Add parent packages into dependency tree

[kubajj]

This PR adds an array of indices of parent nodes to each node of the dependency tree. It is part of my Handling compiler arguments GSoC 2021 project and was separated from #498 for easier review.

[awvwgk]

Thanks, I left a few comments.

I'm not sure whether the parent list should actually be optional in the cache. Without requiring the parent list an existing cache could be loaded and might not contain this information. Instead we should require the data and invalidate the cache if the parent list is not present.

Also, we should add a unit test to make sure the parent list is written and read correctly from the cache.

[kubajj]

I'm not sure whether the parent list should actually be optional in the cache.

I think that as main package does not have any parents, wouldn't this cause an issue?

Without requiring the parent list an existing cache could be loaded and might not contain this information. Instead we should require the data and invalidate the cache if the parent list is not present.

Should I call it with requested=.true. or just check the stat and raise an error if it is not toml_stat%success?

[awvwgk]

I think that as main package does not have any parents, wouldn't this cause an issue?

The list would be empty in this case. For a cyclic dependency the main project might as well have a parent.

Should I call it with requested=.true. or just check the stat and raise an error if it is not toml_stat%success?

This is more difficult, in principle we could just request it but this potentially result in no connections in case of an old cache. Therefore, best is to not request it and create an error if it is not associated. This should automatically invalidate the cache and fpm will recreate it by building the dependency tree again.

[kubajj]

Should I call it with requested=.true. or just check the stat and raise an error if it is not toml_stat%success?

This is more difficult, in principle we could just request it but this potentially result in no connections in case of an old cache. Therefore, best is to not request it and create an error if it is not associated. This should automatically invalidate the cache and fpm will recreate it by building the dependency tree again.

Should this really be an error? I thought that we should just dump the cache and build the dependency tree on our own if the parent table is missing.

[awvwgk]

Generating an error while reading the cache is okay. An error signals an invalid cache file and the caller is responsible for handling the error by invalidating the cache file (i.e. deleting it).

[kubajj]

Generating an error while reading the cache is okay. An error signals an invalid cache file and the caller is responsible for handling the error by invalidating the cache file (i.e. deleting it).

So I should modify the add_project subroutine to delete cache if self%load fails?

[awvwgk]

This should already be the default behaviour:

fpm/src/fpm/dependency.f90

Lines 237 to 240 in d72d953

	if (allocated(self%cache)) then
	call self%load(self%cache, error)
	if (allocated(error)) return
	end if

Right this is not done yet. add_project would be the right place to fix this.

[kubajj]

This should already be the default behaviour:

fpm/src/fpm/dependency.f90

Lines 237 to 240 in d72d953

	if (allocated(self%cache)) then
	call self%load(self%cache, error)
	if (allocated(error)) return
	end if

Right this is not done yet. add_project would be the right place to fix this.

Can I create an empty toml array? I am struggling to pass ci tests for example packages with just the main package.

[awvwgk]

After looking into this for a while I wonder why have to save the parents of each dependency. Saving the dependees of each dependency is unpractical as adding a dependency requires to potentially modify all nodes. Instead saving the dependencies of each dependency, i.e. the child nodes, keeps the information about a dependency local in each node.

In the context of applying profiles we should be able to walk the tree from the main project while applying either the parent profile or overwriting it with a profile from the current project.

cc @LKedward @everythingfunctional

[LKedward]

Apologies for the late notice Sebastian @awvwgk, but are you available tomorrow at 3PM UTC to join our mentor chat? It might be quicker to discuss this altogether.

[awvwgk]

Sure, should be able to join, I'll make a note for tomorrow 3PM UTC/5PM CEST.

[awvwgk]

As discussed I will take over the implementation of the retaining the topology of the dependency tree and implementing a mechanism to transverse the tree later when building up the model (see #355 for more details).