[cmake-developers] Package Config files with COMPONENTS

[Michael Hertling]

On 03/13/2012 06:30 PM, Brad King wrote:
> On 3/13/2012 1:12 PM, Alexander Neundorf wrote:
>> Setting CMP0018 to new would mean that the project supports only those
>> component-aware packages which follow that new convention.
> 
> Yes, that cannot be a requirement to use
> 
>    cmake_minimum_required(VERSION 2.8.8)
> 
> or whatever version of CMake introduces the policy as that would set
> it to NEW implicitly.  We cannot make this a special case because then
> it would simply not be a policy.  It introduces a requirement on the
> CMake version-awareness of a project's dependencies instead of just
> the project.
> 
>  > Some users/developers would prefer if there was a more strict definition of
>  > how the COMPONENTS arguments should be handled.
> 
> My proposal only requires the small change to CMake already under
> discussion and makes it possible for projects to provide package
> configuration files that handle components in a nice way.  It also
> does not preclude the possibility of a future convention.  However,
> any "strict definition" of behavior would *force* every project to
> adapt to it even if it doesn't make sense just because it made
> sense for a few *other* projects.

A further point why Foo_FOUND shouldn't be set by FIND_PACKAGE()
based on the Foo_*_FOUND variables of the mandatory components:

In {FindFoo,FooConfig}.cmake, I use to remove each component from
Foo_FIND_COMPONENTS once it has been processed, so at the end, the
remaining components are the unknown ones. This is quite handy and
makes it easy to provide Foo_*_FOUND==FALSE for the latters.

Even making Foo_FIND_COMPONENTS read-only is not a suitable solution:
IIRC, there is a consensus that {FindFoo,FooConfig}.cmake is expected
to search prerequisite but unrequested components by itself, and such
an unrequested prerequisite is expected to determine Foo_FOUND along
with the other mandatory components. A r/o Foo_FIND_COMPONENTS tends
to be incomplete in this regard.

The conclusion is that Foo_FIND_COMPONENTS is not guaranteed - at
no time - to hold all components intended to determine Foo_FOUND,
so an automatism like the one in 796ad77 is too short-sighted, IMO.

Instead, I'd agree that it should really be left to FooConfig.cmake
completely which value is assigned to Foo_FOUND in the end - unless
FooConfig.cmake doesn't do this or cannot be loaded. There might be
arbitrary findings the author of FooConfig.cmake wants to consider
for this purpose.

Regards,

Michael