Good decompression speed up technique: musttail

[safinaskar]

Hi. This is impressive program! Unfortunately, it seems decompressing speed is not so good as zstd's. So I propose this trick:

https://blog.reverberate.org/2021/04/21/musttail-efficient-interpreters.html

This post may be useful too: https://blog.reverberate.org/2020/05/29/hoares-rebuttal-bubble-sorts-comeback.html

[richox]

thanks for the sharing.
if i don't understand wrong, the musttail tech is used to improve the performance of recursions. this is very helpful for rule based parsing (like protobuf). however orz does not have recursive calls neither on encoding nor decoding. so i don't think it can be used to improve our performance.

[safinaskar]

Okay. Still I like your orz a lot. It is very unfortunate that its decoding speed is less than zstd's.

So, here is my advice. Take your main function. As well as I understand it is LZDecoder::decode. Make sure it never calls other functions. Look at its disassembly (for example, using objdump). Make sure disassembly doesn't contain call (on x86). Inline everything. The function should never panic. If you want to panic, signal error other way instead. By returning some error code or by directly executing illegal instruction ud2 ( https://www.stix.id.au/wiki/UD2_Undefined_Instruction_generated_by_Clang/LLVM ). Don't return Box<dyn Error> on error. Box<dyn Error> means you should allocate error, i. e. to call allocation function. Just return error code instead

[safinaskar]

Several days ago I looked at objdump of zstd's decode function. It doesn't touch stack at all! It has no call, push and pop (except in begin and end). That is perfection