optimized fizzbuzz in C using ymm registers

fizzbuzz-ymm.c

#include <limits.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <immintrin.h>
#include <unistd.h>

#define inline __attribute__((always_inline))

static char tens[48] __attribute__((aligned(64)));
static size_t tenslen = 0;
static int n;

static char outbuf[8000], *outp = outbuf;

static void flush(void)
{
    write(1, outbuf, outp - outbuf);
    outp = outbuf;
}

static void flush_and_exit(void)
{
    flush();
    exit(0);
}

static void plus10(void)
{
    if (tenslen == 0)
        tenslen = 1;

    for (size_t i = 15;; i--) {
        if (__builtin_expect(++tens[i] <= '9', 1))
            return;
        tens[i] = '0';
        if (16 - i == tenslen)
            ++tenslen;
    }
}

static inline void emit30(int maxval)
{
#define EMIT(s1, s2, s3)                                                                                                           \
    do {                                                                                                                           \
        if (++n > maxval)                                                                                                          \
            flush_and_exit();                                                                                                      \
        static const char s[] __attribute__((aligned(64))) =                                                                       \
            "\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0" s1 s2 s3                                              \
            "\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0";                                                      \
        _mm256_storeu_si256((__m256i *)outp, _mm256_or_si256(tens256, _mm256_loadu_si256((__m256i *)(s + 31 - tenslen))));         \
        outp += tenslen + 2;                                                                                                       \
        if (sizeof(s2) != 1) {                                                                                                     \
            if (++n > maxval)                                                                                                      \
                flush_and_exit();                                                                                                  \
            outp += sizeof(s2) - 1;                                                                                                \
            if (sizeof(s3) != 1) {                                                                                                 \
                if (++n > maxval)                                                                                                  \
                    flush_and_exit();                                                                                              \
                outp += sizeof(s3) - 1;                                                                                            \
            }                                                                                                                      \
        }                                                                                                                          \
    } while (0)

    __m256i tens256 = _mm256_loadu_si256((__m256i *)(tens + 16 - tenslen));

    EMIT("1\n", "", "");
    EMIT("2\n", "fizz\n", "");
    EMIT("4\n", "buzz\n", "fizz\n");
    EMIT("7\n", "", "");
    EMIT("8\n", "fizz\n", "buzz\n");

    plus10();
    tens256 = _mm256_loadu_si256((__m256i *)(tens + 16 - tenslen));

    EMIT("1\n", "fizz\n", "");
    EMIT("3\n", "", "");
    EMIT("4\n", "fizzbuzz\n", "");
    EMIT("6\n", "", "");
    EMIT("7\n", "fizz\n", "");
    EMIT("9\n", "buzz\n", "fizz\n");

    plus10();
    tens256 = _mm256_loadu_si256((__m256i *)(tens + 16 - tenslen));

    EMIT("2\n", "", "");
    EMIT("3\n", "fizz\n", "buzz\n");
    EMIT("6\n", "fizz\n", "");
    EMIT("8\n", "", "");
    EMIT("9\n", "fizzbuzz\n", "");

    plus10();
}

int main(int argc, char **argv)
{
    int maxval;

    if (argc < 2 || sscanf(argv[1], "%d", &maxval) != 1) {
        fprintf(stderr, "usage: %s <max-value>\n", argv[0]);
        exit(1);
    }
    memset(tens, '0', 16);

    while (n + 30 < maxval) {
        emit30(INT_MAX);
        if (outp - outbuf >= sizeof(outbuf) - 500)
            flush();
    }
    emit30(maxval);

    return 0;
}