Why the compiler dosen't enforce correct module declarations?

[Bruno Medeiros]

Sorry for the long delay in this posts.

gareis wrote:
> == Quote from Bruno Medeiros (brunodomedeiros+spam at com.gmail)'s article
>> Sean Kelly wrote:
>>> Ary Manzana wrote:
>>>> Shouldn't the compiler say "Wait, you are saying that module one is in
>>>> file main.d, module two is in other.d and module lala.la is in
>>>> dir/some_other.d, this isn't quite well"?
>>> I've decided that this is actually a good thing.  It allows large
>>> modules to be split across multiple files and for different
>>> implementation files to be chosen at compile time when using the
>>> header/source model.  It's also the only way I've found to compile the
>>> implementation of "object.d" for Tango.  Actually naming it "object.d"
>>> caused all sorts of problems.
>>>
>>>
>>> Sean
>> But:
>> A) "It allows large modules to be split across multiple files"
>> What do you mean? Splitting a large module into multiple files implies
>> putting each new file in a different module. What does that have to do
>> with allowing module-file name mismatches?
> 
> The first part is mostly naming convention. Though if you can't possibly imagine
> using one file without having the exact source of another file, they should be
> one module. You could combine the files, but that tends to work poorly after a
> while.
> 
> Still, I'd use imports most of the time, if it were possible.
> 
> For the second part, Mr. Kelly gave an example: his object module couldn't
> appear in object.d without causing issues.
> 

Yes, the object module example is a valid use case for this behavior. 
But it is a very special case (hell, it's *unique*), which doesn't by 
itself justify module-file name mismatches.

>> B) "and for different implementation files to be chosen at compile time
>> when using the header/source model."
>> And why is that a good thing? If I use conditional compilation and
>> define conditional properties at compile time I achieve the same effect.
> 
> You can wrap your entire file in version statements, but then it's difficult to
> work with. If you have multiple people working on the functions for different
> versions, you might get conflicts with cvs / subversion / what-have-you
> (especially with coding style and editor issues). And IDEs might barf on entire
> functions wrapped in version statements and appearing multiple times.
> 

Yes, that is not a good solution.

> You could wrap all the import statements in version statements, but that's ugly
> and doesn't scale that well. Modules that aren't platform-specific or
> feature-specific will end up having to pay attention to platforms and features.
> 

Yes, that is not a good solution either.

> You could have a module that just has a bunch of statements like:
> ---
> version (Windows) import foo.windows;
> version (Linux) {
>    version (Bigendian) {
>       import foo.linux_bigendian;
>    } else {
>       import foo.linux_little_endian;
>    }
>    version (UsingSSL) import foo.linux_ssl;
> }
> ---
> 
> That's also annoying. And contrived.
> 

Like I said before, that is the alternative that I was thinking, and I 
don't see what's bad with it. Why is it annoying and contrived? The only 
issue I see is the one you mention below:

> Still, if you're using makefiles and autotools, conditional compilation of files
> is the easiest way to go. You can compile each file into a separate object file
> for faster recompilation in the face of changes, and you don't have any issues
> with, for instance, foo.windows panicking because it can't use Windows functions
> on Linux.
> 

Indeed, you would need to specify to the build tool (or IDE) that 
certain modules (like the windows one on linux) are not to be compiled. 
But you have to the same thing with makefiles and autotools, so you're 
not any worse. In fact, a smart IDE (or build tool with semantic 
knowledge) might understand the version statements and automatically 
know which modules to compile or not.
But if you're still not convinced, there is perhaps another alternative. 
Use different files for different implementations, but put each file on 
a different source folder, thus allowing each file to correctly match 
the pack+module names to the directory+file structure:


--- src-linux/foo/bar.d ---
module foo.bar;
...
--- src-windows/foo/bar.d ---
module foo.bar;
...
----

This is how it's done on Java programs with platform specific code(like 
SWT), and should work fine for D *and* be IDE-friendly. This should be 
adequate for big conditional variants, like platform, but no so much for 
small optional features, but you can always use a combination of methods 
for the cases where its not adequate.

-- 
Bruno Medeiros - MSc in CS/E student
http://www.prowiki.org/wiki4d/wiki.cgi?BrunoMedeiros#D