Function do_count_chars

crates/wtf8/src/core_str_count.rs:22–91

(s: &Wtf8)

}

fn do_count_chars(s: &Wtf8) -> usize {
    // For correctness, `CHUNK_SIZE` must be:
    //
    // - Less than or equal to 255, otherwise we'll overflow bytes in `counts`.
    // - A multiple of `UNROLL_INNER`, otherwise our `break` inside the
    //   `body.chunks(CHUNK_SIZE)` loop is incorrect.
    //
    // For performance, `CHUNK_SIZE` should be:
    // - Relatively cheap to `/` against (so some simple sum of powers of two).
    // - Large enough to avoid paying for the cost of the `sum_bytes_in_usize`
    //   too often.
    const CHUNK_SIZE: usize = 192;

    // Check the properties of `CHUNK_SIZE` and `UNROLL_INNER` that are required
    // for correctness.
    const _: () = assert!(CHUNK_SIZE < 256);
    const _: () = assert!(CHUNK_SIZE.is_multiple_of(UNROLL_INNER));

    // SAFETY: transmuting `[u8]` to `[usize]` is safe except for size
    // differences which are handled by `align_to`.
    let (head, body, tail) = unsafe { s.as_bytes().align_to::<usize>() };

    // This should be quite rare, and basically exists to handle the degenerate
    // cases where align_to fails (as well as miri under symbolic alignment
    // mode).
    //
    // The `unlikely` helps discourage LLVM from inlining the body, which is
    // nice, as we would rather not mark the `char_count_general_case` function
    // as cold.
    if unlikely(body.is_empty() || head.len() > USIZE_SIZE || tail.len() > USIZE_SIZE) {
        return char_count_general_case(s.as_bytes());
    }

    let mut total = char_count_general_case(head) + char_count_general_case(tail);
    // Split `body` into `CHUNK_SIZE` chunks to reduce the frequency with which
    // we call `sum_bytes_in_usize`.
    for chunk in body.chunks(CHUNK_SIZE) {
        // We accumulate intermediate sums in `counts`, where each byte contains
        // a subset of the sum of this chunk, like a `[u8; size_of::<usize>()]`.
        let mut counts = 0;

        let (unrolled_chunks, remainder) = chunk.as_chunks::<UNROLL_INNER>();
        for unrolled in unrolled_chunks {
            for &word in unrolled {
                // Because `CHUNK_SIZE` is < 256, this addition can't cause the
                // count in any of the bytes to overflow into a subsequent byte.
                counts += contains_non_continuation_byte(word);
            }
        }

        // Sum the values in `counts` (which, again, is conceptually a `[u8;
        // size_of::<usize>()]`), and accumulate the result into `total`.
        total += sum_bytes_in_usize(counts);

        // If there's any data in `remainder`, then handle it. This will only
        // happen for the last `chunk` in `body.chunks()` (because `CHUNK_SIZE`
        // is divisible by `UNROLL_INNER`), so we explicitly break at the end
        // (which seems to help LLVM out).