Why does GCC speak about a dependent base? - c++

This is a continuation of When is a class member visible?
After making the class compile with GCC by moving the declaration of pk_ to the beginning, I tried to use it:
#include <string>
#include <map>
#include <type_traits>
using from_type = std::map<std::string, std::string>;
template<typename PK, size_t N>
struct pf {
public:
PK pk_;
pf(from_type const& a, std::string const& pkn) noexcept(noexcept(fill(pk_, std::string{})))
: map_(a) // GCC 4.8 requires ()s for references
, pk_{ [&]{ fill(pk_, pkn); return pk_; }() }
{
}
template<typename prop_t>
typename std::enable_if<
std::is_integral<typename std::decay<prop_t>::type>::value,
pf<PK, N>>::type const&
fill(prop_t& , std::string const& , prop_t = 0) const noexcept(false);
pf<PK, N> const&
fill(std::string& , std::string const&) const noexcept;
protected:
from_type const& map_;
uint32_t aieee;
};
std::string k;
from_type m;
int i;
std::string s;
static_assert(!noexcept(pf<int , 42>{m, k}), "int could throw");
static_assert( noexcept(pf<std::string, 17>{m, k}), "string shouldn't throw");
clang 4.0, 6.0 and trunk again compiled the program.
GCC was still not happy:
$ g++-99 -Wall -pedantic -Wextra -Wformat=2 -std=c++14 pf.cpp
pf.cpp: In instantiation of ‘pf<PK, N>::pf(const from_type&, const string&) [with PK = int; long unsigned int N = 42; from_type = std::map<std::__cxx11::basic_string<char>, std::__cxx11::basic_string<char> >; std::string = std::__cxx11::basic_string<char>]’:
pf.cpp:38:49: required from here
pf.cpp:12:74: error: no matching function for call to ‘fill(int&, std::__cxx11::basic_string<char>)’
12 | pf(from_type const& a, std::string const& pkn) noexcept(noexcept(fill(pk_, std::string{})))
| ~~~~^~~~~~~~~~~~~~~~~~~~
In file included from /usr/local/lib/gcc/x86_64-pc-linux-gnu/9.0.0/include/c++/bits/char_traits.h:39,
from /usr/local/lib/gcc/x86_64-pc-linux-gnu/9.0.0/include/c++/string:40,
from pf.cpp:1:
/usr/local/lib/gcc/x86_64-pc-linux-gnu/9.0.0/include/c++/bits/stl_algobase.h:742:5: note: candidate: ‘template<class _ForwardIterator, class _Tp> void std::fill(_ForwardIterator, _ForwardIterator, const _Tp&)’
742 | fill(_ForwardIterator __first, _ForwardIterator __last, const _Tp& __value)
| ^~~~
/usr/local/lib/gcc/x86_64-pc-linux-gnu/9.0.0/include/c++/bits/stl_algobase.h:742:5: note: template argument deduction/substitution failed:
pf.cpp:12:74: note: deduced conflicting types for parameter ‘_ForwardIterator’ (‘int’ and ‘std::__cxx11::basic_string<char>’)
12 | pf(from_type const& a, std::string const& pkn) noexcept(noexcept(fill(pk_, std::string{})))
| ~~~~^~~~~~~~~~~~~~~~~~~~
pf.cpp: In instantiation of ‘pf<PK, N>::pf(const from_type&, const string&) [with PK = std::__cxx11::basic_string<char>; long unsigned int N = 17; from_type = std::map<std::__cxx11::basic_string<char>, std::__cxx11::basic_string<char> >; std::string = std::__cxx11::basic_string<char>]’:
pf.cpp:39:49: required from here
pf.cpp:12:74: error: no matching function for call to ‘fill(std::__cxx11::basic_string<char>&, std::__cxx11::basic_string<char>)’
In file included from /usr/local/lib/gcc/x86_64-pc-linux-gnu/9.0.0/include/c++/bits/char_traits.h:39,
from /usr/local/lib/gcc/x86_64-pc-linux-gnu/9.0.0/include/c++/string:40,
from pf.cpp:1:
/usr/local/lib/gcc/x86_64-pc-linux-gnu/9.0.0/include/c++/bits/stl_algobase.h:742:5: note: candidate: ‘template<class _ForwardIterator, class _Tp> void std::fill(_ForwardIterator, _ForwardIterator, const _Tp&)’
742 | fill(_ForwardIterator __first, _ForwardIterator __last, const _Tp& __value)
| ^~~~
/usr/local/lib/gcc/x86_64-pc-linux-gnu/9.0.0/include/c++/bits/stl_algobase.h:742:5: note: template argument deduction/substitution failed:
pf.cpp:12:74: note: candidate expects 3 arguments, 2 provided
12 | pf(from_type const& a, std::string const& pkn) noexcept(noexcept(fill(pk_, std::string{})))
| ~~~~^~~~~~~~~~~~~~~~~~~~
pf.cpp:39:16: error: static assertion failed: string shouldn't throw
39 | static_assert( noexcept(pf<std::string, 17>{m, k}), "string shouldn't throw");
| ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
This was confusing (who said anything about iterators?) until I saw the answer to my first question. The compiler couldn't see the fill members, so it tried the only fill method available to it: the one from <algorithm>, which was accidentally included via
. /usr/local/lib/gcc/x86_64-pc-linux-gnu/9.0.0/include/c++/string
.. /usr/local/lib/gcc/x86_64-pc-linux-gnu/9.0.0/include/c++/bits/char_traits.h
... /usr/local/lib/gcc/x86_64-pc-linux-gnu/9.0.0/include/c++/bits/stl_algobase.h
So I renamed the fill members to fillz (including the one in the noexcept operator), which resulted in:
$ g++-99 -Wall -pedantic -Wextra -Wformat=2 -std=c++14 pf.cpp
pf.cpp: In instantiation of ‘pf<PK, N>::pf(const from_type&, const string&) [with PK = int; long unsigned int N = 42; from_type = std::map<std::__cxx11::basic_string<char>, std::__cxx11::basic_string<char> >; std::string = std::__cxx11::basic_string<char>]’:
pf.cpp:38:49: required from here
pf.cpp:12:75: error: ‘fillz’ was not declared in this scope, and no declarations were found by argument-dependent lookup at the point of instantiation [-fpermissive]
12 | pf(from_type const& a, std::string const& pkn) noexcept(noexcept(fillz(pk_, std::string{})))
| ~~~~~^~~~~~~~~~~~~~~~~~~~
pf.cpp:12:75: note: declarations in dependent base ‘pf<int, 42>’ are not found by unqualified lookup
pf.cpp:12:75: note: use ‘pf::fillz’ instead
pf.cpp:12:75: error: cannot call member function ‘const typename std::enable_if<std::is_integral<typename std::decay<prop_t>::type>::value, pf<PK, N> >::type& pf<PK, N>::fillz(prop_t&, const string&, prop_t) const [with prop_t = int; PK = int; long unsigned int N = 42; typename std::enable_if<std::is_integral<typename std::decay<prop_t>::type>::value, pf<PK, N> >::type = pf<int, 42>; std::string = std::__cxx11::basic_string<char>]’ without object
pf.cpp: In instantiation of ‘pf<PK, N>::pf(const from_type&, const string&) [with PK = std::__cxx11::basic_string<char>; long unsigned int N = 17; from_type = std::map<std::__cxx11::basic_string<char>, std::__cxx11::basic_string<char> >; std::string = std::__cxx11::basic_string<char>]’:
pf.cpp:39:49: required from here
pf.cpp:12:75: error: ‘fillz’ was not declared in this scope, and no declarations were found by argument-dependent lookup at the point of instantiation [-fpermissive]
pf.cpp:12:75: note: declarations in dependent base ‘pf<std::__cxx11::basic_string<char>, 17>’ are not found by unqualified lookup
pf.cpp:12:75: note: use ‘pf::fillz’ instead
pf.cpp:12:75: error: cannot call member function ‘const pf<PK, N>& pf<PK, N>::fillz(std::string&, const string&) const [with PK = std::__cxx11::basic_string<char>; long unsigned int N = 17; std::string = std::__cxx11::basic_string<char>]’ without object
pf.cpp:39:16: error: static assertion failed: string shouldn't throw
39 | static_assert( noexcept(pf<std::string, 17>{m, k}), "string shouldn't throw");
| ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Obviously GCC can still not see the member (now named fillz), but why does it complain about dependent bases? Struct pf has nothing to do with inheritance. Could it be because this kind of visibility problem comes up most often with dependent bases?
In the end, the correct usage turned out to be:
noexcept(noexcept(std::declval<pf&>().fill(std::declval<PK&>(), std::string{})))

At first glance, this seems like it could be more than one bug in GCC.
(All standard quotes are from C++14, as you are compiling with -std=c++14. All bold is added.)
[basic.scope.class]¶1
The potential scope of a name declared in a class consists [...] also of all function bodies, default arguments, exception-specifications, and brace-or-equal-initializers of non-static data members in that class (including such things in nested classes).
So in the noexcept clause of a member function declaration, all members of the class are in scope regardless of the order in which they are declared.
But this isn't enough to convince ourselves that your code is well-formed. We are in a class template, so next we need to consider two-phase lookup:
[temp.dep]¶1
In an expression of the form:
postfix-expression ( expression-listopt )
where the postfix-expression is an unqualified-id, the unqualified-id denotes a dependent name if
(1.1) any of the expressions in the expression-list is a pack expansion,
(1.2) any of the expressions in the expression-list is a type-dependent expression, or
(1.3) if the unqualified-id is a template-id in which any of the template arguments depends on a template parameter.
Such names are unbound and are looked up at the point of the template instantiation in both the context of the template definition and the context of the point of instantiation.
Therefore fill is a dependent name in fill(pk_, std::string{}) because pk_ is type-dependent, so fill will be looked up when the template is instantiated, both in the context of the template itself and the point where it was instantiated. Due to ADL and the fact that one of the arguments is std::string, the std namespace is included in the lookup, leading to the messages about std::fill. But lookup should also find the member function fill, as previously discussed.
Now we have convinced ourselves that the code is well-formed. We could simply file a bug and move on. But let's keep digging to see how well we understand what's happening.
[temp.dep]¶3
In the definition of a class or class template, if a base class depends on a template-parameter, the base class scope is not examined during unqualified name lookup either at the point of definition of the class template or member or during an instantiation of the class template or member.
The error messages seem to indicate that GCC is (wrongly) considering the enclosing class of fill to be a dependent base of fill, and hence applying the wording of [temp.dep]¶3.
The usual way to refer to names in dependent bases is to refer to them through a qualified-id (e.g. pf::fill) or a class member access expression (e.g. this->fill). So what happens when we try either of these approaches as a workaround?
Writing this->fill seems to work in your example code, but as noted by #n.m. in the comments, this is brittle and doesn't work in a more minimal example, yielding the error "invalid use of 'this' at top level".
Writing pf::fill yields the error "cannot call member function without object". This is probably for the same reason that this->fill failed: if this is invalid then it must also be invalid to transform an id-expression into a implicit member access expression.
[expr.prim.general]/3
If a declaration declares a member function or member function template of a class X, the expression this is a prvalue of type "pointer to cv-qualifier-seq X" between the optional cv-qualifer-seq and the end of the function-definition, member-declarator, or declarator.
So this is valid to appear in an exception-specification and GCC is wrong to reject it, and this looks on the face of it to be a bug separate from the one we already diagnosed. This last one is already known as 52869, but the testcase submitted with the bugfix failed to exercise the case in which the noexcept appears before the declaration of the member function it refers to, as in your case. So it's unclear whether your first set of errors are really part of the same bug.

Related

Is this a gcc 11 compiler bug with C++20 concepts, or am I doing something wrong?

I'm experimenting with C++20 concepts, using my own local build of a clone of the GCC 11 source code. I'm getting compilation errors from GCC that seem wrong to me. I've reduced my code triggering the diagnostic to what's below, with the GCC command line to compile it and the full resulting diagnostic output below that.
For focus, the particular section of the diagnostic output that seems wrong to me is
repro.cpp:9:13: note: the required expression ‘flequal(a, b)’ is invalid, because
9 | {flequal(a, b)} -> std::convertible_to<bool>;
| ~~~~~~~^~~~~~
repro.cpp:9:13: error: ‘flequal’ was not declared in this scope, and no declarations were found by argument-dependent lookup at the point of instantiation [-fpermissive]
repro.cpp:22:6: note: ‘template<class auto:1, class auto:2> requires (Flequalable<auto:1>) && (Flequalable<auto:2>) bool flequal(const Bar<auto:1>&, const Bar<auto:2>&)’ declared here, later in the translation unit
22 | bool flequal(Bar<Flequalable auto> const &a, Bar<Flequalable auto> const &b) {
| ^~~~~~~
It says the candidate function was declared later in the translation unit than the point of instantiation, but that is not true, at least going by source line order in my code snippet.
Is this build error wrong, or am I doing something wrong in my use of C++20 concepts?
Note the variant, in a comment at the end of the code, which compiles without any diagnostic.
Code:
#include <concepts>
using std::floating_point;
template< typename T >
concept Flequalable
= floating_point<T>
&& requires(T a, T b) {
{flequal(a, b)} -> std::convertible_to<bool>;
};
template<floating_point T>
bool flequal( T a, T b) {
return true;
}
template<typename T>
struct Bar {
T t;
};
bool flequal(Bar<Flequalable auto> const &a, Bar<Flequalable auto> const &b) {
return true;
}
bool foo() {
Bar<double> a = {2.0};
Bar<double> b = {3.0};
return flequal(a, b); // Causes diagnostic
// return flequal(a.t, b.t); // This works
}
Compilation command-line:
/usr/local/gcc-11-master/bin/g++-11 -c repro.cpp -o repro.o -std=c++20 -fconcepts-diagnostics-depth=5 2> diagnostic.txt
Full GCC diagnostic:
repro.cpp: In function ‘bool foo()’:
repro.cpp:29:24: error: no matching function for call to ‘flequal(Bar<double>&, Bar<double>&)’
29 | return flequal(a, b); // Causes diagnostic
| ^
repro.cpp:13:6: note: candidate: ‘template<class T> requires floating_point<T> bool flequal(T, T)’
13 | bool flequal( T a, T b) {
| ^~~~~~~
repro.cpp:13:6: note: template argument deduction/substitution failed:
repro.cpp:13:6: note: constraints not satisfied
In file included from repro.cpp:1:
/usr/local/gcc-11-master/include/c++/11.0.1/concepts: In substitution of ‘template<class T> requires floating_point<T> bool flequal(T, T) [with T = Bar<double>]’:
repro.cpp:29:24: required from here
/usr/local/gcc-11-master/include/c++/11.0.1/concepts:111:13: required for the satisfaction of ‘floating_point<T>’ [with T = Bar<double>]
/usr/local/gcc-11-master/include/c++/11.0.1/concepts:111:30: note: the expression ‘is_floating_point_v<_Tp> [with _Tp = Bar<double>]’ evaluated to ‘false’
111 | concept floating_point = is_floating_point_v<_Tp>;
| ^~~~~~~~~~~~~~~~~~~~~~~~
repro.cpp:22:6: note: candidate: ‘template<class auto:1, class auto:2> requires (Flequalable<auto:1>) && (Flequalable<auto:2>) bool flequal(const Bar<auto:1>&, const Bar<auto:2>&)’
22 | bool flequal(Bar<Flequalable auto> const &a, Bar<Flequalable auto> const &b) {
| ^~~~~~~
repro.cpp:22:6: note: template argument deduction/substitution failed:
repro.cpp:22:6: note: constraints not satisfied
repro.cpp: In substitution of ‘template<class auto:1, class auto:2> requires (Flequalable<auto:1>) && (Flequalable<auto:2>) bool flequal(const Bar<auto:1>&, const Bar<auto:2>&) [with auto:1 = double; auto:2 = double]’:
repro.cpp:29:24: required from here
repro.cpp:6:9: required for the satisfaction of ‘Flequalable<auto:1>’ [with auto:1 = double]
repro.cpp:8:8: in requirements with ‘T a’, ‘T b’ [with T = double]
repro.cpp:9:13: note: the required expression ‘flequal(a, b)’ is invalid, because
9 | {flequal(a, b)} -> std::convertible_to<bool>;
| ~~~~~~~^~~~~~
repro.cpp:9:13: error: ‘flequal’ was not declared in this scope, and no declarations were found by argument-dependent lookup at the point of instantiation [-fpermissive]
repro.cpp:22:6: note: ‘template<class auto:1, class auto:2> requires (Flequalable<auto:1>) && (Flequalable<auto:2>) bool flequal(const Bar<auto:1>&, const Bar<auto:2>&)’ declared here, later in the translation unit
22 | bool flequal(Bar<Flequalable auto> const &a, Bar<Flequalable auto> const &b) {
| ^~~~~~~
Here's the exact GCC version information:
Using built-in specs.
COLLECT_GCC=/usr/local/gcc-11-master/bin/g++-11
COLLECT_LTO_WRAPPER=/usr/local/gcc-11-master/libexec/gcc/x86_64-pc-linux-gnu/11.0.1/lto-wrapper
Target: x86_64-pc-linux-gnu
Configured with: ../gcc/configure --prefix=/usr/local/gcc-11-master --program-suffix=-11 --enable-libstdcxx-debug : (reconfigured) ../gcc/configure --prefix=/usr/local/gcc-11-master --program-suffix=-11 --enable-libstdcxx-debug
Thread model: posix
Supported LTO compression algorithms: zlib
gcc version 11.0.1 20210304 (experimental) (GCC)
First, Bar<Flequalable auto> is not valid syntax. You can have a top-level variable declared like Concept auto v but you can't nest that within templates. So it really should be:
template <Flequalable T, Flequalable U>
bool flequal(Bar<T> const &a, Bar<U> const &b) {
return true;
}
Now, what happens when we try to call flequal(a, b). We deduce T=double and U=double and then try to evaluate if double satisfies Flequalable. double is floating_point, so we keep going to check the requirement that flequal(a, b) is a valid expression for two doubles.
This is unqualified lookup, so we first do regular unqualified lookup for the name flequal. It's important to keep in mind that this lookup happens from the point of the concept definition, not from the point where it is used. That is, we're looking up from here:
#include <concepts>
using std::floating_point;
template< typename T >
concept Flequalable
= floating_point<T>
&& requires(T a, T b) {
{flequal(a, b)} -> std::convertible_to<bool>; // <== here!
};
Not where the concept is used:
template <Flequalable T, Flequalable U> // <== not here
bool flequal(Bar<T> const &a, Bar<U> const &b) {
return true;
}
Nor where the function template using the concept is invoked:
return flequal(a, b); // <== not here either
We find nothing. There are no declarations of flequal visible from that point. Then, we do argument-dependent lookup. double has no associated namespaces, so this also trivially finds nothing. As such, there are zero candidates, so double does not satisfy Flequalable.
In order to make this work, you have to make sure that unqualified lookup in the concept Flequalable actually finds your flequal function. Just swap the two declarations, and then everything works fine.
This makes it seem like flequal is some kind of customization point, but there are only a fixed number of floating_point types (float, double, long double, and cv-qualified versions thereof) so I'm not entirely sure what the point here is.

C++ Templates: correct way to return a new type

Sorry for the generic title, but I'm unable to focus the problem.
I have a templatized class method that accept an argument pack and provides a new type in return, to hide the details of the implementation. More specifically, the class handles SQLite queries, and the method calls sqlite3_prepare() to prepare the statement before executing the query.
class Table {
...
template <typename ...Ts>
class PreparedStatement { ... };
template <typename ...Ts>
PreparedStatement<Ts...> prepare(std::tuple<Ts...> tuple) {
// do something
return PreparedStatement<Ts...> ( ... );
}
That works well with "normal" types, but the problem occurs when the arguments are declared const:
const Field<int> fld = createField<int>("name");
...
PreparedStatement<decltype(fld)> s = prepare(make_tuple(fld));
The error is the following:
no match for 'operator =' (operand types are PreparedStatenent<const Field<int>> and PreparedStatement<Field<int>>
I suspect the issue is in my declaration of the function, is there a way to fix this issue and make the function more "elegant" ?
NOTE: I know I can fix the issue by manually declare the s variable, but my doubts are on how the method was implemented.
As Many Asked, here's an example:
#include <tuple>
template <typename T>
struct Field {
};
class Table {
public:
template <typename ...Ts>
class PreparedStatement {
public:
PreparedStatement() {};
};
template <typename ...Ts>
PreparedStatement<Ts...> prepare(std::tuple<Ts...> tuple) {
// do something
return PreparedStatement<Ts...> ( );
}
};
int main()
{
Field<int> fld;
Table t;
Table::PreparedStatement<decltype(fld)> p;
p = t.prepare(std::make_tuple(fld));
// here comes the problem
const Field<int> f2;
Table::PreparedStatement<decltype(f2)> p2;
p2 = t.prepare(std::make_tuple(f2));
return 0;
}
and here's the compiler output
main.cpp: In function 'int main()': main.cpp:35:39: error: no match
for 'operator=' (operand types are 'Table::PreparedStatement >' and 'Table::PreparedStatement >')
p2 = t.prepare(std::make_tuple(f2));
^ main.cpp:10:10: note: candidate: constexpr Table::PreparedStatement >&
Table::PreparedStatement >::operator=(const
Table::PreparedStatement >&)
class PreparedStatement {
^~~~~~~~~~~~~~~~~ main.cpp:10:10: note: no known conversion for argument 1 from 'Table::PreparedStatement >'
to 'const Table::PreparedStatement >&'
main.cpp:10:10: note: candidate: constexpr
Table::PreparedStatement >&
Table::PreparedStatement
::operator=(Table::PreparedStatement >&&) main.cpp:10:10: note: no known conversion for argument 1 from
'Table::PreparedStatement >' to
'Table::PreparedStatement >&&'
UPDATE
As many noted, I could use auto to deduce the type, but in some condition auto cannot practically be used. One is, for example, if I need to declare the statement in the Class Context.
So suppose auto is forbidden for some reason. Isn't any other solution available? See the updated code above.
cppreference.com for make_tuple tells us:
template< class... Types >
tuple<VTypes...> make_tuple( Types&&... args );
For each Ti in Types..., the corresponding type Vi in Vtypes... is
std::decay<Ti>::type unless application of std::decay results in
std::reference_wrapper<X> for some type X, in which case the deduced
type is X&.
While std::decay, among other things, removes cv-qualifiers. So your type will be no PreparedStatement<const Field<int>>, but PreparedStatement<Field<int>>.
You can use auto, as manni66 proposed, to avoid such problems.
auto s = prepare(make_tuple(fld));
I could use auto to deduce the type, but in some condition auto cannot practically be used. One is, for example, if I need to declare the statement in the Class Context. So suppose auto is forbidden for some reason. Isn't any other solution available? See the updated code above.
Instead of auto, you can use a decltype expression that take in count the value returned by prepare.
I mean... instead of
Table::PreparedStatement<decltype(f2)> p2;
you can try with
decltype(t.prepare(std::make_tuple(f2))) p2;
or
decltype(std::declval<Table>().prepare(
std::make_tuple(std::declval<Field<int>>()))) p2;
I suppose you can use a similar decltype() also to declare members of your classes.

std::bind on a generic lambda - auto type deduction

Consider the following code:
#include <iostream>
#include <functional>
int main() {
auto run = [](auto&& f, auto&& arg) {
f(std::forward<decltype(arg)>(arg));
};
auto foo = [](int &x) {};
int var;
auto run_foo = std::bind(run, foo, var);
run_foo();
return 0;
}
Which gives the following compilation error when compiled with clang:
$ clang++ -std=c++14 my_test.cpp
my_test.cpp:6:9: error: no matching function for call to object of type 'const (lambda at my_test.cpp:8:16)'
f(std::forward<decltype(arg)>(arg));
^
/usr/bin/../lib64/gcc/x86_64-pc-linux-gnu/6.3.1/../../../../include/c++/6.3.1/functional:998:14: note: in instantiation of function template specialization 'main()::(anonymous class)::operator()<const (lambda at my_test.cpp:8:16) &, const int &>' requested here
= decltype( std::declval<typename enable_if<(sizeof...(_Args) >= 0),
^
/usr/bin/../lib64/gcc/x86_64-pc-linux-gnu/6.3.1/../../../../include/c++/6.3.1/functional:1003:2: note: in instantiation of default argument for 'operator()<>' required here
operator()(_Args&&... __args) const
^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
my_test.cpp:11:12: note: while substituting deduced template arguments into function template 'operator()' [with _Args = <>, _Result = (no value)]
run_foo();
^
my_test.cpp:8:16: note: candidate function not viable: 1st argument ('const int') would lose const qualifier
auto foo = [](int &x) {};
^
my_test.cpp:8:16: note: conversion candidate of type 'void (*)(int &)'
1 error generated.
Why is arg deduced to be const int& instead of just int&?
std::bind documentation says:
Given an object g obtained from an earlier call to bind, when it is
invoked in a function call expression g(u1, u2, ... uM), an invocation
of the stored object takes place, as if by std::invoke(fd,
std::forward(v1), std::forward(v2), ...,
std::forward(vN)), where fd is a value of type std::decay_t the
values and types of the bound arguments v1, v2, ..., vN are determined
as specified below.
...
Otherwise, the
ordinary stored argument arg is passed to the invokable object as
lvalue argument: the argument vn in the std::invoke call above is
simply arg and the corresponding type Vn is T cv &, where cv is the
same cv-qualification as that of g.
But in this case, run_foo is cv-unqualified. What am I missing?
MWE:
#include <functional>
int main() {
int i;
std::bind([] (auto& x) {x = 1;}, i)();
}
[func.bind]/(10.4) states that the cv-qualifiers of the argument passed to the lambda are those of the argument to bind, augmented by the cv-qualifiers of the call wrapper; but there are none, and thus a non-const int should be passed in.
Both libc++ and libstdc++ fail to resolve the call. For libc++, reported as #32856, libstdc++ as #80564. The main problem is that both libraries infer the return type in the signature somehow, looking like this for libstdc++:
// Call as const
template<typename... _Args, typename _Result
= decltype( std::declval<typename enable_if<(sizeof...(_Args) >= 0),
typename add_const<_Functor>::type&>::type>()(
_Mu<_Bound_args>()( std::declval<const _Bound_args&>(),
std::declval<tuple<_Args...>&>() )... ) )>
_Result operator()(_Args&&... __args) const
During template argument deduction as necessitated by overload resolution, the default template argument will be instantiated, which causes a hard error due to our ill-formed assignment inside the closure.
This can be fixed by perhaps a deduced placeholder: remove _Result and its default argument entirely, and declare the return type as decltype(auto). This way, we also get rid of SFINAE which influences overload resolution and thereby induces incorrect behaviour:
#include <functional>
#include <type_traits>
struct A {
template <typename T>
std::enable_if_t<std::is_const<T>{}> operator()(T&) const;
};
int main() {
int i;
std::bind(A{}, i)();
}
This should not compile—as explained above, the argument passed to A::operator() should be non-const because i and the forwarding call wrapper are. However, again, this compiles under libc++ and libstdc++, because their operator()s fall back on const versions after the non-const ones fail under SFINAE.

Function from one library matched to template from another library

I'm working on a C++ project that uses two different libraries: spdlog for logging, and mutils-serialization for serializing objects to bytes (for sending over the network). Both libraries use namespaces properly, but when I attempt to write a program that uses both of them at the same time, my compiler (g++ 6.2) gives me nonsensical errors that seem to indicate it is attempting to instantiate a function template from the spdlog library by using the definition of a function template from the mutils library.
Here's my simple test program:
#include <spdlog/spdlog.h>
#include <spdlog/fmt/ostr.h>
#include "TestSerializableObject.h"
int main(int argc, char** argv) {
auto global_logger = spdlog::rotating_logger_mt("global_logger", "log", 1024 * 1024 * 500, 3);
global_logger->set_pattern("[%H:%M:%S.%e] [%l] %v");
global_logger->set_level(spdlog::level::trace);
std::shared_ptr<spdlog::logger> logger(spdlog::get("global_logger"));
auto message = std::make_shared<messaging::TestSerializableObject>(
1, 2, "A message!");
logger->trace("Received a message: {}", *message);
}
TestSerializableObject is a simple class that implements mutils::ByteRepresentable (the interface that enables serialization and pulls in the mutils-serialization library), and provides an operator<< (which is required for spdlog to be able to log it). I can post the code for it if necessary.
When I compile this with g++ -std=c++14 -I"./src" -I"./libraries" -I"./libraries/mutils/" -L"./libraries/" -O0 -g3 -Wall "src/LibraryCollisionTest.cpp", I get this long, ugly error (don't worry, I'll help you parse it):
In file included from ./libraries/mutils/mutils.hpp:3:0,
from ./libraries/mutils-serialization/SerializationSupport.hpp:2,
from src/TestSerializableObject.h:10,
from src/LibraryCollisionTest.cpp:10:
./libraries/mutils/args-finder.hpp: In instantiation of ‘struct mutils::function_traits<messaging::TestSerializableObject>’:
./libraries/mutils/args-finder.hpp:75:41: required from ‘auto mutils::convert(F) [with F = messaging::TestSerializableObject; ignore = void]’
./libraries/spdlog/fmt/bundled/format.h:1276:46: required from ‘struct fmt::internal::ConvertToInt<messaging::TestSerializableObject>’
./libraries/spdlog/fmt/bundled/format.h:1485:5: required by substitution of ‘template<class T> fmt::internal::MakeValue<Formatter>::MakeValue(const T&, typename fmt::internal::EnableIf<fmt::internal::Not<fmt::internal::ConvertToInt<T>::value>::value, int>::type) [with T = messaging::TestSerializableObject]’
./libraries/spdlog/fmt/bundled/format.h:2465:12: required from ‘static fmt::internal::Value fmt::internal::ArgArray<N, true>::make(const T&) [with Formatter = fmt::BasicFormatter<char>; T = messaging::TestSerializableObject; unsigned int N = 1u]’
./libraries/spdlog/fmt/bundled/format.h:2898:5: required from ‘void fmt::BasicWriter<Char>::write(fmt::BasicCStringRef<CharType>, const Args& ...) [with Args = {messaging::TestSerializableObject}; Char = char]’
./libraries/spdlog/details/logger_impl.h:69:9: required from ‘void spdlog::logger::log(spdlog::level::level_enum, const char*, const Args& ...) [with Args = {messaging::TestSerializableObject}]’
./libraries/spdlog/details/logger_impl.h:127:5: required from ‘void spdlog::logger::trace(const char*, const Args& ...) [with Args = {messaging::TestSerializableObject}]’
src/LibraryCollisionTest.cpp:21:53: required from here
./libraries/mutils/args-finder.hpp:12:37: error: ‘operator()’ is not a member of ‘messaging::TestSerializableObject’
: public function_traits<decltype(&T::operator())>
^~
./libraries/mutils/args-finder.hpp: In instantiation of ‘auto mutils::convert(F) [with F = messaging::TestSerializableObject; ignore = void]’:
./libraries/spdlog/fmt/bundled/format.h:1276:46: required from ‘struct fmt::internal::ConvertToInt<messaging::TestSerializableObject>’
./libraries/spdlog/fmt/bundled/format.h:1485:5: required by substitution of ‘template<class T> fmt::internal::MakeValue<Formatter>::MakeValue(const T&, typename fmt::internal::EnableIf<fmt::internal::Not<fmt::internal::ConvertToInt<T>::value>::value, int>::type) [with T = messaging::TestSerializableObject]’
./libraries/spdlog/fmt/bundled/format.h:2465:12: required from ‘static fmt::internal::Value fmt::internal::ArgArray<N, true>::make(const T&) [with Formatter = fmt::BasicFormatter<char>; T = messaging::TestSerializableObject; unsigned int N = 1u]’
./libraries/spdlog/fmt/bundled/format.h:2898:5: required from ‘void fmt::BasicWriter<Char>::write(fmt::BasicCStringRef<CharType>, const Args& ...) [with Args = {messaging::TestSerializableObject}; Char = char]’
./libraries/spdlog/details/logger_impl.h:69:9: required from ‘void spdlog::logger::log(spdlog::level::level_enum, const char*, const Args& ...) [with Args = {messaging::TestSerializableObject}]’
./libraries/spdlog/details/logger_impl.h:127:5: required from ‘void spdlog::logger::trace(const char*, const Args& ...) [with Args = {messaging::TestSerializableObject}]’
src/LibraryCollisionTest.cpp:21:53: required from here
./libraries/mutils/args-finder.hpp:75:41: error: ‘as_function’ is not a member of ‘mutils::function_traits<messaging::TestSerializableObject>’
return function_traits<F>::as_function(f);
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~^~~
In file included from ./libraries/spdlog/fmt/fmt.h:21:0,
from ./libraries/spdlog/common.h:41,
from ./libraries/spdlog/spdlog.h:12,
from src/LibraryCollisionTest.cpp:8:
./libraries/spdlog/fmt/bundled/format.h: In instantiation of ‘struct fmt::internal::ConvertToInt<messaging::TestSerializableObject>’:
./libraries/spdlog/fmt/bundled/format.h:1485:5: required by substitution of ‘template<class T> fmt::internal::MakeValue<Formatter>::MakeValue(const T&, typename fmt::internal::EnableIf<fmt::internal::Not<fmt::internal::ConvertToInt<T>::value>::value, int>::type) [with T = messaging::TestSerializableObject]’
./libraries/spdlog/fmt/bundled/format.h:2465:12: required from ‘static fmt::internal::Value fmt::internal::ArgArray<N, true>::make(const T&) [with Formatter = fmt::BasicFormatter<char>; T = messaging::TestSerializableObject; unsigned int N = 1u]’
./libraries/spdlog/fmt/bundled/format.h:2898:5: required from ‘void fmt::BasicWriter<Char>::write(fmt::BasicCStringRef<CharType>, const Args& ...) [with Args = {messaging::TestSerializableObject}; Char = char]’
./libraries/spdlog/details/logger_impl.h:69:9: required from ‘void spdlog::logger::log(spdlog::level::level_enum, const char*, const Args& ...) [with Args = {messaging::TestSerializableObject}]’
./libraries/spdlog/details/logger_impl.h:127:5: required from ‘void spdlog::logger::trace(const char*, const Args& ...) [with Args = {messaging::TestSerializableObject}]’
src/LibraryCollisionTest.cpp:21:53: required from here
./libraries/spdlog/fmt/bundled/format.h:1276:38: warning: invalid application of ‘sizeof’ to a void type [-Wpointer-arith]
enum { enable_conversion = sizeof(convert(get<T>())) == sizeof(Yes) };
The key line is here:
./libraries/mutils/args-finder.hpp: In instantiation of ‘auto mutils::convert(F)
[with F = messaging::TestSerializableObject; ignore = void]’:
./libraries/spdlog/fmt/bundled/format.h:1276:46: required from ‘struct
fmt::internal::ConvertToInt<messaging::TestSerializableObject>’
./libraries/spdlog/fmt/bundled/format.h:1485:5: required by substitution of
‘template<class T> fmt::internal::MakeValue<Formatter>::MakeValue(const
T&, typename fmt::internal::EnableIf<fmt::internal::Not<
fmt::internal::ConvertToInt<T>::value>::value, int>::type) [with T =
messaging::TestSerializableObject]’
Somehow, g++ has jumped from expanding a templated function inside the spdlog library, in namespace fmt::internal, to a function template in the mutils library, in namespace mutils, which is clearly not what the spdlog library intended to do! If I look at line 1276 of format.h, it's the one that calls the "convert" function inside this template struct:
template<typename T>
struct ConvertToInt
{
enum { enable_conversion = sizeof(convert(get<T>())) == sizeof(Yes) };
enum { value = ConvertToIntImpl2<T, enable_conversion>::value };
};
A few lines above, sure enough, is the function "convert":
template <typename T>
T &get();
Yes &convert(fmt::ULongLong);
No &convert(...);
These are all inside the namespace fmt::internal, and my IDE agrees that if I want the definition of the function "convert" on line 1276, I should jump to the function "convert" on line 1248. So why does g++ ignore this definition, and instead try to use the definition for mutils::convert(), which isn't even in the right namespace?
Note that clang also fails to compile this program, and makes the same mistake, so I don't think that this is a bug in g++.
This is definitively a bug in spdlog fmtlib, used internally by spdlog.
The problem is described summarily in this FAQ:
What is “Argument-Dependent Lookup” (aka ADL, or “Koenig Lookup”)?
Because messaging::TestSerializableObject inherits from a type in namespace mutils, when convert is called unqualified from inside namespace fmt::internal with a TestSerializableObject, both fmt::internal::convert and mutils::convert are considered in the overload set. Variadic functions always rank last during overload resolution, so the template argument F in the latter is a better match than the ... in the former and mutils::convert is chosen.
This is in no way specific to your code or to mutils – any type with a unary function or function template named convert in the same namespace or a parent namespace is susceptible to this problem.
The fix is to qualify the convert call and change the definition of fmt::internal::ConvertToInt<T>::enable_conversion from
enum { enable_conversion = sizeof(convert(get<T>())) == sizeof(Yes) };
to
enum { enable_conversion = sizeof(internal::convert(get<T>())) == sizeof(Yes) };
In my own code I make a habit of always qualifying all calls to functions inside any internal/detail namespace, even from code inside that same namespace, unless ADL usage is explicitly intended. (N.b. calls don't need to be fully qualified, merely qualified.) I learned this lesson from watching Boost have to deal with this problem the hard way as C++11 was emerging.

How does make_pair know the types of its args?

The definition for make_pair in the MSVC++ "utility" header is:
template<class _Ty1,
class _Ty2> inline
pair<_Ty1, _Ty2> make_pair(_Ty1 _Val1, _Ty2 _Val2)
{ // return pair composed from arguments
return (pair<_Ty1, _Ty2>(_Val1, _Val2));
}
I use make_pair all the time though without putting the argument types in angle brackets:
map<string,int> theMap ;
theMap.insert( make_pair( "string", 5 ) ) ;
Shouldn't I have to tell make_pair that the first argument is std::string and not char* ?
How does it know?
Function template calls can usually avoid explicit template arguments (ie make_pair<…>) by argument deduction, which is defined by C++03 §14.8.2. Excerpt:
When a function template
specialization is referenced, all of
the template arguments must have
values. The values can be either
explicitly specified or, in some
cases, deduced from the use.
The specific rules are a bit complicated, but typically it "just works" as long as you have only one specialization which is generally qualified enough.
Your example uses two steps of deduction and one implicit conversion.
make_pair returns a pair<char const*, int>,
then template<class U, classV> pair<string,int>::pair( pair<U,V> const & ) kicks in with U = char*, V = int and performs member-wise initialization,
invoking string::string(char*).
It doesn't. make_pair generated a pair<char*,int> (or maybe a pair<char const*,int>).
However, if you'll note in the implementation of pair there's a templated copy constructor:
template < typename Other1, typename Other2 >
pair(pair<Other1,Other2>& other)
: first(other.first), second(other.second)
{}
This may be implemented in slightly different ways but amounts to the same thing. Since this constructor is implicit, the compiler attempts to create pair<std::string,int> out of your pair<char*,int> - since the necessary types are convertible this works.
make_pair() exists precisely so that argument type deduction can be used to determine the template parameter types.
See this SO question: Using free function as pseudo-constructors to exploit template parameter deduction
It relies on the fact that the constructor of std::string accepts a const char*.
It doesn't matter if this constructor of std::string is explicit or not. The template deducts the type and uses the copy constructor of pair to convert it. It also doesn't matter whether or not the pair constructor is explicit.
If you turn the constructor of std::string into:
class string
{
public:
string(char* s)
{
}
};
you get this error:
/usr/include/c++/4.3/bits/stl_pair.h: In constructor ‘std::pair<_T1, _T2>::pair(const std::pair<_U1, _U2>&) [with _U1 = const char*, _U2 = int, _T1 = const string, _T2 = int]’:
foo.cpp:27: instantiated from here
/usr/include/c++/4.3/bits/stl_pair.h:106: error: invalid conversion from ‘const char* const’ to ‘char*’
/usr/include/c++/4.3/bits/stl_pair.h:106: error: initializing argument 1 of ‘string::string(char*)’
The constructor looks like this:
template<class _U1, class _U2>
pair(const pair<_U1, _U2>& __p)
: first(__p.first),
second(__p.second) { }
The copy constructor looks like this:
template<class _U1, class _U2>
pair(const pair<_U1, _U2>& __p)
: first(__p.first),
second(__p.second) { }