Consider hinting to the compiler that indices returned from `memchr()` are within bounds of the haystack

[coolreader18]

Currently, indexing into the haystack with the index returned by memchr usually generates a bounds check, because the compiler doesn't know that it's valid. However, if an assert_unchecked(i < haystack.len()) is inserted, the compiler has enough information to elide the check (godbolt).

Pros:

• Reduces bounds-checking in downstream code; this doesn't remove a ton of overhead but could possibly allow for further optimizations that had been inhibited by the branch.

Cons:

• assert_unchecked() is only available since Rust 1.81, so would either have to feature-gate this, do rust version detection, or (most likely) use if !cond { unreachable_unchecked(); } to polyfill it. The last option should be equivalent to assert_unchecked(), but I'm not sure if it's treated exactly the same by the compiler.
• Returning an out-of-bound index becomes UB rather than just a logic error. I assume there was already downstream code relying on this, though, and also it'd probably already be UB internally.

[BurntSushi]

I think this sounds okay to me in principle. But your godbolt example isn't really sufficient to demonstrate that this will work here. Most of the public routines in this crate won't inline, specifically because they are either quite large or because they use ISA extensions that cannot be inlined unless the caller has those same extensions enabled (which is still rare today on x86-64). So if this assertion is put inside of a non-inlined function, I'd guess it wouldn't have the desired effect. That is, your inline(always) in your godbolt example is load bearing.

So doing this correctly is likely possible, but requires care. I'm open to PRs.

Rust 1.81 was released over a year ago, so I'm fine with the MSRV bump here.