Re: [Boost-users] [iostreams] Inconsistency b/w docs and code of stream()

From: Jonathan Turkanis (turkanis_at_[hidden])
Date: 2008-03-05 12:50:55

Sorry it took me so long to reply; I was out of town.

Nitay Joffe wrote:
>> Would you please post the compiler errors?

> Sure, here is sample code that causes the error:
> 
> #include <iosfwd>
> #include <boost/iostreams/categories.hpp>
> #include <boost/iostreams/stream.hpp>
> 
> struct MyDevice {
>   typedef char char_type;
>   typedef boost::iostreams::source_tag category;
> 
>   std::streamsize read(char *s, std::streamsize n) {
>     return n;
>   }
> };
> 
> typedef boost::iostreams::stream<MyDevice> MyStream;
> 
> int main() {
>   MyDevice d;
>   MyStream s(d, (std::streamsize)500);
>   return 0;
> }
> 
> 
> And here are the compile errors:
> 
> g++ a.cc
> k/usr/include/boost/iostreams/stream.hpp: In constructor
> 'boost::iostreams::stream<Device, Tr, Alloc>::stream(U100&, const U0&) [with
> U100 = MyDevice, U0 = std::streamsize, Device = MyDevice, Tr =
> std::char_traits<char>, Alloc = std::allocator<char>]':
> a.cc:18:   instantiated from here
> /usr/include/boost/iostreams/stream.hpp:110: error: no matching function for
> call to 'MyDevice::MyDevice(MyDevice&, const long int&)'
> a.cc:5: note: candidates are: MyDevice::MyDevice()
> a.cc:5: note:                 MyDevice::MyDevice(const MyDevice&)
> make: *** [all] Error 1

Okay, I see the problem now; the compiler is picking up one of the 
templated forwarding constructors instead of the primary constructor, 
since the type of the buffer size is not an exact match.

> I believe the culprit is the file
 > iostreams/detail/push_params.hpp, which defines the buffer and pback size
 > using:
 >
 > #define BOOST_IOSTREAMS_PUSH_PARAMS() \
 >     , int buffer_size = -1 , int pback_size = -1 \
 >     /**/
 >
 > This macro is then used by stream and friends in their constructor
 > declarations. On 64 bit machines std::streamsize is a long int, so the
 > constructor parameters do not match up and compilation fails.

I remember using int here so that users could write

        MyStream(myDevice, 500);

instead of

        MyStream(myDevice, (streamsize) 500);

I guess it didn't occur to me that users would read the documentation 
and cast the parameter to the appropriate type :-)

I documented the parameters as having type streamsize because the 
library uses that type for buffer sizes. I could change the 
documentation to use "int", but I think I'd like to make both types work 
here, so probably it would be best to disable the forwarding 
constructors if the first parameter is an exact match for the template 
parameter T and the second parameter is integral.

I've opened a ticket: 
http://article.gmane.org/gmane.comp.lib.boost.user/33728. Thanks for 
reporting this issue.

 > Thanks,
 > -n

-- 
Jonathan Turkanis
CodeRage
http://www.coderage.com