why this testcase compile failed for gcc.

why this testcase compile failed for gcc.

[zhang qingshan]

/* ------- test case -------------------*/
void x();
template <typename T>
void foo(const T*);

int main() {
   foo(x);
}

GCC 4.5 complains:

a.cpp: In function 'int main()':
a.cpp:6: error: no matching function for call to 'foo(void (&)())'

It seems that, const T * is resolved as void (&)());

IMO, T --> void (), const T --> void (), const T * --> void (*)(), and
it should be leagle.

Thanks

Re: why this testcase compile failed for gcc.

[Ian Lance Taylor]

zhang qingshan <steven.zhang54373@gmail.com> writes:

> /* ------- test case -------------------*/
> void x();
> template <typename T>
> void foo(const T*);
>
> int main() {
>    foo(x);
> }
>
> GCC 4.5 complains:
>
> a.cpp: In function 'int main()':
> a.cpp:6: error: no matching function for call to 'foo(void (&)())'
>
> It seems that, const T * is resolved as void (&)());
>
> IMO, T --> void (), const T --> void (), const T * --> void (*)(), and
> it should be leagle.

You are confusing function pointers with pointers to data objects.  They
are not the same thing in C++.  That is, a function pointer is not a
special type of pointer.

Ian

Re: why this testcase compile failed for gcc.

[Marc Glisse]

On Fri, 18 Feb 2011, Ian Lance Taylor wrote:

> zhang qingshan <steven.zhang54373@gmail.com> writes:
>
>> /* ------- test case -------------------*/
>> void x();
>> template <typename T>
>> void foo(const T*);
>>
>> int main() {
>>    foo(x);
>> }
>>
>> GCC 4.5 complains:
>>
>> a.cpp: In function 'int main()':
>> a.cpp:6: error: no matching function for call to 'foo(void (&)())'
>>
>> It seems that, const T * is resolved as void (&)());
>>
>> IMO, T --> void (), const T --> void (), const T * --> void (*)(), and
>> it should be leagle.
>
> You are confusing function pointers with pointers to data objects.  They
> are not the same thing in C++.  That is, a function pointer is not a
> special type of pointer.

Well it still kind of is. Without the "const", the code compiles fine. But 
the fact that a const qualifier on a function type is ignored seems not to 
apply to template type deduction.

(note that most compilers behave like g++ here, except for clang which 
accepts the original code)

-- 
Marc Glisse

Re: why this testcase compile failed for gcc.

[zhang qingshan]

Thanks, Marc. Please consider this test code:

void f();

template <typename T> void foo(const T);

void Test()
{
  foo(f);
}

int main(Test());

It compiles sucessfully for GCC 4.5. As you know, we are trying to add
cv-qualifier on top of function type while template deduce happens.

On Fri, Feb 18, 2011 at 11:38 PM, Marc Glisse <marc.glisse@inria.fr> wrote:
> On Fri, 18 Feb 2011, Ian Lance Taylor wrote:
>
>> zhang qingshan <steven.zhang54373@gmail.com> writes:
>>
>>> /* ------- test case -------------------*/
>>> void x();
>>> template <typename T>
>>> void foo(const T*);
>>>
>>> int main() {
>>>   foo(x);
>>> }
>>>
>>> GCC 4.5 complains:
>>>
>>> a.cpp: In function 'int main()':
>>> a.cpp:6: error: no matching function for call to 'foo(void (&)())'
>>>
>>> It seems that, const T * is resolved as void (&)());
>>>
>>> IMO, T --> void (), const T --> void (), const T * --> void (*)(), and
>>> it should be leagle.
>>
>> You are confusing function pointers with pointers to data objects.  They
>> are not the same thing in C++.  That is, a function pointer is not a
>> special type of pointer.
>
> Well it still kind of is. Without the "const", the code compiles fine. But
> the fact that a const qualifier on a function type is ignored seems not to
> apply to template type deduction.
>
> (note that most compilers behave like g++ here, except for clang which
> accepts the original code)
>
> --
> Marc Glisse
>

Consider this test case:

void f();
template <typename T> foo(const T&);
void Test()
{
foo(f);
}

Re: why this testcase compile failed for gcc.

[zhang qingshan]

IMO, function pointer is a pointer which its type descriptor is a
function type ...

Any difference with the normal pointer?

On Fri, Feb 18, 2011 at 11:00 PM, Ian Lance Taylor <iant@google.com> wrote:
> zhang qingshan <steven.zhang54373@gmail.com> writes:
>
>> /* ------- test case -------------------*/
>> void x();
>> template <typename T>
>> void foo(const T*);
>>
>> int main() {
>>    foo(x);
>> }
>>
>> GCC 4.5 complains:
>>
>> a.cpp: In function 'int main()':
>> a.cpp:6: error: no matching function for call to 'foo(void (&)())'
>>
>> It seems that, const T * is resolved as void (&)());
>>
>> IMO, T --> void (), const T --> void (), const T * --> void (*)(), and
>> it should be leagle.
>
> You are confusing function pointers with pointers to data objects.  They
> are not the same thing in C++.  That is, a function pointer is not a
> special type of pointer.
>
> Ian
>

Re: why this testcase compile failed for gcc.

[Jonathan Wakely]

On 21 February 2011 06:31, zhang qingshan wrote:
> IMO, function pointer is a pointer which its type descriptor is a
> function type ...
>
> Any difference with the normal pointer?

An lvalue of function type T can be implicitly converted to T*.

That is not true for non-function types.

Re: why this testcase compile failed for gcc.

[Ian Lance Taylor]

zhang qingshan <steven.zhang54373@gmail.com> writes:

> IMO, function pointer is a pointer which its type descriptor is a
> function type ...
>
> Any difference with the normal pointer?

The C standard does not require that a function pointer use the same
representation as any other type of pointer.  In particular, on
processors with Harvard architectures, function pointers and regular
pointers are inherently different.

Ian

Re: why this testcase compile failed for gcc.

[Andi Hellmund]

Hey
> The C standard does not require that a function pointer use the same
> representation as any other type of pointer.  In particular, on
> processors with Harvard architectures, function pointers and regular
> pointers are inherently different.
>    
@Ian: do you have a reference section where this is described. I shortly 
tried to search for a few keywords in the standard, but couldn't find 
the obvious. I would just be interested in the details ...

@Steven: to give you a practical example: HP-UX on Itanium for example 
uses two different representations for data pointers and functions 
pointers: while data pointers are usual 64-bit addresses (purely 
assuming 64-bit compilation mode), function pointers however are 
pointers to a data structure called function description which contains 
two further pointers: (a) the final target address and (b) a so-called 
global data pointer for relative data addressing. Though if you would 
print out a function pointer on HP-UX, you would also get a single 
address, but it is NOT the address of the function to be called. This 
example is not as good as Ian's one, but is somehow similar, I think.

Best regards,
Andi

Re: why this testcase compile failed for gcc.

[Jonathan Wakely]

On 22 February 2011 18:45, Andi Hellmund wrote:
> Hey
>>
>> The C standard does not require that a function pointer use the same
>> representation as any other type of pointer.  In particular, on
>> processors with Harvard architectures, function pointers and regular
>> pointers are inherently different.
>>
>
> @Ian: do you have a reference section where this is described. I shortly
> tried to search for a few keywords in the standard, but couldn't find the
> obvious. I would just be interested in the details ...

6.2.5 paragraph 26 covers representation.

6.3.2.3 paragraphs 1, 7 and 8 cover allowed conversions between
pointer types. A pointer to function type cannot be converted to a
pointer to object or incomplete type.