Optimize sha256 for aarch64

[dgbo]

Hi,

The sha256 is top hotspot for some applications, e.g. Phase1 in filecoin/lotus, this patch optimizes it for aarch64.
It mainly remove the dependency between ldr and add in each four rounds by preloading K constants into NEON registers.
We also take advantages of load/store pair and ld1 instructions to reduce the instructions executed.

Verified with repository sha2 [1], test commands: cargo +nightly test --features "asm", cargo +nightly test --release --features "asm".

On our aarch64 server (core tsv110), we witnessed 11.90% improvements with benches under asm-hashes/sha2:

# before
test bench_compress256 ... bench:          47 ns/iter (+/- 0) = 1361 MB/s
# after
test bench_compress256 ... bench:          42 ns/iter (+/- 0) = 1523 MB/s

[1] https://github.com/dgbo/hashes/tree/master/sha2

[newpavlov]

Thank you!

@linkmauve
Can you please take a look?

[tarcieri]

I can give this a try on an M1

[dgbo]

Thank you for watching this.

BTW, we run lotus-bench of filecoin/lotus, this PR reduce 10.5% execution time of Phase1 in this application.

Regards.

[dgbo]

Ping... Can I get a review for this, is there any suggestions? Thanks. :)

[newpavlov]

@tarcieri
Can I leave this PR to you? Unfortunately I don't know ARM assembly well enough and do not have appropriate hardware to test the code.

[tarcieri]

Yep, been meaning to validate it on an M1. I'll see if I can take a look today.

[tarcieri]

Hrmm, interesting. The sha2-asm crate does not build on the M1 at all, either before or with this change:

   Compiling sha2-asm v0.5.4 (/Users/tony/asm-hashes/sha2)
The following warnings were emitted during compilation:

warning: src/sha256_aarch64.S:64:2: error: ADR/ADRP relocations must be GOT relative
warning:  adrp x2, .K
warning:  ^
warning: src/sha256_aarch64.S:64:2: error: unknown AArch64 fixup kind!
warning:  adrp x2, .K
warning:  ^
warning: src/sha256_aarch64.S:65:2: error: unknown AArch64 fixup kind!
warning:  add x2, x2, :lo12:.K
warning:  ^

error: failed to run custom build command for `sha2-asm v0.5.4 (/Users/tony/asm-hashes/sha2)`

I'd say this PR looks good to merge then, but separately we should track an M1 build failure.

Edit: opened #28 to track this

[dgbo]

Thanks!