
For the sake of experiment, in the latter case you can suppress template argument deduction and specify the template argument explicitly
succeed<int>(i);
That will forcefully specify T as int and lead to the very same error as in the class version for the very same reason.
Similarly, you can "simulate" function's behavior for your class by specifying the template argument as int &
test<int &> t(i);
The same "reference collapsing" rules will make your constructor invocation to compile successfully.

Does C++11 require this lambda to be declared mutable?

Consider this C++11 code:
#include <functional>
#include <cstdlib>
template <typename F>
void test(F &&f) {
auto foo = [f]() {
f();
};
foo();
}
int main() {
test(std::bind(std::puts, "hello"));
return 0;
}
GCC and Clang accept this as valid C++11 code, but Visual Studio 2013 requires the lambda to be declared mutable (auto foo = [f]() mutable { ... }). Otherwise I get this error:
error C3848: expression having type 'const std::_Bind<true,int,int (__cdecl *const )(const char *),const char (&)[6]>' would lose some const-volatile qualifiers in order to call 'int std::_Bind<true,int,int (__cdecl *const )(const char *),const char (&)[6]>::operator ()<>(void)'
Is Visual Studio right to reject this code without mutable, or is it valid C++11?
(Curiously Clang rejects the code if you change std::bind(std::puts, "hello") to std::bind(std::exit, 0) apparently because it considers noreturn to make the function type different; I'm quite sure this is a bug.)
This isn't really about lambdas.
#include <functional>
#include <cstdlib>
int main() {
const auto x = std::bind(std::puts, "hello");
x();
}
This is accepted by GCC, but rejected by MSVC.
The standard is unclear on whether this is valid, IMO. The return value g of std::bind has an unspecified return type for which g(...) is valid and defined in terms of the cv-qualifiers of g, but the standard doesn't actually say that any operator() must be callable for const-qualified objects or references. It strongly implies that this is intended to be valid, because otherwise the reference to g's cv-qualifiers seems useless, but it doesn't actually say it is valid.
Because of that, I think MSVC's behaviour is not what the standard's authors intended, but it may nonetheless conform to what the standard requires.
This looks like a bug in the Visual Studio implementation of bind, returning a type with only a non-const function call operator. It should return a type that forwards all function calls to the bound function object, regardless of its own cv-qualifications.
To summarise the rather opaque language of C++11 20.8.9.1.2, function calls on the result of bind should be forwarded to the bound function object, and so should be allowed if calls on that object would be allowed. So it would be an error if the bound function object weren't callable if const; but here, being a function pointer, it is callable regardless of cv-qualifications.