Qualified constructor revisited

Abstract

This DIP redesigns qualified constructor definitions.

Motivation

From 2.063, qualified constructor is implemented, and unique constructor

concept is added.

https://github.com/D-Programming-Language/dmd/pull/1726

However, current definition is very complex and hard to understand.

http://dlang.org/class#constructors\

> “Constructors can be overloaded with different attributes.” …

The issues in current unique constructor definition are:

• Always requires pure attribute.

• Cannot check whether a constructor is really unique or not, by only

  looking the signature.

  You also need see the types of object fields, and should check they don’t appear in the constructor parameters.

• Constructor qualifier (inout, const, etc) is not directly related to the unique constructor concept.

This situation should be improved.

Description

In the ideal world, objects would have only one of the two qualifiers - mutable or immutable. So, there would be just two constructors:

• “mutable constructor” would support creating mutable object.
• “immutable constructor” would support creating immutable object.

However, in actual D code, there are two wildcard qualifiers, const and inout. Therefore, we would also need additional two ways to:

• construct inout object from inout arguments, inside inout function.
• construct arbitrary qualified object from arbitrary arguments.

For those requests, provide additional two concepts, “inout constructor” and “const constructor”.

Mutable Constructor

If a constructor is unqualified, it will be used for mutable object construction.

struct S {     this(int) { ... } } void main() {     S sm = S(1); }

Immutable Constructor

If a constructor is qualified with immutable, it will be used for immutable object construction.

struct S {     this(int) immutable { ... } } void main() {     immutable S si = immutable S(1); }

Inout Constructor

If a constructor is qualified with inout, and has one or more inout parameters, it will become inout constructor.

struct S {     this(inout int[] a) inout { ... } } void main() {     int[] ma;     immutable int[] ia; \     auto sm = S(ma);            // OK     auto si = immutable S(ia);  // OK \   //auto sm = immutable S(ma);  // NG   //auto si = S(ia);            // NG }

// For classes, zero-arg inout constructor is allowed? // –> To reduce confusion, this DIP disallows such definition.

Const Constructor

If a constructor is qualified with const, it will be used to construct arbitrary qualified objects.

Inside const constuctor, you need to initialize the instance fields by Unique Expression (See DIP49).

struct S {     const int[] arr;     this(int[] a) const {       //this.arr = a;       // NG         this.arr = a.dup;   // OK, array.dup makes unique expression for int[]     } }

After construction finished, the generate object will be unique object - it means that the object owns no reference to the external state. So, it could also be called “unique constructor”, based on the concept.

Advantages with respect to the currently implemented definition:

• pure attribute is not necessary.
• If a constructor can be used for arbitrary qualified object construction, it is qualified with const.

Overloading of qualified constructors

If mutable constructor is defined,

• it is always used for mutable object construction.
• if immutable constructor is not defined, it will be used for const object construction.

If immutable constructor is defined,

• it is always used for immutable object construction.
• if mutable constructor is not defined, it will be used for const object construction.

If inout constructor is defined,

• it is always used for inout object construction.
• if all other constructors, mutable, immutable, and const, are not defined, it will be used for const object construction.

If const constructor is defined,

• it is always used for const object construction.
• if mutable constructor is not deinfed, it will be used for mutable object construction.
• if immutable constructor is not deinfed, it will be used for immutable object construction.
• if inout constructor is not deinfed, it will be used for inout object construction.

Related issues

• Inside constructor, the first assignment of a field should be treated as its initialization.

: This is necessary due to make unique constructor definition simple.

Issue 9665 - Structure constant members can not be initialized if have opAssign

–> Has been fixed in 2.064.

• Requires inout + const qualifier

: Currently, const type qualifier always overrides inout.

But the behavior will sometimes accidentally hurt inout constructor concept.

struct Rebindable(T) {

    this(inout T initializer) inout { ... }

    // Intend to define inout constructor.

}

Rebindable!(const Object) r;

// Currently inout(const(int)) is shrinked to const(int),

// so the constructor would loose inout parameter, and

// will cause "inout constructor should have one or more inout parameters" error.

To resolve the issue, I think we need to fix issue 6930.

Issue 6930 - combined type of immutable(T) and inout(T) should be inout(const(T))

Why ‘const’ constructor will be called to create arbitrary qualified object?

When an object is constructed by const constructor, the object would have either mutable or immutable qualifier. And, const method is always callable on both mutable and immutable object.

struct S {     this(int) const { ... } } void main() {    // const constructor is callable on constructing mutable object    S sm = S(1); \    // const constructor is callable on constructing immutable object    immutable S si = immutable S(1); }

There’s no mutation against widely known “const method” concept.

Rationale

DIP49 is based on the same concept, and they are designed symmetry right now.

Copyright

This document has been placed in the Public Domain.