~bzr-pqm/bzr/bzr.dev

/*

 * diff-delta.c: generate a delta between two buffers

 *

 * This code was greatly inspired by parts of LibXDiff from Davide Libenzi

 * http://www.xmailserver.org/xdiff-lib.html

 *

 * Rewritten for GIT by Nicolas Pitre <nico@cam.org>, (C) 2005-2007

 * Adapted for Bazaar by John Arbash Meinel <john@arbash-meinel.com> (C) 2009

 *

 * This program is free software; you can redistribute it and/or modify

 * it under the terms of the GNU General Public License as published by

 * the Free Software Foundation; either version 2 of the License, or

 * (at your option) any later version.

 *

 * NB: The version in GIT is 'version 2 of the Licence only', however Nicolas

 * has granted permission for use under 'version 2 or later' in private email

 * to Robert Collins and Karl Fogel on the 6th April 2009.

 */

#include <stdio.h>

#include "delta.h"

#include <stdlib.h>

#include <string.h>

#include <assert.h>

/* maximum hash entry list for the same hash bucket */

#define HASH_LIMIT 64

#define RABIN_SHIFT 23

#define RABIN_WINDOW 16

/* The hash map is sized to put 4 entries per bucket, this gives us ~even room

 * for more data. Tweaking this number above 4 doesn't seem to help much,

 * anyway.

 */

#define EXTRA_NULLS 4

static const unsigned int T[256] = {

    0x00000000, 0xab59b4d1, 0x56b369a2, 0xfdeadd73, 0x063f6795, 0xad66d344,

    0x508c0e37, 0xfbd5bae6, 0x0c7ecf2a, 0xa7277bfb, 0x5acda688, 0xf1941259,

    0x0a41a8bf, 0xa1181c6e, 0x5cf2c11d, 0xf7ab75cc, 0x18fd9e54, 0xb3a42a85,

    0x4e4ef7f6, 0xe5174327, 0x1ec2f9c1, 0xb59b4d10, 0x48719063, 0xe32824b2,

    0x1483517e, 0xbfdae5af, 0x423038dc, 0xe9698c0d, 0x12bc36eb, 0xb9e5823a,

    0x440f5f49, 0xef56eb98, 0x31fb3ca8, 0x9aa28879, 0x6748550a, 0xcc11e1db,

    0x37c45b3d, 0x9c9defec, 0x6177329f, 0xca2e864e, 0x3d85f382, 0x96dc4753,

    0x6b369a20, 0xc06f2ef1, 0x3bba9417, 0x90e320c6, 0x6d09fdb5, 0xc6504964,

    0x2906a2fc, 0x825f162d, 0x7fb5cb5e, 0xd4ec7f8f, 0x2f39c569, 0x846071b8,

    0x798aaccb, 0xd2d3181a, 0x25786dd6, 0x8e21d907, 0x73cb0474, 0xd892b0a5,

    0x23470a43, 0x881ebe92, 0x75f463e1, 0xdeadd730, 0x63f67950, 0xc8afcd81,

    0x354510f2, 0x9e1ca423, 0x65c91ec5, 0xce90aa14, 0x337a7767, 0x9823c3b6,

    0x6f88b67a, 0xc4d102ab, 0x393bdfd8, 0x92626b09, 0x69b7d1ef, 0xc2ee653e,

    0x3f04b84d, 0x945d0c9c, 0x7b0be704, 0xd05253d5, 0x2db88ea6, 0x86e13a77,

    0x7d348091, 0xd66d3440, 0x2b87e933, 0x80de5de2, 0x7775282e, 0xdc2c9cff,

    0x21c6418c, 0x8a9ff55d, 0x714a4fbb, 0xda13fb6a, 0x27f92619, 0x8ca092c8,

    0x520d45f8, 0xf954f129, 0x04be2c5a, 0xafe7988b, 0x5432226d, 0xff6b96bc,

    0x02814bcf, 0xa9d8ff1e, 0x5e738ad2, 0xf52a3e03, 0x08c0e370, 0xa39957a1,

    0x584ced47, 0xf3155996, 0x0eff84e5, 0xa5a63034, 0x4af0dbac, 0xe1a96f7d,

    0x1c43b20e, 0xb71a06df, 0x4ccfbc39, 0xe79608e8, 0x1a7cd59b, 0xb125614a,

    0x468e1486, 0xedd7a057, 0x103d7d24, 0xbb64c9f5, 0x40b17313, 0xebe8c7c2,

    0x16021ab1, 0xbd5bae60, 0x6cb54671, 0xc7ecf2a0, 0x3a062fd3, 0x915f9b02,

    0x6a8a21e4, 0xc1d39535, 0x3c394846, 0x9760fc97, 0x60cb895b, 0xcb923d8a,

    0x3678e0f9, 0x9d215428, 0x66f4eece, 0xcdad5a1f, 0x3047876c, 0x9b1e33bd,

    0x7448d825, 0xdf116cf4, 0x22fbb187, 0x89a20556, 0x7277bfb0, 0xd92e0b61,

    0x24c4d612, 0x8f9d62c3, 0x7836170f, 0xd36fa3de, 0x2e857ead, 0x85dcca7c,

    0x7e09709a, 0xd550c44b, 0x28ba1938, 0x83e3ade9, 0x5d4e7ad9, 0xf617ce08,

    0x0bfd137b, 0xa0a4a7aa, 0x5b711d4c, 0xf028a99d, 0x0dc274ee, 0xa69bc03f,

    0x5130b5f3, 0xfa690122, 0x0783dc51, 0xacda6880, 0x570fd266, 0xfc5666b7,

    0x01bcbbc4, 0xaae50f15, 0x45b3e48d, 0xeeea505c, 0x13008d2f, 0xb85939fe,

    0x438c8318, 0xe8d537c9, 0x153feaba, 0xbe665e6b, 0x49cd2ba7, 0xe2949f76,

    0x1f7e4205, 0xb427f6d4, 0x4ff24c32, 0xe4abf8e3, 0x19412590, 0xb2189141,

    0x0f433f21, 0xa41a8bf0, 0x59f05683, 0xf2a9e252, 0x097c58b4, 0xa225ec65,

    0x5fcf3116, 0xf49685c7, 0x033df00b, 0xa86444da, 0x558e99a9, 0xfed72d78,

    0x0502979e, 0xae5b234f, 0x53b1fe3c, 0xf8e84aed, 0x17bea175, 0xbce715a4,

    0x410dc8d7, 0xea547c06, 0x1181c6e0, 0xbad87231, 0x4732af42, 0xec6b1b93,

    0x1bc06e5f, 0xb099da8e, 0x4d7307fd, 0xe62ab32c, 0x1dff09ca, 0xb6a6bd1b,

    0x4b4c6068, 0xe015d4b9, 0x3eb80389, 0x95e1b758, 0x680b6a2b, 0xc352defa,

    0x3887641c, 0x93ded0cd, 0x6e340dbe, 0xc56db96f, 0x32c6cca3, 0x999f7872,

    0x6475a501, 0xcf2c11d0, 0x34f9ab36, 0x9fa01fe7, 0x624ac294, 0xc9137645,

    0x26459ddd, 0x8d1c290c, 0x70f6f47f, 0xdbaf40ae, 0x207afa48, 0x8b234e99,

    0x76c993ea, 0xdd90273b, 0x2a3b52f7, 0x8162e626, 0x7c883b55, 0xd7d18f84,

    0x2c043562, 0x875d81b3, 0x7ab75cc0, 0xd1eee811

};

static const unsigned int U[256] = {

    0x00000000, 0x7eb5200d, 0x5633f4cb, 0x2886d4c6, 0x073e5d47, 0x798b7d4a,

    0x510da98c, 0x2fb88981, 0x0e7cba8e, 0x70c99a83, 0x584f4e45, 0x26fa6e48,

    0x0942e7c9, 0x77f7c7c4, 0x5f711302, 0x21c4330f, 0x1cf9751c, 0x624c5511,

    0x4aca81d7, 0x347fa1da, 0x1bc7285b, 0x65720856, 0x4df4dc90, 0x3341fc9d,

    0x1285cf92, 0x6c30ef9f, 0x44b63b59, 0x3a031b54, 0x15bb92d5, 0x6b0eb2d8,

    0x4388661e, 0x3d3d4613, 0x39f2ea38, 0x4747ca35, 0x6fc11ef3, 0x11743efe,

    0x3eccb77f, 0x40799772, 0x68ff43b4, 0x164a63b9, 0x378e50b6, 0x493b70bb,

    0x61bda47d, 0x1f088470, 0x30b00df1, 0x4e052dfc, 0x6683f93a, 0x1836d937,

    0x250b9f24, 0x5bbebf29, 0x73386bef, 0x0d8d4be2, 0x2235c263, 0x5c80e26e,

    0x740636a8, 0x0ab316a5, 0x2b7725aa, 0x55c205a7, 0x7d44d161, 0x03f1f16c,

    0x2c4978ed, 0x52fc58e0, 0x7a7a8c26, 0x04cfac2b, 0x73e5d470, 0x0d50f47d,

    0x25d620bb, 0x5b6300b6, 0x74db8937, 0x0a6ea93a, 0x22e87dfc, 0x5c5d5df1,

    0x7d996efe, 0x032c4ef3, 0x2baa9a35, 0x551fba38, 0x7aa733b9, 0x041213b4,

    0x2c94c772, 0x5221e77f, 0x6f1ca16c, 0x11a98161, 0x392f55a7, 0x479a75aa,

    0x6822fc2b, 0x1697dc26, 0x3e1108e0, 0x40a428ed, 0x61601be2, 0x1fd53bef,

    0x3753ef29, 0x49e6cf24, 0x665e46a5, 0x18eb66a8, 0x306db26e, 0x4ed89263,

    0x4a173e48, 0x34a21e45, 0x1c24ca83, 0x6291ea8e, 0x4d29630f, 0x339c4302,

    0x1b1a97c4, 0x65afb7c9, 0x446b84c6, 0x3adea4cb, 0x1258700d, 0x6ced5000,

    0x4355d981, 0x3de0f98c, 0x15662d4a, 0x6bd30d47, 0x56ee4b54, 0x285b6b59,

    0x00ddbf9f, 0x7e689f92, 0x51d01613, 0x2f65361e, 0x07e3e2d8, 0x7956c2d5,

    0x5892f1da, 0x2627d1d7, 0x0ea10511, 0x7014251c, 0x5facac9d, 0x21198c90,

    0x099f5856, 0x772a785b, 0x4c921c31, 0x32273c3c, 0x1aa1e8fa, 0x6414c8f7,

    0x4bac4176, 0x3519617b, 0x1d9fb5bd, 0x632a95b0, 0x42eea6bf, 0x3c5b86b2,

    0x14dd5274, 0x6a687279, 0x45d0fbf8, 0x3b65dbf5, 0x13e30f33, 0x6d562f3e,

    0x506b692d, 0x2ede4920, 0x06589de6, 0x78edbdeb, 0x5755346a, 0x29e01467,

    0x0166c0a1, 0x7fd3e0ac, 0x5e17d3a3, 0x20a2f3ae, 0x08242768, 0x76910765,

    0x59298ee4, 0x279caee9, 0x0f1a7a2f, 0x71af5a22, 0x7560f609, 0x0bd5d604,

    0x235302c2, 0x5de622cf, 0x725eab4e, 0x0ceb8b43, 0x246d5f85, 0x5ad87f88,

    0x7b1c4c87, 0x05a96c8a, 0x2d2fb84c, 0x539a9841, 0x7c2211c0, 0x029731cd,

    0x2a11e50b, 0x54a4c506, 0x69998315, 0x172ca318, 0x3faa77de, 0x411f57d3,

    0x6ea7de52, 0x1012fe5f, 0x38942a99, 0x46210a94, 0x67e5399b, 0x19501996,

    0x31d6cd50, 0x4f63ed5d, 0x60db64dc, 0x1e6e44d1, 0x36e89017, 0x485db01a,

    0x3f77c841, 0x41c2e84c, 0x69443c8a, 0x17f11c87, 0x38499506, 0x46fcb50b,

    0x6e7a61cd, 0x10cf41c0, 0x310b72cf, 0x4fbe52c2, 0x67388604, 0x198da609,

    0x36352f88, 0x48800f85, 0x6006db43, 0x1eb3fb4e, 0x238ebd5d, 0x5d3b9d50,

    0x75bd4996, 0x0b08699b, 0x24b0e01a, 0x5a05c017, 0x728314d1, 0x0c3634dc,

    0x2df207d3, 0x534727de, 0x7bc1f318, 0x0574d315, 0x2acc5a94, 0x54797a99,

    0x7cffae5f, 0x024a8e52, 0x06852279, 0x78300274, 0x50b6d6b2, 0x2e03f6bf,

    0x01bb7f3e, 0x7f0e5f33, 0x57888bf5, 0x293dabf8, 0x08f998f7, 0x764cb8fa,

    0x5eca6c3c, 0x207f4c31, 0x0fc7c5b0, 0x7172e5bd, 0x59f4317b, 0x27411176,

    0x1a7c5765, 0x64c97768, 0x4c4fa3ae, 0x32fa83a3, 0x1d420a22, 0x63f72a2f,

    0x4b71fee9, 0x35c4dee4, 0x1400edeb, 0x6ab5cde6, 0x42331920, 0x3c86392d,

    0x133eb0ac, 0x6d8b90a1, 0x450d4467, 0x3bb8646a

};

struct index_entry {

    const unsigned char *ptr;

    const struct source_info *src;

    unsigned int val;

};

struct index_entry_linked_list {

    struct index_entry *p_entry;

    struct index_entry_linked_list *next;

};

struct unpacked_index_entry {

    struct index_entry entry;

    struct unpacked_index_entry *next;

};

struct delta_index {

    unsigned long memsize; /* Total bytes pointed to by this index */

    const struct source_info *last_src; /* Information about the referenced source */

    unsigned int hash_mask; /* val & hash_mask gives the hash index for a given

                               entry */

    unsigned int num_entries; /* The total number of entries in this index */

    struct index_entry *last_entry; /* Pointer to the last valid entry */

    struct index_entry *hash[];

};

static unsigned int

limit_hash_buckets(struct unpacked_index_entry **hash,

                   unsigned int *hash_count, unsigned int hsize,

                   unsigned int entries)

{

    struct unpacked_index_entry *entry;

    unsigned int i;

    /*

     * Determine a limit on the number of entries in the same hash

     * bucket.  This guards us against pathological data sets causing

     * really bad hash distribution with most entries in the same hash

     * bucket that would bring us to O(m*n) computing costs (m and n

     * corresponding to reference and target buffer sizes).

     *

     * Make sure none of the hash buckets has more entries than

     * we're willing to test.  Otherwise we cull the entry list

     * uniformly to still preserve a good repartition across

     * the reference buffer.

     */

    for (i = 0; i < hsize; i++) {

        int acc;

        if (hash_count[i] <= HASH_LIMIT)

            continue;

        /* We leave exactly HASH_LIMIT entries in the bucket */

        entries -= hash_count[i] - HASH_LIMIT;

        entry = hash[i];

        acc = 0;

        /*

         * Assume that this loop is gone through exactly

         * HASH_LIMIT times and is entered and left with

         * acc==0.  So the first statement in the loop

         * contributes (hash_count[i]-HASH_LIMIT)*HASH_LIMIT

         * to the accumulator, and the inner loop consequently

         * is run (hash_count[i]-HASH_LIMIT) times, removing

         * one element from the list each time.  Since acc

         * balances out to 0 at the final run, the inner loop

         * body can't be left with entry==NULL.  So we indeed

         * encounter entry==NULL in the outer loop only.

         */

        do {

            acc += hash_count[i] - HASH_LIMIT;

            if (acc > 0) {

                struct unpacked_index_entry *keep = entry;

                do {

                    entry = entry->next;

                    acc -= HASH_LIMIT;

                } while (acc > 0);

                keep->next = entry->next;

            }

            entry = entry->next;

        } while (entry);

    }

    return entries;

}

static struct delta_index *

pack_delta_index(struct unpacked_index_entry **hash, unsigned int hsize,

                 unsigned int num_entries, struct delta_index *old_index)

{

    unsigned int i, j, hmask, memsize, fit_in_old, copied_count;

    struct unpacked_index_entry *entry;

    struct delta_index *index;

    struct index_entry *packed_entry, **packed_hash, *old_entry, *copy_from;

    struct index_entry null_entry = {0};

    void *mem;

    hmask = hsize - 1;

    // if (old_index) {

    //     fprintf(stderr, "Packing %d entries into %d for total of %d entries"

    //                     " %x => %x\n",

    //                     num_entries - old_index->num_entries,

    //                     old_index->num_entries, num_entries,

    //                     old_index->hash_mask, hmask);

    // } else {

    //     fprintf(stderr, "Packing %d entries into a new index\n",

    //                     num_entries);

    // }

    /* First, see if we can squeeze the new items into the existing structure.

     */

    fit_in_old = 0;

    copied_count = 0;

    if (old_index && old_index->hash_mask == hmask) {

        fit_in_old = 1;

        for (i = 0; i < hsize; ++i) {

            packed_entry = NULL;

            for (entry = hash[i]; entry; entry = entry->next) {

                if (packed_entry == NULL) {

                    /* Find the last open spot */

                    packed_entry = old_index->hash[i + 1];

                    --packed_entry;

                    while (packed_entry >= old_index->hash[i]

                           && packed_entry->ptr == NULL) {

                        --packed_entry;

                    }

                    ++packed_entry;

                }

                if (packed_entry >= old_index->hash[i+1]

                    || packed_entry->ptr != NULL) {

                    /* There are no free spots here :( */

                    fit_in_old = 0;

                    break;

                }

                /* We found an empty spot to put this entry

                 * Copy it over, and remove it from the linked list, just in

                 * case we end up running out of room later.

                 */

                *packed_entry++ = entry->entry;

                assert(entry == hash[i]);

                hash[i] = entry->next;

                copied_count += 1;

                old_index->num_entries++;

            }

            if (!fit_in_old) {

                break;

            }

        }

    }

    if (old_index) {

        if (fit_in_old) {

            // fprintf(stderr, "Fit all %d entries into old index\n",

            //                 copied_count);

            /* No need to allocate a new buffer */

            return NULL;

        } else {

            // fprintf(stderr, "Fit only %d entries into old index,"

            //                 " reallocating\n", copied_count);

        }

    }

    /*

     * Now create the packed index in array form

     * rather than linked lists.

     * Leave a 2-entry gap for inserting more entries between the groups

     */

    memsize = sizeof(*index)

        + sizeof(*packed_hash) * (hsize+1)

        + sizeof(*packed_entry) * (num_entries + hsize * EXTRA_NULLS);

    mem = malloc(memsize);

    if (!mem) {

        return NULL;

    }

    index = mem;

    index->memsize = memsize;

    index->hash_mask = hmask;

    index->num_entries = num_entries;

    if (old_index) {

        if (hmask < old_index->hash_mask) {

            fprintf(stderr, "hash mask was shrunk %x => %x\n",

                            old_index->hash_mask, hmask);

        }

        assert(hmask >= old_index->hash_mask);

    }

    mem = index->hash;

    packed_hash = mem;

    mem = packed_hash + (hsize+1);

    packed_entry = mem;

    for (i = 0; i < hsize; i++) {

        /*

         * Coalesce all entries belonging to one linked list

         * into consecutive array entries.

         */

        packed_hash[i] = packed_entry;

        /* Old comes earlier as a source, so it always comes first in a given

         * hash bucket.

         */

        if (old_index) {

            /* Could we optimize this to use memcpy when hmask ==

             * old_index->hash_mask? Would it make any real difference?

             */

            j = i & old_index->hash_mask;

            copy_from = old_index->hash[j];

            for (old_entry = old_index->hash[j];

                 old_entry < old_index->hash[j + 1] && old_entry->ptr != NULL;

                 old_entry++) {

                if ((old_entry->val & hmask) == i) {

                    *packed_entry++ = *old_entry;

                }

            }

        }

        for (entry = hash[i]; entry; entry = entry->next) {

            *packed_entry++ = entry->entry;

        }

        /* TODO: At this point packed_entry - packed_hash[i] is the number of

         *       records that we have inserted into this hash bucket.

         *       We should *really* consider doing some limiting along the

         *       lines of limit_hash_buckets() to avoid pathological behavior.

         */

        /* Now add extra 'NULL' entries that we can use for future expansion. */

        for (j = 0; j < EXTRA_NULLS; ++j ) {

            *packed_entry++ = null_entry;

        }

    }

    /* Sentinel value to indicate the length of the last hash bucket */

    packed_hash[hsize] = packed_entry;

    if (packed_entry - (struct index_entry *)mem

        != num_entries + hsize*EXTRA_NULLS) {

        fprintf(stderr, "We expected %d entries, but created %d\n",

                num_entries + hsize*EXTRA_NULLS,

                (int)(packed_entry - (struct index_entry*)mem));

    }

    assert(packed_entry - (struct index_entry *)mem

            == num_entries + hsize*EXTRA_NULLS);

    index->last_entry = (packed_entry - 1);

    return index;

}

struct delta_index *

create_delta_index(const struct source_info *src,

                   struct delta_index *old)

{

    unsigned int i, hsize, hmask, num_entries, prev_val, *hash_count;

    unsigned int total_num_entries;

    const unsigned char *data, *buffer;

    struct delta_index *index;

    struct unpacked_index_entry *entry, **hash;

    void *mem;

    unsigned long memsize;

    if (!src->buf || !src->size)

        return NULL;

    buffer = src->buf;

    /* Determine index hash size.  Note that indexing skips the

       first byte to allow for optimizing the Rabin's polynomial

       initialization in create_delta(). */

    num_entries = (src->size - 1)  / RABIN_WINDOW;

    if (old != NULL)

        total_num_entries = num_entries + old->num_entries;

    else

        total_num_entries = num_entries;

    hsize = total_num_entries / 4;

    for (i = 4; (1u << i) < hsize && i < 31; i++);

    hsize = 1 << i;

    hmask = hsize - 1;

    if (old && old->hash_mask > hmask) {

        hmask = old->hash_mask;

        hsize = hmask + 1;

    }

    /* allocate lookup index */

    memsize = sizeof(*hash) * hsize +

          sizeof(*entry) * total_num_entries;

    mem = malloc(memsize);

    if (!mem)

        return NULL;

    hash = mem;

    mem = hash + hsize;

    entry = mem;

    memset(hash, 0, hsize * sizeof(*hash));

    /* allocate an array to count hash num_entries */

    hash_count = calloc(hsize, sizeof(*hash_count));

    if (!hash_count) {

        free(hash);

        return NULL;

    }

    /* then populate the index for the new data */

    prev_val = ~0;

    for (data = buffer + num_entries * RABIN_WINDOW - RABIN_WINDOW;

         data >= buffer;

         data -= RABIN_WINDOW) {

        unsigned int val = 0;

        for (i = 1; i <= RABIN_WINDOW; i++)

            val = ((val << 8) | data[i]) ^ T[val >> RABIN_SHIFT];

        if (val == prev_val) {

            /* keep the lowest of consecutive identical blocks */

            entry[-1].entry.ptr = data + RABIN_WINDOW;

            --num_entries;

            --total_num_entries;

        } else {

            prev_val = val;

            i = val & hmask;

            entry->entry.ptr = data + RABIN_WINDOW;

            entry->entry.val = val;

            entry->entry.src = src;

            entry->next = hash[i];

            hash[i] = entry++;

            hash_count[i]++;

        }

    }

    /* TODO: It would be nice to limit_hash_buckets at a better time. */

    total_num_entries = limit_hash_buckets(hash, hash_count, hsize,

                                           total_num_entries);

    free(hash_count);

    if (old) {

        old->last_src = src;

    }

    index = pack_delta_index(hash, hsize, total_num_entries, old);

    free(hash);

    if (!index) {

        return NULL;

    }

    index->last_src = src;

    return index;

}

/* Take some entries, and put them into a custom hash.

 * @param entries   A list of entries, sorted by position in file

 * @param num_entries   Length of entries

 * @param out_hsize     The maximum size of the hash, the final size will be

 *                      returned here

 */

struct index_entry_linked_list **

_put_entries_into_hash(struct index_entry *entries, unsigned int num_entries,

                       unsigned int hsize)

{

    unsigned int hash_offset, hmask, memsize;

    struct index_entry *entry, *last_entry;

    struct index_entry_linked_list *out_entry, **hash;

    void *mem;

    hmask = hsize - 1;

    memsize = sizeof(*hash) * hsize +

          sizeof(*out_entry) * num_entries;

    mem = malloc(memsize);

    if (!mem)

        return NULL;

    hash = mem;

    mem = hash + hsize;

    out_entry = mem;

    memset(hash, 0, sizeof(*hash)*(hsize+1));

    /* We know that entries are in the order we want in the output, but they

     * aren't "grouped" by hash bucket yet.

     */

    last_entry = entries + num_entries;

    for (entry = entries + num_entries - 1; entry >= entries; --entry) {

        hash_offset = entry->val & hmask;

        out_entry->p_entry = entry;

        out_entry->next = hash[hash_offset];

        /* TODO: Remove entries that have identical vals, or at least filter

         *       the map a little bit.

         * if (hash[i] != NULL) {

         * }

         */

        hash[hash_offset] = out_entry;

        ++out_entry;

    }

    return hash;

}

struct delta_index *

create_index_from_old_and_new_entries(const struct delta_index *old_index,

                                      struct index_entry *entries,

                                      unsigned int num_entries)

{

    unsigned int i, j, hsize, hmask, total_num_entries;

    struct delta_index *index;

    struct index_entry *entry, *packed_entry, **packed_hash;

    struct index_entry *last_entry, null_entry = {0};

    void *mem;

    unsigned long memsize;

    struct index_entry_linked_list *unpacked_entry, **mini_hash;

    /* Determine index hash size.  Note that indexing skips the

       first byte to allow for optimizing the Rabin's polynomial

       initialization in create_delta(). */

    total_num_entries = num_entries + old_index->num_entries;

    hsize = total_num_entries / 4;

    for (i = 4; (1u << i) < hsize && i < 31; i++);

    hsize = 1 << i;

    if (hsize < old_index->hash_mask) {

        /* For some reason, there was a code path that would actually *shrink*

         * the hash size. This screws with some later code, and in general, I

         * think it better to make the hash bigger, rather than smaller. So

         * we'll just force the size here.

         * Possibly done by create_delta_index running into a

         * limit_hash_buckets call, that ended up transitioning across a

         * power-of-2. The cause isn't 100% clear, though.

         */

        hsize = old_index->hash_mask + 1;

    }

    hmask = hsize - 1;

    // fprintf(stderr, "resizing index to insert %d entries into array"

    //                 " with %d entries: %x => %x\n",

    //         num_entries, old_index->num_entries, old_index->hash_mask, hmask);

    memsize = sizeof(*index)

        + sizeof(*packed_hash) * (hsize+1)

        + sizeof(*packed_entry) * (total_num_entries + hsize*EXTRA_NULLS);

    mem = malloc(memsize);

    if (!mem) {

        return NULL;

    }

    index = mem;

    index->memsize = memsize;

    index->hash_mask = hmask;

    index->num_entries = total_num_entries;

    index->last_src = old_index->last_src;

    mem = index->hash;

    packed_hash = mem;

    mem = packed_hash + (hsize+1);

    packed_entry = mem;

    mini_hash = _put_entries_into_hash(entries, num_entries, hsize);

    if (mini_hash == NULL) {

        free(index);

        return NULL;

    }

    last_entry = entries + num_entries;

    for (i = 0; i < hsize; i++) {

        /*

         * Coalesce all entries belonging in one hash bucket

         * into consecutive array entries.

         * The entries in old_index all come before 'entries'.

         */

        packed_hash[i] = packed_entry;

        /* Copy any of the old entries across */

        /* Would we rather use memcpy? */

        if (hmask == old_index->hash_mask) {

            for (entry = old_index->hash[i];

                 entry < old_index->hash[i+1] && entry->ptr != NULL;

                 ++entry) {

                assert((entry->val & hmask) == i);

                *packed_entry++ = *entry;

            }

        } else {

            /* If we resized the index from this action, all of the old values

             * will be found in the previous location, but they will end up

             * spread across the new locations.

             */

            j = i & old_index->hash_mask;

            for (entry = old_index->hash[j];

                 entry < old_index->hash[j+1] && entry->ptr != NULL;

                 ++entry) {

                assert((entry->val & old_index->hash_mask) == j);

                if ((entry->val & hmask) == i) {

                    /* Any entries not picked up here will be picked up on the

                     * next pass.

                     */

                    *packed_entry++ = *entry;

                }

            }

        }

        /* Now see if we need to insert any of the new entries.

         * Note that loop ends up O(hsize*num_entries), so we expect that

         * num_entries is always small.

         * We also help a little bit by collapsing the entry range when the

         * endpoints are inserted. However, an alternative would be to build a

         * quick hash lookup for just the new entries.

         * Testing shows that this list can easily get up to about 100

         * entries, the tradeoff is a malloc, 1 pass over the entries, copying

         * them into a sorted buffer, and a free() when done,

         */

        for (unpacked_entry = mini_hash[i];

             unpacked_entry;

             unpacked_entry = unpacked_entry->next) {

            assert((unpacked_entry->p_entry->val & hmask) == i);

            *packed_entry++ = *(unpacked_entry->p_entry);

        }

        /* Now insert some extra nulls */

        for (j = 0; j < EXTRA_NULLS; ++j) {

            *packed_entry++ = null_entry;

        }

    }

    free(mini_hash);

    /* Sentinel value to indicate the length of the last hash bucket */

    packed_hash[hsize] = packed_entry;

    if ((packed_entry - (struct index_entry *)mem)

        != (total_num_entries + hsize*EXTRA_NULLS)) {

        fprintf(stderr, "We expected %d entries, but created %d\n",

                total_num_entries + hsize*EXTRA_NULLS,

                (int)(packed_entry - (struct index_entry*)mem));

        fflush(stderr);

    }

    assert((packed_entry - (struct index_entry *)mem)

           == (total_num_entries + hsize * EXTRA_NULLS));

    index->last_entry = (packed_entry - 1);

    return index;

}

void

get_text(char buff[128], const unsigned char *ptr)

{

    unsigned int i;

    const unsigned char *start;

    unsigned char cmd;

    start = (ptr-RABIN_WINDOW-1);

    cmd = *(start);

    if (cmd < 0x80) {// This is likely to be an insert instruction

        if (cmd < RABIN_WINDOW) {