sha-1: asm fails to build on i686 and aarch64

[jszwedko]

Apologies, I'm unsure if this is specifically an issue with hashes or maybe has something to do with an interaction with cross, but when I attempt to cross-compile to aarch64-unknown-linux-gnu it seems to chose invalid assembly instructions.

Part of the output of:

   Compiling sha1-asm v0.4.4
The following warnings were emitted during compilation:

warning: src/x64.S: Assembler messages:
warning: src/x64.S:111: Error: unknown mnemonic `movq' -- `movq %rbx,%xmm0'
warning: src/x64.S:112: Error: unknown mnemonic `movq' -- `movq %rbp,%xmm1'
warning: src/x64.S:113: Error: unknown mnemonic `subq' -- `subq $64,%rsp'
warning: src/x64.S:116: Error: unknown mnemonic `movq' -- `movq %rdi,%r8'
warning: src/x64.S:117: Error: unknown mnemonic `movl' -- `movl 0(%rdi),%eax'
warning: src/x64.S:118: Error: unknown mnemonic `movl' -- `movl 4(%rdi),%ebx'
warning: src/x64.S:119: Error: unknown mnemonic `movl' -- `movl 8(%rdi),%ecx'
warning: src/x64.S:120: Error: unknown mnemonic `movl' -- `movl 12(%rdi),%edx'
warning: src/x64.S:121: Error: unknown mnemonic `movl' -- `movl 16(%rdi),%ebp'
warning: src/x64.S:122: Error: unknown mnemonic `movq' -- `movq %rsi,%rdi'

See full output here along with Cargo.toml and src/main.rs here: https://gist.github.com/jszwedko/9d0cfb8db04fdeffb83cae1f1e2acec4

[tarcieri]

That's odd. We do test sha1-asm in CI with cross:

https://github.com/RustCrypto/asm-hashes/blob/master/.github/workflows/sha1.yml#L93-L114

Here's the latest run:

https://github.com/RustCrypto/asm-hashes/runs/2141158457

[tarcieri]

In the sha2 crate, we have a cross-based end-to-end test which uses the asm feature:

https://github.com/RustCrypto/hashes/blob/master/.github/workflows/sha2.yml#L119-L141

We don't presently have a similar test for sha1 though. It may be good to add.

[jszwedko]

👍 it does seem like the sha1 workflow just tests that it can build the crate, but doesn't actually exercise it.

[tarcieri]

I opened RustCrypto/hashes#252 which adds a CI config more like the one for the sha2 crate. We'll see if it can capture this issue at least.

Edit: confirm a failure on aarch64:

https://github.com/RustCrypto/hashes/pull/252/checks?check_run_id=2271544657

But also, it failed to on i686 as well(!):

https://github.com/RustCrypto/hashes/pull/252/checks?check_run_id=2271545400

It did at least succeed on x86_64:

https://github.com/RustCrypto/hashes/pull/252/checks?check_run_id=2271545439

[jszwedko]

@tarcieri nice! Thanks for adding the tests and verifying the failure.

[Byron]

I don't know if it is of relevance here, but sometime there will be more intrinsics for ARM in stdarch which might help to make it work on aarch64.

I am particularly interested in helping this along as this issue prevents gitoxide from outperforming git in the pack indexing phase on Apple Silicon (even though it performs just as well on x86-64).

[tarcieri]

As was mentioned on #5 (comment), the main problem there is the intrinsics aren't stable and only available on nightly.

In the meantime we're using ARM ASM to support stable. We could potentially experiment with nightly-only intrinsics as an alternative to ASM in order to have an implementation ready when it stabilizes.

Perhaps I should open a more general tracking issue about this. Edit: opened RustCrypto/hashes#257 to track ARM intrinsic-based implementations.

[tarcieri]

Since this issue is related to ASM, I'm going to move it to the asm-hashes repo.

[newpavlov]

We decided to deprecate the asm-hashes crates. Future ASM backends will be implemented using inline asm, hopefully the discussed problem will not apply to them.