]> git.ipfire.org Git - thirdparty/git.git/blobdiff - fast-import.c
Improve reuse of sha1_file library within fast-import.
[thirdparty/git.git] / fast-import.c
index 778b8bfdd4aab6697b9955d1ee7ae928c1905956..f0f51a68998222d48b4c3f8e2c5727f0842690f4 100644 (file)
@@ -20,6 +20,7 @@ Format of STDIN stream:
     'committer' sp name '<' email '>' ts tz lf
     commit_msg
     ('from' sp (ref_str | hexsha1 | sha1exp_str | idnum) lf)?
+    ('merge' sp (ref_str | hexsha1 | sha1exp_str | idnum) lf)*
     file_change*
     lf;
   commit_msg ::= data;
@@ -35,7 +36,9 @@ Format of STDIN stream:
     tag_msg;
   tag_msg ::= data;
 
-  reset_branch ::= 'reset' sp ref_str lf;
+  reset_branch ::= 'reset' sp ref_str lf
+    ('from' sp (ref_str | hexsha1 | sha1exp_str | idnum) lf)?
+    lf;
 
      # note: the first idnum in a stream should be 1 and subsequent
      # idnums should not have gaps between values as this will cause
@@ -132,9 +135,10 @@ struct mark_set
 struct last_object
 {
        void *data;
-       unsigned int len;
+       unsigned long len;
+       unsigned long offset;
        unsigned int depth;
-       unsigned char sha1[20];
+       unsigned no_free:1;
 };
 
 struct mem_pool
@@ -157,14 +161,18 @@ struct tree_entry
 {
        struct tree_content *tree;
        struct atom_str* name;
-       unsigned int mode;
-       unsigned char sha1[20];
+       struct tree_entry_ms
+       {
+               unsigned int mode;
+               unsigned char sha1[20];
+       } versions[2];
 };
 
 struct tree_content
 {
        unsigned int entry_capacity; /* must match avail_tree_content */
        unsigned int entry_count;
+       unsigned int delta_depth;
        struct tree_entry *entries[FLEX_ARRAY]; /* more */
 };
 
@@ -191,6 +199,17 @@ struct tag
        unsigned char sha1[20];
 };
 
+struct dbuf
+{
+       void *buffer;
+       size_t capacity;
+};
+
+struct hash_list
+{
+       struct hash_list *next;
+       unsigned char sha1[20];
+};
 
 /* Stats and misc. counters */
 static unsigned long max_depth = 10;
@@ -203,6 +222,7 @@ static unsigned long duplicate_count;
 static unsigned long marks_set_count;
 static unsigned long object_count_by_type[9];
 static unsigned long duplicate_count_by_type[9];
+static unsigned long delta_count_by_type[9];
 
 /* Memory pools */
 static size_t mem_pool_alloc = 2*1024*1024 - sizeof(struct mem_pool);
@@ -215,16 +235,13 @@ static unsigned int atom_cnt;
 static struct atom_str **atom_table;
 
 /* The .pack file being generated */
+static struct packed_git *pack_data;
 static int pack_fd;
 static unsigned long pack_size;
 static unsigned char pack_sha1[20];
-static unsigned char* pack_base;
-static unsigned long pack_moff;
-static unsigned long pack_mlen = 128*1024*1024;
-static unsigned long page_size;
 
 /* Table of objects we've written. */
-static unsigned int object_entry_alloc = 1000;
+static unsigned int object_entry_alloc = 5000;
 static struct object_entry_pool *blocks;
 static struct object_entry *object_table[1 << 16];
 static struct mark_set *marks;
@@ -238,6 +255,8 @@ static unsigned int tree_entry_alloc = 1000;
 static void *avail_tree_entry;
 static unsigned int avail_tree_table_sz = 100;
 static struct avail_tree_content **avail_tree_table;
+static struct dbuf old_tree;
+static struct dbuf new_tree;
 
 /* Branch data */
 static unsigned long max_active_branches = 5;
@@ -253,6 +272,7 @@ static struct tag *last_tag;
 /* Input stream parsing */
 static struct strbuf command_buf;
 static unsigned long next_mark;
+static struct dbuf new_data;
 static FILE* branch_log;
 
 
@@ -277,7 +297,7 @@ static struct object_entry* new_object(unsigned char *sha1)
                alloc_objects(object_entry_alloc);
 
        e = blocks->next_free++;
-       memcpy(e->sha1, sha1, sizeof(e->sha1));
+       hashcpy(e->sha1, sha1);
        return e;
 }
 
@@ -286,7 +306,7 @@ static struct object_entry* find_object(unsigned char *sha1)
        unsigned int h = sha1[0] << 8 | sha1[1];
        struct object_entry *e;
        for (e = object_table[h]; e; e = e->next)
-               if (!memcmp(sha1, e->sha1, sizeof(e->sha1)))
+               if (!hashcmp(sha1, e->sha1))
                        return e;
        return NULL;
 }
@@ -298,7 +318,7 @@ static struct object_entry* insert_object(unsigned char *sha1)
        struct object_entry *p = NULL;
 
        while (e) {
-               if (!memcmp(sha1, e->sha1, sizeof(e->sha1)))
+               if (!hashcmp(sha1, e->sha1))
                        return e;
                p = e;
                e = e->next;
@@ -367,6 +387,17 @@ static char* pool_strdup(const char *s)
        return r;
 }
 
+static void size_dbuf(struct dbuf *b, size_t maxlen)
+{
+       if (b->buffer) {
+               if (b->capacity >= maxlen)
+                       return;
+               free(b->buffer);
+       }
+       b->capacity = ((maxlen / 1024) + 1) * 1024;
+       b->buffer = xmalloc(b->capacity);
+}
+
 static void insert_mark(unsigned long idnum, struct object_entry *oe)
 {
        struct mark_set *s = marks;
@@ -452,6 +483,8 @@ static struct branch* new_branch(const char *name)
        b = pool_calloc(1, sizeof(struct branch));
        b->name = pool_strdup(name);
        b->table_next_branch = branch_table[hc];
+       b->branch_tree.versions[0].mode = S_IFDIR;
+       b->branch_tree.versions[1].mode = S_IFDIR;
        branch_table[hc] = b;
        branch_count++;
        return b;
@@ -486,6 +519,7 @@ static struct tree_content* new_tree_content(unsigned int cnt)
 
        t = (struct tree_content*)f;
        t->entry_count = 0;
+       t->delta_depth = 0;
        return t;
 }
 
@@ -512,6 +546,7 @@ static struct tree_content* grow_tree_content(
 {
        struct tree_content *r = new_tree_content(t->entry_count + amt);
        r->entry_count = t->entry_count;
+       r->delta_depth = t->delta_depth;
        memcpy(r->entries,t->entries,t->entry_count*sizeof(t->entries[0]));
        release_tree_content(t);
        return r;
@@ -558,19 +593,6 @@ static void yread(int fd, void *buffer, size_t length)
        }
 }
 
-static void ywrite(int fd, void *buffer, size_t length)
-{
-       ssize_t ret = 0;
-       while (ret < length) {
-               ssize_t size = xwrite(fd, (char *) buffer + ret, length - ret);
-               if (!size)
-                       die("Write to descriptor %i: end of file", fd);
-               if (size < 0)
-                       die("Write to descriptor %i: %s", fd, strerror(errno));
-               ret += size;
-       }
-}
-
 static size_t encode_header(
        enum object_type type,
        size_t size,
@@ -579,7 +601,7 @@ static size_t encode_header(
        int n = 1;
        unsigned char c;
 
-       if (type < OBJ_COMMIT || type > OBJ_DELTA)
+       if (type < OBJ_COMMIT || type > OBJ_REF_DELTA)
                die("bad type %d", type);
 
        c = (type << 4) | (size & 15);
@@ -616,7 +638,7 @@ static int store_object(
        SHA1_Update(&c, dat, datlen);
        SHA1_Final(sha1, &c);
        if (sha1out)
-               memcpy(sha1out, sha1, sizeof(sha1));
+               hashcpy(sha1out, sha1);
 
        e = insert_object(sha1);
        if (mark)
@@ -642,20 +664,30 @@ static int store_object(
        deflateInit(&s, zlib_compression_level);
 
        if (delta) {
+               unsigned long ofs = e->offset - last->offset;
+               unsigned pos = sizeof(hdr) - 1;
+
+               delta_count_by_type[type]++;
                last->depth++;
                s.next_in = delta;
                s.avail_in = deltalen;
-               hdrlen = encode_header(OBJ_DELTA, deltalen, hdr);
-               ywrite(pack_fd, hdr, hdrlen);
-               ywrite(pack_fd, last->sha1, sizeof(sha1));
-               pack_size += hdrlen + sizeof(sha1);
+
+               hdrlen = encode_header(OBJ_OFS_DELTA, deltalen, hdr);
+               write_or_die(pack_fd, hdr, hdrlen);
+               pack_size += hdrlen;
+
+               hdr[pos] = ofs & 127;
+               while (ofs >>= 7)
+                       hdr[--pos] = 128 | (--ofs & 127);
+               write_or_die(pack_fd, hdr + pos, sizeof(hdr) - pos);
+               pack_size += sizeof(hdr) - pos;
        } else {
                if (last)
                        last->depth = 0;
                s.next_in = dat;
                s.avail_in = datlen;
                hdrlen = encode_header(type, datlen, hdr);
-               ywrite(pack_fd, hdr, hdrlen);
+               write_or_die(pack_fd, hdr, hdrlen);
                pack_size += hdrlen;
        }
 
@@ -665,142 +697,27 @@ static int store_object(
                /* nothing */;
        deflateEnd(&s);
 
-       ywrite(pack_fd, out, s.total_out);
+       write_or_die(pack_fd, out, s.total_out);
        pack_size += s.total_out;
 
        free(out);
        if (delta)
                free(delta);
        if (last) {
-               if (last->data)
+               if (last->data && !last->no_free)
                        free(last->data);
                last->data = dat;
+               last->offset = e->offset;
                last->len = datlen;
-               memcpy(last->sha1, sha1, sizeof(sha1));
        }
        return 0;
 }
 
-static unsigned char* map_pack(unsigned long offset, unsigned int *left)
-{
-       if (offset >= pack_size)
-               die("object offset outside of pack file");
-       if (!pack_base
-                       || offset < pack_moff
-                       || (offset + 20) >= (pack_moff + pack_mlen)) {
-               if (pack_base)
-                       munmap(pack_base, pack_mlen);
-               pack_moff = (offset / page_size) * page_size;
-               pack_base = mmap(NULL,pack_mlen,PROT_READ,MAP_SHARED,
-                       pack_fd,pack_moff);
-               if (pack_base == MAP_FAILED)
-                       die("Failed to map generated pack: %s", strerror(errno));
-               remap_count++;
-       }
-       offset -= pack_moff;
-       if (left)
-               *left = pack_mlen - offset;
-       return pack_base + offset;
-}
-
-static unsigned long unpack_object_header(unsigned long offset,
-       enum object_type *type,
-       unsigned long *sizep)
-{
-       unsigned shift;
-       unsigned char c;
-       unsigned long size;
-
-       c = *map_pack(offset++, NULL);
-       *type = (c >> 4) & 7;
-       size = c & 15;
-       shift = 4;
-       while (c & 0x80) {
-               c = *map_pack(offset++, NULL);
-               size += (c & 0x7f) << shift;
-               shift += 7;
-       }
-       *sizep = size;
-       return offset;
-}
-
-static void *unpack_non_delta_entry(unsigned long o, unsigned long sz)
-{
-       z_stream stream;
-       unsigned char *result;
-
-       result = xmalloc(sz + 1);
-       result[sz] = 0;
-
-       memset(&stream, 0, sizeof(stream));
-       stream.next_in = map_pack(o, &stream.avail_in);
-       stream.next_out = result;
-       stream.avail_out = sz;
-
-       inflateInit(&stream);
-       for (;;) {
-               int st = inflate(&stream, Z_FINISH);
-               if (st == Z_STREAM_END)
-                       break;
-               if (st == Z_OK || st == Z_BUF_ERROR) {
-                       o = stream.next_in - pack_base + pack_moff;
-                       stream.next_in = map_pack(o, &stream.avail_in);
-                       continue;
-               }
-               die("Error %i from zlib during inflate.", st);
-       }
-       inflateEnd(&stream);
-       if (stream.total_out != sz)
-               die("Error after inflate: sizes mismatch");
-       return result;
-}
-
-static void *unpack_entry(unsigned long offset, unsigned long *sizep);
-
-static void *unpack_delta_entry(unsigned long offset,
-       unsigned long delta_size,
-       unsigned long *sizep)
-{
-       struct object_entry *base_oe;
-       unsigned char *base_sha1;
-       void *delta_data, *base, *result;
-       unsigned long base_size, result_size;
-
-       base_sha1 = map_pack(offset, NULL);
-       base_oe = find_object(base_sha1);
-       if (!base_oe)
-               die("I'm broken; I can't find a base I know must be here.");
-       base = unpack_entry(base_oe->offset, &base_size);
-       delta_data = unpack_non_delta_entry(offset + 20, delta_size);
-       result = patch_delta(base, base_size,
-                            delta_data, delta_size,
-                            &result_size);
-       if (!result)
-               die("failed to apply delta");
-       free(delta_data);
-       free(base);
-       *sizep = result_size;
-       return result;
-}
-
-static void *unpack_entry(unsigned long offset, unsigned long *sizep)
+static void *gfi_unpack_entry(unsigned long ofs, unsigned long *sizep)
 {
-       unsigned long size;
-       enum object_type kind;
-
-       offset = unpack_object_header(offset, &kind, &size);
-       switch (kind) {
-       case OBJ_DELTA:
-               return unpack_delta_entry(offset, size, sizep);
-       case OBJ_COMMIT:
-       case OBJ_TREE:
-       case OBJ_BLOB:
-       case OBJ_TAG:
-               *sizep = size;
-               return unpack_non_delta_entry(offset, size);
-       default:
-               die("I created an object I can't read!");
-       }
+       char type[20];
+       pack_data->pack_size = pack_size + 20;
+       return unpack_entry(pack_data, ofs, type, sizep);
 }
 
 static const char *get_mode(const char *str, unsigned int *modep)
@@ -819,6 +736,7 @@ static const char *get_mode(const char *str, unsigned int *modep)
 
 static void load_tree(struct tree_entry *root)
 {
+       unsigned char* sha1 = root->versions[1].sha1;
        struct object_entry *myoe;
        struct tree_content *t;
        unsigned long size;
@@ -826,19 +744,20 @@ static void load_tree(struct tree_entry *root)
        const char *c;
 
        root->tree = t = new_tree_content(8);
-       if (!memcmp(root->sha1, null_sha1, 20))
+       if (is_null_sha1(sha1))
                return;
 
-       myoe = find_object(root->sha1);
+       myoe = find_object(sha1);
        if (myoe) {
                if (myoe->type != OBJ_TREE)
-                       die("Not a tree: %s", sha1_to_hex(root->sha1));
-               buf = unpack_entry(myoe->offset, &size);
+                       die("Not a tree: %s", sha1_to_hex(sha1));
+               t->delta_depth = 0;
+               buf = gfi_unpack_entry(myoe->offset, &size);
        } else {
                char type[20];
-               buf = read_sha1_file(root->sha1, type, &size);
+               buf = read_sha1_file(sha1, type, &size);
                if (!buf || strcmp(type, tree_type))
-                       die("Can't load tree %s", sha1_to_hex(root->sha1));
+                       die("Can't load tree %s", sha1_to_hex(sha1));
        }
 
        c = buf;
@@ -850,56 +769,117 @@ static void load_tree(struct tree_entry *root)
                t->entries[t->entry_count++] = e;
 
                e->tree = NULL;
-               c = get_mode(c, &e->mode);
+               c = get_mode(c, &e->versions[1].mode);
                if (!c)
-                       die("Corrupt mode in %s", sha1_to_hex(root->sha1));
+                       die("Corrupt mode in %s", sha1_to_hex(sha1));
+               e->versions[0].mode = e->versions[1].mode;
                e->name = to_atom(c, strlen(c));
                c += e->name->str_len + 1;
-               memcpy(e->sha1, c, sizeof(e->sha1));
+               hashcpy(e->versions[0].sha1, (unsigned char*)c);
+               hashcpy(e->versions[1].sha1, (unsigned char*)c);
                c += 20;
        }
        free(buf);
 }
 
-static int tecmp (const void *_a, const void *_b)
+static int tecmp0 (const void *_a, const void *_b)
 {
        struct tree_entry *a = *((struct tree_entry**)_a);
        struct tree_entry *b = *((struct tree_entry**)_b);
        return base_name_compare(
-               a->name->str_dat, a->name->str_len, a->mode,
-               b->name->str_dat, b->name->str_len, b->mode);
+               a->name->str_dat, a->name->str_len, a->versions[0].mode,
+               b->name->str_dat, b->name->str_len, b->versions[0].mode);
 }
 
-static void store_tree(struct tree_entry *root)
+static int tecmp1 (const void *_a, const void *_b)
 {
-       struct tree_content *t = root->tree;
+       struct tree_entry *a = *((struct tree_entry**)_a);
+       struct tree_entry *b = *((struct tree_entry**)_b);
+       return base_name_compare(
+               a->name->str_dat, a->name->str_len, a->versions[1].mode,
+               b->name->str_dat, b->name->str_len, b->versions[1].mode);
+}
+
+static void mktree(struct tree_content *t,
+       int v,
+       unsigned long *szp,
+       struct dbuf *b)
+{
+       size_t maxlen = 0;
        unsigned int i;
-       size_t maxlen;
-       char *buf, *c;
+       char *c;
 
-       if (memcmp(root->sha1, null_sha1, 20))
-               return;
+       if (!v)
+               qsort(t->entries,t->entry_count,sizeof(t->entries[0]),tecmp0);
+       else
+               qsort(t->entries,t->entry_count,sizeof(t->entries[0]),tecmp1);
 
-       maxlen = 0;
        for (i = 0; i < t->entry_count; i++) {
-               maxlen += t->entries[i]->name->str_len + 34;
-               if (t->entries[i]->tree)
-                       store_tree(t->entries[i]);
+               if (t->entries[i]->versions[v].mode)
+                       maxlen += t->entries[i]->name->str_len + 34;
        }
 
-       qsort(t->entries, t->entry_count, sizeof(t->entries[0]), tecmp);
-       buf = c = xmalloc(maxlen);
+       size_dbuf(b, maxlen);
+       c = b->buffer;
        for (i = 0; i < t->entry_count; i++) {
                struct tree_entry *e = t->entries[i];
-               c += sprintf(c, "%o", e->mode);
+               if (!e->versions[v].mode)
+                       continue;
+               c += sprintf(c, "%o", e->versions[v].mode);
                *c++ = ' ';
                strcpy(c, e->name->str_dat);
                c += e->name->str_len + 1;
-               memcpy(c, e->sha1, 20);
+               hashcpy((unsigned char*)c, e->versions[v].sha1);
                c += 20;
        }
-       store_object(OBJ_TREE, buf, c - buf, NULL, root->sha1, 0);
-       free(buf);
+       *szp = c - (char*)b->buffer;
+}
+
+static void store_tree(struct tree_entry *root)
+{
+       struct tree_content *t = root->tree;
+       unsigned int i, j, del;
+       unsigned long new_len;
+       struct last_object lo;
+       struct object_entry *le;
+
+       if (!is_null_sha1(root->versions[1].sha1))
+               return;
+
+       for (i = 0; i < t->entry_count; i++) {
+               if (t->entries[i]->tree)
+                       store_tree(t->entries[i]);
+       }
+
+       le = find_object(root->versions[0].sha1);
+       if (!S_ISDIR(root->versions[0].mode) || !le) {
+               lo.data = NULL;
+               lo.depth = 0;
+       } else {
+               mktree(t, 0, &lo.len, &old_tree);
+               lo.data = old_tree.buffer;
+               lo.offset = le->offset;
+               lo.depth = t->delta_depth;
+               lo.no_free = 1;
+       }
+
+       mktree(t, 1, &new_len, &new_tree);
+       store_object(OBJ_TREE, new_tree.buffer, new_len,
+               &lo, root->versions[1].sha1, 0);
+
+       t->delta_depth = lo.depth;
+       for (i = 0, j = 0, del = 0; i < t->entry_count; i++) {
+               struct tree_entry *e = t->entries[i];
+               if (e->versions[1].mode) {
+                       e->versions[0].mode = e->versions[1].mode;
+                       hashcpy(e->versions[0].sha1, e->versions[1].sha1);
+                       t->entries[j++] = e;
+               } else {
+                       release_tree_entry(e);
+                       del++;
+               }
+       }
+       t->entry_count -= del;
 }
 
 static int tree_content_set(
@@ -923,25 +903,26 @@ static int tree_content_set(
                e = t->entries[i];
                if (e->name->str_len == n && !strncmp(p, e->name->str_dat, n)) {
                        if (!slash1) {
-                               if (e->mode == mode && !memcmp(e->sha1, sha1, 20))
+                               if (e->versions[1].mode == mode
+                                               && !hashcmp(e->versions[1].sha1, sha1))
                                        return 0;
-                               e->mode = mode;
-                               memcpy(e->sha1, sha1, 20);
+                               e->versions[1].mode = mode;
+                               hashcpy(e->versions[1].sha1, sha1);
                                if (e->tree) {
                                        release_tree_content_recursive(e->tree);
                                        e->tree = NULL;
                                }
-                               memcpy(root->sha1, null_sha1, 20);
+                               hashclr(root->versions[1].sha1);
                                return 1;
                        }
-                       if (!S_ISDIR(e->mode)) {
+                       if (!S_ISDIR(e->versions[1].mode)) {
                                e->tree = new_tree_content(8);
-                               e->mode = S_IFDIR;
+                               e->versions[1].mode = S_IFDIR;
                        }
                        if (!e->tree)
                                load_tree(e);
                        if (tree_content_set(e, slash1 + 1, sha1, mode)) {
-                               memcpy(root->sha1, null_sha1, 20);
+                               hashclr(root->versions[1].sha1);
                                return 1;
                        }
                        return 0;
@@ -952,17 +933,19 @@ static int tree_content_set(
                root->tree = t = grow_tree_content(t, 8);
        e = new_tree_entry();
        e->name = to_atom(p, n);
+       e->versions[0].mode = 0;
+       hashclr(e->versions[0].sha1);
        t->entries[t->entry_count++] = e;
        if (slash1) {
                e->tree = new_tree_content(8);
-               e->mode = S_IFDIR;
+               e->versions[1].mode = S_IFDIR;
                tree_content_set(e, slash1 + 1, sha1, mode);
        } else {
                e->tree = NULL;
-               e->mode = mode;
-               memcpy(e->sha1, sha1, 20);
+               e->versions[1].mode = mode;
+               hashcpy(e->versions[1].sha1, sha1);
        }
-       memcpy(root->sha1, null_sha1, 20);
+       hashclr(root->versions[1].sha1);
        return 1;
 }
 
@@ -982,15 +965,18 @@ static int tree_content_remove(struct tree_entry *root, const char *p)
        for (i = 0; i < t->entry_count; i++) {
                e = t->entries[i];
                if (e->name->str_len == n && !strncmp(p, e->name->str_dat, n)) {
-                       if (!slash1 || !S_ISDIR(e->mode))
+                       if (!slash1 || !S_ISDIR(e->versions[1].mode))
                                goto del_entry;
                        if (!e->tree)
                                load_tree(e);
                        if (tree_content_remove(e, slash1 + 1)) {
-                               if (!e->tree->entry_count)
-                                       goto del_entry;
-                               memcpy(root->sha1, null_sha1, 20);
-                               return 1;
+                               for (n = 0; n < e->tree->entry_count; n++) {
+                                       if (e->tree->entries[n]->versions[1].mode) {
+                                               hashclr(root->versions[1].sha1);
+                                               return 1;
+                                       }
+                               }
+                               goto del_entry;
                        }
                        return 0;
                }
@@ -998,11 +984,13 @@ static int tree_content_remove(struct tree_entry *root, const char *p)
        return 0;
 
 del_entry:
-       for (i++; i < t->entry_count; i++)
-               t->entries[i-1] = t->entries[i];
-       t->entry_count--;
-       release_tree_entry(e);
-       memcpy(root->sha1, null_sha1, 20);
+       if (e->tree) {
+               release_tree_content_recursive(e->tree);
+               e->tree = NULL;
+       }
+       e->versions[1].mode = 0;
+       hashclr(e->versions[1].sha1);
+       hashclr(root->versions[1].sha1);
        return 1;
 }
 
@@ -1014,7 +1002,7 @@ static void init_pack_header()
        hdr.hdr_version = htonl(2);
        hdr.hdr_entries = 0;
 
-       ywrite(pack_fd, &hdr, sizeof(hdr));
+       write_or_die(pack_fd, &hdr, sizeof(hdr));
        pack_size = sizeof(hdr);
 }
 
@@ -1035,7 +1023,7 @@ static void fixup_header_footer()
 
        cnt = htonl(object_count);
        SHA1_Update(&c, &cnt, 4);
-       ywrite(pack_fd, &cnt, 4);
+       write_or_die(pack_fd, &cnt, 4);
 
        buf = xmalloc(128 * 1024);
        for (;;) {
@@ -1047,14 +1035,14 @@ static void fixup_header_footer()
        free(buf);
 
        SHA1_Final(pack_sha1, &c);
-       ywrite(pack_fd, pack_sha1, sizeof(pack_sha1));
+       write_or_die(pack_fd, pack_sha1, sizeof(pack_sha1));
 }
 
 static int oecmp (const void *_a, const void *_b)
 {
        struct object_entry *a = *((struct object_entry**)_a);
        struct object_entry *b = *((struct object_entry**)_b);
-       return memcmp(a->sha1, b->sha1, sizeof(a->sha1));
+       return hashcmp(a->sha1, b->sha1);
 }
 
 static void write_index(const char *idx_name)
@@ -1109,7 +1097,7 @@ static void dump_branches()
 
        for (i = 0; i < branch_table_sz; i++) {
                for (b = branch_table[i]; b; b = b->table_next_branch) {
-                       lock = lock_any_ref_for_update(b->name, NULL, 0);
+                       lock = lock_any_ref_for_update(b->name, NULL);
                        if (!lock || write_ref_sha1(lock, b->sha1, msg) < 0)
                                die("Can't write %s", b->name);
                }
@@ -1125,7 +1113,7 @@ static void dump_tags()
 
        for (t = first_tag; t; t = t->next_tag) {
                sprintf(path, "refs/tags/%s", t->name);
-               lock = lock_any_ref_for_update(path, NULL, 0);
+               lock = lock_any_ref_for_update(path, NULL);
                if (!lock || write_ref_sha1(lock, t->sha1, msg) < 0)
                        die("Can't write %s", path);
        }
@@ -1283,6 +1271,7 @@ static void file_change_m(struct branch *b)
        if (*p == ':') {
                char *x;
                oe = find_mark(strtoul(p + 1, &x, 10));
+               hashcpy(sha1, oe->sha1);
                p = x;
        } else {
                if (get_sha1_hex(p, sha1))
@@ -1359,8 +1348,10 @@ static void cmd_from(struct branch *b)
        if (b == s)
                die("Can't create a branch from itself: %s", b->name);
        else if (s) {
-               memcpy(b->sha1, s->sha1, 20);
-               memcpy(b->branch_tree.sha1, s->branch_tree.sha1, 20);
+               unsigned char *t = s->branch_tree.versions[1].sha1;
+               hashcpy(b->sha1, s->sha1);
+               hashcpy(b->branch_tree.versions[0].sha1, t);
+               hashcpy(b->branch_tree.versions[1].sha1, t);
        } else if (*from == ':') {
                unsigned long idnum = strtoul(from + 1, NULL, 10);
                struct object_entry *oe = find_mark(idnum);
@@ -1368,18 +1359,21 @@ static void cmd_from(struct branch *b)
                char *buf;
                if (oe->type != OBJ_COMMIT)
                        die("Mark :%lu not a commit", idnum);
-               memcpy(b->sha1, oe->sha1, 20);
-               buf = unpack_entry(oe->offset, &size);
+               hashcpy(b->sha1, oe->sha1);
+               buf = gfi_unpack_entry(oe->offset, &size);
                if (!buf || size < 46)
                        die("Not a valid commit: %s", from);
                if (memcmp("tree ", buf, 5)
-                       || get_sha1_hex(buf + 5, b->branch_tree.sha1))
+                       || get_sha1_hex(buf + 5, b->branch_tree.versions[1].sha1))
                        die("The commit %s is corrupt", sha1_to_hex(b->sha1));
                free(buf);
+               hashcpy(b->branch_tree.versions[0].sha1,
+                       b->branch_tree.versions[1].sha1);
        } else if (!get_sha1(from, b->sha1)) {
-               if (!memcmp(b->sha1, null_sha1, 20))
-                       memcpy(b->branch_tree.sha1, null_sha1, 20);
-               else {
+               if (is_null_sha1(b->sha1)) {
+                       hashclr(b->branch_tree.versions[0].sha1);
+                       hashclr(b->branch_tree.versions[1].sha1);
+               } else {
                        unsigned long size;
                        char *buf;
 
@@ -1388,9 +1382,11 @@ static void cmd_from(struct branch *b)
                        if (!buf || size < 46)
                                die("Not a valid commit: %s", from);
                        if (memcmp("tree ", buf, 5)
-                               || get_sha1_hex(buf + 5, b->branch_tree.sha1))
+                               || get_sha1_hex(buf + 5, b->branch_tree.versions[1].sha1))
                                die("The commit %s is corrupt", sha1_to_hex(b->sha1));
                        free(buf);
+                       hashcpy(b->branch_tree.versions[0].sha1,
+                               b->branch_tree.versions[1].sha1);
                }
        } else
                die("Invalid ref name or SHA1 expression: %s", from);
@@ -1398,6 +1394,48 @@ static void cmd_from(struct branch *b)
        read_next_command();
 }
 
+static struct hash_list* cmd_merge(unsigned int *count)
+{
+       struct hash_list *list = NULL, *n, *e;
+       const char *from, *endp;
+       char *str_uq;
+       struct branch *s;
+
+       *count = 0;
+       while (!strncmp("merge ", command_buf.buf, 6)) {
+               from = strchr(command_buf.buf, ' ') + 1;
+               str_uq = unquote_c_style(from, &endp);
+               if (str_uq) {
+                       if (*endp)
+                               die("Garbage after string in: %s", command_buf.buf);
+                       from = str_uq;
+               }
+
+               n = xmalloc(sizeof(*n));
+               s = lookup_branch(from);
+               if (s)
+                       hashcpy(n->sha1, s->sha1);
+               else if (*from == ':') {
+                       unsigned long idnum = strtoul(from + 1, NULL, 10);
+                       struct object_entry *oe = find_mark(idnum);
+                       if (oe->type != OBJ_COMMIT)
+                               die("Mark :%lu not a commit", idnum);
+                       hashcpy(n->sha1, oe->sha1);
+               } else if (get_sha1(from, n->sha1))
+                       die("Invalid ref name or SHA1 expression: %s", from);
+
+               n->next = NULL;
+               if (list)
+                       e->next = n;
+               else
+                       list = n;
+               e = n;
+               *count++;
+               read_next_command();
+       }
+       return list;
+}
+
 static void cmd_new_commit()
 {
        struct branch *b;
@@ -1408,7 +1446,8 @@ static void cmd_new_commit()
        char *sp;
        char *author = NULL;
        char *committer = NULL;
-       char *body;
+       struct hash_list *merge_list = NULL;
+       unsigned int merge_count;
 
        /* Obtain the branch name from the rest of our command */
        sp = strchr(command_buf.buf, ' ') + 1;
@@ -1439,6 +1478,7 @@ static void cmd_new_commit()
        msg = cmd_data(&msglen);
        read_next_command();
        cmd_from(b);
+       merge_list = cmd_merge(&merge_count);
 
        /* ensure the branch is active/loaded */
        if (!b->branch_tree.tree || !max_active_branches) {
@@ -1461,14 +1501,24 @@ static void cmd_new_commit()
 
        /* build the tree and the commit */
        store_tree(&b->branch_tree);
-       body = xmalloc(97 + msglen
+       hashcpy(b->branch_tree.versions[0].sha1,
+               b->branch_tree.versions[1].sha1);
+       size_dbuf(&new_data, 97 + msglen
+               + merge_count * 49
                + (author
                        ? strlen(author) + strlen(committer)
                        : 2 * strlen(committer)));
-       sp = body;
-       sp += sprintf(sp, "tree %s\n", sha1_to_hex(b->branch_tree.sha1));
-       if (memcmp(b->sha1, null_sha1, 20))
+       sp = new_data.buffer;
+       sp += sprintf(sp, "tree %s\n",
+               sha1_to_hex(b->branch_tree.versions[1].sha1));
+       if (!is_null_sha1(b->sha1))
                sp += sprintf(sp, "parent %s\n", sha1_to_hex(b->sha1));
+       while (merge_list) {
+               struct hash_list *next = merge_list->next;
+               sp += sprintf(sp, "parent %s\n", sha1_to_hex(merge_list->sha1));
+               free(merge_list);
+               merge_list = next;
+       }
        if (author)
                sp += sprintf(sp, "%s\n", author);
        else
@@ -1481,8 +1531,9 @@ static void cmd_new_commit()
        free(committer);
        free(msg);
 
-       store_object(OBJ_COMMIT, body, sp - body, NULL, b->sha1, next_mark);
-       free(body);
+       store_object(OBJ_COMMIT,
+               new_data.buffer, sp - (char*)new_data.buffer,
+               NULL, b->sha1, next_mark);
        b->last_commit = object_count_by_type[OBJ_COMMIT];
 
        if (branch_log) {
@@ -1508,7 +1559,6 @@ static void cmd_new_tag()
        struct branch *s;
        void *msg;
        size_t msglen;
-       char *body;
        struct tag *t;
        unsigned long from_mark = 0;
        unsigned char sha1[20];
@@ -1547,13 +1597,13 @@ static void cmd_new_tag()
 
        s = lookup_branch(from);
        if (s) {
-               memcpy(sha1, s->sha1, 20);
+               hashcpy(sha1, s->sha1);
        } else if (*from == ':') {
                from_mark = strtoul(from + 1, NULL, 10);
                struct object_entry *oe = find_mark(from_mark);
                if (oe->type != OBJ_COMMIT)
                        die("Mark :%lu not a commit", from_mark);
-               memcpy(sha1, oe->sha1, 20);
+               hashcpy(sha1, oe->sha1);
        } else if (!get_sha1(from, sha1)) {
                unsigned long size;
                char *buf;
@@ -1580,8 +1630,8 @@ static void cmd_new_tag()
        msg = cmd_data(&msglen);
 
        /* build the tag object */
-       body = xmalloc(67 + strlen(t->name) + strlen(tagger) + msglen);
-       sp = body;
+       size_dbuf(&new_data, 67+strlen(t->name)+strlen(tagger)+msglen);
+       sp = new_data.buffer;
        sp += sprintf(sp, "object %s\n", sha1_to_hex(sha1));
        sp += sprintf(sp, "type %s\n", type_names[OBJ_COMMIT]);
        sp += sprintf(sp, "tag %s\n", t->name);
@@ -1591,8 +1641,8 @@ static void cmd_new_tag()
        free(tagger);
        free(msg);
 
-       store_object(OBJ_TAG, body, sp - body, NULL, t->sha1, 0);
-       free(body);
+       store_object(OBJ_TAG, new_data.buffer, sp - (char*)new_data.buffer,
+               NULL, t->sha1, 0);
 
        if (branch_log) {
                int need_dq = quote_c_style(t->name, NULL, NULL, 0);
@@ -1630,8 +1680,12 @@ static void cmd_reset_branch()
                        b->branch_tree.tree = NULL;
                }
        }
+       else
+               b = new_branch(sp);
        if (str_uq)
                free(str_uq);
+       read_next_command();
+       cmd_from(b);
 }
 
 static const char fast_import_usage[] =
@@ -1641,14 +1695,13 @@ int main(int argc, const char **argv)
 {
        const char *base_name;
        int i;
-       unsigned long est_obj_cnt = 1000;
+       unsigned long est_obj_cnt = object_entry_alloc;
        char *pack_name;
        char *idx_name;
        struct stat sb;
 
        setup_ident();
        git_config(git_default_config);
-       page_size = getpagesize();
 
        for (i = 1; i < argc; i++) {
                const char *a = argv[i];
@@ -1684,6 +1737,10 @@ int main(int argc, const char **argv)
        if (pack_fd < 0)
                die("Can't create %s: %s", pack_name, strerror(errno));
 
+       pack_data = xcalloc(1, sizeof(*pack_data) + strlen(pack_name) + 2);
+       strcpy(pack_data->pack_name, pack_name);
+       pack_data->pack_fd = pack_fd;
+
        init_pack_header();
        alloc_objects(est_obj_cnt);
        strbuf_init(&command_buf);
@@ -1715,16 +1772,17 @@ int main(int argc, const char **argv)
        dump_branches();
        dump_tags();
        dump_marks();
-       fclose(branch_log);
+       if (branch_log)
+               fclose(branch_log);
 
        fprintf(stderr, "%s statistics:\n", argv[0]);
-       fprintf(stderr, "---------------------------------------------------\n");
+       fprintf(stderr, "---------------------------------------------------------------------\n");
        fprintf(stderr, "Alloc'd objects: %10lu (%10lu overflow  )\n", alloc_count, alloc_count - est_obj_cnt);
-       fprintf(stderr, "Total objects:   %10lu (%10lu duplicates)\n", object_count, duplicate_count);
-       fprintf(stderr, "      blobs  :   %10lu (%10lu duplicates)\n", object_count_by_type[OBJ_BLOB], duplicate_count_by_type[OBJ_BLOB]);
-       fprintf(stderr, "      trees  :   %10lu (%10lu duplicates)\n", object_count_by_type[OBJ_TREE], duplicate_count_by_type[OBJ_TREE]);
-       fprintf(stderr, "      commits:   %10lu (%10lu duplicates)\n", object_count_by_type[OBJ_COMMIT], duplicate_count_by_type[OBJ_COMMIT]);
-       fprintf(stderr, "      tags   :   %10lu (%10lu duplicates)\n", object_count_by_type[OBJ_TAG], duplicate_count_by_type[OBJ_TAG]);
+       fprintf(stderr, "Total objects:   %10lu (%10lu duplicates                  )\n", object_count, duplicate_count);
+       fprintf(stderr, "      blobs  :   %10lu (%10lu duplicates %10lu deltas)\n", object_count_by_type[OBJ_BLOB], duplicate_count_by_type[OBJ_BLOB], delta_count_by_type[OBJ_BLOB]);
+       fprintf(stderr, "      trees  :   %10lu (%10lu duplicates %10lu deltas)\n", object_count_by_type[OBJ_TREE], duplicate_count_by_type[OBJ_TREE], delta_count_by_type[OBJ_TREE]);
+       fprintf(stderr, "      commits:   %10lu (%10lu duplicates %10lu deltas)\n", object_count_by_type[OBJ_COMMIT], duplicate_count_by_type[OBJ_COMMIT], delta_count_by_type[OBJ_COMMIT]);
+       fprintf(stderr, "      tags   :   %10lu (%10lu duplicates %10lu deltas)\n", object_count_by_type[OBJ_TAG], duplicate_count_by_type[OBJ_TAG], delta_count_by_type[OBJ_TAG]);
        fprintf(stderr, "Total branches:  %10lu (%10lu loads     )\n", branch_count, branch_load_count);
        fprintf(stderr, "      marks:     %10u (%10lu unique    )\n", (1 << marks->shift) * 1024, marks_set_count);
        fprintf(stderr, "      atoms:     %10u\n", atom_cnt);
@@ -1732,12 +1790,11 @@ int main(int argc, const char **argv)
        fprintf(stderr, "       pools:    %10lu KiB\n", total_allocd/1024);
        fprintf(stderr, "     objects:    %10lu KiB\n", (alloc_count*sizeof(struct object_entry))/1024);
        fprintf(stderr, "Pack remaps:     %10lu\n", remap_count);
-       fprintf(stderr, "---------------------------------------------------\n");
-
        stat(pack_name, &sb);
        fprintf(stderr, "Pack size:       %10lu KiB\n", (unsigned long)(sb.st_size/1024));
        stat(idx_name, &sb);
        fprintf(stderr, "Index size:      %10lu KiB\n", (unsigned long)(sb.st_size/1024));
+       fprintf(stderr, "---------------------------------------------------------------------\n");
 
        fprintf(stderr, "\n");