Copy construction with inaccessible base class copy c-tor

[Victor Bazarov]

Hello,

Take a look at this program:
-----------------------------------
class B
{
B(const B&);
B& operator=(const B&);
public:
B(int);
};

class C : public B
{
public:
C(int);
};

int main()
{
C c = C(42); // Comeau: no error
B b = B(42); // Comeau: error
}

-----------------------------------
Both 'c' and 'b' are copy-constructed, as I understand it.
B's copy c-tor is inaccessible, so construction of 'b' is
ill-formed. And Comeau (online test drive) flags it such.
But it lets the construction of 'c' through. Should it?

As I understand it, a temporary can be omitted, but its
creation should be possible (12.2/1) as if it weren't omitted.
And since C's copy c-tor cannot be created (12.8/7), the code
that requires (or would require) it is also ill-formed.
Where do I err? Or do I?

Thanks!

V
--
Please remove capital 'A's when replying by e-mail
I do not respond to top-posted replies, please don't ask



[ See http://www.gotw.ca/resources/clcm.htm for info about ]
[ comp.lang.c++.moderated. First time posters: Do this! ]

 

[Arun]

Hello,

Hello,

Take a look at this program:
-----------------------------------
class B
{
B(const B&);
B& operator=(const B&);
public:
B(int);

};

class C : public B
{
public:
C(int);

};

int main()
{
C c = C(42); // Comeau: no error
B b = B(42); // Comeau: error

}

If everything were right the copy constructor would be called.

B's copy c-tor is inaccessible, so construction of 'b' is
ill-formed. And Comeau (online test drive) flags it such.

yes. b cannot be copy constructed.

But it lets the construction of 'c' through. Should it?

NO. c cannot be copy constructed as well. Compiler shall not generate
a copy constructor as the base class copy constructor is private.
SO this is an error too.

 

[Ian Collins]

Hello,

Take a look at this program:
-----------------------------------
class B
{
B(const B&);
B& operator=(const B&);
public:
B(int);
};

class C : public B
{
public:
C(int);
};

int main()
{
C c = C(42); // Comeau: no error
B b = B(42); // Comeau: error
}

-----------------------------------
Both 'c' and 'b' are copy-constructed, as I understand it.
B's copy c-tor is inaccessible, so construction of 'b' is
ill-formed. And Comeau (online test drive) flags it such.
But it lets the construction of 'c' through. Should it?

Sounds logical. gcc rejects both lines, Sun CC only barfs on C c = C(42);

 

[Nicola Musatti]

class B {
B(const B&);
B& operator=(const B&);
public:
B(int);
};

class C : public B {
public:
C(int);
};

int main() {
C c = C(42); // Comeau: no error
} [...]
As I understand it, a temporary can be omitted, but its
creation should be possible (12.2/1) as if it weren't omitted.
And since C's copy c-tor cannot be created (12.8/7), the code
that requires (or would require) it is also ill-formed.
Where do I err? Or do I?

I believe you're right. To me 12.8-7 appears quite clear on the
subject: The copy constructor is defined even if it's use is elided
and as it has an inaccessible base class copy constructor, the program
is ill-formed.

Cheers,
Nicola Musatti

 

[Craig Scott]

Hello,

Take a look at this program:
-----------------------------------
class B
{
B(const B&);
B& operator=(const B&);
public:
B(int);

};

class C : public B
{
public:
C(int);

};

int main()
{
C c = C(42); // Comeau: no error
B b = B(42); // Comeau: error

}

-----------------------------------
Both 'c' and 'b' are copy-constructed, as I understand it.
B's copy c-tor is inaccessible, so construction of 'b' is
ill-formed. And Comeau (online test drive) flags it such.
But it lets the construction of 'c' through. Should it?

As I understand it, a temporary can be omitted, but its
creation should be possible (12.2/1) as if it weren't omitted.
And since C's copy c-tor cannot be created (12.8/7), the code
that requires (or would require) it is also ill-formed.
Where do I err? Or do I?

To my reading, section 12.8.7 (which you mention) is very clear on
this point:

"A program is illformed if the class for which a copy constructor is
implicitly defined has ... a base class with an inaccessible or
ambiguous copy constructor."

This is exactly your situation, since C has an implicit copy
constructor and the copy constructor in its base, B, is inaccessible.
If there is any confusion about whether the copy constructor or the
assignment operator is being used, the assignment operator is subject
to the same restriction (see 12.8.12) and would make the example ill
formed for the same reasons. Thus, any discussion about whether the
compiler is allowed to by-pass the assignment operator and use the
copy constructor instead should reach the same conclusion. For
reference, 12.8.15 discusses how the compiler is allowed to by-pass
the assignment and directly copy-construct in-place, but it should be
noted that it does not *require* an implementation to do that.

 

[Kai-Uwe Bux]

Hello,

Take a look at this program:
-----------------------------------
class B
{
B(const B&);
B& operator=(const B&);
public:
B(int);
};

class C : public B
{
public:
C(int);
};

int main()
{
C c = C(42); // Comeau: no error
B b = B(42); // Comeau: error
}

-----------------------------------
Both 'c' and 'b' are copy-constructed, as I understand it.
B's copy c-tor is inaccessible, so construction of 'b' is
ill-formed. And Comeau (online test drive) flags it such.
But it lets the construction of 'c' through. Should it?

As I understand it, a temporary can be omitted, but its
creation should be possible (12.2/1) as if it weren't omitted.
And since C's copy c-tor cannot be created (12.8/7), the code
that requires (or would require) it is also ill-formed.
Where do I err? Or do I?

You are quite correct. I think this is a bug in Comeau. If I am not
mistaken, a similar (maybe the same) bug already appeared about a year ago
on this list:

http://groups.google.com/group/comp.lang.c++/browse_frm/thread/a455c59346fecc7c/5ddba796ee5478f5


Best

Kai-Uwe Bux

 

[Old Wolf]

class B
{
B(const B&);
B& operator=(const B&);
public:
B(int);

};

class C : public B
{
public:
C(int);

};

int main()
{
C c = C(42); // Comeau: no error
B b = B(42); // Comeau: error
}

The compiler should generate an error because you haven't provided
a body for the constructor of C. I'm assuming that either the
compiler stopped when it got the error on the B line, or it did
generate the error but you failed to report it.

AFAIK the C++ standard only says that compilers must generate a
diagnostic for particular erroneous code, and you reported that
a diagnostic was generated, so the behaviour is conforming. The
standard doesn't require that multiple diagnostics be produced,
even if there are multiple errors.

Try adding in bodies for all functions, and taking out the 'B' line
and see if you get an error. (If there is still no error, it would
be interesting to add further code to see how the compiler is
constructing the B part of the C object).

 

[James Kanze]

The compiler should generate an error because you haven't provided
a body for the constructor of C.

That's undefined behavior, not an error requiring a diagnostic.
Most implementations will only detect it when linking, and of
course, he never got that far.

[...]

Try adding in bodies for all functions, and taking out the 'B' line
and see if you get an error. (If there is still no error, it would
be interesting to add further code to see how the compiler is
constructing the B part of the C object).

The error is, of course, that an accessible copy constructor is
required, but that the compiler cannot generate one for C,
because there isn't one for B. There is no requirement that the
copy constructor actually be called, however, and I know of no
compiler that will call it.

What is doubtlessly occuring is that the Comeau compiler sees
that it can generate the declaration for the copy constructor
for C, and doesn't go any further, since it doesn't actually
need the definition. (The standard makes it clear, however,
that it is a diagnosible error if the compiler is unable to
generate the definition, even if it doesn't use it.)

 

[=?iso-8859-1?q?Daniel_Kr=FCgler?=]

That's undefined behavior, not an error requiring a diagnostic.
Most implementations will only detect it when linking, and of
course, he never got that far.

That's an interesting position, I never realized that. Does that mean,
that essentially every C++ program more or less relies on this kind
of UB? Even the standard provides class definitions, where
some of the usual suspects (copy c'tor, assignment op)
are declared but not defined, e.g. (I take N2134 as reference):

20.5.14.2:
// 20.5.14.2.6, undefined operators:
template<class Function2> bool operator==(const function<Function2>&);
template<class Function2> bool operator!=(const function<Function2>&);

22.1.1.1.2:
facet(const facet&); // not defined
void operator=(const facet&); // not defined

22.1.1.1.3:
void operator=(const id&); // not defined
id(const id&); // not defined

27.4.4:
basic_ios(const basic_ios& ); // not defined
basic_ios& operator=(const basic_ios&); // not defined

27.6.1.1.2:
sentry(const sentry&); // not defined
sentry& operator=(const sentry&); // not defined

27.6.2.3:
sentry(const sentry&); // not defined
sentry& operator=(const sentry&); // not defined

Of course one can argue that the standard is allowed to demand
that and library writers have to ensure it's realization, but that
seems
somewhat unsatisfactory to me.

Is this one of the implicite rules or explicitely spoken out?
(I haven't searched yet)

Greetings,

Daniel

 

[lingtools]

What about this ?
-------------------------
class B
{
protected:
B(const B&);
B& operator=(const B&);
public:
B(int);
};
class C : public B
{
public:
C(int);
};
int main()
{
C c = C(42); //
B b = B(42); //
}

 

[Victor Bazarov]

What about this ?
-------------------------
class B
{
protected:
B(const B&);
B& operator=(const B&);
public:
B(int);
};
class C : public B
{
public:
C(int);
};
int main()
{
C c = C(42); //
B b = B(42); //
}

The construction of 'c' should compile, the construction of
'b' should not, IMO. 'C' has an implicit copy c-tor, which
calls the explicit copy c-tor for 'B', which it can do because
the protected access is granted to 'C'.

The construction of 'b' should NOT compile because it involves
making a copy of the temporary, and 'main' has no access to the
protected copy c-tor in 'B'.

I can see the correct behaviour in Comeau, but I have no way of
verifying this with GNU C++ at this time.

V
--
Please remove capital 'A's when replying by e-mail
I do not respond to top-posted replies, please don't ask



[ See http://www.gotw.ca/resources/clcm.htm for info about ]
[ comp.lang.c++.moderated. First time posters: Do this! ]

 

[Nicola Musatti]

That's an interesting position, I never realized that. Does that mean,
that essentially every C++ program more or less relies on this kind
of UB? Even the standard provides class definitions, where
some of the usual suspects (copy c'tor, assignment op)
are declared but not defined, e.g. (I take N2134 as reference):

20.5.14.2:
// 20.5.14.2.6, undefined operators:
template<class Function2> bool operator==(const function<Function2>&);
template<class Function2> bool operator!=(const function<Function2>&);

I'm not sure whether I misunderstood James or you did. Just to be
sure, in my opinion the following invokes undefined behaviour:

class A {
public:
A();
};

class B : public A {
};

int main() {
B b;
}

A compiler might be able to generate a definition for B::B(), however
this would contain a call to A::A() which cannot be resolved by the
linker. This not very different from the following program:

void f();

int main() {
f();
}

The following, instead, is downright malformed:

class C {
private:
C();
};

class D : public C {
};

int main() {
D d;
}

The compiler is not allowed to generate a definition for D:()
because C::C() is not accessible. In a way malformation kicks in
before undefined behaviour does.

Cheers,
Nicola Musatti

 

[=?iso-8859-1?q?Daniel_Kr=FCgler?=]

I'm not sure whether I misunderstood James or you did. Just to be
sure, in my opinion the following invokes undefined behaviour:

I'm quite sure that you didn't misunderstood James. Don't ask me why,
but I got totally lost in an arcane idea.
Fact is, as you correctly said, that an inaccessible and not defined
class member is (normally) harmless, because access rules apply before
the compiler has finished. What I indeed never consciously realised,
(but this was not what I meant in my previous posting) was that an
used but undefined entity (e.g. usual function or function member)
causes UB, most probably because I always relied on the linker to
recognize such situations. What I want to say is, that according to
the outcome of this discussion even

class T {
T(); // not defined;
public:
static void foo() {
T item;
}
};

int main() {
T::foo();
}

or even simpler

void foo(); // not defined

int main() {
foo();
}

causes UB. I assume that this surprising result
(at least for me) is related to the fact that the
C++ standard does not describe a linker model.
The existence of the linker and the occurence of
a linker error here and there seems so natural
for most of us that I never considered its cause
to be a programmers fault which actually belongs
to the UB family.

Greetings from Bremen,

Daniel Krügler

 

[James Kanze]

That's an interesting position, I never realized that. Does that mean,
that essentially every C++ program more or less relies on this kind
of UB? Even the standard provides class definitions, where
some of the usual suspects (copy c'tor, assignment op)
are declared but not defined, e.g. (I take N2134 as reference):

I'm afraid I wasn't very clear. Old Wolf said "The compiler
should generate an error because you haven't provided a body for
the constructor of C." That's really the only statement I'm
responding to; not providing bodies (or providing multiple
bodies) to a function that is actually used is a violation of
the one definition rule, and as such is undefined behavior.
(Not providing bodies to a function which is not actually used
is perfectly legal. Applying the standard's definition of
"used", i.e. virtual functions are always used, as is a user
declared operator delete, and maybe some other special cases.)

The original code requires a diagnostic (and will require one
even if bodies are provided for all of the user defined
functions) because the compiler is required to generated the
definition of the implicite copy constructor, and can't.

 

[James Kanze]

I'm quite sure that you didn't misunderstood James. Don't ask me why,
but I got totally lost in an arcane idea.

When I reread my statement, I got confused by it myself, so the
problem is on my side.

Fact is, as you correctly said, that an inaccessible and not defined
class member is (normally) harmless, because access rules apply before
the compiler has finished. What I indeed never consciously realised,
(but this was not what I meant in my previous posting) was that an
used but undefined entity (e.g. usual function or function member)
causes UB, most probably because I always relied on the linker to
recognize such situations. What I want to say is, that according to
the outcome of this discussion even

class T {
T(); // not defined;
public:
static void foo() {
T item;
}
};

int main() {
T::foo();

}

or even simpler

void foo(); // not defined

int main() {
foo();

}

causes UB.

That's correct. In practice, in this particular case, you will
certainly get an error from the linker (although I once used a
linker, many, many years ago, that would occasionally use the
address 0, rather than reporting an undefined external).

I assume that this surprising result
(at least for me) is related to the fact that the
C++ standard does not describe a linker model.
The existence of the linker and the occurence of
a linker error here and there seems so natural
for most of us that I never considered its cause
to be a programmers fault which actually belongs
to the UB family.

I think that the real reason is that this is part of a larger
rule, the one definition rule, and there are some violations of
this rule which are not easily detected. If, for example,
instead of providing no definition, you provide two, and you
link through libraries. It wouldn't be too difficult for the
standard to require a diagnostic in the case of a missing
definition; it may not describe the exact linker model, but it
does talk about resolving external symbols, and such, in the
phases of compilation. It's probably not worth addressing now,
however; hopefully, the next version of the standard will
contain some support for dynamic linking, and then, you do sort
of get undefined behavior.

 

[Greg Comeau]

class B
{
B(const B&);
B& operator=(const B&);
public:
B(int);
};

class C : public B
{
public:
C(int);
};

int main()
{
C c = C(42); // Comeau: no error
B b = B(42); // Comeau: error
}

Both 'c' and 'b' are copy-constructed, as I understand it.
B's copy c-tor is inaccessible, so construction of 'b' is
ill-formed. And Comeau (online test drive) flags it such.
But it lets the construction of 'c' through. Should it?

As I understand it, a temporary can be omitted, but its
creation should be possible (12.2/1) as if it weren't omitted.
And since C's copy c-tor cannot be created (12.8/7), the code
that requires (or would require) it is also ill-formed.
Where do I err? Or do I?

Sorry for the delay in responding. I started mulling it and
got caught up in other things. I don't think you err:
The standard seems pretty clear that it needs to be accessible
in your context, even if it can be sidestepped regarding its "need".

Kanze seems to have the answer I would have provided (thanks James
It's a more deeper issue internally and as with all reports
it will be given consideration at some point.
--
Greg Comeau / 20 years of Comeauity! Intel Mac Port now in beta!
Comeau C/C++ ONLINE ==> http://www.comeaucomputing.com/tryitout
World Class Compilers: Breathtaking C++, Amazing C99, Fabulous C90.
Comeau C/C++ with Dinkumware's Libraries... Have you tried it?

[ See http://www.gotw.ca/resources/clcm.htm for info about ]
[ comp.lang.c++.moderated. First time posters: Do this! ]