Support memchr on arm64 too

[bjorn3]

Miri has implemented all vendor intrinsics needed by the memchr crate on x86. It still misses them for arm64 however.

[RalfJung]

Yeah we had a heroic contributor implement a lot of x86 intrinsics.

Do you have a list of intrinsics we'd need for this? Is there some place the semantics of these intrinsics are documented?

[bjorn3]

I believe https://github.com/BurntSushi/memchr/blob/746182171d2e88600652de3e6768cd8aef77e585/src/vector.rs#L292-L460 has all uses of arm64 vendor intrinsics. So

• vdupq_n_u8
• vld1q_u8
• vreinterpretq_u16_u8
• vshrn_n_u16
• vreinterpret_u64_u8
• vget_lane_u64
• vceqq_u8
• vandq_u8
• vorrq_u8
• vreinterpretq_u64_u8
• vgetq_lane_u64
• vpmaxq_u8

Several of these are implemented using the same intrinsics as portable-simd and as such likely already supported. For the remaining ones the core::arch doc pages link to the arm documentation which contains their behavior both as english text and in the form of a function implemented in a custom dsl.

[bjorn3]

Only vpmaxq_u8 needed a miri implementation. There rest seems to use portable simd intrinsics internally.

[RalfJung]

the core::arch doc pages link to the arm documentation

I can't find such a link at https://doc.rust-lang.org/nightly/core/arch/aarch64/fn.vpmaxq_u8.html ?

[bjorn3]

I guess some are missing the link. In any case I've implemented it locally, but I'm getting two weird test failures.