Related
The Eigen3 documentation warns against passing Eigen objects by value, but they only refer to objects being used as function arguments.
Suppose I'm using Eigen 3.4.0 and C++20. If I have a struct with an Eigen member, does this mean I can't std::move a pass-by-value in the constructor? Do I need to pass-by-reference and copy the object? Or is this handled somehow by modern move-semantics?
If I can't std::move Eigen objects in a constructor, does this mean I should explicitly delete the move-constructors from my struct?
For example,
#include <utility>
#include <Eigen/Core>
struct Node {
Eigen::Vector3d position;
double temperature;
// is this constructor safe to use?
Node(Eigen::Vector3d position_, const double temperature_)
: position(std::move(position_)), temperature(temperature_) {}
// or must it be this?
Node(const Eigen::Vector3d& position_, const double temperature_)
: position(position_), temperature(temperature_) {}
// also, should move-constructors be explicitly deleted?
Node(Node&&) = delete;
Node& operator=(Node&&) = delete;
};
There is nothing magic about Eigen objects. Fixed sized types such as Vector3d behave like std::array. Dynamic sized types like VectorXd behave like std::vector.
Pass-by-value for a dynamic sized type typically is a mistake because it usually invokes a copy construction which can be very expensive for large matrices. Pass-by-reference (const, lvalue, rvalue) is almost always the right choice [footnote 1].
Pass-by-value for fixed-sized types can be a benefit because the first few arguments are usually passed in registers (depending on the platform). This avoids spilling values to the stack. However, this doesn't work for Eigen. I assume they declare a destructor, even if they don't need one. That turns any pass-by-value into pass-by-reference to a hidden copy. You can see this in godbolt. This seems like a missed optimization in Eigen.
In conclusion: Use pass-by-reference. Move-construction makes sense for dynamic sized eigen arrays. It makes no difference for fixed sized types.
Footnote 1: A rare exception can be cases were you would need to do a copy anyway inside the function.
Since we have move semantics in C++, nowadays it is usual to do
void set_a(A a) { _a = std::move(a); }
The reasoning is that if a is an rvalue, the copy will be elided and there will be just one move.
But what happens if a is an lvalue? It seems there will be a copy construction and then a move assignment (assuming A has a proper move assignment operator). Move assignments can be costly if the object has too many member variables.
On the other hand, if we do
void set_a(const A& a) { _a = a; }
There will be just one copy assignment. Can we say this way is preferred over the pass-by-value idiom if we will pass lvalues?
Expensive-to-move types are rare in modern C++ usage. If you are concerned about the cost of the move, write both overloads:
void set_a(const A& a) { _a = a; }
void set_a(A&& a) { _a = std::move(a); }
or a perfect-forwarding setter:
template <typename T>
void set_a(T&& a) { _a = std::forward<T>(a); }
that will accept lvalues, rvalues, and anything else implicitly convertible to decltype(_a) without requiring extra copies or moves.
Despite requiring an extra move when setting from an lvalue, the idiom is not bad since (a) the vast majority of types provide constant-time moves and (b) copy-and-swap provides exception safety and near-optimal performance in a single line of code.
But what happens if a is an lvalue? It seems there will be a copy
construction and then a move assignment (assuming A has a proper move
assignment operator). Move assignments can be costly if the object has
too many member variables.
Problem well spotted. I wouldn't go as far as to say that the pass-by-value-and-then-move construct is a bad idiom but it definitely has its potential pitfalls.
If your type is expensive to move and / or moving it is essentially just a copy, then the pass-by-value approach is suboptimal. Examples of such types would include types with a fixed size array as a member: It may be relatively expensive to move and a move is just a copy. See also
Small String Optimization and Move Operations and
"Want speed? Measure." (by Howard Hinnant)
in this context.
The pass-by-value approach has the advantage that you only need to maintain one function but you pay for this with performance. It depends on your application whether this maintenance advantage outweighs the loss in performance.
The pass by lvalue and rvalue reference approach can lead to maintenance headaches quickly if you have multiple arguments. Consider this:
#include <vector>
using namespace std;
struct A { vector<int> v; };
struct B { vector<int> v; };
struct C {
A a;
B b;
C(const A& a, const B& b) : a(a), b(b) { }
C(const A& a, B&& b) : a(a), b(move(b)) { }
C( A&& a, const B& b) : a(move(a)), b(b) { }
C( A&& a, B&& b) : a(move(a)), b(move(b)) { }
};
If you have multiple arguments, you will have a permutation problem. In this very simple example, it is probably still not that bad to maintain these 4 constructors. However, already in this simple case, I would seriously consider using the pass-by-value approach with a single function
C(A a, B b) : a(move(a)), b(move(b)) { }
instead of the above 4 constructors.
So long story short, neither approach is without drawbacks. Make your decisions based on actual profiling information, instead of optimizing prematurely.
The current answers are quite incomplete. Instead, I will try to conclude based on the lists of pros and cons I find.
Short answer
In short, it may be OK, but sometimes bad.
This idiom, namely the unifying interface, has better clarity (both in conceptual design and implementation) compared to forwarding templates or different overloads. It is sometimes used with copy-and-swap (actually, as well as move-and-swap in this case).
Detailed analysis
The pros are:
It needs only one function for each parameter list.
It needs indeed only one, not multiple ordinary overloads (or even 2n overloads when you have n parameters when each one can be unqualified or const-qualified).
Like within a forwarding template, parameters passed by value are compatible with not only const, but volatile, which reduce even more ordinary overloads.
Combined with the bullet above, you don't need 4n overloads to serve to {unqulified, const, const, const volatile} combinations for n parameters.
Compared to a forwarding template, it can be a non-templated function as long as the parameters are not needed to be generic (parameterized through template type parameters). This allows out-of-line definitions instead of template definitions needed to be instantiated for each instance in each translation unit, which can make significant improvement to translation-time performance (typically, during both compiling and linking).
It also makes other overloads (if any) easier to implement.
If you have a forwarding template for a parameter object type T, it may still clash with overloads having a parameter const T& in the same position, because the argument can be a lvalue of type T and the template instantiated with type T& (rather than const T&) for it can be more preferred by the overloading rule when there is no other way to differentiate which is the best overloading candidate. This inconsistency may be quite surprising.
In particular, consider you have forwarding template constructor with one parameter of type P&& in a class C. How many time will you forget to excluded the instance of P&& away from possibly cv-qualified C by SFINAE (e.g. by adding typename = enable_if_t<!is_same<C, decay_t<P>> to the template-parameter-list), to ensure it does not clash with copy/move constructors (even when the latter are explicitly user-provided)?
Since the parameter is passed by value of a non-reference type, it can force the argument be passed as a prvalue. This can make a difference when the argument is of a class literal type. Consider there is such a class with a static constexpr data member declared in some class without an out-of-class definition, when it is used as an argument to a parameter of lvalue reference type, it may eventually fail to link, because it is odr-used and there is no definition of it.
Note since ISO C++ 17 the rules of static constexpr data member have changed to introduce a definition implicitly, so the difference is not significant in this case.
The cons are:
A unifying interface can not replace copy and move constructors where the parameter object type is identical to the class. Otherwise, copy-initialization of the parameter would be infinite recursion, because it will call the unifying constructor, and the constructor then call itself.
As mentioned by other answers, if the cost of copy is not ignorable (cheap and predictable enough), this means you will almost always have the degeneration of performance in the calls when the copy is not needed, because copy-initialization of a unifying passed-by-value parameter unconditionally introduce a copy (either copied-to or moved-to) of the argument unless elided.
Even with mandatory elision since C++17, copy-initialization of a parameter object is still hardly free to be removed away - unless the implementation try very hard to prove the behavior not changed according to as-if rules instead of the dedicated copy elision rules applicable here, which might be sometimes impossible without a whole program analysis.
Likewise, the cost of destruction may not be ignorable as well, particularly when non-trivial subobjects are taken into account (e.g. in cases of containers). The difference is that, it does not only apply to the copy-initialization introduced by the copy construction, but also by the move construction. Making move cheaper than copy in constructors can not improve the situation. The more cost of copy-initialization, the more cost of destruction you have to afford.
A minor shortcoming is that there is no way to tweak the interface in different ways as plural overloads, for example, specifying different noexcept-specifiers for parameters of const& and && qualified types.
OTOH, in this example, unifying interface will usually provide you with noexcept(false) copy + noexcept move if you specifies noexcept, or always noexcept(false) when you specify nothing (or explicit noexcept(false)). (Note in the former case, noexcept does not prevent throwing during copy because that will only occur during evaluation of arguments, which is out of the function body.) There is no further chance to tune them separately.
This is considered minor because it is not frequently needed in reality.
Even if such overloads are used, they are probably confusing by nature: different specifiers may hide subtle but important behavioral differences which are difficult to reason about. Why not different names instead of overloads?
Note the example of noexcept may be particularly problematic since C++17 because noexcept-specification now affect the function type. (Some unexpected compatibility issues can be diagnosed by Clang++ warning.)
Sometimes the unconditional copy is actually useful. Because composition of operations with strong-exception guarantee does not hold the guarantee in nature, a copy can be used as a transactional state holder when the strong-exception guarantee is required and the operation cannot be broken down as sequence of operations with no less strict (no-exception or strong) exception guarantee. (This includes the copy-and-swap idiom, although assignments are not recommended to be unified for other reasons in general, see below.) However, this does not mean the copy is otherwise unacceptable. If the intention of the interface is always to create some object of type T, and the cost of moving T is ignorable, the copy can be moved to the target without unwanted overhead.
Conclusions
So for some given operations, here are suggestions about whether using a unifying interface to replace them:
If not all of the parameter types match the unifying interface, or if there is behavioral difference other than the cost of new copies among operations being unified, there cannot be a unifying interface.
If the following conditions are failed to be fit for all parameters, there cannot be a unifying interface. (But it can still be broken down to different named-functions, delegating one call to another.)
For any parameter of type T, if a copy of each argument is needed for all operations, use unifying.
If both copy and move construction of T have ignorable cost, use unifying.
If the intention of the interface is always to create some object of type T, and the cost of the move construction of T is ignorable, use unifying.
Otherwise, avoid unifying.
Here are some examples need to avoid unifying:
Assignment operations (including assignment to the subobjects thereof, typically with copy-and-swap idiom) for T without ignorable cost in copy and move constructions does not meet the criteria of unifying, because the intention of assignment is not to create (but to replace the content of) the object. The copied object will eventually be destructed, which incurs unnecessary overhead. This is even more obvious for cases of self-assignment.
Insertion of values to a container does not meet the criteria, unless both the copy-initialization and destruction have ignorable cost. If the operation fails (due to the allocation failure, duplicate values or so on) after copy-initialization, the parameters have to be destructed, which incurs unnecessary overhead.
Conditionally creation of object based on parameters will incur the overhead when it does not actually create the object (e.g. std::map::insert_or_assign-like container insertion even in spite of the failure above).
Note the accurate limit of "ignorable" cost is somewhat subjective because it eventually depends on how much cost can be tolerated by the developers and/or the users, and it may vary case by case.
Practically, I (conservatively) assume any trivially copyable and trivailly destructible type whose size is not more than one machine word (like a pointer) qualifying the criteria of ignorable cost in general - if the resulted code actually cost too much in such case, it suggests either a wrong configuration of the build tool is used, or the toolchain is not ready for production.
Do profile if there is any further doubt on performance.
Additional case study
There are some other well-known types preferred to be passed by value or not, depending on the conventions:
Types need to preserve reference values by convention should not be passed by value.
A canonical example is the argument forwarding call wrapper defined in ISO C++, which requires to forward references. Note in the caller position it may also preserve the reference respecting to the ref-qualifier.
An instance of this example is std::bind. See also the resolution of LWG 817.
Some generic code may directly copy some parameters. It may be even without std::move, because the cost of the copy is assumed to be ignorable and a move does not necessarily make it better.
Such parameters include iterators and function objects (except the case of argument forwarding caller wrappers discussed above).
Note the constructor template of std::function (but not the assignment operator template) also uses the pass-by-value functor parameter.
Types presumably having the cost comparable to pass-by-value parameter types with ignorable cost are also preferred to be pass-by-value. (Sometimes they are used as dedicated alternatives.) For example, instances of std::initializer_list and std::basic_string_view are more or less two pointers or a pointer plus a size. This fact makes them cheap enough to be directly passed without using references.
Some types should be better avoided passed by value unless you do need a copy. There are different reasons.
Avoid copy by default, because the copy may be quite expensive, or at least it is not easy to guarantee the copy is cheap without some inspection of the runtime properties of the value being copied. Containers are typical examples in this sort.
Without statically knowing how many elements in a container, it is generally not safe (in the sense of a DoS attack, for example) to be copied.
A nested container (of other containers) will easily make the performance problem of copying worse.
Even empty containers are not guaranteed cheap to be copied. (Strictly speaking, this depends on the concrete implementation of the container, e.g. the existence of the "sentinel" element for some node-based containers... But no, keep it simple, just avoid copying by default.)
Avoid copy by default, even when the performance is totally uninterested, because there can be some unexpected side effects.
In particular, allocator-awared containers and some other types with similar treatment to allocators ("container semantics", in David Krauss' word), should not be passed by value - allocator propagation is just another big semantic worm can.
A few other types conventionally depend. For example, see GotW #91 for shared_ptr instances. (However, not all smart pointers are like that; observer_ptr are more like raw pointers.)
For the general case where the value will be stored, the pass-by-value only is a good compromise-
For the case where you know that only lvalues will be passed (some tightly coupled code) it's unreasonable, unsmart.
For the case where one suspects a speed improvement by providing both, first THINK TWICE, and if that didn't help, MEASURE.
Where the value will not be stored I prefer the pass by reference, because that prevents umpteen needless copy operations.
Finally, if programming could be reduced to unthinking application of rules, we could leave it to robots. So IMHO it's not a good idea to focus so much on rules. Better to focus on what the advantages and costs are, for different situations. Costs include not only speed, but also e.g. code size and clarity. Rules can't generally handle such conflicts of interest.
Pass by value, then move is actually a good idiom for objects that you know are movable.
As you mentioned, if an rvalue is passed, it'll either elide the copy, or be moved, then within the constructor it will be moved.
You could overload the copy constructor and move constructor explicitly, however it gets more complicated if you have more than one parameter.
Consider the example,
class Obj {
public:
Obj(std::vector<int> x, std::vector<int> y)
: X(std::move(x)), Y(std::move(y)) {}
private:
/* Our internal data. */
std::vector<int> X, Y;
}; // Obj
Suppose if you wanted to provide explicit versions, you end up with 4 constructors like so:
class Obj {
public:
Obj(std::vector<int> &&x, std::vector<int> &&y)
: X(std::move(x)), Y(std::move(y)) {}
Obj(std::vector<int> &&x, const std::vector<int> &y)
: X(std::move(x)), Y(y) {}
Obj(const std::vector<int> &x, std::vector<int> &&y)
: X(x), Y(std::move(y)) {}
Obj(const std::vector<int> &x, const std::vector<int> &y)
: X(x), Y(y) {}
private:
/* Our internal data. */
std::vector<int> X, Y;
}; // Obj
As you can see, as you increase the number of parameters, the number of necessary constructors grow in permutations.
If you don't have a concrete type but have a templatized constructor, you can use perfect-forwarding like so:
class Obj {
public:
template <typename T, typename U>
Obj(T &&x, U &&y)
: X(std::forward<T>(x)), Y(std::forward<U>(y)) {}
private:
std::vector<int> X, Y;
}; // Obj
References:
Want Speed? Pass by Value
C++ Seasoning
I am answering myself because I will try to summarize some of the answers. How many moves/copies do we have in each case?
(A) Pass by value and move assignment construct, passing a X parameter. If X is a...
Temporary: 1 move (the copy is elided)
Lvalue: 1 copy 1 move
std::move(lvalue): 2 moves
(B) Pass by reference and copy assignment usual (pre C++11) construct. If X is a...
Temporary: 1 copy
Lvalue: 1 copy
std::move(lvalue): 1 copy
We can assume the three kinds of parameters are equally probable. So every 3 calls we have (A) 4 moves and 1 copy, or (B) 3 copies. I.e., in average, (A) 1.33 moves and 0.33 copies per call or (B) 1 copy per call.
If we come to a situation when our classes consist mostly of PODs, moves are as expensive as copies. So we would have 1.66 copies (or moves) per call to the setter in case (A) and 1 copies in case (B).
We can say that in some circumstances (PODs based types), the pass-by-value-and-then-move construct is a very bad idea. It is 66% slower and it depends on a C++11 feature.
On the other hand, if our classes include containers (which make use of dynamic memory), (A) should be much faster (except if we mostly pass lvalues).
Please, correct me if I'm wrong.
Readability in the declaration:
void foo1( A a ); // easy to read, but unless you see the implementation
// you don't know for sure if a std::move() is used.
void foo2( const A & a ); // longer declaration, but the interface shows
// that no copy is required on calling foo().
Performance:
A a;
foo1( a ); // copy + move
foo2( a ); // pass by reference + copy
Responsibilities:
A a;
foo1( a ); // caller copies, foo1 moves
foo2( a ); // foo2 copies
For typical inline code there is usually no difference when optimized.
But foo2() might do the copy only on certain conditions (e.g. insert into map if key does not exist), whereas for foo1() the copy will always be done.
Imagine we have a trivially-copyable type:
struct Trivial
{
float A{};
int B{};
}
which gets constructed and stored in an std::vector:
class ClientCode
{
std::vector<Trivial> storage{};
...
void some_function()
{
...
Trivial t{};
fill_trivial_from_some_api(t, other_args);
storage.push_back(std::move(t)); // Redundant std::move.
...
}
}
Normally, this is a pointless operation, as the object will be copied anyway.
However, an advantage of keeping the std::move call is that if the Trivial type would be changed to no longer be trivially-copyable, the client code will not silently perform an extra copy operation, but a more appropriate move. (The situation is quite possible in my scenario, where the trivial type is used for managing external resources.)
So my question is whether there any technical downsides to applying the redundant std::move?
However, an advantage of keeping the std::move call is that if the Trivial type would be changed to no longer be trivially-copyable, the client code will not silently perform an extra copy operation, but a more appropriate move.
This is correct and something you should think about.
So my question is whether there any technical downsides to applying the redundant std::move?
Depends on where the moved object is being consumed. In the case of push_back, everything is fine, as push_back has both const T& and T&& overloads that behave intuitively.
Imagine another function that had a T&& overload that has completely different behavior from const T&: the semantics of your code will change with std::move.
Since we have move semantics in C++, nowadays it is usual to do
void set_a(A a) { _a = std::move(a); }
The reasoning is that if a is an rvalue, the copy will be elided and there will be just one move.
But what happens if a is an lvalue? It seems there will be a copy construction and then a move assignment (assuming A has a proper move assignment operator). Move assignments can be costly if the object has too many member variables.
On the other hand, if we do
void set_a(const A& a) { _a = a; }
There will be just one copy assignment. Can we say this way is preferred over the pass-by-value idiom if we will pass lvalues?
Expensive-to-move types are rare in modern C++ usage. If you are concerned about the cost of the move, write both overloads:
void set_a(const A& a) { _a = a; }
void set_a(A&& a) { _a = std::move(a); }
or a perfect-forwarding setter:
template <typename T>
void set_a(T&& a) { _a = std::forward<T>(a); }
that will accept lvalues, rvalues, and anything else implicitly convertible to decltype(_a) without requiring extra copies or moves.
Despite requiring an extra move when setting from an lvalue, the idiom is not bad since (a) the vast majority of types provide constant-time moves and (b) copy-and-swap provides exception safety and near-optimal performance in a single line of code.
But what happens if a is an lvalue? It seems there will be a copy
construction and then a move assignment (assuming A has a proper move
assignment operator). Move assignments can be costly if the object has
too many member variables.
Problem well spotted. I wouldn't go as far as to say that the pass-by-value-and-then-move construct is a bad idiom but it definitely has its potential pitfalls.
If your type is expensive to move and / or moving it is essentially just a copy, then the pass-by-value approach is suboptimal. Examples of such types would include types with a fixed size array as a member: It may be relatively expensive to move and a move is just a copy. See also
Small String Optimization and Move Operations and
"Want speed? Measure." (by Howard Hinnant)
in this context.
The pass-by-value approach has the advantage that you only need to maintain one function but you pay for this with performance. It depends on your application whether this maintenance advantage outweighs the loss in performance.
The pass by lvalue and rvalue reference approach can lead to maintenance headaches quickly if you have multiple arguments. Consider this:
#include <vector>
using namespace std;
struct A { vector<int> v; };
struct B { vector<int> v; };
struct C {
A a;
B b;
C(const A& a, const B& b) : a(a), b(b) { }
C(const A& a, B&& b) : a(a), b(move(b)) { }
C( A&& a, const B& b) : a(move(a)), b(b) { }
C( A&& a, B&& b) : a(move(a)), b(move(b)) { }
};
If you have multiple arguments, you will have a permutation problem. In this very simple example, it is probably still not that bad to maintain these 4 constructors. However, already in this simple case, I would seriously consider using the pass-by-value approach with a single function
C(A a, B b) : a(move(a)), b(move(b)) { }
instead of the above 4 constructors.
So long story short, neither approach is without drawbacks. Make your decisions based on actual profiling information, instead of optimizing prematurely.
The current answers are quite incomplete. Instead, I will try to conclude based on the lists of pros and cons I find.
Short answer
In short, it may be OK, but sometimes bad.
This idiom, namely the unifying interface, has better clarity (both in conceptual design and implementation) compared to forwarding templates or different overloads. It is sometimes used with copy-and-swap (actually, as well as move-and-swap in this case).
Detailed analysis
The pros are:
It needs only one function for each parameter list.
It needs indeed only one, not multiple ordinary overloads (or even 2n overloads when you have n parameters when each one can be unqualified or const-qualified).
Like within a forwarding template, parameters passed by value are compatible with not only const, but volatile, which reduce even more ordinary overloads.
Combined with the bullet above, you don't need 4n overloads to serve to {unqulified, const, const, const volatile} combinations for n parameters.
Compared to a forwarding template, it can be a non-templated function as long as the parameters are not needed to be generic (parameterized through template type parameters). This allows out-of-line definitions instead of template definitions needed to be instantiated for each instance in each translation unit, which can make significant improvement to translation-time performance (typically, during both compiling and linking).
It also makes other overloads (if any) easier to implement.
If you have a forwarding template for a parameter object type T, it may still clash with overloads having a parameter const T& in the same position, because the argument can be a lvalue of type T and the template instantiated with type T& (rather than const T&) for it can be more preferred by the overloading rule when there is no other way to differentiate which is the best overloading candidate. This inconsistency may be quite surprising.
In particular, consider you have forwarding template constructor with one parameter of type P&& in a class C. How many time will you forget to excluded the instance of P&& away from possibly cv-qualified C by SFINAE (e.g. by adding typename = enable_if_t<!is_same<C, decay_t<P>> to the template-parameter-list), to ensure it does not clash with copy/move constructors (even when the latter are explicitly user-provided)?
Since the parameter is passed by value of a non-reference type, it can force the argument be passed as a prvalue. This can make a difference when the argument is of a class literal type. Consider there is such a class with a static constexpr data member declared in some class without an out-of-class definition, when it is used as an argument to a parameter of lvalue reference type, it may eventually fail to link, because it is odr-used and there is no definition of it.
Note since ISO C++ 17 the rules of static constexpr data member have changed to introduce a definition implicitly, so the difference is not significant in this case.
The cons are:
A unifying interface can not replace copy and move constructors where the parameter object type is identical to the class. Otherwise, copy-initialization of the parameter would be infinite recursion, because it will call the unifying constructor, and the constructor then call itself.
As mentioned by other answers, if the cost of copy is not ignorable (cheap and predictable enough), this means you will almost always have the degeneration of performance in the calls when the copy is not needed, because copy-initialization of a unifying passed-by-value parameter unconditionally introduce a copy (either copied-to or moved-to) of the argument unless elided.
Even with mandatory elision since C++17, copy-initialization of a parameter object is still hardly free to be removed away - unless the implementation try very hard to prove the behavior not changed according to as-if rules instead of the dedicated copy elision rules applicable here, which might be sometimes impossible without a whole program analysis.
Likewise, the cost of destruction may not be ignorable as well, particularly when non-trivial subobjects are taken into account (e.g. in cases of containers). The difference is that, it does not only apply to the copy-initialization introduced by the copy construction, but also by the move construction. Making move cheaper than copy in constructors can not improve the situation. The more cost of copy-initialization, the more cost of destruction you have to afford.
A minor shortcoming is that there is no way to tweak the interface in different ways as plural overloads, for example, specifying different noexcept-specifiers for parameters of const& and && qualified types.
OTOH, in this example, unifying interface will usually provide you with noexcept(false) copy + noexcept move if you specifies noexcept, or always noexcept(false) when you specify nothing (or explicit noexcept(false)). (Note in the former case, noexcept does not prevent throwing during copy because that will only occur during evaluation of arguments, which is out of the function body.) There is no further chance to tune them separately.
This is considered minor because it is not frequently needed in reality.
Even if such overloads are used, they are probably confusing by nature: different specifiers may hide subtle but important behavioral differences which are difficult to reason about. Why not different names instead of overloads?
Note the example of noexcept may be particularly problematic since C++17 because noexcept-specification now affect the function type. (Some unexpected compatibility issues can be diagnosed by Clang++ warning.)
Sometimes the unconditional copy is actually useful. Because composition of operations with strong-exception guarantee does not hold the guarantee in nature, a copy can be used as a transactional state holder when the strong-exception guarantee is required and the operation cannot be broken down as sequence of operations with no less strict (no-exception or strong) exception guarantee. (This includes the copy-and-swap idiom, although assignments are not recommended to be unified for other reasons in general, see below.) However, this does not mean the copy is otherwise unacceptable. If the intention of the interface is always to create some object of type T, and the cost of moving T is ignorable, the copy can be moved to the target without unwanted overhead.
Conclusions
So for some given operations, here are suggestions about whether using a unifying interface to replace them:
If not all of the parameter types match the unifying interface, or if there is behavioral difference other than the cost of new copies among operations being unified, there cannot be a unifying interface.
If the following conditions are failed to be fit for all parameters, there cannot be a unifying interface. (But it can still be broken down to different named-functions, delegating one call to another.)
For any parameter of type T, if a copy of each argument is needed for all operations, use unifying.
If both copy and move construction of T have ignorable cost, use unifying.
If the intention of the interface is always to create some object of type T, and the cost of the move construction of T is ignorable, use unifying.
Otherwise, avoid unifying.
Here are some examples need to avoid unifying:
Assignment operations (including assignment to the subobjects thereof, typically with copy-and-swap idiom) for T without ignorable cost in copy and move constructions does not meet the criteria of unifying, because the intention of assignment is not to create (but to replace the content of) the object. The copied object will eventually be destructed, which incurs unnecessary overhead. This is even more obvious for cases of self-assignment.
Insertion of values to a container does not meet the criteria, unless both the copy-initialization and destruction have ignorable cost. If the operation fails (due to the allocation failure, duplicate values or so on) after copy-initialization, the parameters have to be destructed, which incurs unnecessary overhead.
Conditionally creation of object based on parameters will incur the overhead when it does not actually create the object (e.g. std::map::insert_or_assign-like container insertion even in spite of the failure above).
Note the accurate limit of "ignorable" cost is somewhat subjective because it eventually depends on how much cost can be tolerated by the developers and/or the users, and it may vary case by case.
Practically, I (conservatively) assume any trivially copyable and trivailly destructible type whose size is not more than one machine word (like a pointer) qualifying the criteria of ignorable cost in general - if the resulted code actually cost too much in such case, it suggests either a wrong configuration of the build tool is used, or the toolchain is not ready for production.
Do profile if there is any further doubt on performance.
Additional case study
There are some other well-known types preferred to be passed by value or not, depending on the conventions:
Types need to preserve reference values by convention should not be passed by value.
A canonical example is the argument forwarding call wrapper defined in ISO C++, which requires to forward references. Note in the caller position it may also preserve the reference respecting to the ref-qualifier.
An instance of this example is std::bind. See also the resolution of LWG 817.
Some generic code may directly copy some parameters. It may be even without std::move, because the cost of the copy is assumed to be ignorable and a move does not necessarily make it better.
Such parameters include iterators and function objects (except the case of argument forwarding caller wrappers discussed above).
Note the constructor template of std::function (but not the assignment operator template) also uses the pass-by-value functor parameter.
Types presumably having the cost comparable to pass-by-value parameter types with ignorable cost are also preferred to be pass-by-value. (Sometimes they are used as dedicated alternatives.) For example, instances of std::initializer_list and std::basic_string_view are more or less two pointers or a pointer plus a size. This fact makes them cheap enough to be directly passed without using references.
Some types should be better avoided passed by value unless you do need a copy. There are different reasons.
Avoid copy by default, because the copy may be quite expensive, or at least it is not easy to guarantee the copy is cheap without some inspection of the runtime properties of the value being copied. Containers are typical examples in this sort.
Without statically knowing how many elements in a container, it is generally not safe (in the sense of a DoS attack, for example) to be copied.
A nested container (of other containers) will easily make the performance problem of copying worse.
Even empty containers are not guaranteed cheap to be copied. (Strictly speaking, this depends on the concrete implementation of the container, e.g. the existence of the "sentinel" element for some node-based containers... But no, keep it simple, just avoid copying by default.)
Avoid copy by default, even when the performance is totally uninterested, because there can be some unexpected side effects.
In particular, allocator-awared containers and some other types with similar treatment to allocators ("container semantics", in David Krauss' word), should not be passed by value - allocator propagation is just another big semantic worm can.
A few other types conventionally depend. For example, see GotW #91 for shared_ptr instances. (However, not all smart pointers are like that; observer_ptr are more like raw pointers.)
For the general case where the value will be stored, the pass-by-value only is a good compromise-
For the case where you know that only lvalues will be passed (some tightly coupled code) it's unreasonable, unsmart.
For the case where one suspects a speed improvement by providing both, first THINK TWICE, and if that didn't help, MEASURE.
Where the value will not be stored I prefer the pass by reference, because that prevents umpteen needless copy operations.
Finally, if programming could be reduced to unthinking application of rules, we could leave it to robots. So IMHO it's not a good idea to focus so much on rules. Better to focus on what the advantages and costs are, for different situations. Costs include not only speed, but also e.g. code size and clarity. Rules can't generally handle such conflicts of interest.
Pass by value, then move is actually a good idiom for objects that you know are movable.
As you mentioned, if an rvalue is passed, it'll either elide the copy, or be moved, then within the constructor it will be moved.
You could overload the copy constructor and move constructor explicitly, however it gets more complicated if you have more than one parameter.
Consider the example,
class Obj {
public:
Obj(std::vector<int> x, std::vector<int> y)
: X(std::move(x)), Y(std::move(y)) {}
private:
/* Our internal data. */
std::vector<int> X, Y;
}; // Obj
Suppose if you wanted to provide explicit versions, you end up with 4 constructors like so:
class Obj {
public:
Obj(std::vector<int> &&x, std::vector<int> &&y)
: X(std::move(x)), Y(std::move(y)) {}
Obj(std::vector<int> &&x, const std::vector<int> &y)
: X(std::move(x)), Y(y) {}
Obj(const std::vector<int> &x, std::vector<int> &&y)
: X(x), Y(std::move(y)) {}
Obj(const std::vector<int> &x, const std::vector<int> &y)
: X(x), Y(y) {}
private:
/* Our internal data. */
std::vector<int> X, Y;
}; // Obj
As you can see, as you increase the number of parameters, the number of necessary constructors grow in permutations.
If you don't have a concrete type but have a templatized constructor, you can use perfect-forwarding like so:
class Obj {
public:
template <typename T, typename U>
Obj(T &&x, U &&y)
: X(std::forward<T>(x)), Y(std::forward<U>(y)) {}
private:
std::vector<int> X, Y;
}; // Obj
References:
Want Speed? Pass by Value
C++ Seasoning
I am answering myself because I will try to summarize some of the answers. How many moves/copies do we have in each case?
(A) Pass by value and move assignment construct, passing a X parameter. If X is a...
Temporary: 1 move (the copy is elided)
Lvalue: 1 copy 1 move
std::move(lvalue): 2 moves
(B) Pass by reference and copy assignment usual (pre C++11) construct. If X is a...
Temporary: 1 copy
Lvalue: 1 copy
std::move(lvalue): 1 copy
We can assume the three kinds of parameters are equally probable. So every 3 calls we have (A) 4 moves and 1 copy, or (B) 3 copies. I.e., in average, (A) 1.33 moves and 0.33 copies per call or (B) 1 copy per call.
If we come to a situation when our classes consist mostly of PODs, moves are as expensive as copies. So we would have 1.66 copies (or moves) per call to the setter in case (A) and 1 copies in case (B).
We can say that in some circumstances (PODs based types), the pass-by-value-and-then-move construct is a very bad idea. It is 66% slower and it depends on a C++11 feature.
On the other hand, if our classes include containers (which make use of dynamic memory), (A) should be much faster (except if we mostly pass lvalues).
Please, correct me if I'm wrong.
Readability in the declaration:
void foo1( A a ); // easy to read, but unless you see the implementation
// you don't know for sure if a std::move() is used.
void foo2( const A & a ); // longer declaration, but the interface shows
// that no copy is required on calling foo().
Performance:
A a;
foo1( a ); // copy + move
foo2( a ); // pass by reference + copy
Responsibilities:
A a;
foo1( a ); // caller copies, foo1 moves
foo2( a ); // foo2 copies
For typical inline code there is usually no difference when optimized.
But foo2() might do the copy only on certain conditions (e.g. insert into map if key does not exist), whereas for foo1() the copy will always be done.
It seems to me there should be four variants of boost::optional
optional<Foo> => holds a mutable Foo and can be reassigned after initialization
optional<Foo const> const => holds a const Foo and can't be reassigned after initialization
optional<Foo> const => (should?) hold a mutable Foo but can't be reassigned after initialization
optional<Foo const> => (should?) hold a const Foo and can be reassigned after initialization
The first 2 cases work as expected. But the optional<Foo> const dereferences to a const Foo, and the optional<Foo const> doesn't allow reassignment after initialization (as touched upon in this question).
The reassignment of the const value types is specifically what I ran into, and the error is:
/usr/include/boost/optional/optional.hpp:486: error: passing ‘const Foo’ as ‘this’ argument of ‘Foo& Foo::operator=(const Foo&)’ discards qualifiers [-fpermissive]
And it happens here:
void assign_value(argument_type val,is_not_reference_tag) { get_impl() = val; }
After construction, the implementation uses the assignment operator for the type you parameterized the optional with. It obviously doesn't want a left-hand operand which is a const value. But why shouldn't you be able to reset a non-const optional to a new const value, such as in this case:
optional<Foo const> theFoo (maybeGetFoo());
while (someCondition) {
// do some work involving calling some methods on theFoo
// ...but they should only be const ones
theFoo = maybeGetFoo();
}
Some Questions:
Am I right that wanting this is conceptually fine, and not being able to do it is just a fluke in the implementation?
If I don't edit the boost sources, what would be a clean way to implement logic like in the loop above without scrapping boost::optional altogether?
If this does make sense and I were to edit the boost::optional source (which I've already had to do to make it support movable types, though I suspect they'll be doing that themselves soon) then what minimally invasive changes might do the trick?
So basically the problem seems to be related to this note in the documentation for optional& optional<T (not a ref)>::operator= ( T const& rhs ):
Notes: If *this was initialized, T's assignment operator is used, otherwise, its copy-constructor is used.
That is, suppose you have boost::optional<const Foo> theFoo;. Since a default constructed boost::optional<> is empty, the statement:
theFoo=defaultFoo;
should mean "copy construct defaultFoo into theFoos internal storage." Since there's nothing already in that internal storage, this makes sense, even if the internal storage is supposed to house a const Foo. Once finished, theFoo will not be empty.
Once theFoo contains a value, the statement
theFoo=defaultFoo;
should mean "assign defaultFoo into the object in theFoos internal storage." But theFoos internal storage isn't assignable (as it is const), and so this should raise a (compile time?) error.
Unfortunately, you'll notice the last two statements are identical, but conceptually require different compile time behavior. There's nothing to let the compiler tell the difference between the two, though.
Particularly in the scenario you're describing, it might make more sense to define boost::optional<...>'s assignment operator to instead have the semantics:
If *this was initialized, its current contents are first destroyed. Then T's copy-constructor is used.
After all, it's entirely possible to invoke T's assignment operator if that's what you really want to do, by saying *theFoo = rhs.
(1) One's opinion on what the behavior "should" be depends on whether optionals are "a container for zero or one objects of an arbitrary type" or "a thin proxy for a type, with an added feature". The existing code uses the latter idea, and by doing so, it removes half of the "four different behaviors" in the list. This reduces the complexity, and keeps you from unintentionally introducing inefficient usages.
(2) For any Foo type whose values are copyable, one can easily switch between mutable and immutable optionals by making a new one. So in the given case, you'd get it as mutable briefly and then copy it into an immutable value.
optional<Foo> theMutableFoo (maybeGetFoo());
while (someCondition) {
optional<Foo const> theFoo (theMutableFoo);
// do some work involving calling some methods on theFoo
// ...but they should only be const ones
// ...therefore, just don't use theMutableFoo in here!
theMutableFoo = maybeGetFoo();
}
Given the model that it's a "thin proxy" for a type, this is exactly the same kind of thing you would have to do if the type were not wrapped in an optional. An ordinary const value type needs the same treatment in such situations.
(3) One would have to follow up on the information given by #Andrzej to find out. But such an implementation change would probably not perform better than creating a new optional every time (as in the loop above). It's probably best to accept the existing design.
Sources: #R.MartinhoFernandez, #KerrekSB
To answer your three questions:
The implementation follows the design philosophy that optional's assignment uses T's assignment; and in this sense, the implementation is fine.
optional was designed with possible extensions in mind. Optional is only an interface for the underlying class optional_base. Instead of using optional you can derive your own class from optional_base. optional_base has a protected member construct, which does almost what you need. You will need a new member, say reset(T) that firsts clears the optional_base and then calls construct().
Alternatively, you can add member reset(T) to optional. This would be the least intrusive change.
You could also try the reference implementation of optional from this proposal.