Re: Improved linking mechanism

From: David Abrahams (dave_at_[hidden])
Date: 2003-10-20 10:00:37

Vladimir Prus <ghost_at_[hidden]> writes:

> David Abrahams wrote:
>
>>
>> I never understood the meaning of <use>. I still don't.
>
> <use>some_lib
>
> means: "grab usage requirements" from some_lib. Nothing else. "some_lib" won't 
> be linked to the target which contains this property.

OK, that's very logical, thanks. It's not in the docs, AFAICT. I
believe we need an index of built-in features and their meaning. I
think these occur in several layers, and features like <use> are in
the core layer because they have a fundamental effect on the
operation of the build system.

>> > Now that you've asked, I am not sure bubble-up linking should be in
>> > docs.
>>
>> Definitely not. Documenting it as a separate concept in linking.html
>> on equal footing with "automatic" linking threw me off, for sure.
>
> Ok, I see.
>
>> Especially if the ability to write
>>
>> lib a : a.cpp : <use>b : : <library>b ;
>>
>> simply "falls out" of other information you're going to document, it
>> would be better for people to discover it that way.
>
> I don't understand what's "discove it that way" means. Sorry.

I mean, users can read the documentation about the meaning of <use>
and <library>, and discover that combining them in that way produces
the result they want. On the other hand...

>> I think you *should* document the behavior of lib targets in the
>> sources of other libraries, and doing it in terms of the line above
>> is perfectly fine.
>
> Okay.

...if you do the above, it will be very evident, so there's not much
to "discover" ;-)

>> IMO it also makes sense for this functionality to be somehow looked up
>> as a property of the "b" target rather than being special-cased.
>
> The current idea is this: you're trying to create static lib "a" and
> have "b" in sources. The generator for "a" cannot consume "b", so
> "b" is returned together with whatever targets are produced for "a".

Right, and... dynamic libraries can be consumed directly, so *they*
just do the right thing as well?

> The only special-casing I see here is that before "a" is
> constructed, all <library>properties are converted into sources, so
> that they are passed to the generator. This approach seems logical
> for me.

Fair enough.

I guess you might also want a way of saying which kinds of targets
each target type "expects" to fail to consume, so that we can report
unintentional consumption failures. For example:

expect-not-consumed STATIC_LIB : STATIC_LIB ;

or something similar.

-- 
Dave Abrahams
Boost Consulting
www.boost-consulting.com