#include <stdint.h>
#include <immintrin.h>
#include <sys/types.h>

#define MODE_URL        1
#define MODE_RAW        2
#define MODE_JSON       8

#define as_m32v(v)      (*(uint32_t *)(v))
#define as_m64v(v)      (*(uint64_t *)(v))

#define as_m128p(v)     ((__m128i *)(v))
#define as_m256p(v)     ((__m256i *)(v))

#define as_m8c(v)       ((const uint8_t *)(v))
#define as_m128c(v)     ((const __m128i *)(v))
#define as_m256c(v)     ((const __m256i *)(v))

#define always_inline   inline __attribute__((always_inline)) 

struct slice_t {
    char * buf;
    size_t len;
    size_t cap;
};

/** Encoder Helper Functions **/

static const char TabEncodeCharsetStd[64] = "ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz0123456789+/";
static const char TabEncodeCharsetURL[64] = "ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz0123456789-_";

static const uint8_t VecEncodeShuffles[32] = {
    1, 0, 2, 1, 4, 3, 5, 4, 7, 6, 8, 7, 10, 9, 11, 10,
    1, 0, 2, 1, 4, 3, 5, 4, 7, 6, 8, 7, 10, 9, 11, 10,
};

static const uint8_t VecEncodeCharsetStd[32] = {
    'a' - 26, '0' - 52, '0' - 52, '0' - 52, '0' - 52, '0' - 52, '0' - 52, '0' - 52,
    '0' - 52, '0' - 52, '0' - 52, '+' - 62, '/' - 63, 'A'     ,        0,        0,
    'a' - 26, '0' - 52, '0' - 52, '0' - 52, '0' - 52, '0' - 52, '0' - 52, '0' - 52,
    '0' - 52, '0' - 52, '0' - 52, '+' - 62, '/' - 63, 'A'     ,        0,        0,
};

static const uint8_t VecEncodeCharsetURL[32] = {
    'a' - 26, '0' - 52, '0' - 52, '0' - 52, '0' - 52, '0' - 52, '0' - 52, '0' - 52,
    '0' - 52, '0' - 52, '0' - 52, '-' - 62, '_' - 63, 'A'     ,        0,        0,
    'a' - 26, '0' - 52, '0' - 52, '0' - 52, '0' - 52, '0' - 52, '0' - 52, '0' - 52,
    '0' - 52, '0' - 52, '0' - 52, '-' - 62, '_' - 63, 'A'     ,        0,        0,
};

static always_inline __m256i encode_avx2(__m128i v0, __m128i v1, const uint8_t *tab) {
    __m256i vv = _mm256_set_m128i    (v1, v0);
    __m256i sh = _mm256_loadu_si256  (as_m256c(VecEncodeShuffles));
    __m256i in = _mm256_shuffle_epi8 (vv, sh);
    __m256i t0 = _mm256_and_si256    (in, _mm256_set1_epi32(0x0fc0fc00));
    __m256i t1 = _mm256_mulhi_epu16  (t0, _mm256_set1_epi32(0x04000040));
    __m256i t2 = _mm256_and_si256    (in, _mm256_set1_epi32(0x003f03f0));
    __m256i t3 = _mm256_mullo_epi16  (t2, _mm256_set1_epi32(0x01000010));
    __m256i vi = _mm256_or_si256     (t1, t3);
    __m256i s0 = _mm256_cmpgt_epi8   (_mm256_set1_epi8(26), vi);
    __m256i s1 = _mm256_and_si256    (_mm256_set1_epi8(13), s0);
    __m256i s2 = _mm256_loadu_si256  (as_m256c(tab));
    __m256i r0 = _mm256_subs_epu8    (vi, _mm256_set1_epi8(51));
    __m256i r1 = _mm256_or_si256     (r0, s1);
    __m256i r2 = _mm256_shuffle_epi8 (s2, r1);
    __m256i r3 = _mm256_add_epi8     (vi, r2);
    return r3;
}

/** Function Implementations **/

static always_inline void do_b64encode(struct slice_t *out, const struct slice_t *src, int mode) {
    char *          ob = out->buf + out->len;
    char *          op = out->buf + out->len;
    const char *    ip = src->buf;
    const char *    ie = src->buf + src->len;
    const char *    st = TabEncodeCharsetStd;
    const uint8_t * vt = VecEncodeCharsetStd;

    /* check for empty string */
    if (src->len == 0) {
        return;
    }

    /* check for URL encoding */
    if (mode & MODE_URL) {
        st = TabEncodeCharsetURL;
        vt = VecEncodeCharsetURL;
    }

    /* SIMD 24 bytes loop, but the SIMD instruction will load 4 bytes
     * past the end, so it's safe only if there are 28 bytes or more left */
#ifdef __AVX2__
    while (ip <= ie - 28) {
        __m128i v0 = _mm_loadu_si128 (as_m128c(ip));
        __m128i v1 = _mm_loadu_si128 (as_m128c(ip + 12));
        __m256i vv = encode_avx2     (v0, v1, vt);

        /* store the result, and advance buffer pointers */
        _mm256_storeu_si256(as_m256p(op), vv);
        op += 32;
        ip += 24;
    }

    /* can do one more 24 bytes round, but needs special handling */
    if (ip <= ie - 24) {
        __m128i v0 = _mm_loadu_si128 (as_m128c(ip));
        __m128i v1 = _mm_loadu_si128 (as_m128c(ip + 8));
        __m128i v2 = _mm_srli_si128  (v1, 4);
        __m256i vv = encode_avx2     (v0, v2, vt);

        /* store the result, and advance buffer pointers */
        _mm256_storeu_si256(as_m256p(op), vv);
        op += 32;
        ip += 24;
    }
#endif

    /* no more bytes */
    if (ip == ie) {
        out->len += op - ob;
        return;
    }

    /* handle the remaining bytes with scalar code (with 4 bytes load) */
    while (ip <= ie - 4) {
        uint32_t v0 = __builtin_bswap32(*(const uint32_t *)ip);
        uint8_t  v1 = (v0 >> 26) & 0x3f;
        uint8_t  v2 = (v0 >> 20) & 0x3f;
        uint8_t  v3 = (v0 >> 14) & 0x3f;
        uint8_t  v4 = (v0 >>  8) & 0x3f;

        /* encode the characters, and move to next block */
        ip += 3;
        *op++ = st[v1];
        *op++ = st[v2];
        *op++ = st[v3];
        *op++ = st[v4];
    }

    /* load the last bytes */
    size_t   dp = ie - ip;
    uint32_t v0 = (uint32_t)(uint8_t)ip[0] << 16;

#define B2 v0 |= (uint32_t)(uint8_t)ip[2]
#define B1 v0 |= (uint32_t)(uint8_t)ip[1] << 8

#define R4 *op++ = st[(v0 >>  0) & 0x3f]
#define R3 *op++ = st[(v0 >>  6) & 0x3f]
#define R2 *op++ = st[(v0 >> 12) & 0x3f]
#define R1 *op++ = st[(v0 >> 18) & 0x3f]

#define NB { out->len += op - ob; }
#define PD { if ((mode & MODE_RAW) == 0) { *op++ = '='; } }

    /* encode the last few bytes */
    switch (dp) {
        case 3  : B2; B1; R1; R2; R3; R4; NB; break;
        case 2  :     B1; R1; R2; R3; PD; NB; break;
        case 1  :         R1; R2; PD; PD; NB; break;
        default :                         NB; break;
    }

#undef PD
#undef NB
#undef R1
#undef R2
#undef R3
#undef R4
#undef B1
#undef B2
}

/** Decoder Helper Functions **/

static const uint8_t VecPacking[32] = {
    2, 1, 0, 6, 5, 4, 10, 9, 8, 14, 13, 12, 128, 128, 128, 128,
    2, 1, 0, 6, 5, 4, 10, 9, 8, 14, 13, 12, 128, 128, 128, 128
};

static const uint8_t VecDecodeBits[32] = {
    0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
    0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00
};

static const uint8_t VecDecodeTableStd[128] = {
    0x00, 0x00, 0x13, 0x04, 0xbf, 0xbf, 0xb9, 0xb9, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
    0x00, 0x00, 0x13, 0x04, 0xbf, 0xbf, 0xb9, 0xb9, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
    0xa8, 0xf8, 0xf8, 0xf8, 0xf8, 0xf8, 0xf8, 0xf8, 0xf8, 0xf8, 0xf0, 0x54, 0x50, 0x50, 0x50, 0x54,
    0xa8, 0xf8, 0xf8, 0xf8, 0xf8, 0xf8, 0xf8, 0xf8, 0xf8, 0xf8, 0xf0, 0x54, 0x50, 0x50, 0x50, 0x54,
    0x2f, 0x2f, 0x2f, 0x2f, 0x2f, 0x2f, 0x2f, 0x2f, 0x2f, 0x2f, 0x2f, 0x2f, 0x2f, 0x2f, 0x2f, 0x2f,
    0x2f, 0x2f, 0x2f, 0x2f, 0x2f, 0x2f, 0x2f, 0x2f, 0x2f, 0x2f, 0x2f, 0x2f, 0x2f, 0x2f, 0x2f, 0x2f,
    0x10, 0x10, 0x10, 0x10, 0x10, 0x10, 0x10, 0x10, 0x10, 0x10, 0x10, 0x10, 0x10, 0x10, 0x10, 0x10,
    0x10, 0x10, 0x10, 0x10, 0x10, 0x10, 0x10, 0x10, 0x10, 0x10, 0x10, 0x10, 0x10, 0x10, 0x10, 0x10
};

static const uint8_t VecDecodeTableURL[128] = {
    0x00, 0x00, 0x11, 0x04, 0xbf, 0xbf, 0xb9, 0xb9, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
    0x00, 0x00, 0x11, 0x04, 0xbf, 0xbf, 0xb9, 0xb9, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
    0xa8, 0xf8, 0xf8, 0xf8, 0xf8, 0xf8, 0xf8, 0xf8, 0xf8, 0xf8, 0xf0, 0x50, 0x50, 0x54, 0x50, 0x70,
    0xa8, 0xf8, 0xf8, 0xf8, 0xf8, 0xf8, 0xf8, 0xf8, 0xf8, 0xf8, 0xf0, 0x50, 0x50, 0x54, 0x50, 0x70,
    0x5f, 0x5f, 0x5f, 0x5f, 0x5f, 0x5f, 0x5f, 0x5f, 0x5f, 0x5f, 0x5f, 0x5f, 0x5f, 0x5f, 0x5f, 0x5f,
    0x5f, 0x5f, 0x5f, 0x5f, 0x5f, 0x5f, 0x5f, 0x5f, 0x5f, 0x5f, 0x5f, 0x5f, 0x5f, 0x5f, 0x5f, 0x5f,
    0xe0, 0xe0, 0xe0, 0xe0, 0xe0, 0xe0, 0xe0, 0xe0, 0xe0, 0xe0, 0xe0, 0xe0, 0xe0, 0xe0, 0xe0, 0xe0,
    0xe0, 0xe0, 0xe0, 0xe0, 0xe0, 0xe0, 0xe0, 0xe0, 0xe0, 0xe0, 0xe0, 0xe0, 0xe0, 0xe0, 0xe0, 0xe0,
};

static const uint8_t VecDecodeCharsetStd[256] = {
    0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
    0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
    0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,   62, 0xff, 0xff, 0xff,   63,
      52,   53,   54,   55,   56,   57,   58,   59,   60,   61, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
    0xff,    0,    1,    2,    3,    4,    5,    6,    7,    8,    9,   10,   11,   12,   13,   14,
      15,   16,   17,   18,   19,   20,   21,   22,   23,   24,   25, 0xff, 0xff, 0xff, 0xff, 0xff,
    0xff,   26,   27,   28,   29,   30,   31,   32,   33,   34,   35,   36,   37,   38,   39,   40,
      41,   42,   43,   44,   45,   46,   47,   48,   49,   50,   51, 0xff, 0xff, 0xff, 0xff, 0xff,
    0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
    0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
    0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
    0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
    0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
    0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
    0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
    0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff
};

static const uint8_t VecDecodeCharsetURL[256] = {
    0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
    0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
    0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,   62, 0xff, 0xff,
      52,   53,   54,   55,   56,   57,   58,   59,   60,   61, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
    0xff,    0,    1,    2,    3,    4,    5,    6,    7,    8,    9,   10,   11,   12,   13,   14,
      15,   16,   17,   18,   19,   20,   21,   22,   23,   24,   25, 0xff, 0xff, 0xff, 0xff,   63,
    0xff,   26,   27,   28,   29,   30,   31,   32,   33,   34,   35,   36,   37,   38,   39,   40,
      41,   42,   43,   44,   45,   46,   47,   48,   49,   50,   51, 0xff, 0xff, 0xff, 0xff, 0xff,
    0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
    0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
    0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
    0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
    0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
    0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
    0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
    0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff
};

static always_inline void memcopy_24(char *dp, const uint8_t *sp) {
    *(uint64_t *)(dp +  0) = *(const uint64_t *)(sp +  0);
    *(uint64_t *)(dp +  8) = *(const uint64_t *)(sp +  8);
    *(uint64_t *)(dp + 16) = *(const uint64_t *)(sp + 16);
}

static always_inline __m256i decode_avx2(__m256i v0, int *pos, const uint8_t *tab) {
    __m256i v1 = _mm256_srli_epi32           (v0, 4);
    __m256i vl = _mm256_and_si256            (v0, _mm256_set1_epi8(0x0f));
    __m256i vh = _mm256_and_si256            (v1, _mm256_set1_epi8(0x0f));
    __m256i st = _mm256_loadu_si256          (as_m256c(tab));
    __m256i mt = _mm256_loadu_si256          (as_m256c(tab + 32));
    __m256i et = _mm256_loadu_si256          (as_m256c(tab + 64));
    __m256i rt = _mm256_loadu_si256          (as_m256c(tab + 96));
    __m256i pt = _mm256_loadu_si256          (as_m256c(VecPacking));
    __m256i bt = _mm256_loadu_si256          (as_m256c(VecDecodeBits));
    __m256i sh = _mm256_shuffle_epi8         (st, vh);
    __m256i eq = _mm256_cmpeq_epi8           (v0, et);
    __m256i sv = _mm256_blendv_epi8          (sh, rt, eq);
    __m256i bm = _mm256_shuffle_epi8         (mt, vl);
    __m256i bv = _mm256_shuffle_epi8         (bt, vh);
    __m256i mr = _mm256_and_si256            (bm, bv);
    __m256i nm = _mm256_cmpeq_epi8           (mr, _mm256_setzero_si256());
    __m256i sr = _mm256_add_epi8             (v0, sv);
    __m256i r0 = _mm256_and_si256            (sr, _mm256_set1_epi8(0x3f));
    __m256i r1 = _mm256_maddubs_epi16        (r0, _mm256_set1_epi32(0x01400140));
    __m256i r2 = _mm256_madd_epi16           (r1, _mm256_set1_epi32(0x00011000));
    __m256i r3 = _mm256_shuffle_epi8         (r2, pt);
    __m256i r4 = _mm256_permutevar8x32_epi32 (r3, _mm256_setr_epi32(0, 1, 2, 4, 5, 6, 3, 7));
    int64_t mp = _mm256_movemask_epi8        (nm);
    int32_t np = __builtin_ctzll             (mp | 0xffffffff00000000);
    return (*pos = np), r4;
}


#define ALL_01h     (~0ul / 255)
#define ALL_7fh     (ALL_01h * 127)
#define ALL_80h     (ALL_01h * 128)

static always_inline uint32_t hasless(uint32_t x, uint8_t n) {
    return (x - ALL_01h * n) & ~x & ALL_80h;
}

static always_inline uint32_t hasmore(uint32_t x, uint8_t n) {
    return (x + ALL_01h * (127 - n) | x) & ALL_80h;
}

static always_inline uint32_t hasbetween(uint32_t x, uint8_t m, uint8_t n) {
    return (ALL_01h * (127 + n) - (x & ALL_7fh) & ~x & (x & ALL_7fh) + ALL_01h * (127 - m)) & ALL_80h;
}

#undef ALL_01h
#undef ALL_7fh
#undef ALL_80h

static always_inline char unhex16_is(const uint8_t *s) {
    uint32_t v = *(uint32_t *)s;
    return !(hasless(v, '0') || hasmore(v, 'f') || hasbetween(v, '9', 'A') || hasbetween(v, 'F', 'a'));
}

static always_inline uint32_t unhex16_fast(const uint8_t *s) {
    uint32_t a = __builtin_bswap32(*(uint32_t *)s);
    uint32_t b = 9 * ((~a & 0x10101010) >> 4) + (a & 0x0f0f0f0f);
    uint32_t c = (b >> 4) | b;
    uint32_t d = ((c >> 8) & 0xff00) | (c & 0x00ff);
    return d;
}

static always_inline uint8_t unescape_asc(const uint8_t * ie, const uint8_t ** ipp) {
    const uint8_t * ee = (*ipp) + 1;
    uint32_t ch = 0xff; 
    /* check eof */
    if (ee > ie) {
        return 0xff;
    }
    switch (ee[-1]) {
        case 'r': ch = '\r'; break;
        case 'n': ch = '\n'; break;
        case '/': ch = '/'; break;
        case 'u': /* neee more 4 bytes */
        if (ie - ee >= 4 && unhex16_is(ee)) { 
            ch = unhex16_fast(ee);
            /* if not ascii, as 0xff */
            ch = ch < 128 ? ch : 0xff;
            ee += 4;
        }
        break;
    }
    *ipp = ee;
    return ch;
}

/* Return 0 if success, otherwise return the error position + 1 */
static always_inline int64_t decode_block(
    const uint8_t *  ie,
    const uint8_t ** ipp,
    char **          opp,
    const uint8_t *  tab,
    int              mode
) {
    int      nb = 0;
    uint32_t v0 = 0;

    /* buffer pointers */
    char *          op = *opp;
    const uint8_t * ip = *ipp;
    uint8_t id = 0;
    uint8_t ch = 0;
    int pad = 0;

#define may_unescape() { if (ch == '\\' && (mode & MODE_JSON)) ch = unescape_asc(ie, &ip); }
#define skip_newlines() { if (ch == '\r' || ch == '\n') continue; }

    /* load up to 4 characters */
    while (ip < ie && nb < 4) {
        ch = *ip++;
        may_unescape();
        skip_newlines();

        /* lookup the index, and check for invalid characters */
        if ((id = tab[ch]) == 0xff) {
            if ((mode & MODE_RAW) || ch != '=' || nb < 2) goto error;
            pad++; goto tail;
        }

        /* decode the character */
        v0 = (v0 << 6) | id;
        nb++;
    }

    if (nb == 0) {
        /* update the pointers */
        *ipp = ip;
        return 0;
    }

    /* check eof, MODE_STD need paddings */
    if (ip >= ie && nb != 4) {
        if (!(mode & MODE_RAW) || nb == 1) goto error;
    }

decode:
    v0 <<= 6 * (4 - nb); 
    /* ends with eof or 4 characters, decode into output */
    switch (nb) {
        case 4: op[2] = (v0 >>  0) & 0xff;
        case 3: op[1] = (v0 >>  8) & 0xff;
        case 2: op[0] = (v0 >> 16) & 0xff;
    }

    /* update the pointers */
    *ipp = ip;
    *opp = op + nb - 1;
    return 0;

tail:
    /* loop for more paddings */
    while (ip < ie) {
        ch = *ip++;
        may_unescape();
        skip_newlines();
        if (ch != '=') goto error;
        if (++pad + nb > 4) goto error;
    }
    goto decode;
#undef may_unescape
#undef skip_newlines

error:
    /* update eof error position */
    if (ip == ie) ip++;
    return ip - *ipp;


}

static always_inline ssize_t do_b64decode(struct slice_t *out, const char *src, size_t nb, int mode) {
    int     ep;
    __m256i vv;
    int64_t dv;
    uint8_t buf[32] = {0};

    /* check for empty input */
    if (nb == 0) {
        return 0;
    }

    /* output buffer */
    char *ob = out->buf + out->len;
    char *op = out->buf + out->len;
    char *oe = out->buf + out->cap;

    /* input buffer */
    const uint8_t *dt = VecDecodeTableStd;
    const uint8_t *st = VecDecodeCharsetStd;
    const uint8_t *ib = (const uint8_t *)src;
    const uint8_t *ip = (const uint8_t *)src;
    const uint8_t *ie = (const uint8_t *)src + nb;

    /* check for URL encoding */
    if (mode & MODE_URL) {
        dt = VecDecodeTableURL;
        st = VecDecodeCharsetURL;
    }

#ifdef USE_AVX2
    /* decode every 32 bytes, the final round should be handled separately, because the
     * SIMD instruction performs 32-byte store, and it might store past the end of the
     * output buffer */
    while ((ip <= ie - 32) && (op <= oe - 32)) {
        vv = _mm256_loadu_si256(as_m256c(ip));
        vv = decode_avx2(vv, &ep, dt);

        /* check for invalid characters (or '=' paddings) */
        if (ep < 32) {
            if ((dv = decode_block(ie, &ip, &op, st, mode)) != 0) {
                return ib - ip - dv;
            } else {
                continue;
            }
        }

        _mm256_storeu_si256(as_m256p(op), vv);

        /* move to next block */
        ip += 32;
        op += 24;
    }
#endif

scalar:
    /* handle the remaining bytes with scalar code (8 byte loop) */
    while (ip <= ie - 8 && op <= oe - 8) {
        uint8_t v0 = st[ip[0]];
        uint8_t v1 = st[ip[1]];
        uint8_t v2 = st[ip[2]];
        uint8_t v3 = st[ip[3]];
        uint8_t v4 = st[ip[4]];
        uint8_t v5 = st[ip[5]];
        uint8_t v6 = st[ip[6]];
        uint8_t v7 = st[ip[7]];

        /* check for invalid bytes */
        if ((v0 | v1 | v2 | v3 | v4 | v5 | v6 | v7) == 0xff) {
            if ((dv = decode_block(ie, &ip, &op, st, mode)) != 0) {
                return ib - ip - dv;
            } else {
                continue;
            }
        }

        /* construct the characters */
        uint64_t vv = __builtin_bswap64(
            ((uint64_t)v0 << 58) |
            ((uint64_t)v1 << 52) |
            ((uint64_t)v2 << 46) |
            ((uint64_t)v3 << 40) |
            ((uint64_t)v4 << 34) |
            ((uint64_t)v5 << 28) |
            ((uint64_t)v6 << 22) |
            ((uint64_t)v7 << 16)
        );

        /* store the result, and move to next block */
        as_m64v(op) = vv;
        ip += 8;
        op += 6;
    }

    /* handle the remaining bytes with scalar code (4 byte loop) */
    while (ip <= ie - 4 && op <= oe - 4) {
        uint8_t v0 = st[ip[0]];
        uint8_t v1 = st[ip[1]];
        uint8_t v2 = st[ip[2]];
        uint8_t v3 = st[ip[3]];

        /* check for invalid bytes */
        if ((v0 | v1 | v2 | v3) == 0xff) {
            if ((dv = decode_block(ie, &ip, &op, st, mode)) != 0) {
                return ib - ip - dv;
            } else {
                continue;
            }
        }

        /* construct the characters */
        uint32_t vv = __builtin_bswap32(
            ((uint32_t)v0 << 26) |
            ((uint32_t)v1 << 20) |
            ((uint32_t)v2 << 14) |
            ((uint32_t)v3 <<  8)
        );

        /* store the result, and move to next block */
        as_m32v(op) = vv;
        ip += 4;
        op += 3;
    }

    /* decode the last few bytes */
    while (ip < ie) {
        if ((dv = decode_block(ie, &ip, &op, st, mode)) != 0) {
            return ib - ip - dv;
        }
    }

    /* update the result length */
    out->len += op - ob;
    return op - ob;
}