Temporary objects as constructor arguments?

[Andy Buckley]

Hi,

A friend and I have recently had trouble getting code to compile when
using temporary objects in constructors. A minimal example is below:

This code fragment is used to construct seven ROD objects and run
their doSomething() method (which is trivial).

As far as we can tell, all seven methods should successfully construct a
ROD object using temporary objects as constructor parameters, but using
g++ 3.2.3 only the first 4 actually succeed. The remainder seem to be
treated as constructing/declaring(?) a function pointer and so the attempt
to call a method fails for these.

Confusingly, some of the approaches which fail under g++ do compile
successfully using MicroSoft Visual C++ 6.0: can anyone enlighten us as to
whether this is a failing in our understanding of the language or a bug in
g++? (probably the first!)

// Minimal test case:

#include <iostream>

typedef unsigned int U;

class T {
public:
T() {};
T( const U & ) { }
};

class ROD {
public:
ROD(const T &) { }
void doSomething() const {
std::cout << "Hello -- I am a ROD" << std::endl;
}
};

class CC {
public:
void doSomethingElse () {

ROD rod1 = ROD( T( t() ) ) ; // does what I want -- makes a ROD
ROD rod2 ( T( this->t() ) ) ; // does what I want -- makes a ROD
ROD rod3 ( *(new T() ) ) ; // does what I want -- makes a ROD
ROD rod4 ( T( 0 ) ) ; // does what I want -- makes a ROD

rod1.doSomethin g(); // succeeds rod2.doSomethin g();
// succeeds rod3.doSomethin g(); // succeeds
rod4.doSomethin g(); // succeeds

/*
ROD rod5 ( T( t() ) ) ; // doesn't do what I want (*) ROD
rod6 ( ROD( T( t() ) )); // doesn't do what I want (*) ROD
rod7 ( T() ) ; // doesn't do what I want (*)

rod5.doSomethin g(); // doesn't compile
rod6.doSomethin g(); // doesn't compile
rod7.doSomethin g(); // doesn't compile
*/

// (*) At these statements, funtion pointers seem to be declared :(
}
T t() const { return T(0); }
};

int main () {
CC cc;
cc.doSomethingE lse();
return 0;
}

[Victor Bazarov]

Andy Buckley wrote:

A friend and I have recently had trouble getting code to compile when
using temporary objects in constructors. A minimal example is below:

This code fragment is used to construct seven ROD objects and run
their doSomething() method (which is trivial).

As far as we can tell, all seven methods should successfully construct a
ROD object using temporary objects as constructor parameters, but using
g++ 3.2.3 only the first 4 actually succeed. The remainder seem to be
treated as constructing/declaring(?) a function pointer and so the attempt
to call a method fails for these.
The last two are not declarations at all. The fifth is a declaration.
See below.
Confusingly, some of the approaches which fail under g++ do compile
successfully using MicroSoft Visual C++ 6.0: can anyone enlighten us as to
whether this is a failing in our understanding of the language or a bug in
g++? (probably the first!)
It's a bug in your understanding of the language.

// Minimal test case:

#include <iostream>

typedef unsigned int U;

class T {
public:
T() {};
T( const U & ) { }
};

class ROD {
public:
ROD(const T &) { }
void doSomething() const {
std::cout << "Hello -- I am a ROD" << std::endl;
}
};

class CC {
public:
void doSomethingElse () {

ROD rod1 = ROD( T( t() ) ) ; // does what I want -- makes a ROD
ROD rod2 ( T( this->t() ) ) ; // does what I want -- makes a ROD
ROD rod3 ( *(new T() ) ) ; // does what I want -- makes a ROD
ROD rod4 ( T( 0 ) ) ; // does what I want -- makes a ROD

rod1.doSomethin g(); // succeeds rod2.doSomethin g();
// succeeds rod3.doSomethin g(); // succeeds
rod4.doSomethin g(); // succeeds

/*
ROD rod5 ( T( t() ) ) ; // doesn't do what I want (*) ROD
It's a declaration. See FAQ 10.2 and search the Google Groups for
similar problems people have been asking about since Adam.
rod6 ( ROD( T( t() ) )); // doesn't do what I want (*) ROD
Huh? You begin a statement with it, but 'rod6' is undefined.
rod7 ( T() ) ; // doesn't do what I want (*)
'rod7' is undefined as well.

Did you miss "ROD " in front of the two last statements? It seems
to suggest that instead of copying and pasting the code from your
C++ source file, you typed it in. Who knows how many more mistakes
you've made while doing that...

rod5.doSomethin g(); // doesn't compile
rod6.doSomethin g(); // doesn't compile
rod7.doSomethin g(); // doesn't compile
*/

// (*) At these statements, funtion pointers seem to be declared :(
}
T t() const { return T(0); }
};

int main () {
CC cc;
cc.doSomethingE lse();
return 0;
}

V

[Andy Buckley]

On Wed, 25 Aug 2004 18:27:23 +0100, Victor Bazarov wrote:

Andy Buckley wrote:

As far as we can tell, all seven methods should successfully construct
a ROD object using temporary objects as constructor parameters, but
using g++ 3.2.3 only the first 4 actually succeed. The remainder seem
to be treated as constructing/declaring(?) a function pointer and so
the attempt to call a method fails for these.

The last two are not declarations at all. The fifth is a declaration.

ROD rod5 ( T( t() ) ) ; // doesn't work

It's a declaration. See FAQ 10.2 and search the Google Groups for
similar problems people have been asking about since Adam.

Have looked at FAQ 10.2 and while rod5 is definitely being treated as a
declaration of a function which returns a ROD, I'm confies as to why:
the parentheses contain a real (if temporary) instantiation of a T object
rather than a type declaration. FAQ 10.2 only seems to describe the case
analogous to saying

ROD rod5();

which I can understand as declaring a function. Equivalently,

ROD rod5( T t( U u() ) );

according to the C++ function declaration grammar matches the definition
of a set of nested function declarations (however crazy the idea might
be). But when t() should return a concrete instatiation of a T object, I
would have expected this to fail to match the function declaration.

A clearer repost of the code without the copy-construction and
with the erroneous line-wrapping hopefully sorted is below. If you can
spare a moment to explain just what's going on, in particular on the
highlighted line, we'd both be very happy! Thanks.
// Reposted code:

#include <iostream>

typedef unsigned int U;

class T {
public:
T( const U & ) { }
};

class ROD {
public:
ROD(const T &) { }
void doSomething() const {
std::cout << "Hello -- I am a ROD" << std::endl;
}
};

class CC {
public:
void doSomethingElse () {

ROD rod5( T( u() ) ); // *** line of interest
rod5.doSomethin g(); // doesn't compile since it
// thinks rod5 is a function ptr
// of type ROD ()(T (*)())
}

U u() const { return 0; }
};

int main () {
CC cc;
cc.doSomethingE lse();
return 0;
}

[Victor Bazarov]

Andy Buckley wrote:

On Wed, 25 Aug 2004 18:27:23 +0100, Victor Bazarov wrote:

Andy Buckley wrote:

As far as we can tell, all seven methods should successfully construct
a ROD object using temporary objects as constructor parameters, but
using g++ 3.2.3 only the first 4 actually succeed. The remainder seem
to be treated as constructing/declaring(?) a function pointer and so
the attempt to call a method fails for these.

The last two are not declarations at all. The fifth is a declaration.

ROD rod5 ( T( t() ) ) ; // doesn't work

It's a declaration. See FAQ 10.2 and search the Google Groups for
similar problems people have been asking about since Adam.

Have looked at FAQ 10.2 and while rod5 is definitely being treated as a
declaration of a function which returns a ROD, I'm confies as to why:
the parentheses contain a real (if temporary) instantiation of a T object
rather than a type declaration. FAQ 10.2 only seems to describe the case
analogous to saying

ROD rod5();

which I can understand as declaring a function. Equivalently,

ROD rod5( T t( U u() ) );

according to the C++ function declaration grammar matches the definition
of a set of nested function declarations (however crazy the idea might
be). But when t() should return a concrete instatiation of a T object,

Why?

K k(blah)

is either an object declaration or a function declaration depending on
what 'blah' is. If 'blah' is another declaration (or is empty), then 'k'
is a function that takes one argument (or nothing) and returns K.
I
would have expected this to fail to match the function declaration.
Well, that's usually due to one's lack of experience with declarations.
No big deal, we are all learning.

A clearer repost of the code without the copy-construction and
with the erroneous line-wrapping hopefully sorted is below. If you can
spare a moment to explain just what's going on, in particular on the
highlighted line, we'd both be very happy! Thanks.
// Reposted code:

#include <iostream>

typedef unsigned int U;

class T {
public:
T( const U & ) { }
};

class ROD {
public:
ROD(const T &) { }
void doSomething() const {
std::cout << "Hello -- I am a ROD" << std::endl;
}
};

class CC {
public:
void doSomethingElse () {

ROD rod5( T( u() ) ); // *** line of interest
rod5.doSomethin g(); // doesn't compile since it
// thinks rod5 is a function ptr
// of type ROD ()(T (*)())
Yes, that's the behaviour according to the Standard.

First of all, anything that can be interpreted as a declaration, shall
be interpreted as a declaration. That's the rule. It resolves ambiguity
that otherwise exists in these particular cases.

Second, if it's a declaration, how do you interpret it? Begin from what
looks like a variable name, 'rod5'. It has a parenthesis right next to
it. If what's inside the parentheses is not an expressions, it is very
likely another declaration, then 'rod5' is a function. So, let's try to
interpret what's inside the parentheses.

You seem to understand the FAQ 10.2, so, I'll build my explanation based
on that. Imagine that instead of

int a;

you have

int (a);

Does it change anything? No. The parentheses surrounding the variable
name are superfluous in that case, and change nothing. So

int (a);

is equivalent to

int a;

.. If that's so, then I can always remove the parentheses that follow the
type name and the declaration will mean the same, right? Now, if I say

int(a());

what is it? Nothing else but

int a();

(after removing the top-level parentheses). Now replace 'int' with 'T'
and 'a' with 'u'. What do you get?

T(u());

What is it? A declaration of 'u', that is a function that takes no args
and returns an value of type T. Now simply put that declaration inside
the parentheses of a function declaration and you get

ROD rod5(T(u()));

which is equivalent to

ROD rod5( T u() );

which means 'rod5' is a function that takes one argument of type [pointer
to function that takes no arguments and returns T] and returns ROD.

* * *

Now, the question is, "how do we fix this"? This is one way:

ROD rod5 = ROD(T(u()));

which would generate an error because 'u' is undefined. You probably
meant

ROD rod5 = ROD(T(U()));

Another way is to use real objects of type U and T:

U u;
T t(u);
ROD rod5(t);
}

U u() const { return 0; }
};

int main () {
CC cc;
cc.doSomethingE lse();
return 0;
}

Victor

[Christopher Lester]

Hi Victor,

(by the way, I'm the friend Andy referred to)

First I'd like to thank you for your most recent and fullest answer
to this thread. It has been very helpful in clarifying our original
misunderstandin g, and I've definitely learned some things.

If nothing else happens to this thread I will consider it usefully
closed.

Nevertheless, I don't think any harm will come if I try to paraphrase
what you have already said (a) to make sure that we have understood
everything correctly, and (b) to check that a small (almost
throw-away) remark you made near the foot of the last message is
indeed nothing more than a simple slip -- and not something more
important.

So here goes my paraphrasing:

The first and most important thing you said (I'll call it (1)) is:

First of all, anything that can be interpreted as a declaration,
shall be interpreted as a declaration. That's the rule. It
resolves ambiguity that otherwise exists in these particular cases.
The above comment particularly relevant to the line I have
labelled (3) below:

class CC {
void doSomething();
U u(); // (2)
};
void CC::doSomething () {
ROD rod5(T(u())); // (3)
};

which was found within the definition of a method of a class CC that
ALSO DEFINED A METHOD u() (labelled (2) above). [ In the above
fragment I have separated the definition of the doSomething method
from its declaration for reasons that will hopefully become clearer
shortly. ]

The other thing you carefuly explained (but which I don't copy here)
was to explain how a line like (3) could correctly be interpreted IN
THE ABSENCE of a declaration such as (2) as a function declaration.

Our original mistake was to believe that the declaration of u() in
(2) would (within the scope of definitions of methods of CC such as
in location (3)) be sufficient to prevent it later being interpreted
as a declaration for something else.

Your comment (1) tells us WHY we were wrong:

Your comment (1) tells us that although (2) declared our "intention"
(whatever that means!) for how we wanted "u()" to be interpreted, the
mere fact that it COULD be interpreted in (3) as a function
declaration was sufficient to REQUIRE it to be interpreted so. Fair
enough. There must, of course, be a rule.

So if everything I have said above is correct, then I am now a happy
man as I now understand the language a bit better!

Thanks Victor.

I only re-iterate the above because of the throw-away remark I
alluded to earlier that we found at the end of your last mail which
casts just a tiny doubt in my mind, but which probably signifies
nothing much:
Now, the question is, "how do we fix this"? This is one way:
ROD rod5 = ROD(T(u())); // (4)
which would generate an error because 'u' is undefined.

Now you can see why I (in this posting) separated the implementation
of CC::doSomething () from its declaration in the class definition --
because this way I think it's easier to see that 'u' (or at least
CC::u) really IS defined at the point at which the compiler tries to
attack the internals of doSomething. Or at least we intended it to
be defined! As far as I know, the status of the definition of CC::u
would not be affected by moving the implementation of doSomething to
within the class body (even if doSomthing happened to be declared
ahead of u()), and so I presume that what I've said here also applies
equally to what we expected from our last posting.

So I'm going to assume that you simply didn't spot the u()
declaration in our CC class (an understandable mistake to make given
how we managed to mess up the original posting in more than one
way!), and that as a result the status of (4) moves from failing to
fix the problem, to actually being a good problem fix!

Which is a good thing!

So again, thank you for your time.

Christopher

[Victor Bazarov]

"Christophe r Lester" <le****@hep.phy .cam.ac.uk> wrote...

[...]
You got it right.
I only re-iterate the above because of the throw-away remark I
alluded to earlier that we found at the end of your last mail which
casts just a tiny doubt in my mind, but which probably signifies
nothing much:

Now, the question is, "how do we fix this"? This is one way:
ROD rod5 = ROD(T(u())); // (4)
which would generate an error because 'u' is undefined.
Now you can see why I (in this posting) separated the implementation
of CC::doSomething () from its declaration in the class definition --
because this way I think it's easier to see that 'u' (or at least
CC::u) really IS defined at the point at which the compiler tries to
attack the internals of doSomething. Or at least we intended it to
be defined! As far as I know, the status of the definition of CC::u
would not be affected by moving the implementation of doSomething to
within the class body (even if doSomthing happened to be declared
ahead of u()), and so I presume that what I've said here also applies
equally to what we expected from our last posting.

So I'm going to assume that you simply didn't spot the u()
declaration in our CC class (an understandable mistake to make given
how we managed to mess up the original posting in more than one
way!), and that as a result the status of (4) moves from failing to
fix the problem, to actually being a good problem fix!

That's the ticket. I didn't pay too close attention to the fact
that 'u' was another member function. I thought (incorrectly) that
"u()" was a slip-up by Andy, who meant to write "U()".

Which is a good thing!

Depends on how you look at it, don't it?

V