simd int and string parsing on aarch64

[samuelcolvin]

Current state of benchmarks:

(this isn't reflect in codspeed as doesn't reflect the SIMD logic)

[codecov]

Codecov Report

Attention: Patch coverage is 89.32039% with 33 lines in your changes are missing coverage. Please review.

Files	Patch %	Lines
src/simd_aarch64.rs	78.57%	30 Missing ⚠️
src/string_decoder.rs	97.82%	2 Missing ⚠️
src/number_decoder.rs	98.70%	1 Missing ⚠️

📢 Thoughts on this report? Let us know!

[codspeed-hq]

CodSpeed Performance Report

Merging #65 will improve performances by 16.05%

_{Comparing int-simd (fc7570a) with main (37f9138)}

Summary

⚡ 3 improvements
✅ 53 untouched benchmarks

🆕 3 new benchmarks

Benchmarks breakdown

	Benchmark	main	int-simd	Change
⚡	bigints_array_jiter_iter	833.8 µs	733.8 µs	+13.62%
⚡	massive_ints_array_jiter_iter	4.3 ms	3.7 ms	+16.05%
⚡	massive_ints_array_jiter_value	5.2 ms	4.7 ms	+11.16%
🆕	short_numbers_jiter_value	N/A	853.9 µs	N/A
🆕	short_numbers_serde_value	N/A	2.8 ms	N/A
🆕	python_parse_massive_ints_array	N/A	7.2 ms	N/A

[davidhewitt]

Heh I have a feeling I know what I'm going to be doing in the future 👀

[samuelcolvin]

No, I'll do it, just need to be in the office, and shouldn't work on it right now.

[davidhewitt]

Is the conclusion here that it's always faster to do the first 18 bytes before launching to SIMD? Does this differ from strings?

[samuelcolvin]

the point is that there are a lot of 1,2,3,4 character integers (mabye 90%+), whereas they're somewhat less common with strings. This improved the performance with those ints, and didn't make the performance any worse for 15 character ints.

Still, maybe it's worth parsing the first 5 characters of a string via a naive iter.

[samuelcolvin]

@davidhewitt I think this is ready to merge.

[davidhewitt]

Looks great, with just one thought...