Automatic using ::rtl::OUString etc.

[Stephan Bergmann]

On 09/18/2012 12:43 PM, Norbert Thiebaud wrote:
> On Tue, Sep 18, 2012 at 3:38 AM, Stephan Bergmann <sbergman at redhat.com> wrote:
>> It turns out that one drawback of this is with header files that are careful
>> to only declare the incomplete type via
>>
>>    namespace rtl { class OUString; }
>>
>> (instead of including rtl/ustring.hxx) if that is all they need.  They would
>> either need to continue using "rtl::OUString" (instead of just "OUString"),
>> or need to duplicate the using declaration (which could be considered
>> breaking of encapsulation), or include rtl/ustring.hxx instead.
>
> here is (see bottom of the message) a list of cxx file that do
> _not_include ustring.hxx (linux build... so there may be a handfull of
> platform specific others)
> excluding sal and autodoc, that is 236 files... out of 8013 files...
> so 97% of the cxx file of the project already include ustring.hxx (and
> therefore due to the inclusion chaine also include at a minimum:
>   * sal/types.h
>   *  sal/config.h
>   *  sal/macros.h
>   *  sal/detail/log.h
>   *  sal/log.hxx
>   * rtl/ustring.h
>   *  osl/interlck.h
>   *  rtl/string.h
>   *   rtl/textcvt.h
>   *    rtl/textenc.h
>   * rtl/string.hxx
>   *  osl/diagnose.h
>   *  rtl/memory.h
>   *  rtl/stringutils.hxx
>
> so... out of the 163 header  ( git grep -l "namespace rtl *{ *class
> OU" | grep -v OUStringBuffer | wc -l )  that use that namespace rtl {
> class OUString; } construct... most of the time they are trying to
> avoid a #include that happen anyway.

But the relevant data would rather be how many of those cxx files that 
do include rtl/ustring.hxx do so unnecessarily (and are thus needlessly 
recompiled when rtl/ustring.hxx changes -- which it happens to do a lot 
these days).

Granted, the entities declared in rtl/ustring.hxx are used so widely 
that the number of cxx files that do not need to include it is probably 
rather small.  Still, I consider it good practice in general to use the 
idiom of only declaring incomplete types in headers where appropriate. 
(And blame not following that practice as one reason for needlessly long 
rebuild times.)

Stephan