Operator Overloading
Operator overloading is accomplished by rewriting operators whose operands are class or struct objects into calls to specially named member functions. No additional syntax is used.
- Unary Operator Overloading
- Cast Operator Overloading
- Binary Operator Overloading
- Overloading the Comparison Operators
- Function Call Operator Overloading
- Assignment Operator Overloading
- Op Assignment Operator Overloading
- Array Indexing and Slicing Operators Overloading
- Forwarding
Unary Operator Overloading
op | rewrite |
---|---|
-e | e.opUnary!("-")() |
+e | e.opUnary!("+")() |
~e | e.opUnary!("~")() |
*e | e.opUnary!("*")() |
++e | e.opUnary!("++")() |
--e | e.opUnary!("--")() |
For example, in order to overload the - (negation) operator for struct S, and no other operator:
struct S { int m; int opUnary(string s)() if (s == "-") { return -m; } } int foo(S s) { return -s; }
Postincrement e++ and Postdecrement e-- Operators
These are not directly overloadable, but instead are rewritten in terms of the ++e and --e prefix operators:
op | rewrite |
---|---|
e-- | (auto t = e, --e, t) |
e++ | (auto t = e, ++e, t) |
Overloading Index Unary Operators
op | rewrite |
---|---|
-a[b_{1}, b_{2}, ... b_{n}] | a.opIndexUnary!("-")(b_{1}, b_{2}, ... b_{n}) |
+a[b_{1}, b_{2}, ... b_{n}] | a.opIndexUnary!("+")(b_{1}, b_{2}, ... b_{n}) |
~a[b_{1}, b_{2}, ... b_{n}] | a.opIndexUnary!("~")(b_{1}, b_{2}, ... b_{n}) |
*a[b_{1}, b_{2}, ... b_{n}] | a.opIndexUnary!("*")(b_{1}, b_{2}, ... b_{n}) |
++a[b_{1}, b_{2}, ... b_{n}] | a.opIndexUnary!("++")(b_{1}, b_{2}, ... b_{n}) |
--a[b_{1}, b_{2}, ... b_{n}] | a.opIndexUnary!("--")(b_{1}, b_{2}, ... b_{n}) |
Overloading Slice Unary Operators
op | rewrite |
---|---|
-a[i..j] | a.opIndexUnary!("-")(a.opSlice(i, j)) |
+a[i..j] | a.opIndexUnary!("+")(a.opSlice(i, j)) |
~a[i..j] | a.opIndexUnary!("~")(a.opSlice(i, j)) |
*a[i..j] | a.opIndexUnary!("*")(a.opSlice(i, j)) |
++a[i..j] | a.opIndexUnary!("++")(a.opSlice(i, j)) |
--a[i..j] | a.opIndexUnary!("--")(a.opSlice(i, j)) |
-a[ ] | a.opIndexUnary!("-")() |
+a[ ] | a.opIndexUnary!("+")() |
~a[ ] | a.opIndexUnary!("~")() |
*a[ ] | a.opIndexUnary!("*")() |
++a[ ] | a.opIndexUnary!("++")() |
--a[ ] | a.opIndexUnary!("--")() |
For backward compatibility, if the above rewrites fail and opSliceUnary is defined, then the rewrites a.opSliceUnary!(op)(a, i, j) and a.opSliceUnary!(op) are tried instead, respectively.
Cast Operator Overloading
op | rewrite |
---|---|
cast(type ) e | e.opCast!(type)() |
Boolean Operations
Notably absent from the list of overloaded unary operators is the ! logical negation operator. More obscurely absent is a unary operator to convert to a bool result. Instead, these are covered by a rewrite to:
opCast!(bool)(e)
So,
if (e) => if (e.opCast!(bool)) if (!e) => if (!e.opCast!(bool))
etc., whenever a bool result is expected. This only happens, however, for instances of structs. Class references are converted to bool by checking to see if the class reference is null or not.
Binary Operator Overloading
The following binary operators are overloadable:
+ | - | * | / | % | ^^ | & |
| | ^ | << | >> | >>> | ~ | in |
The expression:
a
b
is rewritten as both:
a.opBinary!("$(METACODE op)")(b) b.opBinaryRight!("$(METACODE op)")(a)
and the one with the ‘better’ match is selected. It is an error for both to equally match.
Operator overloading for a number of operators can be done at the same time. For example, if only the + or - operators are supported:
T opBinary(string op)(T rhs) { static if (op == "+") return data + rhs.data; else static if (op == "-") return data - rhs.data; else static assert(0, "Operator "~op~" not implemented"); }
To do them all en masse:
T opBinary(string op)(T rhs) { return mixin("data "~op~" rhs.data"); }
Overloading the Comparison Operators
D allows overloading of the comparison operators ==, !=, <, <=, >=, > via two functions, opEquals and opCmp.
The equality and inequality operators are treated separately because while practically all user-defined types can be compared for equality, only a subset of types have a meaningful ordering. For example, while it makes sense to determine if two RGB color vectors are equal, it is not meaningful to say that one color is greater than another, because colors do not have an ordering. Thus, one would define opEquals for a Color type, but not opCmp.
Furthermore, even with orderable types, the order relation may not be linear. For example, one may define an ordering on sets via the subset relation, such that x < y is true if x is a (strict) subset of y. If x and y are disjoint sets, then neither x < y nor y < x holds, but that does not imply that x == y. Thus, it is insufficient to determine equality purely based on opCmp alone. For this reason, opCmp is only used for the inequality operators <, <=, >=, and >. The equality operators == and != always employ opEquals instead.
Therefore, it is the programmer's responsibility to ensure that opCmp and opEquals are consistent with each other. If opEquals is not specified, the compiler provides a default version that does member-wise comparison. If this suffices, one may define only opCmp to customize the behaviour of the inequality operators. But if not, then a custom version of opEquals should be defined as well, in order to preserve consistent semantics between the two kinds of comparison operators.
Finally, if the user-defined type is to be used as a key in the built-in associative arrays, then the programmer must ensure that the semantics of opEquals and toHash are consistent. If not, the associative array may not work in the expected manner.
Overloading == and !=
Expressions of the form a != b are rewritten as !(a == b).
Given a == b :
- If a and b are both class objects, then the expression is rewritten as:
.object.opEquals(a, b)
and that function is implemented as:
bool opEquals(Object a, Object b) { if (a is b) return true; if (a is null || b is null) return false; if (typeid(a) == typeid(b)) return a.opEquals(b); return a.opEquals(b) && b.opEquals(a); }
- Otherwise the expressions a.opEquals(b) and b.opEquals(a) are tried. If both resolve to the same opEquals function, then the expression is rewritten to be a.opEquals(b).
- If one is a better match than the other, or one compiles and the other does not, the first is selected.
- Otherwise, an error results.
If overridding Object.opEquals() for classes, the class member function signature should look like:
class C { override bool opEquals(Object o) { ... } }
If structs declare an opEquals member function for the identity comparison, it could have several forms, such as:
struct S { // lhs should be mutable object bool opEquals(const S s) { ... } // for r-values (e.g. temporaries) bool opEquals(ref const S s) { ... } // for l-values (e.g. variables) // both hand side can be const object bool opEquals(const S s) const { ... } // for r-values (e.g. temporaries) }
Alternatively, you can declare a single templated opEquals function with an auto ref parameter:
struct S { // for l-values and r-values, // with converting both hand side implicitly to const bool opEquals()(auto ref const S s) const { ... } }
Overloading <, <=, >, and >=
Comparison operations are rewritten as follows:
comparison | rewrite 1 | rewrite 2 |
---|---|---|
a < b | a.opCmp(b) < 0 | b.opCmp(a) > 0 |
a <= b | a.opCmp(b) <= 0 | b.opCmp(a) >= 0 |
a >b | a.opCmp(b) > 0 | b.opCmp(a) < 0 |
a >= b | a.opCmp(b) >= 0 | b.opCmp(a) <= 0 |
Both rewrites are tried. If only one compiles, that one is taken. If they both resolve to the same function, the first rewrite is done. If they resolve to different functions, the best matching one is used. If they both match the same, but are different functions, an ambiguity error results.
If overriding Object.opCmp() for classes, the class member function signature should look like:
class C { override int opCmp(Object o) { ... } }
If structs declare an opCmp member function, it should have the following form:
struct S { int opCmp(ref const S s) const { ... } }
Note that opCmp is only used for the inequality operators; expressions like a == b always uses opEquals. If opCmp is defined but opEquals isn't, the compiler will supply a default version of opEquals that performs member-wise comparison. If this member-wise comparison is not consistent with the user-defined opCmp, then it is up to the programmer to supply an appropriate version of opEquals. Otherwise, inequalities like a <= b will behave inconsistently with equalities like a == b.
Function Call Operator Overloading f()
The function call operator, (), can be overloaded by declaring a function named opCall:
struct F { int opCall(); int opCall(int x, int y, int z); } void test() { F f; int i; i = f(); // same as i = f.opCall(); i = f(3,4,5); // same as i = f.opCall(3,4,5); }
In this way a struct or class object can behave as if it were a function.
Note that merely declaring opCall automatically disables struct literal syntax. To avoid the limitation, you need to also declare a constructor so that it takes priority over opCall in Type(...) syntax.
struct Multiplier { int factor; this(int num) { factor = num; } int opCall(int value) { return value * factor; } } void test() { Multiplier m = Multiplier(10); // invoke constructor int result = m(5); // invoke opCall assert(result == 50); }
Static opCall
static opCall also works as expected for a function call operator with type names.
struct Double { static int opCall(int x) { return x * 2; } } void test() { int i = Double(2); assert(i == 4); }
Mixing struct constructors and static opCall is not allowed.
struct S { this(int i) {} static S opCall() // disallowed due to constructor { return S.init; } }
Note: static opCall can be used to simulate struct constructors with no arguments, but this is not recommended practice. Instead, the preferred solution is to use a factory function to create struct instances.
Assignment Operator Overloading
The assignment operator = can be overloaded if the left hand side is a struct aggregate, and opAssign is a member function of that aggregate.
For struct types, operator overloading for the identity assignment is allowed.struct S { // identiy assignment, allowed. void opAssign(S rhs); // not identity assignment, also allowed. void opAssign(int); } S s; s = S(); // Rewritten to s.opAssign(S()); s = 1; // Rewritten to s.opAssign(1);However for class types, identity assignment is not allowed. All class types have reference semantics, so identity assignment by default rebinds the left-hand-side to the argument at the right, and this is not overridable.
class C { // If X is the same type as C or the type which is // implicitly convertible to C, then opAssign would // accept identity assignment, which is disallowed. // C opAssign(...); // C opAssign(X); // C opAssign(X, ...); // C opAssign(X ...); // C opAssign(X, U = defaultValue, etc.); // not an identity assignment - allowed void opAssign(int); } C c = new C(); c = new C(); // Rebinding referencee c = 1; // Rewritten to c.opAssign(1);
Index Assignment Operator Overloading
If the left hand side of an assignment is an index operation on a struct or class instance, it can be overloaded by providing an opIndexAssign member function. Expressions of the form a[b_{1}, b_{2}, ... b_{n}] = c are rewritten as a.opIndexAssign(c,b_{1}, b_{2}, ... b_{n}).
struct A { int opIndexAssign(int value, size_t i1, size_t i2); } void test() { A a; a[i,3] = 7; // same as a.opIndexAssign(7,i,3); }
Slice Assignment Operator Overloading
If the left hand side of an assignment is a slice operation on a struct or class instance, it can be overloaded by implementing an opIndexAssign member function that takes the return value of the opSlice function as parameter(s). Expressions of the form a[i..j] = c are rewritten as a.opIndexAssign(c, a.opSlice(i, j)), and a[] = c as a.opIndexAssign(c).
See Array Indexing and Slicing Operators Overloading for more details.
struct A { int opIndexAssign(int v); // overloads a[] = v int opIndexAssign(int v, size_t[2] x); // overloads a[i .. j] = v int[2] opSlice(size_t x, size_t y); // overloads i .. j } void test() { A a; int v; a[] = v; // same as a.opIndexAssign(v); a[3..4] = v; // same as a.opIndexAssign(v, a.opSlice(3,4)); }
For backward compatibility, if rewriting a[i..j] as a.opIndexAssign(a.opSlice(i, j)) fails to compile, the legacy rewrite opSliceAssign(c, i, j) is used instead.
Op Assignment Operator Overloading
The following op assignment operators are overloadable:
+= | -= | *= | /= | %= | ^^= | &= |
|= | ^= | <<= | >>= | >>>= | ~= |
The expression:
a
= b
is rewritten as:
a.opOpAssign!("$(METACODE op)")(b)
Index Op Assignment Operator Overloading
If the left hand side of an op= is an index expression on a struct or class instance and opIndexOpAssign is a member:
a[] = c
it is rewritten as:
a.opIndexOpAssign!("$(METACODE op)")(c, )
Slice Op Assignment Operator Overloading
If the left hand side of an op= is a slice expression on a struct or class instance and opIndexOpAssign is a member:
a[] = c
it is rewritten as:
a.opIndexOpAssign!("$(METACODE op)")(c, a.opSlice( ))
and
a[]
= c
it is rewritten as:
a.opIndexOpAssign!("$(METACODE op)")(c)
For backward compatibility, if the above rewrites fail and opSliceOpAssign is defined, then the rewrites a.opSliceOpAssign(c, i, j) and a.opSliceOpAssign(c) are tried, respectively.
Array Indexing and Slicing Operators Overloading
The array indexing and slicing operators are overloaded by implementing the opIndex, opSlice, and opDollar methods. These may be combined to implement multidimensional arrays.
Index Operator Overloading
Expressions of the form arr[b_{1}, b_{2}, ... b_{n}] are translated into arr.opIndex(b_{1}, b_{2}, ... b_{n}). For example:
struct A { int opIndex(size_t i1, size_t i2, size_t i3); } void test() { A a; int i; i = a[5,6,7]; // same as i = a.opIndex(5,6,7); }
In this way a struct or class object can behave as if it were an array.
If an index expression can be rewritten using opIndexAssign or opIndexOpAssign, those are preferred over opIndex.
Slice Operator Overloading
Overloading the slicing operator means overloading expressions like a[] or a[i..j], where the expressions inside the square brackets contain slice expressions of the form i..j.
To overload a[], simply define opIndex with no parameters:
struct S { int[] impl; int[] opIndex() { return impl[]; } } void test() { auto s = S([1,2,3]); auto t = s[]; // calls s.opIndex() assert(t == [1,2,3]); }
To overload array indexing of the form a[i..j, ...], two steps are needed. First, the expressions of the form i..j are translated via opSlice into user-defined objects that encapsulate the endpoints i and j. Then these user-defined objects are passed to opIndex to perform the actual slicing. This design was chosen in order to support mixed indexing and slicing in multidimensional arrays; for example, in translating expressions like arr[1, 2..3, 4].
More precisely, an expression of the form arr[b_{1}, b_{2}, ... b_{n}] is translated into arr.opIndex(c_{1}, c_{2}, ... c_{n}). Each argument b_{i} can be either a single expression, in which case it is passed directly as the corresponding argument c_{i} to opIndex; or it can be a slice expression of the form x_{i}..y_{i}, in which case the corresponding argument c_{i} to opIndex is arr.opSlice!i(x_{i}, y_{i}). Namely:
op | rewrite |
---|---|
arr[1, 2, 3] | arr.opIndex(1, 2, 3) |
arr[1..2, 3..4, 5..6] | arr.opIndex(arr.opSlice!0(1,2), arr.opSlice!1(3,4), arr.opSlice!2(5,6)) |
arr[1, 2..3, 4] | arr.opIndex(1, arr.opSlice!1(2,3), 4) |
Similar translations are done for assignment operators involving slicing, for example:
op | rewrite |
---|---|
arr[1, 2..3, 4] = c | arr.opIndexAssign(c, 1, arr.opSlice!1(2, 3), 4) |
arr[2, 3..4] += c | arr.opIndexOpAssign!"+"(c, 2, arr.opSlice!1(2, 3)) |
The intention is that opSlice!i should return a user-defined object that represents an interval of indices along the i'th dimension of the array. This object is then passed to opIndex to perform the actual slicing operation. If only one-dimensional slicing is desired, opSlice may be declared without the compile-time parameter i.
Note that in all cases, arr is only evaluated once. Thus, an expression like getArray()[1, 2..3, $-1]=c has the effect of:
auto __tmp = getArray();
__tmp.opIndexAssign(c, 1, __tmp.opSlice!1(2,3), __tmp.opDollar!2 - 1);
where the initial function call to getArray is only executed once.
For backward compatibility, a[] and a[i..j] can also be overloaded by implementing opSlice() with no arguments and opSlice(i, j) with two arguments, respectively. This only applies for one-dimensional slicing, and dates from when D did not have full support for multidimensional arrays. This usage of opSlice is discouraged.
Dollar Operator Overloading
Within the arguments to array index and slicing operators, $ gets translated to opDollar!i, where i is the position of the expression $ appears in. For example:
op | rewrite |
---|---|
arr[$-1, $-2, 3] | arr.opIndex(arr.opDollar!0 - 1, arr.opDollar!1 - 2, 3) |
arr[1, 2, 3..$] | arr.opIndex(1, 2, arr.opSlice!2(3, arr.opDollar!2)) |
The intention is that opDollar!i should return the length of the array along its i'th dimension, or a user-defined object representing the end of the array along that dimension, that is understood by opSlice and opIndex.
struct Rectangle { int width, height; int[][] impl; this(int w, int h) { width = w; height = h; impl = new int[w][h]; } int opIndex(size_t i1, size_t i2) { return impl[i1][i2]; } int opDollar(size_t pos)() { static if (pos==0) return width; else return height; } } void test() { auto r = Rectangle(10,20); int i = r[$-1, 0]; // same as: r.opIndex(r.opDollar!0, 0), // which is r.opIndex(r.width-1, 0) int j = r[0, $-1]; // same as: r.opIndex(0, r.opDollar!1) // which is r.opIndex(0, r.height-1) }
As the above example shows, a different compile-time argument is passed to opDollar depending on which argument it appears in. A $ appearing in the first argument gets translated to opDollar!0, a $ appearing in the second argument gets translated to opDollar!1, and so on. Thus, the appropriate value for $ can be returned to implement multidimensional arrays.
Note that opDollar!i is only evaluated once for each i where $ occurs in the corresponding position in the indexing operation. Thus, an expression like arr[$-sqrt($), 0, $-1] has the effect of:
auto __tmp1 = arr.opDollar!0; auto __tmp2 = arr.opDollar!2; arr.opIndex(__tmp1 - sqrt(__tmp1), 0, __tmp2 - 1);
If opIndex is declared with only one argument, the compile-time argument to opDollar may be omitted. In this case, it is illegal to use $ inside an array indexing expression with more than one argument.
Forwarding
Member names not found in a class or struct can be forwarded to a template function named opDispatch for resolution.
import std.stdio; struct S { void opDispatch(string s, T)(T i) { writefln("S.opDispatch('%s', %s)", s, i); } } class C { void opDispatch(string s)(int i) { writefln("C.opDispatch('%s', %s)", s, i); } } struct D { template opDispatch(string s) { enum int opDispatch = 8; } } void main() { S s; s.opDispatch!("hello")(7); s.foo(7); auto c = new C(); c.foo(8); D d; writefln("d.foo = %s", d.foo); assert(d.foo == 8); }