The constant booleans in the backends should be evaluated at compile-time... but there's a slight chance that I could write a test for the ChaCha20Rng that fails on aarch64 with neon. The original add64 macro does what the name suggests, it adds vectors in a 64-bit fashion, meaning that it could perform addition incorrectly when the counter should be 32 bits and the counter is exceeding the u32::MAX.
This shouldn't be an issue with the ChaCha20 cipher since cipher panics when the counter will exceed u32::MAX, but ChaCha20Rng wraps around and is currently supposed to use a 32-bit counter. I'm going to write one more test to see if I can get it to fail, and if it fails, then I'll probably need to change add64 to accept a boolean for performing 32-bit or 64-bit addition.
Addresses #334
The constant booleans in the backends should be evaluated at compile-time... but there's a slight chance that I could write a test for the
ChaCha20Rngthat fails onaarch64with neon. The originaladd64macro does what the name suggests, it adds vectors in a 64-bit fashion, meaning that it could perform addition incorrectly when the counter should be 32 bits and the counter is exceeding theu32::MAX.This shouldn't be an issue with the
ChaCha20cipher sincecipherpanics when the counter will exceedu32::MAX, butChaCha20Rngwraps around and is currently supposed to use a 32-bit counter. I'm going to write one more test to see if I can get it to fail, and if it fails, then I'll probably need to changeadd64to accept a boolean for performing 32-bit or 64-bit addition.