Make Obj trivial and add a separate ObjCollectionParent type

[tgoyne]

Adding a base class and virtual functions to Obj turned out to have a measurable performance impact due to that it inhibits a lot of optimizations that are possible for trivial types (10-15% in the obj-c read benchmarks, which includes some SDK overhead so the slowdown in the core part is bigger). Moving the implementation of CollectionParent to a separate adaptor type fixes this, but requires some code movement.

I found the logic around when collections need to update difficult to follow and so attempted to simplify it. update_if_needed() did meaningfully different things on collections and Obj, with Obj::update_if_needed_with_status() being equivalent to Collection::update_if_needed(). I renamed these to make Obj match collection and use checked_update_if_needed() for the version that throws when detached. A lot of the updating logic was duplicated between the various collections, so I consolidated the parts which logically should always be identical.

All of the tests passed with clearly incorrect handling of m_parent_version (such as never setting it at all on ObjCollectionParent), which made it hard to tell what it was even needed for. I shifted it to only ever being incremented in Obj and then propagated downwards, which makes it harder to mask bugs. Making it uint32_t cuts the size of Obj by 8 bytes on 64-bit platforms since there's otherwise 7 bytes of trailing padding; it's unclear if this is actually useful.

[coveralls-official]

Pull Request Test Coverage Report for Build thomas.goyne_232

Details

• 496 of 559 (88.73%) changed or added relevant lines in 21 files are covered.
• 848 unchanged lines in 34 files lost coverage.
• Overall coverage increased (+0.9%) to 91.787%

---

Changes Missing Coverage	Covered Lines	Changed/Added Lines	%
src/realm/collection.hpp	18	19	94.74%
src/realm/dictionary.cpp	13	14	92.86%
src/realm/obj.cpp	99	100	99.0%
src/realm/object-store/impl/results_notifier.cpp	9	10	90.0%
src/realm/collection_parent.cpp	49	52	94.23%
src/realm/obj.hpp	41	47	87.23%
test/test_list.cpp	152	166	91.57%
src/realm/object-store/impl/collection_notifier.cpp	29	65	44.62%

Files with Coverage Reduction	New Missed Lines	%
src/realm/collection.hpp	1	87.28%
src/realm/list.cpp	1	89.17%
src/realm/list.hpp	1	85.13%
src/realm/obj.hpp	1	90.1%
src/realm/sync/network/network.cpp	1	89.86%
src/realm/table_ref.hpp	1	55.81%
src/realm/set.hpp	2	78.28%
src/realm/sync/network/http.hpp	2	82.41%
test/test_lang_bind_helper.cpp	2	94.24%
src/realm/obj.cpp	3	91.21%

Totals
Change from base Build 2115:	0.9%
Covered Lines:	242761
Relevant Lines:	264484

---

💛 - Coveralls

[jedelbo]

This PR contains 4 different commits, where only the last one is about making Obj trivial, so perhaps this should have been handled in 4 different commits. I will try to address each commit individually:

Less data:
Looks good.

Data race:
Looks good.

Consolidate ...:
You are right that this area has been quite confusing - for me at least since #4057. You are right that it is only neccesary to increment the parent_version in Obj. I think a central case to test is when 2 collection accesors share a common parent collection. I have made an update to the test in #7446. Feel free to merge and/or modify.
Otherwise fine.

ObjCollectionParent:
Looks good

[jedelbo]

Seems like the only reservation I have left is if the test should be updated as indicated in #7446

[tgoyne]

I've expanded the tests to include the scenario tested in #7446 plus a few other things. This ended up revealing a pre-existing bug (accessors for collections with depth > 1 are considered not up to date after mutating via that accessor, and so update every time) that I'll fix in a subsequent PR.