Instance Slices (+29%)

[krakow10]

Push instances into a regular ol' Vec, and keep track of the inserted range in TypeInfo. This means that all instances of a specific class are in a contiguous slice, and are also in the correct iteration order expected by the decode_prop_chunk loops. No more instances_by_ref.get_mut(referent).unwrap(). A HashMap tracks which i32 ref id corresponds to which instance using indices.

There's probably some more optimizations hiding in DeserializerState::finish. The algorithm is currently using Vec::swap_remove to remove the elements from the instances list. (swap_remove removed in 695e88e) The instances_to_construct machinery can probably be combined with the instances_by_ref and PRNT chunk decode in some clever way to do even less work.

I wrote an unsafe version that avoids swap_remove here: 3df205b but there doesn't seem to be a significant impact. I wrote a safe version using a similar technique. Theoretically it should be faster than the swap_remove method, but I haven't put it to the test.

Performance

I observe a 29% improvement in throughput in the Miner's Haven deserialize benchmark.

Lore

The instances could be owned by type_info directly, I was able to run type_ids in parallel with rayon (branch) using this method. I got a 1% speedup!!! Wow!!! Probably because it spent the entire time waiting for 36000 Parts to decode while every other class only has <2000 instances. The InstanceKey technique would still work, but removing each instance in finish() would be triple indirection (key, type_id, index) instead of double indirection (key, index).

I was using the indexmap crate for the Vec + HashMap abomination several times, but it's better to implement separately because of the separable mutable + immutable aliasing over the Vec + HashMap parts.

This has three previous incarnations:

• v1 krakow10/rbx-dom@categorize-chunks...categorize-chunks2: Store instances in TypeInfo. The problem with this is it's based on the nasty categorize-chunks branch. At least I found the performance I was looking for.
• v2 krakow10/rbx-dom@master...type_info-instances: better, but the design is recollecting instances twice - once into instances_by_ref and then again into the dom.
• v3 krakow10/rbx-dom@typestate-de...instance-slices: A breakthrough design using a Vec to back the instances to avoid recollecting twice. I thought the typestate pattern would make it easier to implement, but it turns out it was my past failings (v1 and v2) that were calling out for typestate.