/* inftrees.c -- generate Huffman trees for efficient decoding

 * Copyright (C) 1995-2024 Mark Adler

 * For conditions of distribution and use, see copyright notice in zlib.h

 */

#include "zbuild.h"

#include "zmemory.h"

#include "zutil.h"

#include "inftrees.h"

#include "inflate_p.h"

#include "fallback_builtins.h"

#if defined(__SSE2__)

#  include "arch/x86/x86_intrins.h"

#elif defined(__ARM_NEON) || defined(__ARM_NEON__)

#  include "arch/arm/neon_intrins.h"

#elif defined(__ALTIVEC__)

#  include "arch/power/power_intrins.h"

#endif

const char PREFIX(inflate_copyright)[] = " inflate 1.3.1 Copyright 1995-2024 Mark Adler ";

/*

  If you use the zlib library in a product, an acknowledgment is welcome

  in the documentation of your product. If for some reason you cannot

  include such an acknowledgment, I would appreciate that you keep this

  copyright string in the executable of your product.

 */

/* Count number of codes for each code length. */

static inline void count_lengths(uint16_t *lens, int codes, uint16_t *count) {

    /* IBM...made some weird choices for VSX/VMX. Basically vec_ld has an inherent

     * endianness but we don't want to force VSX to be needed */

    static const ALIGNED_(16) uint8_t one[256] = {

        1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,

        0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,

        0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,

        0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,

        0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,

        0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,

        0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0,

        0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0,

        0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0,

        0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0,

        0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0,

        0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0,

        0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0,

        0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0,

        0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0,

        0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1

    };

#if defined(__ALTIVEC__)

    vector unsigned char s1 = vec_splat_u8(0);

    vector unsigned char s2 = vec_splat_u8(0);

    if (codes & 1) {

        s1 = vec_ld(16 * lens[0], one);

        --codes;

        ++lens;

    }

    while (codes) {

        s1 = vec_add(s1, vec_ld(16 * lens[0], one));

        s2 = vec_add(s2, vec_ld(16 * lens[1], one));

        codes -= 2;

        lens += 2;

    }

    vector unsigned short sum_lo = vec_add(vec_unpackh(s1), vec_unpackh(s2));

    vector unsigned short sum_hi = vec_add(vec_unpackl(s1), vec_unpackl(s2));

    vec_st(sum_lo, 0, &count[0]);

    vec_st(sum_hi, 0, &count[8]);

#elif defined(__ARM_NEON) || defined(__ARM_NEON__)

    int sym;

    uint8x16_t s1 = vdupq_n_u8(0);

    uint8x16_t s2 = vdupq_n_u8(0);

    if (codes & 1) {

        s1 = vld1q_u8(&one[16 * lens[0]]);

    }

    for (sym = codes & 1; sym < codes; sym += 2) {

        s1 = vaddq_u8(s1, vld1q_u8(&one[16 * lens[sym]]));

        s2 = vaddq_u8(s2, vld1q_u8(&one[16 * lens[sym+1]]));

    }

    vst1q_u16(&count[0], vaddl_u8(vget_low_u8(s1), vget_low_u8(s2)));

    vst1q_u16(&count[8], vaddl_u8(vget_high_u8(s1), vget_high_u8(s2)));

#elif defined(__SSE2__)

    int sym;

    __m128i s1 = _mm_setzero_si128();

    __m128i s2 = _mm_setzero_si128();

    if (codes & 1) {

        s1 = _mm_load_si128((const __m128i*)&one[16 * lens[0]]);

    }

    for (sym = codes & 1; sym < codes; sym += 2) {

        s1 = _mm_add_epi8(s1, _mm_load_si128((const __m128i*)&one[16 * lens[sym]]));  // vaddq_u8

        s2 = _mm_add_epi8(s2, _mm_load_si128((const __m128i*)&one[16 * lens[sym+1]]));

    }

#  if defined(__AVX2__)

    __m256i w1 = _mm256_cvtepu8_epi16(s1);

    __m256i w2 = _mm256_cvtepu8_epi16(s2);

    __m256i sum = _mm256_add_epi16(w1, w2);

    _mm256_storeu_si256((__m256i*)&count[0], sum);

#  else

    __m128i zero = _mm_setzero_si128();

    __m128i s1_lo = _mm_unpacklo_epi8(s1, zero);

    __m128i s2_lo = _mm_unpacklo_epi8(s2, zero);

    __m128i sum_lo = _mm_add_epi16(s1_lo, s2_lo);

    _mm_storeu_si128((__m128i*)&count[0], sum_lo);

    __m128i s1_hi = _mm_unpackhi_epi8(s1, zero);

    __m128i s2_hi = _mm_unpackhi_epi8(s2, zero);

    __m128i sum_hi = _mm_add_epi16(s1_hi, s2_hi);

    _mm_storeu_si128((__m128i*)&count[8], sum_hi);

#  endif

#else

    int len, sym;

    for (len = 0; len <= MAX_BITS; len++)

        count[len] = 0;

    for (sym = 0; sym < codes; sym++)

        count[lens[sym]]++;

    Z_UNUSED(one);

#endif

}

/*

   Build a set of tables to decode the provided canonical Huffman code.

   The code lengths are lens[0..codes-1].  The result starts at *table,

   whose indices are 0..2^bits-1.  work is a writable array of at least

   lens shorts, which is used as a work area.  type is the type of code

   to be generated, CODES, LENS, or DISTS.  On return, zero is success,

   -1 is an invalid code, and +1 means that ENOUGH isn't enough.  table

   on return points to the next available entry's address.  bits is the

   requested root table index bits, and on return it is the actual root

   table index bits.  It will differ if the request is greater than the

   longest code or if it is less than the shortest code.

 */

int Z_INTERNAL zng_inflate_table(codetype type, uint16_t *lens, unsigned codes,

                                 code * *table, unsigned *bits, uint16_t *work) {

    unsigned len;                 /* a code's length in bits */

    unsigned sym;                 /* index of code symbols */

    unsigned min, max;            /* minimum and maximum code lengths */

    unsigned root;                /* number of index bits for root table */

    unsigned curr;                /* number of index bits for current table */

    unsigned drop;                /* code bits to drop for sub-table */

    int left;                     /* number of prefix codes available */

    unsigned used;                /* code entries in table used */

    uint16_t rhuff;               /* Reversed huffman code */

    unsigned huff;                /* Huffman code */

    unsigned incr;                /* for incrementing code, index */

    unsigned fill;                /* index for replicating entries */

    unsigned low;                 /* low bits for current root entry */

    unsigned mask;                /* mask for low root bits */

    code here;                    /* table entry for duplication */

    code *next;                   /* next available space in table */

    const uint16_t *base = NULL;  /* base value table to use */

    const uint16_t *extra = NULL; /* extra bits table to use */

    unsigned match = 0;           /* use base and extra for symbol >= match */

    uint16_t ALIGNED_(16) count[MAX_BITS+1]; /* number of codes of each length */

    uint16_t offs[MAX_BITS+1];    /* offsets in table for each length */

    static const uint16_t lbase[31] = { /* Length codes 257..285 base */

        3, 4, 5, 6, 7, 8, 9, 10, 11, 13, 15, 17, 19, 23, 27, 31,

        35, 43, 51, 59, 67, 83, 99, 115, 131, 163, 195, 227, 258, 0, 0};

    static const uint16_t lext[31] = { /* Length codes 257..285 extra */

        16, 16, 16, 16, 16, 16, 16, 16, 17, 17, 17, 17, 18, 18, 18, 18,

        19, 19, 19, 19, 20, 20, 20, 20, 21, 21, 21, 21, 16, 203, 77};

    static const uint16_t dbase[32] = { /* Distance codes 0..29 base */

        1, 2, 3, 4, 5, 7, 9, 13, 17, 25, 33, 49, 65, 97, 129, 193,

        257, 385, 513, 769, 1025, 1537, 2049, 3073, 4097, 6145,

        8193, 12289, 16385, 24577, 0, 0};

    static const uint16_t dext[32] = { /* Distance codes 0..29 extra */

        16, 16, 16, 16, 17, 17, 18, 18, 19, 19, 20, 20, 21, 21, 22, 22,

        23, 23, 24, 24, 25, 25, 26, 26, 27, 27,

        28, 28, 29, 29, 64, 64};

    /*

       Process a set of code lengths to create a canonical Huffman code.  The

       code lengths are lens[0..codes-1].  Each length corresponds to the

       symbols 0..codes-1.  The Huffman code is generated by first sorting the

       symbols by length from short to long, and retaining the symbol order

       for codes with equal lengths.  Then the code starts with all zero bits

       for the first code of the shortest length, and the codes are integer

       increments for the same length, and zeros are appended as the length

       increases.  For the deflate format, these bits are stored backwards

       from their more natural integer increment ordering, and so when the

       decoding tables are built in the large loop below, the integer codes

       are incremented backwards.

       This routine assumes, but does not check, that all of the entries in

       lens[] are in the range 0..MAXBITS.  The caller must assure this.

       1..MAXBITS is interpreted as that code length.  zero means that that

       symbol does not occur in this code.

       The codes are sorted by computing a count of codes for each length,

       creating from that a table of starting indices for each length in the

       sorted table, and then entering the symbols in order in the sorted

       table.  The sorted table is work[], with that space being provided by

       the caller.

       The length counts are used for other purposes as well, i.e. finding

       the minimum and maximum length codes, determining if there are any

       codes at all, checking for a valid set of lengths, and looking ahead

       at length counts to determine sub-table sizes when building the

       decoding tables.

     */

    /* accumulate lengths for codes (assumes lens[] all in 0..MAXBITS) */

    count_lengths(lens, codes, count);

    /* bound code lengths, force root to be within code lengths */

    root = *bits;

    for (max = MAX_BITS; max >= 1; max--)

        if (count[max] != 0) break;

    root = MIN(root, max);

    if (UNLIKELY(max == 0)) {           /* no symbols to code at all */

        here.op = (unsigned char)64;    /* invalid code marker */

        here.bits = (unsigned char)1;

        here.val = (uint16_t)0;

        *(*table)++ = here;             /* make a table to force an error */

        *(*table)++ = here;

        *bits = 1;

        return 0;     /* no symbols, but wait for decoding to report error */

    }

    for (min = 1; min < max; min++)

        if (count[min] != 0) break;

    root = MAX(root, min);

    /* check for an over-subscribed or incomplete set of lengths */

    left = 1;

    for (len = 1; len <= MAX_BITS; len++) {

        left <<= 1;

        left -= count[len];

        if (left < 0) return -1;        /* over-subscribed */

    }

    if (left > 0 && (type == CODES || max != 1))

        return -1;                      /* incomplete set */

    /* generate offsets into symbol table for each length for sorting */

    offs[1] = 0;

    for (len = 1; len < MAX_BITS; len++)

        offs[len + 1] = offs[len] + count[len];

    /* sort symbols by length, by symbol order within each length */

    for (sym = 0; sym < codes; sym++)

        if (lens[sym] != 0) work[offs[lens[sym]]++] = (uint16_t)sym;

    /*

       Create and fill in decoding tables.  In this loop, the table being

       filled is at next and has curr index bits.  The code being used is huff

       with length len.  That code is converted to an index by dropping drop

       bits off of the bottom.  For codes where len is less than drop + curr,

       those top drop + curr - len bits are incremented through all values to

       fill the table with replicated entries.

       root is the number of index bits for the root table.  When len exceeds

       root, sub-tables are created pointed to by the root entry with an index

       of the low root bits of huff.  This is saved in low to check for when a

       new sub-table should be started.  drop is zero when the root table is

       being filled, and drop is root when sub-tables are being filled.

       When a new sub-table is needed, it is necessary to look ahead in the

       code lengths to determine what size sub-table is needed.  The length

       counts are used for this, and so count[] is decremented as codes are

       entered in the tables.

       used keeps track of how many table entries have been allocated from the

       provided *table space.  It is checked for LENS and DIST tables against

       the constants ENOUGH_LENS and ENOUGH_DISTS to guard against changes in

       the initial root table size constants.  See the comments in inftrees.h

       for more information.

       sym increments through all symbols, and the loop terminates when

       all codes of length max, i.e. all codes, have been processed.  This

       routine permits incomplete codes, so another loop after this one fills

       in the rest of the decoding tables with invalid code markers.

     */

    /* The packed table fill below reads and writes each entry as a single uint32_t. */

    Assert(sizeof(code) == 4, "code entry must be exactly 4 bytes");

    /* set up for code type */

    switch (type) {

    case CODES:

        base = extra = work;    /* dummy value--not used */

        match = 20;

        break;

    case LENS:

        base = lbase;

        extra = lext;

        match = 257;

        break;

    case DISTS:

        base = dbase;

        extra = dext;

    }

    /* initialize state for loop */

    rhuff = 0;                  /* starting code, reversed */

    huff = 0;                   /* starting code */

    sym = 0;                    /* starting code symbol */

    len = min;                  /* starting code length */

    next = *table;              /* current table to fill in */

    curr = root;                /* current table index bits */

    drop = 0;                   /* current bits to drop from code for index */

    low = (unsigned)(-1);       /* trigger new sub-table when len > root */

    used = 1U << root;          /* use root table entries */

    mask = used - 1;            /* mask for comparing low */

    /* check available table space */

    if ((type == LENS && used > ENOUGH_LENS) ||

        (type == DISTS && used > ENOUGH_DISTS))

        return 1;

    /* process all codes and make table entries */

    for (;;) {

        /* create table entry */

        here.bits = (unsigned char)(len - drop);

        if (LIKELY(work[sym] >= match)) {

            unsigned op = extra[work[sym] - match];

            here.op = COMBINE_OP(op, here.bits);

            here.bits = COMBINE_BITS(here.bits, op);

            here.val = base[work[sym] - match];

        } else if (work[sym] + 1U < match) {

            here.op = (unsigned char)0;

            here.val = work[sym];

        } else {

            here.op = (unsigned char)(32 + 64);         /* end of block */

            here.val = 0;

        }

        /* replicate for those indices with low len bits equal to huff */

        incr = 1U << (len - drop);

        fill = 1U << curr;

        min = fill;                 /* save offset to next table */

        {

            /* `here` is a 4-byte struct that the compiler writes one field at a time.

               Read it as a packed uint32_t so each slot is filled by a single store. */

            uint32_t here_u32 = zng_memread_4(&here);

            code *base_ptr = next + (huff >> drop);

            unsigned step4 = incr << 2;

            /* Unroll by four so the slot addresses are independent. */

            while (fill >= step4) {

                zng_memwrite_4(&base_ptr[fill - 1 * incr], here_u32);

                zng_memwrite_4(&base_ptr[fill - 2 * incr], here_u32);

                zng_memwrite_4(&base_ptr[fill - 3 * incr], here_u32);

                zng_memwrite_4(&base_ptr[fill - 4 * incr], here_u32);

                fill -= step4;

            }

            while (fill != 0) {

                fill -= incr;

                zng_memwrite_4(&base_ptr[fill], here_u32);

            }

        }

        /* backwards increment the len-bit code huff */

        rhuff = (uint16_t)(rhuff + (0x8000u >> (len - 1)));

        huff = zng_bitreverse16(rhuff);

        /* go to next symbol, update count, len */

        sym++;

        if (--(count[len]) == 0) {

            if (len == max)

                break;

            len = lens[work[sym]];

        }

        /* create new sub-table if needed */

        if (len > root && (huff & mask) != low) {

            /* if first time, transition to sub-tables */

            if (drop == 0)

                drop = root;

            /* increment past last table */

            next += min;            /* here min is 1 << curr */

            /* determine length of next table */

            curr = len - drop;

            left = (int)(1 << curr);

            while (curr + drop < max) {

                left -= count[curr + drop];

                if (left <= 0)

                    break;

                curr++;

                left <<= 1;

            }

            /* check for enough space */

            used += 1U << curr;

            if ((type == LENS && used > ENOUGH_LENS) || (type == DISTS && used > ENOUGH_DISTS))

                return 1;

            /* point entry in root table to sub-table */

            low = huff & mask;

            (*table)[low].op = (unsigned char)curr;

            (*table)[low].bits = (unsigned char)root;

            (*table)[low].val = (uint16_t)(next - *table);

        }

    }

    /* fill in remaining table entry if code is incomplete (guaranteed to have

       at most one remaining entry, since if the code is incomplete, the

       maximum code length that was allowed to get this far is one bit) */

    if (UNLIKELY(huff != 0)) {

        here.op = (unsigned char)64;            /* invalid code marker */

        here.bits = (unsigned char)(len - drop);

        here.val = (uint16_t)0;

        next[huff] = here;

    }

    /* set return parameters */

    *table += used;

    *bits = root;

    return 0;

}