fix: merge namespaces when collecting results

[Nerixyz]

I'm not 100% sure if this is the correct approach. I would guess that something like this already happens somewhere, but I couldn't find anything.

When using multiple sources (→ multiple extraction jobs), then only the first one will be "visible", because its global namespace/global symbol will be the only one that gets merged with info_.merge(info). All other global namespaces will be discarded, because there's already one present. Their members will still be in the corpus, but there's no path from the global symbol to them.

Because of ODR, only namespaces have to be merged.

[cppalliance-bot]

An automated preview of the documentation is available at https://896.mrdocs.prtest2.cppalliance.org/index.html

[alandefreitas]

This looks very weird to me.

When using multiple sources (→ multiple extraction jobs), then only the first one will be "visible"

If by "visible" you mean copied into info_, then no. info_.merge(info); is called on all infos for all TUs. Any new symbols will be added there. The loop is only for symbols that couldn't be included in the info_ because they're duplicates. Then we have the merge algorithm for this.

its global namespace/global symbol will be the only one that gets merged with info_.merge(info)

Not really. The second, third, ... is supposed to be merged with merge(**it, std::move(*other));.

Their members will still be in the corpus, but there's no path from the global symbol to them.

merge(**it, std::move(*other)); is supposed to create this "path" (include the members in the well... members of the namespace). Even if merge(**it, std::move(*other)); didn't do that, I'm not sure how this casting would change things unless the implementation of merge is wrong.

[Nerixyz]

Even if merge(**it, std::move(*other)); didn't do that, I'm not sure how this casting would change things unless the implementation of merge is wrong.

Then the implementation of merge might be wrong. In the execution context, merge(Info& I, Info&& Other) is called, which doesn't perform any downcasting, it only upcasts to SourceInfo. merge(NamespaceInfo& I, NamespaceInfo&& Other) is never called from there. In fact, it would merge the infos (if called).

I think the intention was to call the most specific merge function. One could either turn merge into a virtual function or use visit:

visit(**it, [&]<typename T>(T& target) {
    auto *source = dynamic_cast<T*>(other.get());
    MRDOCS_ASSERT(source);
    merge(target, std::move(*source));
});

^ this might be the intention here.

[alandefreitas]

Exactly!

[cppalliance-bot]

An automated preview of the documentation is available at https://896.mrdocs.prtest2.cppalliance.org/index.html

[alandefreitas]

Now, this is good to go, right? I'll merge this one first because whatever I merge, the other needs to be rebased, and this one is more important.

[Nerixyz]

Thank you!