I have what is essentially class containing a std::map where the values are shared_ptrs wrapping a container which holds different types. Skeleton code follows:
// Just a basic example class
class MyClass {
public:
explicit MyClass(int i) : mI(i) {}
bool operator==(const MyClass& rhs) { return mI == rhs.mI; }
private:
int mI;
};
// A class into which key value pairs can be added where the value
// can be of a different type.
class MultipleTypeMap {
public:
template <typename T>
void AddObject(const std::string& key, const T object) {
auto ptr = make_shared<B<MyClass>>(std::move(object));
mSharedPtrMap.insert(pair<string, shared_ptr<A>>("key", ptr));
}
// ...
private:
class A {
public:
virtual ~A() = default;
};
template<typename T>
class B : public A {
public:
explicit B(const T& t) : item(t) {}
const T item;
};
map<string, shared_ptr<A>> mSharedPtrMap;
};
int main() {
MyClass m(1);
MultipleTypeMap multiMap;
multiMap.AddObject("test", m);
MyClass n(1);
MultipleTypeMap multiMap2;
multiMap2.AddObject("test", n);
if (multiMap == multiMap2) {
cout << "Equal" << endl;
}
return 0;
}
How should a generic == operator of MultipleTypeMap be written so that it compares the contents of mSharedPtrMap by checking that both the lhs and rhs objects have the same number of keys, the same keys and the same objects where same means that the == operator of the keys / objects evaluates to true?
If you type erase (and later on don't know which type you previously stored), then all the functionality must be provided by the base class interface. So, we need a virtual operator== in A that is implemented in each B.
Here is an implementation:
class MultipleTypeMap {
public:
template <typename T>
void AddObject(const std::string& key, T object) {
auto ptr = std::make_unique<B<T>>(std::move(object));
mMap.emplace(key, std::move(ptr));
}
// ...
bool operator==(const MultipleTypeMap& other) const
{
// Sizes must be equal.
if (mMap.size() != other.mMap.size())
return false;
// Sizes are equal, check keys and values in order.
auto itOther = other.mMap.begin();
for (auto it = mMap.begin(); it != mMap.end(); ++it, ++itOther)
{
if (it->first != itOther->first)
return false;
if (*it->second != *itOther->second)
return false;
}
// No differences found.
return true;
}
bool operator!=(const MultipleTypeMap& rhs) const { return !(*this == rhs); }
private:
class A {
public:
virtual ~A() = default;
virtual bool operator==(const A& other) const = 0;
bool operator!=(const A& other) const { return !(*this == other); }
};
template<typename T>
class B : public A
{
public:
explicit B(const T& t) : item(t) {}
bool operator==(const A& other) const override
{
const B<T>* otherB = dynamic_cast<const B<T>*>(&other);
// If the cast fails, types are different.
if (!otherB)
return false;
// Note: The above is probably slow, consider storing (on construction)
// and checking typeids instead.
// Check item equality.
return item == otherB->item;
}
const T item;
};
std::map<std::string, std::unique_ptr<A>> mMap;
};
Demo with tests
Note: I didn't fix every inconsistency in the original code. (Do you want to move or copy-construct your T? Why store const objects when your MyClass comparison operator is not const?)
Related
I have this piece of code:
class ISerializable
{
public:
virtual bool operator==(const ISerializable* /*value*/) const { return false;};
virtual bool operator!=(const ISerializable* /*value*/) const { return true;};
};
class Point2I : public ISerializable
{
public:
bool operator==(const Point2I& value)
{
return (x == value.x && y == value.y);
}
bool operator!=(const Point2I& value)
{
return !(*this == value);
}
public:
int x;
int y;
};
class Coordinate : public ISerializable
{
public:
virtual bool operator==(const Coordinate& value) const;
virtual bool operator!=(const Coordinate& value) const;
};
It is causing me -Woverloaded-virtual warning on gcc compiler.
I understand this warning due to that function declaration in Point2I hides virtual functions from ISerializable.
But I am not sure if just missing const in Point2I can cause this warning.
Can you please help me understand if it is const which is causing this warning or something else? Warning description from gcc didn't mention anything specifically.
Update:
I found another class Coordinate in my code base which was already overriding this and gcc not throwing warning for this. Only difference in Point2I and Coordinate is I didn't declare it virtual with const in Point2I. It appears just const is hiding base class declaration.
if it is const which is causing this warning or something else?
I'd say that it's something else, namely that you are not actually overriding the base class methods, even if you add const.
The argument const ISerializable* is not the same as const Point2I&.
One solution could be to override the base class methods, using const ISerializable& as the argument, and cast in the overridden methods:
class ISerializable {
public:
// add virtual destructor if you need to delete objects through
// base class pointers later:
virtual ~ISerializable() = default;
virtual bool operator==(const ISerializable&) const { return false; }
virtual bool operator!=(const ISerializable&) const { return true; }
};
class Point2I : public ISerializable {
public:
bool operator==(const ISerializable& value) const override {
auto rhs = dynamic_cast<const Point2I*>(&value);
// rhs will be nullptr if the cast above fails
return rhs && (x == rhs->x && y == rhs->y);
}
bool operator!=(const ISerializable& value) const override {
return !(*this == value);
}
private:
int x = 0;
int y = 0;
};
Example usage:
#include <iostream>
class Foo : public ISerializable { // another ISerializable
public:
};
int main() {
Point2I a, b;
std::cout << (a == b) << '\n'; // true - using Point2I::operator==
Foo f;
std::cout << (a == f) << '\n'; // false - using Point2I::operator==
std::cout << (f == a) << '\n'; // false - using ISerializable::operator==
// this makes the default implementation in ISerializable utterly confusing:
std::cout << (f == f) << '\n'; // false - using ISerializable::operator==
}
Another possible solution could be using CRTP but this would not work if you want to compare different types derived from ISerializable<T>:
template<class T>
class ISerializable {
public:
virtual ~ISerializable() = default;
virtual bool operator==(const T&) const = 0;
virtual bool operator!=(const T&) const = 0;
};
class Point2I : public ISerializable<Point2I> {
public:
bool operator==(const Point2I& value) const override {
return (x == value.x && y == value.y);
}
bool operator!=(const Point2I& value) const override {
return !(*this == value);
}
public:
int x;
int y;
};
There are two problems.
The first one is different types of parameters
In these functions the parameters have the pointer type const ISerializable*
virtual bool operator==(const ISerializable* /*value*/) const { return false;};
virtual bool operator!=(const ISerializable* /*value*/) const { return true;};
and in these functions the parameters have the referenced type const Point2I&
bool operator==(const Point2I& value)
{
return (x == value.x && y == value.y);
}
bool operator!=(const Point2I& value)
{
return !(*this == value);
}
The second one is that the first functions are constant member functions while the second functions are not constant member functions.
I want to create efficient and easy to use value type.
Base of the Value is a boost::variant (and std::variant in the future), but I'm new in it.
And I have some questions:
In the code below, is it necessary to use recursive variant?
Is it possible to not inherit from the boost::variant? Maybe more efficient way exists?
Do you have any comments or suggestions on the code below (it's not fully completed code, but only a draft)?
class Value;
typedef std::string String;
typedef std::vector<char> BinData;
typedef String URL;
typedef unsigned long long UID;
TSW_STRONG_TYPEDEF(std::time_t, Time)
typedef std::vector<Value> ValueArray;
typedef std::vector<String> StringArray;
//typedef std::pair<String, Value> NameValue;
typedef std::list<Value> ValueList;
typedef std::list<String> StringList;
typedef std::map<String, String> StringStringMap;
typedef std::map<String, Value> NameValueMap;
struct monostate
{
monostate() = default;
};
constexpr bool operator<(monostate, monostate) noexcept { return false; }
constexpr bool operator>(monostate, monostate) noexcept { return false; }
constexpr bool operator<=(monostate, monostate) noexcept { return true; }
constexpr bool operator>=(monostate, monostate) noexcept { return true; }
constexpr bool operator==(monostate, monostate) noexcept { return true; }
constexpr bool operator!=(monostate, monostate) noexcept { return false; }
typedef monostate Null;
class Object
{
public:
Object() = delete;
Object(const Object &other) = default;
Object(Object &&other);
Object(const String &name);
Object(String &&name);
Object(const String &name, const NameValueMap &fields);
Object(String &&name, const NameValueMap &fields);
Object(const String &name, NameValueMap &&fields);
Object(String &&name, NameValueMap &&fields);
Object &operator=(const Object &other) = default;
Object &operator=(Object &&other);
public:
const String &get_name() const;
const NameValueMap &get_fields() const;
public:
bool operator<(const Object &other) const noexcept;
bool operator>(const Object &other) const noexcept;
bool operator<=(const Object &other) const noexcept;
bool operator>=(const Object &other) const noexcept;
bool operator==(const Object &other) const noexcept;
bool operator!=(const Object &other) const noexcept;
private:
String name_;
NameValueMap fields_;
};
enum class ValueType
{
Undefined, Null, Array, BinData, Boolean, DoubleNumber, Int64Number, String, Time, Object
};
// Types ordnung need to be same with ValueType ordnung.
/// Base for the Value class
typedef boost::variant<monostate, Null, ValueArray, BinData, bool, double, int64_t, String, Time, Object> ValueBase;
/**
* #brief The Value class, implements common framework value.
*
* This class is a container, which can store multiple values, including Values containers.
*
* #note
* Class based on a variant class. It may be either boost::variant or std::variant in C++17 and higher.
*/
class Value : public ValueBase
{
public:
using ValueBase::ValueBase;
Value() = default;
Value(const String::value_type *v) : ValueBase(String(v)) {}
public:
bool is_array() const { return static_cast<ValueType>(which()) == ValueType::Array; }
bool is_bool() const { return static_cast<ValueType>(which()) == ValueType::Boolean; }
bool is_bindata() const { return static_cast<ValueType>(which()) == ValueType::BinData; }
bool is_double() const { return static_cast<ValueType>(which()) == ValueType::DoubleNumber; }
bool is_int64() const { return static_cast<ValueType>(which()) == ValueType::Int64Number; }
bool is_null() const { return static_cast<ValueType>(which()) == ValueType::Null; }
bool is_object() const { return static_cast<ValueType>(which()) == ValueType::Object; }
bool is_string() const { return static_cast<ValueType>(which()) == ValueType::String; }
bool is_time() const { return static_cast<ValueType>(which()) == ValueType::Time; }
bool is_undefined() const { return static_cast<ValueType>(which()) == ValueType::Undefined; }
public:
bool as_bool() const { return as<bool>(); }
BinData &as_bindata() { return as<BinData>(); }
double as_double() const { return as<double>(); }
int64_t as_int64() const { return as<int64_t>(); }
Object &as_object() { return as<Object>(); }
String &as_string() { return as<String>(); }
Time &as_time() { return as<Time>(); }
ValueArray &as_array() { return as<ValueArray>(); }
public:
ValueType value_type() const { return static_cast<ValueType>(which()); }
public:
template <typename T>
const T& as() const { return boost::get<T>(*this); }
template <typename T>
T& as() { return boost::get<T>(*this); }
template <typename T>
const T& as(const T& default_value) const { return type() == typeid(T) ? boost::get<T>(*this) : default_value; }
template <typename T>
T& as(const T& default_value) { return type() == typeid(T) ? boost::get<T>(*this) : default_value; }
template <typename T> boost::optional<T> as_optional() { return boost::make_optional(type() == typeid(T), as<T>()); }
public:
bool operator==(const ValueBase &other) const { return ValueBase::operator==(other); }
bool operator<(const ValueBase &other) const { return ValueBase::operator<(other); }
bool operator>(const ValueBase &other) const { return !((*this) < other || (*this) == other); }
bool operator<=(const ValueBase &other) const { return ((*this) < other || (*this) == other); }
bool operator>=(const ValueBase &other) const { return !((*this) < other); }
bool operator!=(const ValueBase &other) const { return !((*this) == other); }
private:
// Force compile error, prevent Variant(bool) to be called
Value(void *) = delete;
};
Looks okay to me.
You can do without recursive variants IFF your standard library implementation allows instantiation of the container classes for incomplete types.
I'd note that since everything is tied to the base-class publicly, there is nothing about the implementation that can change without also changing the (binary) interface. Therefore I'd surely implement all the members inline so that the compiler will optimize them even without LTO.
It is not clear to me what to_X members do (possibly just a<X> but possibly something else depending on can_convert()?). If it's just a wrap around as_<> I'd rename them as_X() etc.
You might also want to add optional<>-like members like
template <typename T> T const& get_value_or(T const& default_value) const;
And possibly
template <typename T> optional<T> get() const;
// with boost optional you can prevent a copy²:
template <typename T> optional<T const&> get() const;
This enables code like:
if (auto& s = value.get<String>()) {
std::cout << "The string value is '" << *s << "'\n";
} else {
std::cout << "Value has no string value\n";
}
¹ this is not - yet - standard specified. You can always use Boost Container instead which promises this, as well as non-allocating construction
² just make sure you do not allow the operation on a rvalue, to remove a predictable class of errors, so e.g.
template <typename T> optional<T const&> get()&& = delete;
I just build a mini program to understand how this will work because i need this for something a bit more difficult but i can't make this work.
I think i need to define operator overload but i dont know how because they are two objects of set<set<a>>
If you compile you will see a big error where it notice that he can't compare myset == myset2 and i think it will say same for operator != and =
#include <set>
using namespace std;
class a{
private:
int a_;
public:
int get_a() const{ return a_; }
void set_a(int aux){ a_=aux;}
bool operator < (const a& t) const{
return this->get_a() < t.get_a();
}
};
class b{
private:
set<set<a> > b_;
public:
void set_(set<a> aux){ b_.insert(aux); }
//Overload operators?
};
int main(){
b myset;
b myset2;
set<a> subset1;
set<a> subset2;
a myint;
myint.set_a(1);
subset1.insert(myint);
myint.set_a(2);
subset1.insert(myint);
myint.set_a(3);
subset1.insert(myint);
myint.set_a(5);
subset2.insert(myint);
myint.set_a(6);
subset2.insert(myint);
myint.set_a(7);
subset2.insert(myint);
myset.set_(subset1);
myset.set_(subset2);
myset2.set_(subset1);
myset2.set_(subset2);
if(myset == myset2){
cout << "They are equal" << endl;
}
if(myset != myset2){
cout << "They are different" << endl;
}
b myset3;
myset3 = myset2; //Copy one into other
}
In order for your code to work you need to specify following operators (note: they are not created by default)
class a{
private:
int a_;
public:
int get_a() const{ return a_; }
void set_a(int aux){ a_=aux;}
/* needed for set insertion */
bool operator < (const a& other) const {
return this->get_a() < other.get_a();
}
/* needed for set comparison */
bool operator == (const a& other) const {
return this->get_a() == other.get_a();
}
};
class b{
private:
set<set<a> > b_;
public:
void set_(set<a> aux){ b_.insert(aux); }
/* needed, because myset == myset2 is called later in the code */
bool operator == (const b& other) const {
return this->b_ == other.b_;
}
/* needed, because myset != myset2 is called later in the code */
bool operator != (const b& other) const {
return !(*this == other);
}
};
You should also take a look at http://en.cppreference.com/w/cpp/container/set and see what other operators std::set uses internally on its elements.
No operator (except for the default operator=(const T&) and operator=(T&&)) is generated by the compiler by default. You should define them explicitly:
class b{
private:
set<set<a> > b_;
public:
void set_(set<a> aux){ b_.insert(aux); }
//Overload operators?
bool operator==(const b& other) const {
return b_ == other.b_;
}
bool operator!=(const b& other) const {
return b_ != other.b_;
}
};
However, this only does not solve the case. Although comparison operators are already defined for std::set<T>, they only work if there are operators for T. So, in this case, you have to define operator== and operator!= for your a class in the same manner as I showed you with b class.
I have a templated class MatchBase with a function for the operator == as such
template<typename Element>
class MatchBase{
virtual bool operator ==(const MatchBase<Element>& m) const{
if(_v1 == m.getFirst() && _v2 == m.getSecond()){
return true;
}
return false;
}
I know have a daughter class Match that is template specialized. The class Place used for the specialization does not have an operator== to do the comparison. Thus I'm trying to override the operator== to work with the class Place.
On the things I have tried :
class Match : public betterGraph::MatchBase<graphmatch::Place>{
public :
Match(const graphmatch::Place& v, const graphmatch::Place& vv) :
betterGraph::MatchBase<graphmatch::Place>(v, vv)
{};
virtual bool operator ==(const Match& m) const{
if(_v1.mass_center == m.getFirst().mass_center && _v2.mass_center == m.getSecond().mass_center){
return true;
}
return false;
}
};
I also tried
virtual bool operator ==(const betterGraph::MatchBase<graphmatch::Place>& m) const{
if(_v1.mass_center == m.getFirst().mass_center && _v2.mass_center == m.getSecond().mass_center){
return true;
}
return false;
}
But I always hit an error of the type :
error: no match for ‘operator==’ (operand types are ‘const AASS::graphmatch::Place’ and ‘const AASS::graphmatch::Place’)
if(_v1 == m.getFirst() && _v2 == m.getSecond()){
Because it tries to compile the method from the Base class.
Is there any way for me to override this function of the base class in the daughter class ? I've read the question here but here it's the method that is specialized while my class is specialized so I don't see how to do a forward declaration :/.
The function may be virtual but it's still initialized when you inherit your base class.
This is essential as you might write something like this:
MatchBase<Place> test = Match(p1,p2);
MatchBase<Place> is the base class of Match yet they are not the same.
Calling MatchBase<Place>::operator==() will still call the function defined in your template base class.
You have now multiple option:
- make the function in the base class a pure virtual
- implement Place::operator==()
- pass a comperator as argument to your base class as argument
The first two should be clear (if not please ask). For the third this might be a one possible way to do it:
template<typename Element, typename Less = std::less<Element>>
class MatchBase {
protected:
Element _v1;
Element _v2;
public:
MatchBase(const Element& v, const Element& vv) : _v1(v), _v2(vv)
{}
virtual bool operator ==(const MatchBase<Element, Less>& m) const {
Less less;
bool v1Equal = !less(_v1, m.getFirst()) && !less(m.getFirst(), _v1);
bool v2Equal = !less(_v2, m.getSecond()) && !less(m.getSecond(), _v2);
return v1Equal && v2Equal;
}
const Element& getFirst() const { return _v1; }
const Element& getSecond() const { return _v2; }
};
struct Place
{
int mass_center;
};
struct PlaceLess
{
bool operator()(const Place& p1, const Place& p2)
{
return p1.mass_center < p2.mass_center;
};
};
class Match : public MatchBase <Place, PlaceLess>
{
public:
Match(const Place& v, const Place& vv) :
MatchBase<Place, PlaceLess>(v, vv)
{};
};
Another way might be to specialize std::less<T> in this context. So you won't need to pass it as template parameter.
template<typename Element>
class MatchBase {
protected:
Element _v1;
Element _v2;
public:
MatchBase(const Element& v, const Element& vv) : _v1(v), _v2(vv)
{}
virtual bool operator ==(const MatchBase<Element>& m) const {
std::less<Element> less;
bool v1Equal = !less(_v1, m.getFirst()) && !less(m.getFirst(), _v1);
bool v2Equal = !less(_v2, m.getSecond()) && !less(m.getSecond(), _v2);
return v1Equal && v2Equal;
}
const Element& getFirst() const { return _v1; }
const Element& getSecond() const { return _v2; }
};
struct Place
{
int mass_center;
};
template<>
struct std::less<Place>
{
bool operator()(const Place& p1, const Place& p2)
{
return p1.mass_center < p2.mass_center;
};
};
class Match : public MatchBase <Place>
{
public:
Match(const Place& v, const Place& vv) :
MatchBase<Place>(v, vv)
{};
};
Of course you could merge these ways so you might override the Less template parameter if needed.
If you don't plan on using predefined types (thinking of int, std::string, etc...) you could also make sure that the class passed as Element must inherit a class/struct that enforces that operator== is implemented:
template <typename T>
struct IComparable
{
virtual bool operator==(const T& other) const = 0;
};
template<typename Element>
class MatchBase {
static_assert(std::is_base_of<IComparable<Element>, Element>::value, "Element must implement comparable");
protected:
Element _v1;
Element _v2;
public:
MatchBase(const Element& v, const Element& vv) : _v1(v), _v2(vv)
{}
virtual bool operator ==(const MatchBase<Element>& m) const {
return _v1 == m._v1 && _v2 == m._v2;
}
};
struct Place : public IComparable<Place>
{
int mass_center;
bool operator==(const Place& other) const
{
return mass_center == other.mass_center;
};
};
class Match : public MatchBase <Place>
{
public:
Match(const Place& v, const Place& vv) :
MatchBase<Place>(v, vv)
{};
};
I have a type hierarchy, and I'm not sure of a clean / good way to implement operator< and operator==.
Essentially, I already have this:
class Parent {
public:
virtual ~Parent() {}
};
class A : public Parent { int data; };
class B : public Parent { double data; };
class C : public Parent { std::string data; };
bool operator==(A const & lhs, A const & rhs) { return lhs.data == rhs.data; }
bool operator< (A const & lhs, A const & rhs) { return lhs.data < rhs.data; }
bool operator==(B const & lhs, B const & rhs) { return lhs.data == rhs.data; }
bool operator< (B const & lhs, B const & rhs) { return lhs.data < rhs.data; }
bool operator==(C const & lhs, C const & rhs) { return lhs.data == rhs.data; }
bool operator< (C const & lhs, C const & rhs) { return lhs.data < rhs.data; }
What I'd like to implement as well, is this:
bool operator==(Parent const & lhs, Parent const & rhs) { ... }
bool operator< (Parent const & lhs, Parent const & rhs) { ... }
I've currently implemented it by doing:
bool operator==(Parent const & lhs, Parent const & rhs) {
try {
return dynamic_cast<A const &>(lhs) == dynamic_cast<A const &>(rhs);
} catch(std::bad_cast const & e) {
}
try {
return dynamic_cast<B const &>(lhs) == dynamic_cast<B const &>(rhs);
} catch(std::bad_cast const & e) {
}
try {
return dynamic_cast<C const &>(lhs) == dynamic_cast<C const &>(rhs);
} catch(std::bad_cast const & e) {
}
assert(typeid(lhs) != typeid(rhs));
return false;
}
But this just seems awful. Is there a cleaner way of going about this?
For comparisons of complex types, you may find Double Dispatch useful.
If your types are very simple, it is sometimes effective to roll them all into one. In the example of 3 unsigned variants, it would likely be better to just use one type to accommodate all sizes, and to avoid dynamic dispatch and more complicated graphs of types.
Applied to original question; where A, B, and C all used unsigned types:
well, one quick and dirty approach would be:
class Parent {
protected:
virtual ~Parent() {}
public:
bool operator<(const Parent& pOther) const {
return this->as_uint64() < pOther.as_uint64();
}
// ...
private:
// using a type which accommodates all values
virtual uint64_t as_uint64() const = 0;
};
and then deriving from Parent would take the form:
class A : public Parent {
// ...
private:
virtual uint64_t as_uint64() const { return this->data; }
private:
uint16_t data;
};
then Parent could simply define all comparators, and all Parent types would be comparable.
Use a virtual comparator for single dispatch and dynamic_cast for type casting:
class ABC_base {
public:
virtual ~ABC_base() {}
bool operator < (ABC_base const & rhs) const {
return this->comparator(rhs) < 0;
}
protected:
virtual int comparator (ABC_base const &) = 0;
};
class ABC : public ABC_base {
protected:
virtual int comparator(ABC_base const & rhs) const {
try {
return my_comparator(dynamic_cast<ABC const&>(rhs));
// Run-time cast failed - use double dispatch as fallback
} catch (std::bad_cast&) {
return -rhs.comparator(*this);
}
}
private:
int my_comparator(ABC const & rhs) const {
if (data < rhs.data)
return -1;
if (data == rhs.data)
return 0;
if (data > rhs.data)
return 1;
}
T data;
};
Here's how the code works:
The base class's operator < is called, which uses dynamic lookup to find the comparator. It checks the returned value to see if it's lesser.
The derived class's comparator attempts to downcast the base class reference so that comparison can be done on the derived class's members.
Why the base class reference, instead of using the derived class reference?
Virtual dispatch would not work otherwise due to incorrect function signature.
Should the downcast succeed, it calls the non-virtual private comparator. Otherwise, it uses virtual dispatch again to do (rhs ? *this) and negates the result to compensate for the inverted ordering.
Why not have the cast and comparison in the one virtual function? It will make the code messier since the function will do two things: casting and comparing. Hence, there's a private comparator function. Should you want to use the base function in a derived class, along the lines of class ABC_der : public ABC, call ABC::comparator(static_cast<ABC const&>(rhs)). The use of Base:: forces static dispatch so you don't have to expose the helper comparison function.
Right now, this and rhs are of the same type, so we can finally do the actual comparison. A chain of if statements is used to return a value conformant to Java's Comparable and C's qsort() semantics.