spaceskip.c

void next(void) {
    // Compilers will generate something like CMP r1, ' '; LEA r2, [r3+1]; CMOVEQ r3, r2, which has 2 cycle latency.
    input += *input == ' '; // The most common intertoken spacing is 0 or 1 (e.g. look at this file), so use a branchless skip.
start:
    switch (*input) {
    case '\r':
        input++;
        if (*input != '\n') goto start;
        // Default-predicted fallthrough case for \r\n on Windows.
    case '\n': {
        input++;
        line++;
        int len;
        // Do a branchless SIMD skip of up to 16 indentation spaces after a newline.
        _m128i chars = _mm_loadu_si128((_m_m128i *)input);
        BitScanForward(&len, ~_mm_movemask_epi8(_mm_cmpeq_epi8(chars, _mm_set1_epi8(' '))));
        input += len;
        goto start;
    }
    case ' ':
    case '\t':
    case '\v':
        // This path should be rare: more than 1 intertoken space, more than 16 spaces after newline, or rare whitespace character.
        do input++; while (isspace(*input)); // Probably not libc isspace but you get the idea: just eat the length mispredict.
        goto start;
    // ...
    }
}