Generator performance improvement

[faddiv]

Hi @latonz,

I'm considering doing some performance improvement on the generation. Are you interested in it?
What my gut feeling, and some performance tests showed, is that the AttributeDataAccessor has the biggest impact on the performance. So I thought I would improve that.
If you already have some idea about improvements, I would like to hear it.
My main idea is to replace the reflections with mapper methods, re-using the property mappings that are there.
Also, I would add caching on this level too, so the complicated processing doesn't need to be repeated. WDYT?

[latonz]

That sounds great! I’ve also noticed that AttributeDataAccessor shows up in profiles, so focusing there makes sense. Replacing reflection with mapper-generated accessors and reusing existing property mappings sounds like a solid direction.
The only part I’m unsure about is how those methods could be generated efficiently — especially since the attribute access patterns can vary quite a bit. Do you already have an idea in mind for how to structure that generation?
There are already some benchmarks in the benchmarks directory that should help measure the impact of these changes. Adding caching on that level also sounds good, as long as incremental builds are handled correctly.

[faddiv]

I thought of creating the mappings by hand. (Or should mapperly depend on itself?)
The first step was just wrapping the generic method calls with concrete type mappings, then replacing them one by one with manual mapping. The pattern emerges, and the already written pieces are reused.