scrypt is extremely slow compared to native C bindings

[koraykoska]

I benchmarked CryptoSwift scrypt vs this library: https://github.com/greymass/swift-scrypt which is a C binding to libscrypt.

libscrypt scrypt seems to be 4 times faster than CryptoSwift scrypt.

Both have been release compiled with optimizations.

After time profiling I found out which code segments are the issue:

This function takes up about 90% of the total time:
https://github.com/krzyzanowskim/CryptoSwift/blob/main/Sources/CryptoSwift/Scrypt.swift#L206

More specifically, this rotateLeft uses up again more than 90% of the salsa20_8_typed function:
https://github.com/krzyzanowskim/CryptoSwift/blob/main/Sources/CryptoSwift/Utils.swift#L18

The fun bit is that the 1 line of code in rotateLeft is basically exactly the same in the C code in libscrypt.

Has anyone an idea what's causing this slowdown? It's actually so slow that even the JS implementation is faster. Something is up with it but I don't know what it is. Right now it's not usable in production.