Add 64-bit fixslice AES implementation

[peterdettman]

• protoype: replaces 32-bit in lib.rs, impls.rs

Not familiar enough with rust to attempt configuration settings.

[tarcieri]

@peterdettman it'd be good to preserve both the 32-bit and 64-bit backends, but select which one to use based on the host CPU. You can do something like:

 #[cfg(target_pointer_width = "32")] 
mod fixslice32;
 #[cfg(target_pointer_width = "64")] 
mod fixslice64;

If you're having trouble, the cfg-if crate can help:

https://github.com/alexcrichton/cfg-if

[tarcieri]

Benchmarks look very nice:

test aes128_encrypt8 ... bench:         626 ns/iter (+/- 81) = 204 MB/s
test aes192_encrypt8 ... bench:         742 ns/iter (+/- 76) = 172 MB/s
test aes256_encrypt8 ... bench:         851 ns/iter (+/- 162) = 150 MB/s

[peterdettman]

@tarcieri Yeah, I certainly want both to coexist, but I'd really prefer to leave the cfg stuff to you. It's not just cfg either; the 64-bit version does 4 blocks at a time (instead of 2) and somehow the encrypt_block(s) methods in impl.rs will need to know how many blocks the inner implementation is handling.

[tarcieri]

@peterdettman okay, if you'd like I can try to handle the gating.

How about renaming the existing implementation to fixslice32.rs and then removing the mod fixslice32 directive for it, and I can do an additional pass to wire it back up again?

It's not just cfg either; the 64-bit version does 4 blocks at a time (instead of 2) and somehow the encrypt_block(s) methods in impl.rs will need to know how many blocks the inner implementation is handling.

Can you make them all 8 for now and add a TODO to change ParBlocks in the next release? It's a breaking change.

I think there's also some discussion to be had there. Perhaps we should try to align all of the implementations exposed via the aes crate to be 4-blocks (which is also the ideal number for AES-NI on modern Intel archs).

Edit: opened an issue about changing ParBlocks in the next breaking release: #181

[peterdettman]

OK, I've renamed fixslice.rs to fixslice32.rs and lib.rs now switches b/w them on cfg(target_pointer_width = "64"). The fixslice implementations now have a FIXSLICE_BLOCKS constant that impls.rs uses to break up the 8 blocks into 2 or 4 chunks respectively.

[tarcieri]

Since this is the same for all impls it doesn't need to be a macro parameter

[tarcieri]

You can use the FIXSLICE_BLOCKS constant directly here:

Suggested change

	let mut blocks = [Block::default(); $fixslice_blocks];
	let mut blocks = [Block::default(); FIXSLICE_BLOCKS];

[tarcieri]

...also here:

Suggested change

	for chunk in blocks.chunks_mut($fixslice_blocks) {
	for chunk in blocks.chunks_mut(FIXSLICE_BLOCKS) {

[tarcieri]

@peterdettman looks great aside from one nit!

After landing this, I can expand the CI config to run the tests on 32-bit architectures as well as add some cross-based testing on ARM. That way we can ensure CI tests both the 32-bit and 64-bit backends.

[peterdettman]

Fixed nits and rebased.

[peterdettman]

@tarcieri I’d be interested in seeing some arm (32/64) benchmarks for these fixslice implementations.

[tarcieri]

@peterdettman I'll be looking into benchmarking it on Cortex-A and Cortex-M (32-bit).

Somewhere around here I have some aarch64 devices (RasPi) I could benchmark it on too.