Normally constant members of a structure can be initialized in constructor.
However this is not possible if they have overloading of assignment operator.
This restricts usage of complex numbers from std.complex since they are structures with opAssign.
-------------------
struct Foo1 // opAssign is a function
{
int value;
void opAssign(int src) { value = src; }
}
struct Foo2 // opAssign is a template
{
int value;
void opAssign()(int src) { value = src; }
}
struct Boo
{
const Foo1 f1;
const Foo2 f2;
this(int src)
{
f1 = src; // Error!
f2 = src; // Error!
}
}
void main() {}
-------------------
test.d(20): Error: mutable method test.Foo1.opAssign is not callable using a const object
test.d(21): Error: template test.Foo2.opAssign does not match any function template declaration. Candidates are:
test.d(10): test.Foo2.opAssign()(int src)
test.d(21): Error: template test.Foo2.opAssign()(int src) cannot deduce template function from argument types !()(int)
-------------------
Comment #1 by k.hara.pg — 2013-03-09T08:01:49Z
This is unfixable problem, if I'm not mistaken.
Currently, compiler always considers opAssign operator overloading for all field "assignment" in constructor.
struct S {
T field;
this(...) { field = xxx; } // If T has opAssign, it is called.
}
But for non-mutable field, opAssign invocation is not legal, because it may break const correctness.
T* p;
struct T {
void opAssign(int n) { ...; p = &this; }
}
struct S {
immutable T field;
this(...) { field = 1; // invoke T.opAssign (currently not allowed)
/* now global p holds mutable pointer to immutable T object! */
}
}
I have no answer for this issue... So, assigned to Andrei.
Comment #2 by andrej.mitrovich — 2013-03-09T09:07:15Z
(In reply to comment #1)
> This is unfixable problem, if I'm not mistaken.
What if opAssign is const/inout?
Comment #3 by k.hara.pg — 2013-03-09T09:14:18Z
(In reply to comment #2)
> (In reply to comment #1)
> > This is unfixable problem, if I'm not mistaken.
>
> What if opAssign is const/inout?
It would be invoked, but you cannot do any meaningful operation in it.
Comment #4 by andrej.mitrovich — 2013-03-09T09:16:40Z
(In reply to comment #3)
> (In reply to comment #2)
> > (In reply to comment #1)
> > > This is unfixable problem, if I'm not mistaken.
> >
> > What if opAssign is const/inout?
>
> It would be invoked, but you cannot do any meaningful operation in it.
I see. Is there any use for a const opAssign?
Comment #5 by rswhite4 — 2013-03-09T09:32:56Z
(In reply to comment #3)
> (In reply to comment #2)
> > (In reply to comment #1)
> > > This is unfixable problem, if I'm not mistaken.
> >
> > What if opAssign is const/inout?
>
> It would be invoked, but you cannot do any meaningful operation in it.
You can:
import std.stdio;
import std.c.string : memcpy;
struct Foo1 // opAssign is a function
{
int value;
void opAssign(int src) const {
int* ptr = cast(int*) &this.value;
*ptr = src;
}
}
struct Boo
{
const Foo1 f1;
this(int src)
{
f1 = src; // Error!
}
}
void main() {
Boo b = Boo(42);
writeln(b.f1.value); // prints 42
}
Comment #6 by maximzms — 2013-03-09T09:47:58Z
(In reply to comment #1)
> But for non-mutable field, opAssign invocation is not legal, because it may
> break const correctness.
>
> T* p;
> struct T {
> void opAssign(int n) { ...; p = &this; }
> }
> struct S {
> immutable T field;
> this(...) { field = 1; // invoke T.opAssign (currently not allowed)
> /* now global p holds mutable pointer to immutable T object! */
> }
> }
Is there any way to break const correctness if opAssign is pure?
If not, then changes should be allowed for pure opAssign (e.g. with implicit cast field to mutable).
Comment #7 by k.hara.pg — 2013-03-09T10:29:11Z
(In reply to comment #5)
[snip]
To enforce breaking type system by language users is not good at all.
When talking about language semantics, unsafe operation like cast must not appear. Language user can break type system explicitly, but compiler must not do it.
(In reply to comment #6)
> Is there any way to break const correctness if opAssign is pure?
'pure' attribute does not affect constancy.
> If not, then changes should be allowed for pure opAssign (e.g. with implicit
> cast field to mutable).
opAssign should modify its 'this' object for the meaningful operation, but mutable method cannot call on non-const object, even inside constructor. So this problem is unfixable.
Comment #8 by maxim — 2013-03-09T11:02:57Z
(In reply to comment #0)
> Normally constant members of a structure can be initialized in constructor.
> However this is not possible if they have overloading of assignment operator.
>
> This restricts usage of complex numbers from std.complex since they are
> structures with opAssign.
>
> <skipped>
You can do:
f2.value = src;
which doesn't require opAssign as well as doesn't scale as structure grows. This does not work either when structure has private members.
Comment #9 by andrei — 2013-03-09T17:42:11Z
This is an insufficiency in D's design. I think we should approach this the same way as super() invocation and construction of qualified objects:
1. When a constructor is entered, the object is in a "raw" state.
2. While in this raw state, the object is not considered initialized so it can't be passed to functions etc. Its const/immutable members (be they by their own qualifier or propagated from the object) are also raw and can't be passed out.
3. This raw state lasts until super() has been called EXACTLY ONCE and each qualified field has been assigned to EXACTLY ONCE.
4. At that point the object has become "cooked" and all restrictions are lifted.
Kenji, I recall we discussed this design in the context of const and immutable constructors, and you were favorable to it. How about we extend the same design for initializing qualified members in all contexts?
We can reuse the same primitive flow analysis that's now used for super().
Comment #10 by andrei — 2013-03-09T17:43:08Z
Assigned to Kenji :o).
Comment #11 by k.hara.pg — 2013-03-09T18:27:27Z
(In reply to comment #9)
Thanks for your nice reply! I have recalled description written in TDPL.
> Kenji, I recall we discussed this design in the context of const and immutable
> constructors, and you were favorable to it. How about we extend the same design
> for initializing qualified members in all contexts?
I think it is possible.
Now I completely understood. We should need some control flow analysis that is similar to calling super() for the process of initializing object field.
(In reply to comment #10)
> Assigned to Kenji :o).
I have received your request!
Comment #12 by andrei — 2013-03-09T18:32:32Z
@Kenji: that's exactly right. Just in case it isn't obvious, that only assignment for qualified fields while in raw state is not an opAssign call, it's a constructor call.
Comment #13 by maxim — 2013-03-10T01:53:08Z
(In reply to comment #9)
> This is an insufficiency in D's design. I think we should approach this the
> same way as super() invocation and construction of qualified objects
This is close to const vs postblit problem. The problem appeared due to unthoughtful decision to allow const objects mutation during construction. But it is too late now.
> 3. This raw state lasts until super() has been called EXACTLY ONCE and each
> qualified field has been assigned to EXACTLY ONCE.
I am afraid such limitation would hurt programming in cases when raw object is touched more than once.
By the way, what is about example in comment 1 when non-const opAssign stores mutable pointer to immutable data?
> 4. At that point the object has become "cooked" and all restrictions are
> lifted.
>
> Kenji, I recall we discussed this design in the context of const and immutable
> constructors, and you were favorable to it. How about we extend the same design
> for initializing qualified members in all contexts?
>
> We can reuse the same primitive flow analysis that's now used for super().
I think it would be better to have special qualifier for such situation. Anyway we implicitly have it. Problem with flow analysis is that callee may have no idea who is caller due to separate compilation.
Comment #14 by andrei — 2013-03-10T12:33:50Z
(In reply to comment #13)
> (In reply to comment #9)
> > This is an insufficiency in D's design. I think we should approach this the
> > same way as super() invocation and construction of qualified objects
>
> This is close to const vs postblit problem. The problem appeared due to
> unthoughtful decision to allow const objects mutation during construction. But
> it is too late now.
I disagree it's too late, seeing as I actually propose a solution. Also, this is probably not the best forum to air criticisms about the competence and state of mind of the language designers.
> > 3. This raw state lasts until super() has been called EXACTLY ONCE and each
> > qualified field has been assigned to EXACTLY ONCE.
>
> I am afraid such limitation would hurt programming in cases when raw object is
> touched more than once.
I agree it would disallow some correct code, but it's not a strong limitation to ask for exactly one initialization.
> By the way, what is about example in comment 1 when non-const opAssign stores
> mutable pointer to immutable data?
I'm not sure what you mean. What about it? This can't occur in the proposed raw/cooked design.
> > 4. At that point the object has become "cooked" and all restrictions are
> > lifted.
> >
> > Kenji, I recall we discussed this design in the context of const and immutable
> > constructors, and you were favorable to it. How about we extend the same design
> > for initializing qualified members in all contexts?
> >
> > We can reuse the same primitive flow analysis that's now used for super().
>
> I think it would be better to have special qualifier for such situation. Anyway
> we implicitly have it. Problem with flow analysis is that callee may have no
> idea who is caller due to separate compilation.
I disagree. A new qualifier would complicate the language much more than a typechecking rule for constructors. BTW the raw/cooked design is proven - it's been already used in a number of papers and languages.
Comment #15 by maxim — 2013-03-10T13:16:15Z
(In reply to comment #14)
> (In reply to comment #13)
> > (In reply to comment #9)
> > > This is an insufficiency in D's design. I think we should approach this the
> > > same way as super() invocation and construction of qualified objects
> >
> > This is close to const vs postblit problem. The problem appeared due to
> > unthoughtful decision to allow const objects mutation during construction. But
> > it is too late now.
>
> I disagree it's too late, seeing as I actually propose a solution. Also, this
> is probably not the best forum to air criticisms about the competence and state
> of mind of the language designers.
The purpose was not to insult language designers but to point that affecting one feature may break its interaction with other features.
> > By the way, what is about example in comment 1 when non-const opAssign stores
> > mutable pointer to immutable data?
>
> I'm not sure what you mean. What about it? This can't occur in the proposed
> raw/cooked design.
>
From what I understood, the proposal is to relax(remove temporarily) constness of members during ctor invocation which allows code like below:
T* p;
struct T {
void opAssign(int n) { ...; p = &this; }
}
struct S {
immutable T field;
this(...) { field = 1; // invoke T.opAssign (currently not allowed)
/* now global p holds mutable pointer to immutable T object! */
}
}
Or I misunderstood completely and you really proposing to call something like super().
Comment #16 by k.hara.pg — 2013-03-10T20:40:34Z
(In reply to comment #15)
> From what I understood, the proposal is to relax(remove temporarily) constness
> of members during ctor invocation which allows code like below:
>
> T* p;
> struct T {
> void opAssign(int n) { ...; p = &this; }
> }
> struct S {
> immutable T field;
> this(...) { field = 1; // invoke T.opAssign (currently not allowed)
> /* now global p holds mutable pointer to immutable T object! */
> }
> }
>
> Or I misunderstood completely and you really proposing to call something like
> super().
In the raw/cooked design, you cannot call mutable T.opAssign from immutable field, even inside constructor.
Instead, you should use whole object "assignment" for the field.
struct S {
immutable T field;
this(...) {
field = immutable(T)(...); // T's literal or constructor call
// This _looks like_ "assignment" but in practice
// it would be treated as initializing. So opAssign is not invoked.
}
}
Comment #17 by andrei — 2013-05-05T19:00:14Z
So Kenji, did you fix this bug in trunk?
Comment #18 by k.hara.pg — 2013-05-07T04:25:59Z
(In reply to comment #17)
> So Kenji, did you fix this bug in trunk?
In local I've tried to fix this, but it had broke Phobos compilation by the lack of proper postblit feature. Sorry...
Comment #19 by k.hara.pg — 2013-10-10T05:14:26Z
*** Issue 11204 has been marked as a duplicate of this issue. ***
Comment #20 by samukha — 2013-10-10T07:07:59Z
I see there was a discussion in bug 9665. Kenji, if you are fixing this, please don't forget about postblit, which is a constructor and should have similar semantics. That is:
struct S {
@disable this();
this(int x) {
}
static int postblit;
this(this) {
postblit = true;
}
static int assigned;
void opAssign(S s) {
assigned = true;
}
}
class A {
S s;
this() {
auto s1 = S(1);
s = s1;
assert(!S.assigned);
assert(S.postblit);
}
}
void main() {
auto a = new A;
}
Copy-construction should be performed, not assignment.
Comment #21 by samukha — 2013-10-10T07:10:42Z
(In reply to comment #20)
"I see there was a discussion in bug 9665." - ignore this sentence.