perf: remove loop from str::floor_char_boundary

[overlookmotel]

#144472 made str::floor_char_boundary a const function, but in doing so introduced a loop. This is unnecessary because the next UTF-8 character boundary can only be within a 4-byte range.

This produces excessive code for e.g. str.floor_char_boundary(20), because the loop is unrolled.

https://godbolt.org/z/5f3YsM6oK

This PR replaces the loop with 3 checks in series.

In addition to reducing code size in some cases, it also removes bounds checks from calling code when following floor_char_boundary with a call to String::truncate (which I assume might be a common pattern).

Notes

• I tried using index.unchecked_sub(1), but found it doesn't remove bounds checks, whereas index.checked_sub(1).unwrap_unchecked() does. Surprising!

• The assert_uncheckeds are required to elide checks from code following the floor_char_boundary call e.g.:

let index = string.floor_char_boundary(20);
string.truncate(index);

If this PR is accepted, I'll follow up with a similar PR for ceil_char_boundary.

Very happy to receive feedback and amend.

[rustbot]

r? @scottmcm

rustbot has assigned @scottmcm.
They will have a look at your PR within the next two weeks and either review your PR or reassign to another reviewer.

Use r? to explicitly pick a reviewer

[clarfonthey]

Rather than manually unrolling the code here, it might be worth considering if the original implementation can be made const instead:

rust/library/core/src/str/mod.rs

Lines 244 to 247 in 03b17b1

	let lower_bound = index.saturating_sub(3);
	let new_index = self.as_bytes()[lower_bound..=index]
	.iter()
	.rposition(|b| b.is_utf8_char_boundary());

I had this exact thought in the original implementation, which is why I explicitly only checked the four following bytes for a character boundary. Note that I haven't fully scrutinised the code and yours may be better regardless, but I figured I'd mention this for some extra context. I would expect LLVM to automatically unroll a 4-max-iteration loop in most contexts.