Diff between two blobs should take mode changes into account now.
[gitweb.git] / builtin-rev-list.c
index c2db5a5b037babf9020353d9b11dc348915b6c1b..c0329dcecdbb4775b2c438a0bf6429a5ee72b18e 100644 (file)
@@ -35,8 +35,10 @@ static const char rev_list_usage[] =
 "    --header | --pretty\n"
 "    --abbrev=nr | --no-abbrev\n"
 "    --abbrev-commit\n"
+"    --left-right\n"
 "  special purpose:\n"
-"    --bisect"
+"    --bisect\n"
+"    --bisect-vars"
 ;
 
 static struct rev_info revs;
@@ -111,6 +113,10 @@ static void show_object(struct object_array_entry *p)
         * confuse downstream git-pack-objects very badly.
         */
        const char *ep = strchr(p->name, '\n');
+
+       if (p->item->type == OBJ_BLOB && !has_sha1_file(p->item->sha1))
+               die("missing blob object '%s'", sha1_to_hex(p->item->sha1));
+
        if (ep) {
                printf("%s %.*s\n", sha1_to_hex(p->item->sha1),
                       (int) (ep - p->name),
@@ -168,38 +174,273 @@ static void clear_distance(struct commit_list *list)
        }
 }
 
-static struct commit_list *find_bisection(struct commit_list *list)
+#define DEBUG_BISECT 0
+
+static inline int weight(struct commit_list *elem)
 {
-       int nr, closest;
-       struct commit_list *p, *best;
+       return *((int*)(elem->item->util));
+}
 
-       nr = 0;
-       p = list;
-       while (p) {
-               if (!revs.prune_fn || (p->item->object.flags & TREECHANGE))
-                       nr++;
-               p = p->next;
+static inline void weight_set(struct commit_list *elem, int weight)
+{
+       *((int*)(elem->item->util)) = weight;
+}
+
+static int count_interesting_parents(struct commit *commit)
+{
+       struct commit_list *p;
+       int count;
+
+       for (count = 0, p = commit->parents; p; p = p->next) {
+               if (p->item->object.flags & UNINTERESTING)
+                       continue;
+               count++;
        }
-       closest = 0;
-       best = list;
+       return count;
+}
+
+static inline int halfway(struct commit_list *p, int distance, int nr)
+{
+       /*
+        * Don't short-cut something we are not going to return!
+        */
+       if (revs.prune_fn && !(p->item->object.flags & TREECHANGE))
+               return 0;
+       if (DEBUG_BISECT)
+               return 0;
+       /*
+        * 2 and 3 are halfway of 5.
+        * 3 is halfway of 6 but 2 and 4 are not.
+        */
+       distance *= 2;
+       switch (distance - nr) {
+       case -1: case 0: case 1:
+               return 1;
+       default:
+               return 0;
+       }
+}
+
+#if !DEBUG_BISECT
+#define show_list(a,b,c,d) do { ; } while (0)
+#else
+static void show_list(const char *debug, int counted, int nr,
+                     struct commit_list *list)
+{
+       struct commit_list *p;
+
+       fprintf(stderr, "%s (%d/%d)\n", debug, counted, nr);
 
        for (p = list; p; p = p->next) {
-               int distance;
+               struct commit_list *pp;
+               struct commit *commit = p->item;
+               unsigned flags = commit->object.flags;
+               enum object_type type;
+               unsigned long size;
+               char *buf = read_sha1_file(commit->object.sha1, &type, &size);
+               char *ep, *sp;
+
+               fprintf(stderr, "%c%c%c ",
+                       (flags & TREECHANGE) ? 'T' : ' ',
+                       (flags & UNINTERESTING) ? 'U' : ' ',
+                       (flags & COUNTED) ? 'C' : ' ');
+               if (commit->util)
+                       fprintf(stderr, "%3d", weight(p));
+               else
+                       fprintf(stderr, "---");
+               fprintf(stderr, " %.*s", 8, sha1_to_hex(commit->object.sha1));
+               for (pp = commit->parents; pp; pp = pp->next)
+                       fprintf(stderr, " %.*s", 8,
+                               sha1_to_hex(pp->item->object.sha1));
+
+               sp = strstr(buf, "\n\n");
+               if (sp) {
+                       sp += 2;
+                       for (ep = sp; *ep && *ep != '\n'; ep++)
+                               ;
+                       fprintf(stderr, " %.*s", (int)(ep - sp), sp);
+               }
+               fprintf(stderr, "\n");
+       }
+}
+#endif /* DEBUG_BISECT */
+
+/*
+ * zero or positive weight is the number of interesting commits it can
+ * reach, including itself.  Especially, weight = 0 means it does not
+ * reach any tree-changing commits (e.g. just above uninteresting one
+ * but traversal is with pathspec).
+ *
+ * weight = -1 means it has one parent and its distance is yet to
+ * be computed.
+ *
+ * weight = -2 means it has more than one parent and its distance is
+ * unknown.  After running count_distance() first, they will get zero
+ * or positive distance.
+ */
+
+static struct commit_list *find_bisection(struct commit_list *list,
+                                         int *reaches, int *all)
+{
+       int n, nr, on_list, counted, distance;
+       struct commit_list *p, *best, *next, *last;
+       int *weights;
 
-               if (revs.prune_fn && !(p->item->object.flags & TREECHANGE))
+       show_list("bisection 2 entry", 0, 0, list);
+
+       /*
+        * Count the number of total and tree-changing items on the
+        * list, while reversing the list.
+        */
+       for (nr = on_list = 0, last = NULL, p = list;
+            p;
+            p = next) {
+               unsigned flags = p->item->object.flags;
+
+               next = p->next;
+               if (flags & UNINTERESTING)
                        continue;
+               p->next = last;
+               last = p;
+               if (!revs.prune_fn || (flags & TREECHANGE))
+                       nr++;
+               on_list++;
+       }
+       list = last;
+       show_list("bisection 2 sorted", 0, nr, list);
+
+       *all = nr;
+       weights = xcalloc(on_list, sizeof(int*));
+       counted = 0;
+
+       for (n = 0, p = list; p; p = p->next) {
+               struct commit *commit = p->item;
+               unsigned flags = commit->object.flags;
+
+               p->item->util = &weights[n++];
+               switch (count_interesting_parents(commit)) {
+               case 0:
+                       if (!revs.prune_fn || (flags & TREECHANGE)) {
+                               weight_set(p, 1);
+                               counted++;
+                               show_list("bisection 2 count one",
+                                         counted, nr, list);
+                       }
+                       /*
+                        * otherwise, it is known not to reach any
+                        * tree-changing commit and gets weight 0.
+                        */
+                       break;
+               case 1:
+                       weight_set(p, -1);
+                       break;
+               default:
+                       weight_set(p, -2);
+                       break;
+               }
+       }
 
+       show_list("bisection 2 initialize", counted, nr, list);
+
+       /*
+        * If you have only one parent in the resulting set
+        * then you can reach one commit more than that parent
+        * can reach.  So we do not have to run the expensive
+        * count_distance() for single strand of pearls.
+        *
+        * However, if you have more than one parents, you cannot
+        * just add their distance and one for yourself, since
+        * they usually reach the same ancestor and you would
+        * end up counting them twice that way.
+        *
+        * So we will first count distance of merges the usual
+        * way, and then fill the blanks using cheaper algorithm.
+        */
+       for (p = list; p; p = p->next) {
+               if (p->item->object.flags & UNINTERESTING)
+                       continue;
+               n = weight(p);
+               if (n != -2)
+                       continue;
                distance = count_distance(p);
                clear_distance(list);
+               weight_set(p, distance);
+
+               /* Does it happen to be at exactly half-way? */
+               if (halfway(p, distance, nr)) {
+                       p->next = NULL;
+                       *reaches = distance;
+                       free(weights);
+                       return p;
+               }
+               counted++;
+       }
+
+       show_list("bisection 2 count_distance", counted, nr, list);
+
+       while (counted < nr) {
+               for (p = list; p; p = p->next) {
+                       struct commit_list *q;
+                       unsigned flags = p->item->object.flags;
+
+                       if (0 <= weight(p))
+                               continue;
+                       for (q = p->item->parents; q; q = q->next) {
+                               if (q->item->object.flags & UNINTERESTING)
+                                       continue;
+                               if (0 <= weight(q))
+                                       break;
+                       }
+                       if (!q)
+                               continue;
+
+                       /*
+                        * weight for p is unknown but q is known.
+                        * add one for p itself if p is to be counted,
+                        * otherwise inherit it from q directly.
+                        */
+                       if (!revs.prune_fn || (flags & TREECHANGE)) {
+                               weight_set(p, weight(q)+1);
+                               counted++;
+                               show_list("bisection 2 count one",
+                                         counted, nr, list);
+                       }
+                       else
+                               weight_set(p, weight(q));
+
+                       /* Does it happen to be at exactly half-way? */
+                       distance = weight(p);
+                       if (halfway(p, distance, nr)) {
+                               p->next = NULL;
+                               *reaches = distance;
+                               free(weights);
+                               return p;
+                       }
+               }
+       }
+
+       show_list("bisection 2 counted all", counted, nr, list);
+
+       /* Then find the best one */
+       counted = -1;
+       best = list;
+       for (p = list; p; p = p->next) {
+               unsigned flags = p->item->object.flags;
+
+               if (revs.prune_fn && !(flags & TREECHANGE))
+                       continue;
+               distance = weight(p);
                if (nr - distance < distance)
                        distance = nr - distance;
-               if (distance > closest) {
+               if (distance > counted) {
                        best = p;
-                       closest = distance;
+                       counted = distance;
+                       *reaches = weight(p);
                }
        }
        if (best)
                best->next = NULL;
+       free(weights);
        return best;
 }
 
@@ -225,6 +466,7 @@ int cmd_rev_list(int argc, const char **argv, const char *prefix)
        struct commit_list *list;
        int i;
        int read_from_stdin = 0;
+       int bisect_show_vars = 0;
 
        git_config(git_default_config);
        init_revisions(&revs, prefix);
@@ -247,6 +489,11 @@ int cmd_rev_list(int argc, const char **argv, const char *prefix)
                        bisect_list = 1;
                        continue;
                }
+               if (!strcmp(arg, "--bisect-vars")) {
+                       bisect_list = 1;
+                       bisect_show_vars = 1;
+                       continue;
+               }
                if (!strcmp(arg, "--stdin")) {
                        if (read_from_stdin++)
                                die("--stdin given twice?");
@@ -285,8 +532,39 @@ int cmd_rev_list(int argc, const char **argv, const char *prefix)
        if (revs.tree_objects)
                mark_edges_uninteresting(revs.commits, &revs, show_edge);
 
-       if (bisect_list)
-               revs.commits = find_bisection(revs.commits);
+       if (bisect_list) {
+               int reaches = reaches, all = all;
+
+               revs.commits = find_bisection(revs.commits, &reaches, &all);
+               if (bisect_show_vars) {
+                       int cnt;
+                       if (!revs.commits)
+                               return 1;
+                       /*
+                        * revs.commits can reach "reaches" commits among
+                        * "all" commits.  If it is good, then there are
+                        * (all-reaches) commits left to be bisected.
+                        * On the other hand, if it is bad, then the set
+                        * to bisect is "reaches".
+                        * A bisect set of size N has (N-1) commits further
+                        * to test, as we already know one bad one.
+                        */
+                       cnt = all-reaches;
+                       if (cnt < reaches)
+                               cnt = reaches;
+                       printf("bisect_rev=%s\n"
+                              "bisect_nr=%d\n"
+                              "bisect_good=%d\n"
+                              "bisect_bad=%d\n"
+                              "bisect_all=%d\n",
+                              sha1_to_hex(revs.commits->item->object.sha1),
+                              cnt - 1,
+                              all - reaches - 1,
+                              reaches - 1,
+                              all);
+                       return 0;
+               }
+       }
 
        traverse_commit_list(&revs, show_commit, show_object);