Fix compaction on nodes

[tenderlove]

This PR is an attempt to fix compaction issues from #2578. We've split the PR in to 3 commits.

The first commit consistently sets a mark function on the xml node. We need this because we need to provide a static struct to TypedData_Wrap_Struct.

The second commit introduces a helper function for unwrapping xml nodes. Unfortunately, xml nodes use a union type, and subclasses want to unwrap the type in to different things, so we had to extract a function that would unwrap the nodes correctly.

The third commit actually fixes compaction by adding a callback for updating references.

Fixes #2578

[flavorjones]

Note: this should probably be backported to v1.13.x in a patch release. I'll do that in a separate PR once this is on main.

[flavorjones]

Sorry for being slow to merge this. I will work on it (and the backport?) this week.

[stevecheckoway]

This looks good. You should even be able to get rid of the casts, but they're not doing any harm.

Edit: I meant the approval to mean the updated callback signatures looked good. I did notice one of the tests shows a memory leak reported by Valgrind.

[flavorjones]

I rebased and made some minor changes.

[flavorjones]

I've added a commit 1c74ba1, which I would like to squash, which uses a macro Noko_Node_Get_Struct to replace the previous Data_Get_Struct macro and noko_xml_node_unwrap() function.

This results in a smaller diff, since Noko_Node_Get_Struct has the same parameters as Data_Get_Struct. It should apply cleanly to the v1.13.x branch when I backport.

The resulting combined diff is here: https://github.com/sparklemotion/nokogiri/compare/b370fbd290343e93f20e6bf12522c5ba3f2f5caf..1c74ba18fccc4e809d62c62f2f8ee536aae3ee94

@tenderlove @eightbitraptor would love a quick review when you have a moment.

[flavorjones]

Added one more commit onto the end to remove RBIMPL_CAST which blew up on windows.

[flavorjones]

@tenderlove I've got one last question for you, which is "why are we using DATA_PTR instead of TypedData_Get_Struct?"

I think I know the answer (but would like confirmation): the alternative of using TypedData_Get_Struct would require creating a new rb_data_type_t for each of the "subclasses" of struct xmlNode, and that's a lot of code for little gain; and DATA_PTR has the advantage of being much less code, and the lack of type safety is OK because libxml2 seems to take care of this reasonably well.

Does that sound about right?

[flavorjones]

The memcheck failure is because ruby_memcheck needs to be told to ignore memory allocated by rb_ary_push, will take care of that tomorrow.

[tenderlove]

Does that sound about right?

Yes, about.

The problem I found is that we have so many objects that inherit from Node. In order to unwrap the object in a method on node, we'd have to check the type of self on every call in order to understand which struct we need to pass to TypedData_Get_Struct.

I'd prefer to use the "typed" unwrappers everywhere, but I couldn't figure out a "proper" way to do that with C objects that use inheritance (without adding tons of checks).

[flavorjones]

@tenderlove Thanks. This is going to go green so I'm going to clean up commit history (squash some of my changes) and let it go through CI one more time.