md5: add x86_64 and AArch64 asm backends - #917

The original work was placed in the public domain: animetosho/md5-optimisation#4

For maintainers, the original work changed license to Apache2.0/MIT .

I think it'd be helpful if this were split up into separate PRs for AArch64 vs x86(_64). Even #447 went by the wayside because these kinds of ASM backends are hard-to-review, even when adapted from seemingly mature sources of ASM. Though while this ASM seems to have a decent rationale, I'm not finding a whole lot in the way of a test suite.

Proptests which compare the Rust reference implementation with the ASM implementation would be nice. Also from what I can tell CI isn't even testing aarch64 whatsoever, just ensuring it compiles.

It would also be helpful if these changes were justified with benchmarks.

Edit: I looked for formally verified ASM for MD5 and wasn't able to find any

I looked for formally verified ASM for MD5 and wasn't able to find any

Author of https://github.com/animetosho/md5-optimisation said most of them are included to OpenSSL. It might be able to reference OpenSSL's md5.

CI isn't even testing aarch64 whatsoever, just ensuring it compiles.

Does this help?
https://github.com/oech3/rustcrypto-hashes/commit/823b8757ecb94d7e0033eb6234879911b7b24430

That probably works, though you can look at polyval for how we've otherwise set up the matrix: https://github.com/RustCrypto/universal-hashes/blob/7232d5e23fd2d24ae3ce7b8944e6840d7655db6e/.github/workflows/polyval.yml#L36-L80