Is Clarkhat Best Hat?

Y/N: Y

nnCompatTrampoline is a beautiful name for a shared library

yeah that doesn't make any sense to me. I'm probably missing something here.

For enums (sum types) what you want is called a prism, for structs (product types) it's a lens. There's a lot of options here, just depends on how much type-safety/convenience you want.

Here are some options I found:

- docs.rs/serde_json/l...
- use the existing API and ? away the fallible parts between .as_object etc. calls
- docs.rs/json-patch/l...
- docs.rs/smart_access... is the freshest lens alternative I can find. May need to add serde_json integration.

The thing lens would make easier is it's designed to enable setters in addition to getters. I can't see anything about "setters" in serde_json_path so that might be the gotcha.

You could try using something like lens for this but it isn't likely to be easier than serde_json_path.

You can still ask for the borrow checker's assistance in unsafe code in some instances as well. I agree with you, It's a lot better than writing C.

I also didn't feel like I had a very chunky test input for the benchmark and I feel like I could've done more on the file input & parsing side. I wanted to avoid the redundant iteration of the .lines() for width & height.

After I applied the optimizations they were both at 155 microseconds, but the recursive one is maybe at risk of blowing the stack on a larger problem input. I could've taken it further but ran out of gas/time. Would've entailed getting rid of all the hashmaps, etc.

iterative version was spilling registers more between loop iterations in the assembly as well

memory access in the recursive version might've been more linear and localized, iterative was more scattershot initially. row major vs. column major impacts this.

The iterative version was eating it harder on the Vec reallocation initially I think. It seemed like it had more calls out to the hashing stuff as well but I'm not confident in that.

Pre-allocating the data structures made the biggest diff but I got iterative from 280 -> 155 micros on 12900K.

changing point's field type to i32 instead of isize improved both down to around 150-160 microseconds

row-major indexing in iterative's get_unchecked improved it 3.7%, neck-in-neck now.