On Tue, 30 Jun 2009 11:00:47 +0200, Kristian Nielsen
<knielsen@stripped> wrote:
> Tor Didriksen <Tor.Didriksen@stripped> writes:
>
>> On Mon, 29 Jun 2009 12:55:45 +0200, Kristian Nielsen
>> <knielsen@stripped> wrote:
>
>>> Mats, could you describe what the benefits are of using forward
>>> declarations
>>> of structs/classes in place of #including the definition?
>
>> Please see
>>
> http://google-styleguide.googlecode.com/svn/trunk/cppguide.xml#Header_File_Dependencies
>> for an explanation.
>>
>> If a class 'Bar' is implemented in terms of another class 'Foo', or
>> uses it's interface in some way,
>> it will typically store a pointer to it (in its private section).
>> By forward declaring the class
>> class Foo;
>> rather than
>> #include "foo.h"
>> you can often drastically cut down on the number of header files
>> included.
>
> Ok, so to reduce compile time.
>
>> The savings in compile time can be quite substantial.
>> I remember doing this on a project sometime in the 90's cutting down
>> compile/build time from a day (yes 24 hours) to about one hour.
>
> If you expect this kind of improvements on today's computers you may be
> disappointed. CPUs are _much_ faster, C++ compilation can be parallelised
> across multiple cores (as opposed to eg. linking), and I suspect
> compilers
> these days spend proportionally less time on parsing #include compared
> to code
> generation.
>
> Just for fun, I did some quick benchmarking on building the latest mysql
> 5.1
> bzr tree on a cheap 2-year old quad-core. First, compiling one file in
> sql/
> vs. compiling all:
>
> `touch sql/mysqld.cc ; make -j5` -> 11.35 seconds, 121% CPU.
>
> `touch sql/*.cc ; make -j5` -> 120.87 seconds, 352% CPU.
Our numbers are pretty much the same aren't they?
Roughly an order of magnitude difference.
I don't mind waiting 10 seconds for a build.
With two minutes, I'll take up juggling again :-)
-- didrik
>
> So that is a factor of only 10 between compiling one and all C++
> sources. Of
> course that is still something if we can get down to only one (or a few)
> files
> #including something vs. all doing so.
>
> Second, I tried adding #include <elf.h> (random 100k bytes, 2.5k lines
> file)
> at the start of every .cc file in sql/, and timing `make -j5` after
> `make clean`:
>
> Without #include <elf.h>: 5 min 14 sec
>
> With #include <elf.h>: 5 min 17 sec
>
> So not much difference there (may even be less than the statistical
> noise).
>
> Anyway, thanks for the explanation,
>
> - Kristian.
>
