Code

improve diff-delta with sparse and/or repetitive data
authorNicolas Pitre <nico@cam.org>
Wed, 3 May 2006 03:31:00 +0000 (23:31 -0400)
committerJunio C Hamano <junkio@cox.net>
Wed, 3 May 2006 04:32:39 +0000 (21:32 -0700)
It is useless to preserve multiple hash entries for consecutive blocks
with the same hash.  Keeping only the first one will allow for matching
the longest string of identical bytes while subsequent blocks will only
allow for shorter matches.  The backward matching code will match the
end of it as necessary.

This improves both performances (no repeated string compare with long
successions of identical bytes, or even small group of bytes), as well
as compression (less likely to need random hash bucket entry culling),
especially with sparse files.

With well behaved data sets this patch doesn't change much.

Signed-off-by: Nicolas Pitre <nico@cam.org>
Signed-off-by: Junio C Hamano <junkio@cox.net>
diff-delta.c

index 45df786b0a099e03bcca57f40047be04635e130d..c61887518874d550f66d2d52bd7559d023fcb176 100644 (file)
@@ -136,11 +136,12 @@ struct delta_index {
 
 struct delta_index * create_delta_index(const void *buf, unsigned long bufsize)
 {
-       unsigned int i, hsize, hmask, entries, *hash_count;
+       unsigned int i, hsize, hmask, entries, prev_val, *hash_count;
        const unsigned char *data, *buffer = buf;
        struct delta_index *index;
        struct index_entry *entry, **hash;
        void *mem;
+       unsigned long memsize;
 
        if (!buf || !bufsize)
                return NULL;
@@ -155,9 +156,10 @@ struct delta_index * create_delta_index(const void *buf, unsigned long bufsize)
        hmask = hsize - 1;
 
        /* allocate lookup index */
-       mem = malloc(sizeof(*index) +
-                    sizeof(*hash) * hsize +
-                    sizeof(*entry) * entries);
+       memsize = sizeof(*index) +
+                 sizeof(*hash) * hsize +
+                 sizeof(*entry) * entries;
+       mem = malloc(memsize);
        if (!mem)
                return NULL;
        index = mem;
@@ -179,18 +181,26 @@ struct delta_index * create_delta_index(const void *buf, unsigned long bufsize)
        }
 
        /* then populate the index */
-       data = buffer + entries * RABIN_WINDOW - RABIN_WINDOW;
-       while (data >= buffer) {
+       prev_val = ~0;
+       for (data = buffer + 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];
-               i = val & hmask;
-               entry->ptr = data + RABIN_WINDOW;
-               entry->val = val;
-               entry->next = hash[i];
-               hash[i] = entry++;
-               hash_count[i]++;
-               data -= RABIN_WINDOW;
+               if (val == prev_val) {
+                       /* keep the lowest of consecutive identical blocks */
+                       entry[-1].ptr = data + RABIN_WINDOW;
+               } else {
+                       prev_val = val;
+                       i = val & hmask;
+                       entry->ptr = data + RABIN_WINDOW;
+                       entry->val = val;
+                       entry->next = hash[i];
+                       hash[i] = entry++;
+                       hash_count[i]++;
+                       entries--;
+               }
        }
 
        /*
@@ -220,6 +230,10 @@ struct delta_index * create_delta_index(const void *buf, unsigned long bufsize)
        }
        free(hash_count);
 
+       /* If we didn't use all hash entries, free the unused memory. */
+       if (entries)
+               index = realloc(index, memsize - entries * sizeof(*entry));
+
        return index;
 }