read-cache.c: remove the_* from index_has_changes()
[git] / merge-recursive.c
1 /*
2  * Recursive Merge algorithm stolen from git-merge-recursive.py by
3  * Fredrik Kuivinen.
4  * The thieves were Alex Riesen and Johannes Schindelin, in June/July 2006
5  */
6 #include "cache.h"
7 #include "config.h"
8 #include "advice.h"
9 #include "lockfile.h"
10 #include "cache-tree.h"
11 #include "object-store.h"
12 #include "repository.h"
13 #include "commit.h"
14 #include "blob.h"
15 #include "builtin.h"
16 #include "tree-walk.h"
17 #include "diff.h"
18 #include "diffcore.h"
19 #include "tag.h"
20 #include "alloc.h"
21 #include "unpack-trees.h"
22 #include "string-list.h"
23 #include "xdiff-interface.h"
24 #include "ll-merge.h"
25 #include "attr.h"
26 #include "merge-recursive.h"
27 #include "dir.h"
28 #include "submodule.h"
29 #include "revision.h"
30 #include "commit-reach.h"
31
32 struct path_hashmap_entry {
33         struct hashmap_entry e;
34         char path[FLEX_ARRAY];
35 };
36
37 static int path_hashmap_cmp(const void *cmp_data,
38                             const void *entry,
39                             const void *entry_or_key,
40                             const void *keydata)
41 {
42         const struct path_hashmap_entry *a = entry;
43         const struct path_hashmap_entry *b = entry_or_key;
44         const char *key = keydata;
45
46         if (ignore_case)
47                 return strcasecmp(a->path, key ? key : b->path);
48         else
49                 return strcmp(a->path, key ? key : b->path);
50 }
51
52 static unsigned int path_hash(const char *path)
53 {
54         return ignore_case ? strihash(path) : strhash(path);
55 }
56
57 static struct dir_rename_entry *dir_rename_find_entry(struct hashmap *hashmap,
58                                                       char *dir)
59 {
60         struct dir_rename_entry key;
61
62         if (dir == NULL)
63                 return NULL;
64         hashmap_entry_init(&key, strhash(dir));
65         key.dir = dir;
66         return hashmap_get(hashmap, &key, NULL);
67 }
68
69 static int dir_rename_cmp(const void *unused_cmp_data,
70                           const void *entry,
71                           const void *entry_or_key,
72                           const void *unused_keydata)
73 {
74         const struct dir_rename_entry *e1 = entry;
75         const struct dir_rename_entry *e2 = entry_or_key;
76
77         return strcmp(e1->dir, e2->dir);
78 }
79
80 static void dir_rename_init(struct hashmap *map)
81 {
82         hashmap_init(map, dir_rename_cmp, NULL, 0);
83 }
84
85 static void dir_rename_entry_init(struct dir_rename_entry *entry,
86                                   char *directory)
87 {
88         hashmap_entry_init(entry, strhash(directory));
89         entry->dir = directory;
90         entry->non_unique_new_dir = 0;
91         strbuf_init(&entry->new_dir, 0);
92         string_list_init(&entry->possible_new_dirs, 0);
93 }
94
95 static struct collision_entry *collision_find_entry(struct hashmap *hashmap,
96                                                     char *target_file)
97 {
98         struct collision_entry key;
99
100         hashmap_entry_init(&key, strhash(target_file));
101         key.target_file = target_file;
102         return hashmap_get(hashmap, &key, NULL);
103 }
104
105 static int collision_cmp(void *unused_cmp_data,
106                          const struct collision_entry *e1,
107                          const struct collision_entry *e2,
108                          const void *unused_keydata)
109 {
110         return strcmp(e1->target_file, e2->target_file);
111 }
112
113 static void collision_init(struct hashmap *map)
114 {
115         hashmap_init(map, (hashmap_cmp_fn) collision_cmp, NULL, 0);
116 }
117
118 static void flush_output(struct merge_options *o)
119 {
120         if (o->buffer_output < 2 && o->obuf.len) {
121                 fputs(o->obuf.buf, stdout);
122                 strbuf_reset(&o->obuf);
123         }
124 }
125
126 static int err(struct merge_options *o, const char *err, ...)
127 {
128         va_list params;
129
130         if (o->buffer_output < 2)
131                 flush_output(o);
132         else {
133                 strbuf_complete(&o->obuf, '\n');
134                 strbuf_addstr(&o->obuf, "error: ");
135         }
136         va_start(params, err);
137         strbuf_vaddf(&o->obuf, err, params);
138         va_end(params);
139         if (o->buffer_output > 1)
140                 strbuf_addch(&o->obuf, '\n');
141         else {
142                 error("%s", o->obuf.buf);
143                 strbuf_reset(&o->obuf);
144         }
145
146         return -1;
147 }
148
149 static struct tree *shift_tree_object(struct repository *repo,
150                                       struct tree *one, struct tree *two,
151                                       const char *subtree_shift)
152 {
153         struct object_id shifted;
154
155         if (!*subtree_shift) {
156                 shift_tree(&one->object.oid, &two->object.oid, &shifted, 0);
157         } else {
158                 shift_tree_by(&one->object.oid, &two->object.oid, &shifted,
159                               subtree_shift);
160         }
161         if (oideq(&two->object.oid, &shifted))
162                 return two;
163         return lookup_tree(repo, &shifted);
164 }
165
166 static struct commit *make_virtual_commit(struct repository *repo,
167                                           struct tree *tree,
168                                           const char *comment)
169 {
170         struct commit *commit = alloc_commit_node(repo);
171
172         set_merge_remote_desc(commit, comment, (struct object *)commit);
173         commit->maybe_tree = tree;
174         commit->object.parsed = 1;
175         return commit;
176 }
177
178 /*
179  * Since we use get_tree_entry(), which does not put the read object into
180  * the object pool, we cannot rely on a == b.
181  */
182 static int oid_eq(const struct object_id *a, const struct object_id *b)
183 {
184         if (!a && !b)
185                 return 2;
186         return a && b && oideq(a, b);
187 }
188
189 enum rename_type {
190         RENAME_NORMAL = 0,
191         RENAME_VIA_DIR,
192         RENAME_ADD,
193         RENAME_DELETE,
194         RENAME_ONE_FILE_TO_ONE,
195         RENAME_ONE_FILE_TO_TWO,
196         RENAME_TWO_FILES_TO_ONE
197 };
198
199 struct rename_conflict_info {
200         enum rename_type rename_type;
201         struct diff_filepair *pair1;
202         struct diff_filepair *pair2;
203         const char *branch1;
204         const char *branch2;
205         struct stage_data *dst_entry1;
206         struct stage_data *dst_entry2;
207         struct diff_filespec ren1_other;
208         struct diff_filespec ren2_other;
209 };
210
211 /*
212  * Since we want to write the index eventually, we cannot reuse the index
213  * for these (temporary) data.
214  */
215 struct stage_data {
216         struct {
217                 unsigned mode;
218                 struct object_id oid;
219         } stages[4];
220         struct rename_conflict_info *rename_conflict_info;
221         unsigned processed:1;
222 };
223
224 static inline void setup_rename_conflict_info(enum rename_type rename_type,
225                                               struct diff_filepair *pair1,
226                                               struct diff_filepair *pair2,
227                                               const char *branch1,
228                                               const char *branch2,
229                                               struct stage_data *dst_entry1,
230                                               struct stage_data *dst_entry2,
231                                               struct merge_options *o,
232                                               struct stage_data *src_entry1,
233                                               struct stage_data *src_entry2)
234 {
235         int ostage1 = 0, ostage2;
236         struct rename_conflict_info *ci;
237
238         /*
239          * When we have two renames involved, it's easiest to get the
240          * correct things into stage 2 and 3, and to make sure that the
241          * content merge puts HEAD before the other branch if we just
242          * ensure that branch1 == o->branch1.  So, simply flip arguments
243          * around if we don't have that.
244          */
245         if (dst_entry2 && branch1 != o->branch1) {
246                 setup_rename_conflict_info(rename_type,
247                                            pair2,      pair1,
248                                            branch2,    branch1,
249                                            dst_entry2, dst_entry1,
250                                            o,
251                                            src_entry2, src_entry1);
252                 return;
253         }
254
255         ci = xcalloc(1, sizeof(struct rename_conflict_info));
256         ci->rename_type = rename_type;
257         ci->pair1 = pair1;
258         ci->branch1 = branch1;
259         ci->branch2 = branch2;
260
261         ci->dst_entry1 = dst_entry1;
262         dst_entry1->rename_conflict_info = ci;
263         dst_entry1->processed = 0;
264
265         assert(!pair2 == !dst_entry2);
266         if (dst_entry2) {
267                 ci->dst_entry2 = dst_entry2;
268                 ci->pair2 = pair2;
269                 dst_entry2->rename_conflict_info = ci;
270         }
271
272         /*
273          * For each rename, there could have been
274          * modifications on the side of history where that
275          * file was not renamed.
276          */
277         if (rename_type == RENAME_ADD ||
278             rename_type == RENAME_TWO_FILES_TO_ONE) {
279                 ostage1 = o->branch1 == branch1 ? 3 : 2;
280
281                 ci->ren1_other.path = pair1->one->path;
282                 oidcpy(&ci->ren1_other.oid, &src_entry1->stages[ostage1].oid);
283                 ci->ren1_other.mode = src_entry1->stages[ostage1].mode;
284         }
285
286         if (rename_type == RENAME_TWO_FILES_TO_ONE) {
287                 ostage2 = ostage1 ^ 1;
288
289                 ci->ren2_other.path = pair2->one->path;
290                 oidcpy(&ci->ren2_other.oid, &src_entry2->stages[ostage2].oid);
291                 ci->ren2_other.mode = src_entry2->stages[ostage2].mode;
292         }
293 }
294
295 static int show(struct merge_options *o, int v)
296 {
297         return (!o->call_depth && o->verbosity >= v) || o->verbosity >= 5;
298 }
299
300 __attribute__((format (printf, 3, 4)))
301 static void output(struct merge_options *o, int v, const char *fmt, ...)
302 {
303         va_list ap;
304
305         if (!show(o, v))
306                 return;
307
308         strbuf_addchars(&o->obuf, ' ', o->call_depth * 2);
309
310         va_start(ap, fmt);
311         strbuf_vaddf(&o->obuf, fmt, ap);
312         va_end(ap);
313
314         strbuf_addch(&o->obuf, '\n');
315         if (!o->buffer_output)
316                 flush_output(o);
317 }
318
319 static void output_commit_title(struct merge_options *o, struct commit *commit)
320 {
321         struct merge_remote_desc *desc;
322
323         strbuf_addchars(&o->obuf, ' ', o->call_depth * 2);
324         desc = merge_remote_util(commit);
325         if (desc)
326                 strbuf_addf(&o->obuf, "virtual %s\n", desc->name);
327         else {
328                 strbuf_add_unique_abbrev(&o->obuf, &commit->object.oid,
329                                          DEFAULT_ABBREV);
330                 strbuf_addch(&o->obuf, ' ');
331                 if (parse_commit(commit) != 0)
332                         strbuf_addstr(&o->obuf, _("(bad commit)\n"));
333                 else {
334                         const char *title;
335                         const char *msg = get_commit_buffer(commit, NULL);
336                         int len = find_commit_subject(msg, &title);
337                         if (len)
338                                 strbuf_addf(&o->obuf, "%.*s\n", len, title);
339                         unuse_commit_buffer(commit, msg);
340                 }
341         }
342         flush_output(o);
343 }
344
345 static int add_cacheinfo(struct merge_options *o,
346                          unsigned int mode, const struct object_id *oid,
347                          const char *path, int stage, int refresh, int options)
348 {
349         struct index_state *istate = o->repo->index;
350         struct cache_entry *ce;
351         int ret;
352
353         ce = make_cache_entry(istate, mode, oid ? oid : &null_oid, path, stage, 0);
354         if (!ce)
355                 return err(o, _("add_cacheinfo failed for path '%s'; merge aborting."), path);
356
357         ret = add_index_entry(istate, ce, options);
358         if (refresh) {
359                 struct cache_entry *nce;
360
361                 nce = refresh_cache_entry(istate, ce,
362                                           CE_MATCH_REFRESH | CE_MATCH_IGNORE_MISSING);
363                 if (!nce)
364                         return err(o, _("add_cacheinfo failed to refresh for path '%s'; merge aborting."), path);
365                 if (nce != ce)
366                         ret = add_index_entry(istate, nce, options);
367         }
368         return ret;
369 }
370
371 static void init_tree_desc_from_tree(struct tree_desc *desc, struct tree *tree)
372 {
373         parse_tree(tree);
374         init_tree_desc(desc, tree->buffer, tree->size);
375 }
376
377 static int unpack_trees_start(struct merge_options *o,
378                               struct tree *common,
379                               struct tree *head,
380                               struct tree *merge)
381 {
382         int rc;
383         struct tree_desc t[3];
384         struct index_state tmp_index = { NULL };
385
386         memset(&o->unpack_opts, 0, sizeof(o->unpack_opts));
387         if (o->call_depth)
388                 o->unpack_opts.index_only = 1;
389         else
390                 o->unpack_opts.update = 1;
391         o->unpack_opts.merge = 1;
392         o->unpack_opts.head_idx = 2;
393         o->unpack_opts.fn = threeway_merge;
394         o->unpack_opts.src_index = o->repo->index;
395         o->unpack_opts.dst_index = &tmp_index;
396         o->unpack_opts.aggressive = !merge_detect_rename(o);
397         setup_unpack_trees_porcelain(&o->unpack_opts, "merge");
398
399         init_tree_desc_from_tree(t+0, common);
400         init_tree_desc_from_tree(t+1, head);
401         init_tree_desc_from_tree(t+2, merge);
402
403         rc = unpack_trees(3, t, &o->unpack_opts);
404         cache_tree_free(&o->repo->index->cache_tree);
405
406         /*
407          * Update o->repo->index to match the new results, AFTER saving a copy
408          * in o->orig_index.  Update src_index to point to the saved copy.
409          * (verify_uptodate() checks src_index, and the original index is
410          * the one that had the necessary modification timestamps.)
411          */
412         o->orig_index = *o->repo->index;
413         *o->repo->index = tmp_index;
414         o->unpack_opts.src_index = &o->orig_index;
415
416         return rc;
417 }
418
419 static void unpack_trees_finish(struct merge_options *o)
420 {
421         discard_index(&o->orig_index);
422         clear_unpack_trees_porcelain(&o->unpack_opts);
423 }
424
425 struct tree *write_tree_from_memory(struct merge_options *o)
426 {
427         struct tree *result = NULL;
428         struct index_state *istate = o->repo->index;
429
430         if (unmerged_index(istate)) {
431                 int i;
432                 fprintf(stderr, "BUG: There are unmerged index entries:\n");
433                 for (i = 0; i < istate->cache_nr; i++) {
434                         const struct cache_entry *ce = istate->cache[i];
435                         if (ce_stage(ce))
436                                 fprintf(stderr, "BUG: %d %.*s\n", ce_stage(ce),
437                                         (int)ce_namelen(ce), ce->name);
438                 }
439                 BUG("unmerged index entries in merge-recursive.c");
440         }
441
442         if (!istate->cache_tree)
443                 istate->cache_tree = cache_tree();
444
445         if (!cache_tree_fully_valid(istate->cache_tree) &&
446             cache_tree_update(istate, 0) < 0) {
447                 err(o, _("error building trees"));
448                 return NULL;
449         }
450
451         result = lookup_tree(o->repo, &istate->cache_tree->oid);
452
453         return result;
454 }
455
456 static int save_files_dirs(const struct object_id *oid,
457                            struct strbuf *base, const char *path,
458                            unsigned int mode, int stage, void *context)
459 {
460         struct path_hashmap_entry *entry;
461         int baselen = base->len;
462         struct merge_options *o = context;
463
464         strbuf_addstr(base, path);
465
466         FLEX_ALLOC_MEM(entry, path, base->buf, base->len);
467         hashmap_entry_init(entry, path_hash(entry->path));
468         hashmap_add(&o->current_file_dir_set, entry);
469
470         strbuf_setlen(base, baselen);
471         return (S_ISDIR(mode) ? READ_TREE_RECURSIVE : 0);
472 }
473
474 static void get_files_dirs(struct merge_options *o, struct tree *tree)
475 {
476         struct pathspec match_all;
477         memset(&match_all, 0, sizeof(match_all));
478         read_tree_recursive(tree, "", 0, 0, &match_all, save_files_dirs, o);
479 }
480
481 static int get_tree_entry_if_blob(const struct object_id *tree,
482                                   const char *path,
483                                   struct object_id *hashy,
484                                   unsigned int *mode_o)
485 {
486         int ret;
487
488         ret = get_tree_entry(tree, path, hashy, mode_o);
489         if (S_ISDIR(*mode_o)) {
490                 oidcpy(hashy, &null_oid);
491                 *mode_o = 0;
492         }
493         return ret;
494 }
495
496 /*
497  * Returns an index_entry instance which doesn't have to correspond to
498  * a real cache entry in Git's index.
499  */
500 static struct stage_data *insert_stage_data(const char *path,
501                 struct tree *o, struct tree *a, struct tree *b,
502                 struct string_list *entries)
503 {
504         struct string_list_item *item;
505         struct stage_data *e = xcalloc(1, sizeof(struct stage_data));
506         get_tree_entry_if_blob(&o->object.oid, path,
507                                &e->stages[1].oid, &e->stages[1].mode);
508         get_tree_entry_if_blob(&a->object.oid, path,
509                                &e->stages[2].oid, &e->stages[2].mode);
510         get_tree_entry_if_blob(&b->object.oid, path,
511                                &e->stages[3].oid, &e->stages[3].mode);
512         item = string_list_insert(entries, path);
513         item->util = e;
514         return e;
515 }
516
517 /*
518  * Create a dictionary mapping file names to stage_data objects. The
519  * dictionary contains one entry for every path with a non-zero stage entry.
520  */
521 static struct string_list *get_unmerged(struct index_state *istate)
522 {
523         struct string_list *unmerged = xcalloc(1, sizeof(struct string_list));
524         int i;
525
526         unmerged->strdup_strings = 1;
527
528         for (i = 0; i < istate->cache_nr; i++) {
529                 struct string_list_item *item;
530                 struct stage_data *e;
531                 const struct cache_entry *ce = istate->cache[i];
532                 if (!ce_stage(ce))
533                         continue;
534
535                 item = string_list_lookup(unmerged, ce->name);
536                 if (!item) {
537                         item = string_list_insert(unmerged, ce->name);
538                         item->util = xcalloc(1, sizeof(struct stage_data));
539                 }
540                 e = item->util;
541                 e->stages[ce_stage(ce)].mode = ce->ce_mode;
542                 oidcpy(&e->stages[ce_stage(ce)].oid, &ce->oid);
543         }
544
545         return unmerged;
546 }
547
548 static int string_list_df_name_compare(const char *one, const char *two)
549 {
550         int onelen = strlen(one);
551         int twolen = strlen(two);
552         /*
553          * Here we only care that entries for D/F conflicts are
554          * adjacent, in particular with the file of the D/F conflict
555          * appearing before files below the corresponding directory.
556          * The order of the rest of the list is irrelevant for us.
557          *
558          * To achieve this, we sort with df_name_compare and provide
559          * the mode S_IFDIR so that D/F conflicts will sort correctly.
560          * We use the mode S_IFDIR for everything else for simplicity,
561          * since in other cases any changes in their order due to
562          * sorting cause no problems for us.
563          */
564         int cmp = df_name_compare(one, onelen, S_IFDIR,
565                                   two, twolen, S_IFDIR);
566         /*
567          * Now that 'foo' and 'foo/bar' compare equal, we have to make sure
568          * that 'foo' comes before 'foo/bar'.
569          */
570         if (cmp)
571                 return cmp;
572         return onelen - twolen;
573 }
574
575 static void record_df_conflict_files(struct merge_options *o,
576                                      struct string_list *entries)
577 {
578         /* If there is a D/F conflict and the file for such a conflict
579          * currently exists in the working tree, we want to allow it to be
580          * removed to make room for the corresponding directory if needed.
581          * The files underneath the directories of such D/F conflicts will
582          * be processed before the corresponding file involved in the D/F
583          * conflict.  If the D/F directory ends up being removed by the
584          * merge, then we won't have to touch the D/F file.  If the D/F
585          * directory needs to be written to the working copy, then the D/F
586          * file will simply be removed (in make_room_for_path()) to make
587          * room for the necessary paths.  Note that if both the directory
588          * and the file need to be present, then the D/F file will be
589          * reinstated with a new unique name at the time it is processed.
590          */
591         struct string_list df_sorted_entries = STRING_LIST_INIT_NODUP;
592         const char *last_file = NULL;
593         int last_len = 0;
594         int i;
595
596         /*
597          * If we're merging merge-bases, we don't want to bother with
598          * any working directory changes.
599          */
600         if (o->call_depth)
601                 return;
602
603         /* Ensure D/F conflicts are adjacent in the entries list. */
604         for (i = 0; i < entries->nr; i++) {
605                 struct string_list_item *next = &entries->items[i];
606                 string_list_append(&df_sorted_entries, next->string)->util =
607                                    next->util;
608         }
609         df_sorted_entries.cmp = string_list_df_name_compare;
610         string_list_sort(&df_sorted_entries);
611
612         string_list_clear(&o->df_conflict_file_set, 1);
613         for (i = 0; i < df_sorted_entries.nr; i++) {
614                 const char *path = df_sorted_entries.items[i].string;
615                 int len = strlen(path);
616                 struct stage_data *e = df_sorted_entries.items[i].util;
617
618                 /*
619                  * Check if last_file & path correspond to a D/F conflict;
620                  * i.e. whether path is last_file+'/'+<something>.
621                  * If so, record that it's okay to remove last_file to make
622                  * room for path and friends if needed.
623                  */
624                 if (last_file &&
625                     len > last_len &&
626                     memcmp(path, last_file, last_len) == 0 &&
627                     path[last_len] == '/') {
628                         string_list_insert(&o->df_conflict_file_set, last_file);
629                 }
630
631                 /*
632                  * Determine whether path could exist as a file in the
633                  * working directory as a possible D/F conflict.  This
634                  * will only occur when it exists in stage 2 as a
635                  * file.
636                  */
637                 if (S_ISREG(e->stages[2].mode) || S_ISLNK(e->stages[2].mode)) {
638                         last_file = path;
639                         last_len = len;
640                 } else {
641                         last_file = NULL;
642                 }
643         }
644         string_list_clear(&df_sorted_entries, 0);
645 }
646
647 struct rename {
648         struct diff_filepair *pair;
649         /*
650          * Purpose of src_entry and dst_entry:
651          *
652          * If 'before' is renamed to 'after' then src_entry will contain
653          * the versions of 'before' from the merge_base, HEAD, and MERGE in
654          * stages 1, 2, and 3; dst_entry will contain the respective
655          * versions of 'after' in corresponding locations.  Thus, we have a
656          * total of six modes and oids, though some will be null.  (Stage 0
657          * is ignored; we're interested in handling conflicts.)
658          *
659          * Since we don't turn on break-rewrites by default, neither
660          * src_entry nor dst_entry can have all three of their stages have
661          * non-null oids, meaning at most four of the six will be non-null.
662          * Also, since this is a rename, both src_entry and dst_entry will
663          * have at least one non-null oid, meaning at least two will be
664          * non-null.  Of the six oids, a typical rename will have three be
665          * non-null.  Only two implies a rename/delete, and four implies a
666          * rename/add.
667          */
668         struct stage_data *src_entry;
669         struct stage_data *dst_entry;
670         unsigned add_turned_into_rename:1;
671         unsigned processed:1;
672 };
673
674 static int update_stages(struct merge_options *opt, const char *path,
675                          const struct diff_filespec *o,
676                          const struct diff_filespec *a,
677                          const struct diff_filespec *b)
678 {
679
680         /*
681          * NOTE: It is usually a bad idea to call update_stages on a path
682          * before calling update_file on that same path, since it can
683          * sometimes lead to spurious "refusing to lose untracked file..."
684          * messages from update_file (via make_room_for path via
685          * would_lose_untracked).  Instead, reverse the order of the calls
686          * (executing update_file first and then update_stages).
687          */
688         int clear = 1;
689         int options = ADD_CACHE_OK_TO_ADD | ADD_CACHE_SKIP_DFCHECK;
690         if (clear)
691                 if (remove_file_from_index(opt->repo->index, path))
692                         return -1;
693         if (o)
694                 if (add_cacheinfo(opt, o->mode, &o->oid, path, 1, 0, options))
695                         return -1;
696         if (a)
697                 if (add_cacheinfo(opt, a->mode, &a->oid, path, 2, 0, options))
698                         return -1;
699         if (b)
700                 if (add_cacheinfo(opt, b->mode, &b->oid, path, 3, 0, options))
701                         return -1;
702         return 0;
703 }
704
705 static void update_entry(struct stage_data *entry,
706                          struct diff_filespec *o,
707                          struct diff_filespec *a,
708                          struct diff_filespec *b)
709 {
710         entry->processed = 0;
711         entry->stages[1].mode = o->mode;
712         entry->stages[2].mode = a->mode;
713         entry->stages[3].mode = b->mode;
714         oidcpy(&entry->stages[1].oid, &o->oid);
715         oidcpy(&entry->stages[2].oid, &a->oid);
716         oidcpy(&entry->stages[3].oid, &b->oid);
717 }
718
719 static int remove_file(struct merge_options *o, int clean,
720                        const char *path, int no_wd)
721 {
722         int update_cache = o->call_depth || clean;
723         int update_working_directory = !o->call_depth && !no_wd;
724
725         if (update_cache) {
726                 if (remove_file_from_index(o->repo->index, path))
727                         return -1;
728         }
729         if (update_working_directory) {
730                 if (ignore_case) {
731                         struct cache_entry *ce;
732                         ce = index_file_exists(o->repo->index, path, strlen(path),
733                                                ignore_case);
734                         if (ce && ce_stage(ce) == 0 && strcmp(path, ce->name))
735                                 return 0;
736                 }
737                 if (remove_path(path))
738                         return -1;
739         }
740         return 0;
741 }
742
743 /* add a string to a strbuf, but converting "/" to "_" */
744 static void add_flattened_path(struct strbuf *out, const char *s)
745 {
746         size_t i = out->len;
747         strbuf_addstr(out, s);
748         for (; i < out->len; i++)
749                 if (out->buf[i] == '/')
750                         out->buf[i] = '_';
751 }
752
753 static char *unique_path(struct merge_options *o, const char *path, const char *branch)
754 {
755         struct path_hashmap_entry *entry;
756         struct strbuf newpath = STRBUF_INIT;
757         int suffix = 0;
758         size_t base_len;
759
760         strbuf_addf(&newpath, "%s~", path);
761         add_flattened_path(&newpath, branch);
762
763         base_len = newpath.len;
764         while (hashmap_get_from_hash(&o->current_file_dir_set,
765                                      path_hash(newpath.buf), newpath.buf) ||
766                (!o->call_depth && file_exists(newpath.buf))) {
767                 strbuf_setlen(&newpath, base_len);
768                 strbuf_addf(&newpath, "_%d", suffix++);
769         }
770
771         FLEX_ALLOC_MEM(entry, path, newpath.buf, newpath.len);
772         hashmap_entry_init(entry, path_hash(entry->path));
773         hashmap_add(&o->current_file_dir_set, entry);
774         return strbuf_detach(&newpath, NULL);
775 }
776
777 /**
778  * Check whether a directory in the index is in the way of an incoming
779  * file.  Return 1 if so.  If check_working_copy is non-zero, also
780  * check the working directory.  If empty_ok is non-zero, also return
781  * 0 in the case where the working-tree dir exists but is empty.
782  */
783 static int dir_in_way(struct index_state *istate, const char *path,
784                       int check_working_copy, int empty_ok)
785 {
786         int pos;
787         struct strbuf dirpath = STRBUF_INIT;
788         struct stat st;
789
790         strbuf_addstr(&dirpath, path);
791         strbuf_addch(&dirpath, '/');
792
793         pos = index_name_pos(istate, dirpath.buf, dirpath.len);
794
795         if (pos < 0)
796                 pos = -1 - pos;
797         if (pos < istate->cache_nr &&
798             !strncmp(dirpath.buf, istate->cache[pos]->name, dirpath.len)) {
799                 strbuf_release(&dirpath);
800                 return 1;
801         }
802
803         strbuf_release(&dirpath);
804         return check_working_copy && !lstat(path, &st) && S_ISDIR(st.st_mode) &&
805                 !(empty_ok && is_empty_dir(path));
806 }
807
808 /*
809  * Returns whether path was tracked in the index before the merge started,
810  * and its oid and mode match the specified values
811  */
812 static int was_tracked_and_matches(struct merge_options *o, const char *path,
813                                    const struct object_id *oid, unsigned mode)
814 {
815         int pos = index_name_pos(&o->orig_index, path, strlen(path));
816         struct cache_entry *ce;
817
818         if (0 > pos)
819                 /* we were not tracking this path before the merge */
820                 return 0;
821
822         /* See if the file we were tracking before matches */
823         ce = o->orig_index.cache[pos];
824         return (oid_eq(&ce->oid, oid) && ce->ce_mode == mode);
825 }
826
827 /*
828  * Returns whether path was tracked in the index before the merge started
829  */
830 static int was_tracked(struct merge_options *o, const char *path)
831 {
832         int pos = index_name_pos(&o->orig_index, path, strlen(path));
833
834         if (0 <= pos)
835                 /* we were tracking this path before the merge */
836                 return 1;
837
838         return 0;
839 }
840
841 static int would_lose_untracked(struct merge_options *o, const char *path)
842 {
843         struct index_state *istate = o->repo->index;
844
845         /*
846          * This may look like it can be simplified to:
847          *   return !was_tracked(o, path) && file_exists(path)
848          * but it can't.  This function needs to know whether path was in
849          * the working tree due to EITHER having been tracked in the index
850          * before the merge OR having been put into the working copy and
851          * index by unpack_trees().  Due to that either-or requirement, we
852          * check the current index instead of the original one.
853          *
854          * Note that we do not need to worry about merge-recursive itself
855          * updating the index after unpack_trees() and before calling this
856          * function, because we strictly require all code paths in
857          * merge-recursive to update the working tree first and the index
858          * second.  Doing otherwise would break
859          * update_file()/would_lose_untracked(); see every comment in this
860          * file which mentions "update_stages".
861          */
862         int pos = index_name_pos(istate, path, strlen(path));
863
864         if (pos < 0)
865                 pos = -1 - pos;
866         while (pos < istate->cache_nr &&
867                !strcmp(path, istate->cache[pos]->name)) {
868                 /*
869                  * If stage #0, it is definitely tracked.
870                  * If it has stage #2 then it was tracked
871                  * before this merge started.  All other
872                  * cases the path was not tracked.
873                  */
874                 switch (ce_stage(istate->cache[pos])) {
875                 case 0:
876                 case 2:
877                         return 0;
878                 }
879                 pos++;
880         }
881         return file_exists(path);
882 }
883
884 static int was_dirty(struct merge_options *o, const char *path)
885 {
886         struct cache_entry *ce;
887         int dirty = 1;
888
889         if (o->call_depth || !was_tracked(o, path))
890                 return !dirty;
891
892         ce = index_file_exists(o->unpack_opts.src_index,
893                                path, strlen(path), ignore_case);
894         dirty = verify_uptodate(ce, &o->unpack_opts) != 0;
895         return dirty;
896 }
897
898 static int make_room_for_path(struct merge_options *o, const char *path)
899 {
900         int status, i;
901         const char *msg = _("failed to create path '%s'%s");
902
903         /* Unlink any D/F conflict files that are in the way */
904         for (i = 0; i < o->df_conflict_file_set.nr; i++) {
905                 const char *df_path = o->df_conflict_file_set.items[i].string;
906                 size_t pathlen = strlen(path);
907                 size_t df_pathlen = strlen(df_path);
908                 if (df_pathlen < pathlen &&
909                     path[df_pathlen] == '/' &&
910                     strncmp(path, df_path, df_pathlen) == 0) {
911                         output(o, 3,
912                                _("Removing %s to make room for subdirectory\n"),
913                                df_path);
914                         unlink(df_path);
915                         unsorted_string_list_delete_item(&o->df_conflict_file_set,
916                                                          i, 0);
917                         break;
918                 }
919         }
920
921         /* Make sure leading directories are created */
922         status = safe_create_leading_directories_const(path);
923         if (status) {
924                 if (status == SCLD_EXISTS)
925                         /* something else exists */
926                         return err(o, msg, path, _(": perhaps a D/F conflict?"));
927                 return err(o, msg, path, "");
928         }
929
930         /*
931          * Do not unlink a file in the work tree if we are not
932          * tracking it.
933          */
934         if (would_lose_untracked(o, path))
935                 return err(o, _("refusing to lose untracked file at '%s'"),
936                            path);
937
938         /* Successful unlink is good.. */
939         if (!unlink(path))
940                 return 0;
941         /* .. and so is no existing file */
942         if (errno == ENOENT)
943                 return 0;
944         /* .. but not some other error (who really cares what?) */
945         return err(o, msg, path, _(": perhaps a D/F conflict?"));
946 }
947
948 static int update_file_flags(struct merge_options *o,
949                              const struct object_id *oid,
950                              unsigned mode,
951                              const char *path,
952                              int update_cache,
953                              int update_wd)
954 {
955         int ret = 0;
956
957         if (o->call_depth)
958                 update_wd = 0;
959
960         if (update_wd) {
961                 enum object_type type;
962                 void *buf;
963                 unsigned long size;
964
965                 if (S_ISGITLINK(mode)) {
966                         /*
967                          * We may later decide to recursively descend into
968                          * the submodule directory and update its index
969                          * and/or work tree, but we do not do that now.
970                          */
971                         update_wd = 0;
972                         goto update_index;
973                 }
974
975                 buf = read_object_file(oid, &type, &size);
976                 if (!buf)
977                         return err(o, _("cannot read object %s '%s'"), oid_to_hex(oid), path);
978                 if (type != OBJ_BLOB) {
979                         ret = err(o, _("blob expected for %s '%s'"), oid_to_hex(oid), path);
980                         goto free_buf;
981                 }
982                 if (S_ISREG(mode)) {
983                         struct strbuf strbuf = STRBUF_INIT;
984                         if (convert_to_working_tree(o->repo->index, path, buf, size, &strbuf)) {
985                                 free(buf);
986                                 size = strbuf.len;
987                                 buf = strbuf_detach(&strbuf, NULL);
988                         }
989                 }
990
991                 if (make_room_for_path(o, path) < 0) {
992                         update_wd = 0;
993                         goto free_buf;
994                 }
995                 if (S_ISREG(mode) || (!has_symlinks && S_ISLNK(mode))) {
996                         int fd;
997                         if (mode & 0100)
998                                 mode = 0777;
999                         else
1000                                 mode = 0666;
1001                         fd = open(path, O_WRONLY | O_TRUNC | O_CREAT, mode);
1002                         if (fd < 0) {
1003                                 ret = err(o, _("failed to open '%s': %s"),
1004                                           path, strerror(errno));
1005                                 goto free_buf;
1006                         }
1007                         write_in_full(fd, buf, size);
1008                         close(fd);
1009                 } else if (S_ISLNK(mode)) {
1010                         char *lnk = xmemdupz(buf, size);
1011                         safe_create_leading_directories_const(path);
1012                         unlink(path);
1013                         if (symlink(lnk, path))
1014                                 ret = err(o, _("failed to symlink '%s': %s"),
1015                                           path, strerror(errno));
1016                         free(lnk);
1017                 } else
1018                         ret = err(o,
1019                                   _("do not know what to do with %06o %s '%s'"),
1020                                   mode, oid_to_hex(oid), path);
1021         free_buf:
1022                 free(buf);
1023         }
1024 update_index:
1025         if (!ret && update_cache)
1026                 if (add_cacheinfo(o, mode, oid, path, 0, update_wd,
1027                                   ADD_CACHE_OK_TO_ADD))
1028                         return -1;
1029         return ret;
1030 }
1031
1032 static int update_file(struct merge_options *o,
1033                        int clean,
1034                        const struct object_id *oid,
1035                        unsigned mode,
1036                        const char *path)
1037 {
1038         return update_file_flags(o, oid, mode, path, o->call_depth || clean, !o->call_depth);
1039 }
1040
1041 /* Low level file merging, update and removal */
1042
1043 struct merge_file_info {
1044         struct object_id oid;
1045         unsigned mode;
1046         unsigned clean:1,
1047                  merge:1;
1048 };
1049
1050 static int merge_3way(struct merge_options *o,
1051                       mmbuffer_t *result_buf,
1052                       const struct diff_filespec *one,
1053                       const struct diff_filespec *a,
1054                       const struct diff_filespec *b,
1055                       const char *branch1,
1056                       const char *branch2,
1057                       const int extra_marker_size)
1058 {
1059         mmfile_t orig, src1, src2;
1060         struct ll_merge_options ll_opts = {0};
1061         char *base_name, *name1, *name2;
1062         int merge_status;
1063
1064         ll_opts.renormalize = o->renormalize;
1065         ll_opts.extra_marker_size = extra_marker_size;
1066         ll_opts.xdl_opts = o->xdl_opts;
1067
1068         if (o->call_depth) {
1069                 ll_opts.virtual_ancestor = 1;
1070                 ll_opts.variant = 0;
1071         } else {
1072                 switch (o->recursive_variant) {
1073                 case MERGE_RECURSIVE_OURS:
1074                         ll_opts.variant = XDL_MERGE_FAVOR_OURS;
1075                         break;
1076                 case MERGE_RECURSIVE_THEIRS:
1077                         ll_opts.variant = XDL_MERGE_FAVOR_THEIRS;
1078                         break;
1079                 default:
1080                         ll_opts.variant = 0;
1081                         break;
1082                 }
1083         }
1084
1085         if (strcmp(a->path, b->path) ||
1086             (o->ancestor != NULL && strcmp(a->path, one->path) != 0)) {
1087                 base_name = o->ancestor == NULL ? NULL :
1088                         mkpathdup("%s:%s", o->ancestor, one->path);
1089                 name1 = mkpathdup("%s:%s", branch1, a->path);
1090                 name2 = mkpathdup("%s:%s", branch2, b->path);
1091         } else {
1092                 base_name = o->ancestor == NULL ? NULL :
1093                         mkpathdup("%s", o->ancestor);
1094                 name1 = mkpathdup("%s", branch1);
1095                 name2 = mkpathdup("%s", branch2);
1096         }
1097
1098         read_mmblob(&orig, &one->oid);
1099         read_mmblob(&src1, &a->oid);
1100         read_mmblob(&src2, &b->oid);
1101
1102         merge_status = ll_merge(result_buf, a->path, &orig, base_name,
1103                                 &src1, name1, &src2, name2,
1104                                 o->repo->index, &ll_opts);
1105
1106         free(base_name);
1107         free(name1);
1108         free(name2);
1109         free(orig.ptr);
1110         free(src1.ptr);
1111         free(src2.ptr);
1112         return merge_status;
1113 }
1114
1115 static int find_first_merges(struct repository *repo,
1116                              struct object_array *result, const char *path,
1117                              struct commit *a, struct commit *b)
1118 {
1119         int i, j;
1120         struct object_array merges = OBJECT_ARRAY_INIT;
1121         struct commit *commit;
1122         int contains_another;
1123
1124         char merged_revision[42];
1125         const char *rev_args[] = { "rev-list", "--merges", "--ancestry-path",
1126                                    "--all", merged_revision, NULL };
1127         struct rev_info revs;
1128         struct setup_revision_opt rev_opts;
1129
1130         memset(result, 0, sizeof(struct object_array));
1131         memset(&rev_opts, 0, sizeof(rev_opts));
1132
1133         /* get all revisions that merge commit a */
1134         xsnprintf(merged_revision, sizeof(merged_revision), "^%s",
1135                   oid_to_hex(&a->object.oid));
1136         repo_init_revisions(repo, &revs, NULL);
1137         rev_opts.submodule = path;
1138         /* FIXME: can't handle linked worktrees in submodules yet */
1139         revs.single_worktree = path != NULL;
1140         setup_revisions(ARRAY_SIZE(rev_args)-1, rev_args, &revs, &rev_opts);
1141
1142         /* save all revisions from the above list that contain b */
1143         if (prepare_revision_walk(&revs))
1144                 die("revision walk setup failed");
1145         while ((commit = get_revision(&revs)) != NULL) {
1146                 struct object *o = &(commit->object);
1147                 if (in_merge_bases(b, commit))
1148                         add_object_array(o, NULL, &merges);
1149         }
1150         reset_revision_walk();
1151
1152         /* Now we've got all merges that contain a and b. Prune all
1153          * merges that contain another found merge and save them in
1154          * result.
1155          */
1156         for (i = 0; i < merges.nr; i++) {
1157                 struct commit *m1 = (struct commit *) merges.objects[i].item;
1158
1159                 contains_another = 0;
1160                 for (j = 0; j < merges.nr; j++) {
1161                         struct commit *m2 = (struct commit *) merges.objects[j].item;
1162                         if (i != j && in_merge_bases(m2, m1)) {
1163                                 contains_another = 1;
1164                                 break;
1165                         }
1166                 }
1167
1168                 if (!contains_another)
1169                         add_object_array(merges.objects[i].item, NULL, result);
1170         }
1171
1172         object_array_clear(&merges);
1173         return result->nr;
1174 }
1175
1176 static void print_commit(struct commit *commit)
1177 {
1178         struct strbuf sb = STRBUF_INIT;
1179         struct pretty_print_context ctx = {0};
1180         ctx.date_mode.type = DATE_NORMAL;
1181         format_commit_message(commit, " %h: %m %s", &sb, &ctx);
1182         fprintf(stderr, "%s\n", sb.buf);
1183         strbuf_release(&sb);
1184 }
1185
1186 static int merge_submodule(struct merge_options *o,
1187                            struct object_id *result, const char *path,
1188                            const struct object_id *base, const struct object_id *a,
1189                            const struct object_id *b)
1190 {
1191         struct commit *commit_base, *commit_a, *commit_b;
1192         int parent_count;
1193         struct object_array merges;
1194
1195         int i;
1196         int search = !o->call_depth;
1197
1198         /* store a in result in case we fail */
1199         oidcpy(result, a);
1200
1201         /* we can not handle deletion conflicts */
1202         if (is_null_oid(base))
1203                 return 0;
1204         if (is_null_oid(a))
1205                 return 0;
1206         if (is_null_oid(b))
1207                 return 0;
1208
1209         if (add_submodule_odb(path)) {
1210                 output(o, 1, _("Failed to merge submodule %s (not checked out)"), path);
1211                 return 0;
1212         }
1213
1214         if (!(commit_base = lookup_commit_reference(o->repo, base)) ||
1215             !(commit_a = lookup_commit_reference(o->repo, a)) ||
1216             !(commit_b = lookup_commit_reference(o->repo, b))) {
1217                 output(o, 1, _("Failed to merge submodule %s (commits not present)"), path);
1218                 return 0;
1219         }
1220
1221         /* check whether both changes are forward */
1222         if (!in_merge_bases(commit_base, commit_a) ||
1223             !in_merge_bases(commit_base, commit_b)) {
1224                 output(o, 1, _("Failed to merge submodule %s (commits don't follow merge-base)"), path);
1225                 return 0;
1226         }
1227
1228         /* Case #1: a is contained in b or vice versa */
1229         if (in_merge_bases(commit_a, commit_b)) {
1230                 oidcpy(result, b);
1231                 if (show(o, 3)) {
1232                         output(o, 3, _("Fast-forwarding submodule %s to the following commit:"), path);
1233                         output_commit_title(o, commit_b);
1234                 } else if (show(o, 2))
1235                         output(o, 2, _("Fast-forwarding submodule %s"), path);
1236                 else
1237                         ; /* no output */
1238
1239                 return 1;
1240         }
1241         if (in_merge_bases(commit_b, commit_a)) {
1242                 oidcpy(result, a);
1243                 if (show(o, 3)) {
1244                         output(o, 3, _("Fast-forwarding submodule %s to the following commit:"), path);
1245                         output_commit_title(o, commit_a);
1246                 } else if (show(o, 2))
1247                         output(o, 2, _("Fast-forwarding submodule %s"), path);
1248                 else
1249                         ; /* no output */
1250
1251                 return 1;
1252         }
1253
1254         /*
1255          * Case #2: There are one or more merges that contain a and b in
1256          * the submodule. If there is only one, then present it as a
1257          * suggestion to the user, but leave it marked unmerged so the
1258          * user needs to confirm the resolution.
1259          */
1260
1261         /* Skip the search if makes no sense to the calling context.  */
1262         if (!search)
1263                 return 0;
1264
1265         /* find commit which merges them */
1266         parent_count = find_first_merges(o->repo, &merges, path,
1267                                          commit_a, commit_b);
1268         switch (parent_count) {
1269         case 0:
1270                 output(o, 1, _("Failed to merge submodule %s (merge following commits not found)"), path);
1271                 break;
1272
1273         case 1:
1274                 output(o, 1, _("Failed to merge submodule %s (not fast-forward)"), path);
1275                 output(o, 2, _("Found a possible merge resolution for the submodule:\n"));
1276                 print_commit((struct commit *) merges.objects[0].item);
1277                 output(o, 2, _(
1278                        "If this is correct simply add it to the index "
1279                        "for example\n"
1280                        "by using:\n\n"
1281                        "  git update-index --cacheinfo 160000 %s \"%s\"\n\n"
1282                        "which will accept this suggestion.\n"),
1283                        oid_to_hex(&merges.objects[0].item->oid), path);
1284                 break;
1285
1286         default:
1287                 output(o, 1, _("Failed to merge submodule %s (multiple merges found)"), path);
1288                 for (i = 0; i < merges.nr; i++)
1289                         print_commit((struct commit *) merges.objects[i].item);
1290         }
1291
1292         object_array_clear(&merges);
1293         return 0;
1294 }
1295
1296 static int merge_mode_and_contents(struct merge_options *o,
1297                                    const struct diff_filespec *one,
1298                                    const struct diff_filespec *a,
1299                                    const struct diff_filespec *b,
1300                                    const char *filename,
1301                                    const char *branch1,
1302                                    const char *branch2,
1303                                    const int extra_marker_size,
1304                                    struct merge_file_info *result)
1305 {
1306         if (o->branch1 != branch1) {
1307                 /*
1308                  * It's weird getting a reverse merge with HEAD on the bottom
1309                  * side of the conflict markers and the other branch on the
1310                  * top.  Fix that.
1311                  */
1312                 return merge_mode_and_contents(o, one, b, a,
1313                                                filename,
1314                                                branch2, branch1,
1315                                                extra_marker_size, result);
1316         }
1317
1318         result->merge = 0;
1319         result->clean = 1;
1320
1321         if ((S_IFMT & a->mode) != (S_IFMT & b->mode)) {
1322                 result->clean = 0;
1323                 if (S_ISREG(a->mode)) {
1324                         result->mode = a->mode;
1325                         oidcpy(&result->oid, &a->oid);
1326                 } else {
1327                         result->mode = b->mode;
1328                         oidcpy(&result->oid, &b->oid);
1329                 }
1330         } else {
1331                 if (!oid_eq(&a->oid, &one->oid) && !oid_eq(&b->oid, &one->oid))
1332                         result->merge = 1;
1333
1334                 /*
1335                  * Merge modes
1336                  */
1337                 if (a->mode == b->mode || a->mode == one->mode)
1338                         result->mode = b->mode;
1339                 else {
1340                         result->mode = a->mode;
1341                         if (b->mode != one->mode) {
1342                                 result->clean = 0;
1343                                 result->merge = 1;
1344                         }
1345                 }
1346
1347                 if (oid_eq(&a->oid, &b->oid) || oid_eq(&a->oid, &one->oid))
1348                         oidcpy(&result->oid, &b->oid);
1349                 else if (oid_eq(&b->oid, &one->oid))
1350                         oidcpy(&result->oid, &a->oid);
1351                 else if (S_ISREG(a->mode)) {
1352                         mmbuffer_t result_buf;
1353                         int ret = 0, merge_status;
1354
1355                         merge_status = merge_3way(o, &result_buf, one, a, b,
1356                                                   branch1, branch2,
1357                                                   extra_marker_size);
1358
1359                         if ((merge_status < 0) || !result_buf.ptr)
1360                                 ret = err(o, _("Failed to execute internal merge"));
1361
1362                         if (!ret &&
1363                             write_object_file(result_buf.ptr, result_buf.size,
1364                                               blob_type, &result->oid))
1365                                 ret = err(o, _("Unable to add %s to database"),
1366                                           a->path);
1367
1368                         free(result_buf.ptr);
1369                         if (ret)
1370                                 return ret;
1371                         result->clean = (merge_status == 0);
1372                 } else if (S_ISGITLINK(a->mode)) {
1373                         result->clean = merge_submodule(o, &result->oid,
1374                                                         one->path,
1375                                                         &one->oid,
1376                                                         &a->oid,
1377                                                         &b->oid);
1378                 } else if (S_ISLNK(a->mode)) {
1379                         switch (o->recursive_variant) {
1380                         case MERGE_RECURSIVE_NORMAL:
1381                                 oidcpy(&result->oid, &a->oid);
1382                                 if (!oid_eq(&a->oid, &b->oid))
1383                                         result->clean = 0;
1384                                 break;
1385                         case MERGE_RECURSIVE_OURS:
1386                                 oidcpy(&result->oid, &a->oid);
1387                                 break;
1388                         case MERGE_RECURSIVE_THEIRS:
1389                                 oidcpy(&result->oid, &b->oid);
1390                                 break;
1391                         }
1392                 } else
1393                         BUG("unsupported object type in the tree");
1394         }
1395
1396         if (result->merge)
1397                 output(o, 2, _("Auto-merging %s"), filename);
1398
1399         return 0;
1400 }
1401
1402 static int handle_rename_via_dir(struct merge_options *o,
1403                                  struct diff_filepair *pair,
1404                                  const char *rename_branch,
1405                                  const char *other_branch)
1406 {
1407         /*
1408          * Handle file adds that need to be renamed due to directory rename
1409          * detection.  This differs from handle_rename_normal, because
1410          * there is no content merge to do; just move the file into the
1411          * desired final location.
1412          */
1413         const struct diff_filespec *dest = pair->two;
1414
1415         if (!o->call_depth && would_lose_untracked(o, dest->path)) {
1416                 char *alt_path = unique_path(o, dest->path, rename_branch);
1417
1418                 output(o, 1, _("Error: Refusing to lose untracked file at %s; "
1419                                "writing to %s instead."),
1420                        dest->path, alt_path);
1421                 /*
1422                  * Write the file in worktree at alt_path, but not in the
1423                  * index.  Instead, write to dest->path for the index but
1424                  * only at the higher appropriate stage.
1425                  */
1426                 if (update_file(o, 0, &dest->oid, dest->mode, alt_path))
1427                         return -1;
1428                 free(alt_path);
1429                 return update_stages(o, dest->path, NULL,
1430                                      rename_branch == o->branch1 ? dest : NULL,
1431                                      rename_branch == o->branch1 ? NULL : dest);
1432         }
1433
1434         /* Update dest->path both in index and in worktree */
1435         if (update_file(o, 1, &dest->oid, dest->mode, dest->path))
1436                 return -1;
1437         return 0;
1438 }
1439
1440 static int handle_change_delete(struct merge_options *o,
1441                                 const char *path, const char *old_path,
1442                                 const struct object_id *o_oid, int o_mode,
1443                                 const struct object_id *changed_oid,
1444                                 int changed_mode,
1445                                 const char *change_branch,
1446                                 const char *delete_branch,
1447                                 const char *change, const char *change_past)
1448 {
1449         char *alt_path = NULL;
1450         const char *update_path = path;
1451         int ret = 0;
1452
1453         if (dir_in_way(o->repo->index, path, !o->call_depth, 0) ||
1454             (!o->call_depth && would_lose_untracked(o, path))) {
1455                 update_path = alt_path = unique_path(o, path, change_branch);
1456         }
1457
1458         if (o->call_depth) {
1459                 /*
1460                  * We cannot arbitrarily accept either a_sha or b_sha as
1461                  * correct; since there is no true "middle point" between
1462                  * them, simply reuse the base version for virtual merge base.
1463                  */
1464                 ret = remove_file_from_index(o->repo->index, path);
1465                 if (!ret)
1466                         ret = update_file(o, 0, o_oid, o_mode, update_path);
1467         } else {
1468                 /*
1469                  * Despite the four nearly duplicate messages and argument
1470                  * lists below and the ugliness of the nested if-statements,
1471                  * having complete messages makes the job easier for
1472                  * translators.
1473                  *
1474                  * The slight variance among the cases is due to the fact
1475                  * that:
1476                  *   1) directory/file conflicts (in effect if
1477                  *      !alt_path) could cause us to need to write the
1478                  *      file to a different path.
1479                  *   2) renames (in effect if !old_path) could mean that
1480                  *      there are two names for the path that the user
1481                  *      may know the file by.
1482                  */
1483                 if (!alt_path) {
1484                         if (!old_path) {
1485                                 output(o, 1, _("CONFLICT (%s/delete): %s deleted in %s "
1486                                        "and %s in %s. Version %s of %s left in tree."),
1487                                        change, path, delete_branch, change_past,
1488                                        change_branch, change_branch, path);
1489                         } else {
1490                                 output(o, 1, _("CONFLICT (%s/delete): %s deleted in %s "
1491                                        "and %s to %s in %s. Version %s of %s left in tree."),
1492                                        change, old_path, delete_branch, change_past, path,
1493                                        change_branch, change_branch, path);
1494                         }
1495                 } else {
1496                         if (!old_path) {
1497                                 output(o, 1, _("CONFLICT (%s/delete): %s deleted in %s "
1498                                        "and %s in %s. Version %s of %s left in tree at %s."),
1499                                        change, path, delete_branch, change_past,
1500                                        change_branch, change_branch, path, alt_path);
1501                         } else {
1502                                 output(o, 1, _("CONFLICT (%s/delete): %s deleted in %s "
1503                                        "and %s to %s in %s. Version %s of %s left in tree at %s."),
1504                                        change, old_path, delete_branch, change_past, path,
1505                                        change_branch, change_branch, path, alt_path);
1506                         }
1507                 }
1508                 /*
1509                  * No need to call update_file() on path when change_branch ==
1510                  * o->branch1 && !alt_path, since that would needlessly touch
1511                  * path.  We could call update_file_flags() with update_cache=0
1512                  * and update_wd=0, but that's a no-op.
1513                  */
1514                 if (change_branch != o->branch1 || alt_path)
1515                         ret = update_file(o, 0, changed_oid, changed_mode, update_path);
1516         }
1517         free(alt_path);
1518
1519         return ret;
1520 }
1521
1522 static int handle_rename_delete(struct merge_options *o,
1523                                 struct diff_filepair *pair,
1524                                 const char *rename_branch,
1525                                 const char *delete_branch)
1526 {
1527         const struct diff_filespec *orig = pair->one;
1528         const struct diff_filespec *dest = pair->two;
1529
1530         if (handle_change_delete(o,
1531                                  o->call_depth ? orig->path : dest->path,
1532                                  o->call_depth ? NULL : orig->path,
1533                                  &orig->oid, orig->mode,
1534                                  &dest->oid, dest->mode,
1535                                  rename_branch, delete_branch,
1536                                  _("rename"), _("renamed")))
1537                 return -1;
1538
1539         if (o->call_depth)
1540                 return remove_file_from_index(o->repo->index, dest->path);
1541         else
1542                 return update_stages(o, dest->path, NULL,
1543                                      rename_branch == o->branch1 ? dest : NULL,
1544                                      rename_branch == o->branch1 ? NULL : dest);
1545 }
1546
1547 static struct diff_filespec *filespec_from_entry(struct diff_filespec *target,
1548                                                  struct stage_data *entry,
1549                                                  int stage)
1550 {
1551         struct object_id *oid = &entry->stages[stage].oid;
1552         unsigned mode = entry->stages[stage].mode;
1553         if (mode == 0 || is_null_oid(oid))
1554                 return NULL;
1555         oidcpy(&target->oid, oid);
1556         target->mode = mode;
1557         return target;
1558 }
1559
1560 static int handle_file_collision(struct merge_options *o,
1561                                  const char *collide_path,
1562                                  const char *prev_path1,
1563                                  const char *prev_path2,
1564                                  const char *branch1, const char *branch2,
1565                                  const struct object_id *a_oid,
1566                                  unsigned int a_mode,
1567                                  const struct object_id *b_oid,
1568                                  unsigned int b_mode)
1569 {
1570         struct merge_file_info mfi;
1571         struct diff_filespec null, a, b;
1572         char *alt_path = NULL;
1573         const char *update_path = collide_path;
1574
1575         /*
1576          * It's easiest to get the correct things into stage 2 and 3, and
1577          * to make sure that the content merge puts HEAD before the other
1578          * branch if we just ensure that branch1 == o->branch1.  So, simply
1579          * flip arguments around if we don't have that.
1580          */
1581         if (branch1 != o->branch1) {
1582                 return handle_file_collision(o, collide_path,
1583                                              prev_path2, prev_path1,
1584                                              branch2, branch1,
1585                                              b_oid, b_mode,
1586                                              a_oid, a_mode);
1587         }
1588
1589         /*
1590          * In the recursive case, we just opt to undo renames
1591          */
1592         if (o->call_depth && (prev_path1 || prev_path2)) {
1593                 /* Put first file (a_oid, a_mode) in its original spot */
1594                 if (prev_path1) {
1595                         if (update_file(o, 1, a_oid, a_mode, prev_path1))
1596                                 return -1;
1597                 } else {
1598                         if (update_file(o, 1, a_oid, a_mode, collide_path))
1599                                 return -1;
1600                 }
1601
1602                 /* Put second file (b_oid, b_mode) in its original spot */
1603                 if (prev_path2) {
1604                         if (update_file(o, 1, b_oid, b_mode, prev_path2))
1605                                 return -1;
1606                 } else {
1607                         if (update_file(o, 1, b_oid, b_mode, collide_path))
1608                                 return -1;
1609                 }
1610
1611                 /* Don't leave something at collision path if unrenaming both */
1612                 if (prev_path1 && prev_path2)
1613                         remove_file(o, 1, collide_path, 0);
1614
1615                 return 0;
1616         }
1617
1618         /* Remove rename sources if rename/add or rename/rename(2to1) */
1619         if (prev_path1)
1620                 remove_file(o, 1, prev_path1,
1621                             o->call_depth || would_lose_untracked(o, prev_path1));
1622         if (prev_path2)
1623                 remove_file(o, 1, prev_path2,
1624                             o->call_depth || would_lose_untracked(o, prev_path2));
1625
1626         /*
1627          * Remove the collision path, if it wouldn't cause dirty contents
1628          * or an untracked file to get lost.  We'll either overwrite with
1629          * merged contents, or just write out to differently named files.
1630          */
1631         if (was_dirty(o, collide_path)) {
1632                 output(o, 1, _("Refusing to lose dirty file at %s"),
1633                        collide_path);
1634                 update_path = alt_path = unique_path(o, collide_path, "merged");
1635         } else if (would_lose_untracked(o, collide_path)) {
1636                 /*
1637                  * Only way we get here is if both renames were from
1638                  * a directory rename AND user had an untracked file
1639                  * at the location where both files end up after the
1640                  * two directory renames.  See testcase 10d of t6043.
1641                  */
1642                 output(o, 1, _("Refusing to lose untracked file at "
1643                                "%s, even though it's in the way."),
1644                        collide_path);
1645                 update_path = alt_path = unique_path(o, collide_path, "merged");
1646         } else {
1647                 /*
1648                  * FIXME: It's possible that the two files are identical
1649                  * and that the current working copy happens to match, in
1650                  * which case we are unnecessarily touching the working
1651                  * tree file.  It's not a likely enough scenario that I
1652                  * want to code up the checks for it and a better fix is
1653                  * available if we restructure how unpack_trees() and
1654                  * merge-recursive interoperate anyway, so punting for
1655                  * now...
1656                  */
1657                 remove_file(o, 0, collide_path, 0);
1658         }
1659
1660         /* Store things in diff_filespecs for functions that need it */
1661         memset(&a, 0, sizeof(struct diff_filespec));
1662         memset(&b, 0, sizeof(struct diff_filespec));
1663         null.path = a.path = b.path = (char *)collide_path;
1664         oidcpy(&null.oid, &null_oid);
1665         null.mode = 0;
1666         oidcpy(&a.oid, a_oid);
1667         a.mode = a_mode;
1668         a.oid_valid = 1;
1669         oidcpy(&b.oid, b_oid);
1670         b.mode = b_mode;
1671         b.oid_valid = 1;
1672
1673         if (merge_mode_and_contents(o, &null, &a, &b, collide_path,
1674                                     branch1, branch2, o->call_depth * 2, &mfi))
1675                 return -1;
1676         mfi.clean &= !alt_path;
1677         if (update_file(o, mfi.clean, &mfi.oid, mfi.mode, update_path))
1678                 return -1;
1679         if (!mfi.clean && !o->call_depth &&
1680             update_stages(o, collide_path, NULL, &a, &b))
1681                 return -1;
1682         free(alt_path);
1683         /*
1684          * FIXME: If both a & b both started with conflicts (only possible
1685          * if they came from a rename/rename(2to1)), but had IDENTICAL
1686          * contents including those conflicts, then in the next line we claim
1687          * it was clean.  If someone cares about this case, we should have the
1688          * caller notify us if we started with conflicts.
1689          */
1690         return mfi.clean;
1691 }
1692
1693 static int handle_rename_add(struct merge_options *o,
1694                              struct rename_conflict_info *ci)
1695 {
1696         /* a was renamed to c, and a separate c was added. */
1697         struct diff_filespec *a = ci->pair1->one;
1698         struct diff_filespec *c = ci->pair1->two;
1699         char *path = c->path;
1700         char *prev_path_desc;
1701         struct merge_file_info mfi;
1702
1703         int other_stage = (ci->branch1 == o->branch1 ? 3 : 2);
1704
1705         output(o, 1, _("CONFLICT (rename/add): "
1706                "Rename %s->%s in %s.  Added %s in %s"),
1707                a->path, c->path, ci->branch1,
1708                c->path, ci->branch2);
1709
1710         prev_path_desc = xstrfmt("version of %s from %s", path, a->path);
1711         if (merge_mode_and_contents(o, a, c, &ci->ren1_other, prev_path_desc,
1712                                     o->branch1, o->branch2,
1713                                     1 + o->call_depth * 2, &mfi))
1714                 return -1;
1715         free(prev_path_desc);
1716
1717         return handle_file_collision(o,
1718                                      c->path, a->path, NULL,
1719                                      ci->branch1, ci->branch2,
1720                                      &mfi.oid, mfi.mode,
1721                                      &ci->dst_entry1->stages[other_stage].oid,
1722                                      ci->dst_entry1->stages[other_stage].mode);
1723 }
1724
1725 static char *find_path_for_conflict(struct merge_options *o,
1726                                     const char *path,
1727                                     const char *branch1,
1728                                     const char *branch2)
1729 {
1730         char *new_path = NULL;
1731         if (dir_in_way(o->repo->index, path, !o->call_depth, 0)) {
1732                 new_path = unique_path(o, path, branch1);
1733                 output(o, 1, _("%s is a directory in %s adding "
1734                                "as %s instead"),
1735                        path, branch2, new_path);
1736         } else if (would_lose_untracked(o, path)) {
1737                 new_path = unique_path(o, path, branch1);
1738                 output(o, 1, _("Refusing to lose untracked file"
1739                                " at %s; adding as %s instead"),
1740                        path, new_path);
1741         }
1742
1743         return new_path;
1744 }
1745
1746 static int handle_rename_rename_1to2(struct merge_options *o,
1747                                      struct rename_conflict_info *ci)
1748 {
1749         /* One file was renamed in both branches, but to different names. */
1750         struct merge_file_info mfi;
1751         struct diff_filespec other;
1752         struct diff_filespec *add;
1753         struct diff_filespec *one = ci->pair1->one;
1754         struct diff_filespec *a = ci->pair1->two;
1755         struct diff_filespec *b = ci->pair2->two;
1756         char *path_desc;
1757
1758         output(o, 1, _("CONFLICT (rename/rename): "
1759                "Rename \"%s\"->\"%s\" in branch \"%s\" "
1760                "rename \"%s\"->\"%s\" in \"%s\"%s"),
1761                one->path, a->path, ci->branch1,
1762                one->path, b->path, ci->branch2,
1763                o->call_depth ? _(" (left unresolved)") : "");
1764
1765         path_desc = xstrfmt("%s and %s, both renamed from %s",
1766                             a->path, b->path, one->path);
1767         if (merge_mode_and_contents(o, one, a, b, path_desc,
1768                                     ci->branch1, ci->branch2,
1769                                     o->call_depth * 2, &mfi))
1770                 return -1;
1771         free(path_desc);
1772
1773         if (o->call_depth) {
1774                 /*
1775                  * FIXME: For rename/add-source conflicts (if we could detect
1776                  * such), this is wrong.  We should instead find a unique
1777                  * pathname and then either rename the add-source file to that
1778                  * unique path, or use that unique path instead of src here.
1779                  */
1780                 if (update_file(o, 0, &mfi.oid, mfi.mode, one->path))
1781                         return -1;
1782
1783                 /*
1784                  * Above, we put the merged content at the merge-base's
1785                  * path.  Now we usually need to delete both a->path and
1786                  * b->path.  However, the rename on each side of the merge
1787                  * could also be involved in a rename/add conflict.  In
1788                  * such cases, we should keep the added file around,
1789                  * resolving the conflict at that path in its favor.
1790                  */
1791                 add = filespec_from_entry(&other, ci->dst_entry1, 2 ^ 1);
1792                 if (add) {
1793                         if (update_file(o, 0, &add->oid, add->mode, a->path))
1794                                 return -1;
1795                 }
1796                 else
1797                         remove_file_from_index(o->repo->index, a->path);
1798                 add = filespec_from_entry(&other, ci->dst_entry2, 3 ^ 1);
1799                 if (add) {
1800                         if (update_file(o, 0, &add->oid, add->mode, b->path))
1801                                 return -1;
1802                 }
1803                 else
1804                         remove_file_from_index(o->repo->index, b->path);
1805         } else {
1806                 /*
1807                  * For each destination path, we need to see if there is a
1808                  * rename/add collision.  If not, we can write the file out
1809                  * to the specified location.
1810                  */
1811                 add = filespec_from_entry(&other, ci->dst_entry1, 2 ^ 1);
1812                 if (add) {
1813                         if (handle_file_collision(o, a->path,
1814                                                   NULL, NULL,
1815                                                   ci->branch1, ci->branch2,
1816                                                   &mfi.oid, mfi.mode,
1817                                                   &add->oid, add->mode) < 0)
1818                                 return -1;
1819                 } else {
1820                         char *new_path = find_path_for_conflict(o, a->path,
1821                                                                 ci->branch1,
1822                                                                 ci->branch2);
1823                         if (update_file(o, 0, &mfi.oid, mfi.mode, new_path ? new_path : a->path))
1824                                 return -1;
1825                         free(new_path);
1826                         if (update_stages(o, a->path, NULL, a, NULL))
1827                                 return -1;
1828                 }
1829
1830                 add = filespec_from_entry(&other, ci->dst_entry2, 3 ^ 1);
1831                 if (add) {
1832                         if (handle_file_collision(o, b->path,
1833                                                   NULL, NULL,
1834                                                   ci->branch1, ci->branch2,
1835                                                   &add->oid, add->mode,
1836                                                   &mfi.oid, mfi.mode) < 0)
1837                                 return -1;
1838                 } else {
1839                         char *new_path = find_path_for_conflict(o, b->path,
1840                                                                 ci->branch2,
1841                                                                 ci->branch1);
1842                         if (update_file(o, 0, &mfi.oid, mfi.mode, new_path ? new_path : b->path))
1843                                 return -1;
1844                         free(new_path);
1845                         if (update_stages(o, b->path, NULL, NULL, b))
1846                                 return -1;
1847                 }
1848         }
1849
1850         return 0;
1851 }
1852
1853 static int handle_rename_rename_2to1(struct merge_options *o,
1854                                      struct rename_conflict_info *ci)
1855 {
1856         /* Two files, a & b, were renamed to the same thing, c. */
1857         struct diff_filespec *a = ci->pair1->one;
1858         struct diff_filespec *b = ci->pair2->one;
1859         struct diff_filespec *c1 = ci->pair1->two;
1860         struct diff_filespec *c2 = ci->pair2->two;
1861         char *path = c1->path; /* == c2->path */
1862         char *path_side_1_desc;
1863         char *path_side_2_desc;
1864         struct merge_file_info mfi_c1;
1865         struct merge_file_info mfi_c2;
1866
1867         output(o, 1, _("CONFLICT (rename/rename): "
1868                "Rename %s->%s in %s. "
1869                "Rename %s->%s in %s"),
1870                a->path, c1->path, ci->branch1,
1871                b->path, c2->path, ci->branch2);
1872
1873         path_side_1_desc = xstrfmt("version of %s from %s", path, a->path);
1874         path_side_2_desc = xstrfmt("version of %s from %s", path, b->path);
1875         if (merge_mode_and_contents(o, a, c1, &ci->ren1_other, path_side_1_desc,
1876                                     o->branch1, o->branch2,
1877                                     1 + o->call_depth * 2, &mfi_c1) ||
1878             merge_mode_and_contents(o, b, &ci->ren2_other, c2, path_side_2_desc,
1879                                     o->branch1, o->branch2,
1880                                     1 + o->call_depth * 2, &mfi_c2))
1881                 return -1;
1882         free(path_side_1_desc);
1883         free(path_side_2_desc);
1884
1885         return handle_file_collision(o, path, a->path, b->path,
1886                                      ci->branch1, ci->branch2,
1887                                      &mfi_c1.oid, mfi_c1.mode,
1888                                      &mfi_c2.oid, mfi_c2.mode);
1889 }
1890
1891 /*
1892  * Get the diff_filepairs changed between o_tree and tree.
1893  */
1894 static struct diff_queue_struct *get_diffpairs(struct merge_options *o,
1895                                                struct tree *o_tree,
1896                                                struct tree *tree)
1897 {
1898         struct diff_queue_struct *ret;
1899         struct diff_options opts;
1900
1901         repo_diff_setup(o->repo, &opts);
1902         opts.flags.recursive = 1;
1903         opts.flags.rename_empty = 0;
1904         opts.detect_rename = merge_detect_rename(o);
1905         /*
1906          * We do not have logic to handle the detection of copies.  In
1907          * fact, it may not even make sense to add such logic: would we
1908          * really want a change to a base file to be propagated through
1909          * multiple other files by a merge?
1910          */
1911         if (opts.detect_rename > DIFF_DETECT_RENAME)
1912                 opts.detect_rename = DIFF_DETECT_RENAME;
1913         opts.rename_limit = o->merge_rename_limit >= 0 ? o->merge_rename_limit :
1914                             o->diff_rename_limit >= 0 ? o->diff_rename_limit :
1915                             1000;
1916         opts.rename_score = o->rename_score;
1917         opts.show_rename_progress = o->show_rename_progress;
1918         opts.output_format = DIFF_FORMAT_NO_OUTPUT;
1919         diff_setup_done(&opts);
1920         diff_tree_oid(&o_tree->object.oid, &tree->object.oid, "", &opts);
1921         diffcore_std(&opts);
1922         if (opts.needed_rename_limit > o->needed_rename_limit)
1923                 o->needed_rename_limit = opts.needed_rename_limit;
1924
1925         ret = xmalloc(sizeof(*ret));
1926         *ret = diff_queued_diff;
1927
1928         opts.output_format = DIFF_FORMAT_NO_OUTPUT;
1929         diff_queued_diff.nr = 0;
1930         diff_queued_diff.queue = NULL;
1931         diff_flush(&opts);
1932         return ret;
1933 }
1934
1935 static int tree_has_path(struct tree *tree, const char *path)
1936 {
1937         struct object_id hashy;
1938         unsigned int mode_o;
1939
1940         return !get_tree_entry(&tree->object.oid, path,
1941                                &hashy, &mode_o);
1942 }
1943
1944 /*
1945  * Return a new string that replaces the beginning portion (which matches
1946  * entry->dir), with entry->new_dir.  In perl-speak:
1947  *   new_path_name = (old_path =~ s/entry->dir/entry->new_dir/);
1948  * NOTE:
1949  *   Caller must ensure that old_path starts with entry->dir + '/'.
1950  */
1951 static char *apply_dir_rename(struct dir_rename_entry *entry,
1952                               const char *old_path)
1953 {
1954         struct strbuf new_path = STRBUF_INIT;
1955         int oldlen, newlen;
1956
1957         if (entry->non_unique_new_dir)
1958                 return NULL;
1959
1960         oldlen = strlen(entry->dir);
1961         newlen = entry->new_dir.len + (strlen(old_path) - oldlen) + 1;
1962         strbuf_grow(&new_path, newlen);
1963         strbuf_addbuf(&new_path, &entry->new_dir);
1964         strbuf_addstr(&new_path, &old_path[oldlen]);
1965
1966         return strbuf_detach(&new_path, NULL);
1967 }
1968
1969 static void get_renamed_dir_portion(const char *old_path, const char *new_path,
1970                                     char **old_dir, char **new_dir)
1971 {
1972         char *end_of_old, *end_of_new;
1973         int old_len, new_len;
1974
1975         *old_dir = NULL;
1976         *new_dir = NULL;
1977
1978         /*
1979          * For
1980          *    "a/b/c/d/e/foo.c" -> "a/b/some/thing/else/e/foo.c"
1981          * the "e/foo.c" part is the same, we just want to know that
1982          *    "a/b/c/d" was renamed to "a/b/some/thing/else"
1983          * so, for this example, this function returns "a/b/c/d" in
1984          * *old_dir and "a/b/some/thing/else" in *new_dir.
1985          *
1986          * Also, if the basename of the file changed, we don't care.  We
1987          * want to know which portion of the directory, if any, changed.
1988          */
1989         end_of_old = strrchr(old_path, '/');
1990         end_of_new = strrchr(new_path, '/');
1991
1992         if (end_of_old == NULL || end_of_new == NULL)
1993                 return;
1994         while (*--end_of_new == *--end_of_old &&
1995                end_of_old != old_path &&
1996                end_of_new != new_path)
1997                 ; /* Do nothing; all in the while loop */
1998         /*
1999          * We've found the first non-matching character in the directory
2000          * paths.  That means the current directory we were comparing
2001          * represents the rename.  Move end_of_old and end_of_new back
2002          * to the full directory name.
2003          */
2004         if (*end_of_old == '/')
2005                 end_of_old++;
2006         if (*end_of_old != '/')
2007                 end_of_new++;
2008         end_of_old = strchr(end_of_old, '/');
2009         end_of_new = strchr(end_of_new, '/');
2010
2011         /*
2012          * It may have been the case that old_path and new_path were the same
2013          * directory all along.  Don't claim a rename if they're the same.
2014          */
2015         old_len = end_of_old - old_path;
2016         new_len = end_of_new - new_path;
2017
2018         if (old_len != new_len || strncmp(old_path, new_path, old_len)) {
2019                 *old_dir = xstrndup(old_path, old_len);
2020                 *new_dir = xstrndup(new_path, new_len);
2021         }
2022 }
2023
2024 static void remove_hashmap_entries(struct hashmap *dir_renames,
2025                                    struct string_list *items_to_remove)
2026 {
2027         int i;
2028         struct dir_rename_entry *entry;
2029
2030         for (i = 0; i < items_to_remove->nr; i++) {
2031                 entry = items_to_remove->items[i].util;
2032                 hashmap_remove(dir_renames, entry, NULL);
2033         }
2034         string_list_clear(items_to_remove, 0);
2035 }
2036
2037 /*
2038  * See if there is a directory rename for path, and if there are any file
2039  * level conflicts for the renamed location.  If there is a rename and
2040  * there are no conflicts, return the new name.  Otherwise, return NULL.
2041  */
2042 static char *handle_path_level_conflicts(struct merge_options *o,
2043                                          const char *path,
2044                                          struct dir_rename_entry *entry,
2045                                          struct hashmap *collisions,
2046                                          struct tree *tree)
2047 {
2048         char *new_path = NULL;
2049         struct collision_entry *collision_ent;
2050         int clean = 1;
2051         struct strbuf collision_paths = STRBUF_INIT;
2052
2053         /*
2054          * entry has the mapping of old directory name to new directory name
2055          * that we want to apply to path.
2056          */
2057         new_path = apply_dir_rename(entry, path);
2058
2059         if (!new_path) {
2060                 /* This should only happen when entry->non_unique_new_dir set */
2061                 if (!entry->non_unique_new_dir)
2062                         BUG("entry->non_unqiue_dir not set and !new_path");
2063                 output(o, 1, _("CONFLICT (directory rename split): "
2064                                "Unclear where to place %s because directory "
2065                                "%s was renamed to multiple other directories, "
2066                                "with no destination getting a majority of the "
2067                                "files."),
2068                        path, entry->dir);
2069                 clean = 0;
2070                 return NULL;
2071         }
2072
2073         /*
2074          * The caller needs to have ensured that it has pre-populated
2075          * collisions with all paths that map to new_path.  Do a quick check
2076          * to ensure that's the case.
2077          */
2078         collision_ent = collision_find_entry(collisions, new_path);
2079         if (collision_ent == NULL)
2080                 BUG("collision_ent is NULL");
2081
2082         /*
2083          * Check for one-sided add/add/.../add conflicts, i.e.
2084          * where implicit renames from the other side doing
2085          * directory rename(s) can affect this side of history
2086          * to put multiple paths into the same location.  Warn
2087          * and bail on directory renames for such paths.
2088          */
2089         if (collision_ent->reported_already) {
2090                 clean = 0;
2091         } else if (tree_has_path(tree, new_path)) {
2092                 collision_ent->reported_already = 1;
2093                 strbuf_add_separated_string_list(&collision_paths, ", ",
2094                                                  &collision_ent->source_files);
2095                 output(o, 1, _("CONFLICT (implicit dir rename): Existing "
2096                                "file/dir at %s in the way of implicit "
2097                                "directory rename(s) putting the following "
2098                                "path(s) there: %s."),
2099                        new_path, collision_paths.buf);
2100                 clean = 0;
2101         } else if (collision_ent->source_files.nr > 1) {
2102                 collision_ent->reported_already = 1;
2103                 strbuf_add_separated_string_list(&collision_paths, ", ",
2104                                                  &collision_ent->source_files);
2105                 output(o, 1, _("CONFLICT (implicit dir rename): Cannot map "
2106                                "more than one path to %s; implicit directory "
2107                                "renames tried to put these paths there: %s"),
2108                        new_path, collision_paths.buf);
2109                 clean = 0;
2110         }
2111
2112         /* Free memory we no longer need */
2113         strbuf_release(&collision_paths);
2114         if (!clean && new_path) {
2115                 free(new_path);
2116                 return NULL;
2117         }
2118
2119         return new_path;
2120 }
2121
2122 /*
2123  * There are a couple things we want to do at the directory level:
2124  *   1. Check for both sides renaming to the same thing, in order to avoid
2125  *      implicit renaming of files that should be left in place.  (See
2126  *      testcase 6b in t6043 for details.)
2127  *   2. Prune directory renames if there are still files left in the
2128  *      the original directory.  These represent a partial directory rename,
2129  *      i.e. a rename where only some of the files within the directory
2130  *      were renamed elsewhere.  (Technically, this could be done earlier
2131  *      in get_directory_renames(), except that would prevent us from
2132  *      doing the previous check and thus failing testcase 6b.)
2133  *   3. Check for rename/rename(1to2) conflicts (at the directory level).
2134  *      In the future, we could potentially record this info as well and
2135  *      omit reporting rename/rename(1to2) conflicts for each path within
2136  *      the affected directories, thus cleaning up the merge output.
2137  *   NOTE: We do NOT check for rename/rename(2to1) conflicts at the
2138  *         directory level, because merging directories is fine.  If it
2139  *         causes conflicts for files within those merged directories, then
2140  *         that should be detected at the individual path level.
2141  */
2142 static void handle_directory_level_conflicts(struct merge_options *o,
2143                                              struct hashmap *dir_re_head,
2144                                              struct tree *head,
2145                                              struct hashmap *dir_re_merge,
2146                                              struct tree *merge)
2147 {
2148         struct hashmap_iter iter;
2149         struct dir_rename_entry *head_ent;
2150         struct dir_rename_entry *merge_ent;
2151
2152         struct string_list remove_from_head = STRING_LIST_INIT_NODUP;
2153         struct string_list remove_from_merge = STRING_LIST_INIT_NODUP;
2154
2155         hashmap_iter_init(dir_re_head, &iter);
2156         while ((head_ent = hashmap_iter_next(&iter))) {
2157                 merge_ent = dir_rename_find_entry(dir_re_merge, head_ent->dir);
2158                 if (merge_ent &&
2159                     !head_ent->non_unique_new_dir &&
2160                     !merge_ent->non_unique_new_dir &&
2161                     !strbuf_cmp(&head_ent->new_dir, &merge_ent->new_dir)) {
2162                         /* 1. Renamed identically; remove it from both sides */
2163                         string_list_append(&remove_from_head,
2164                                            head_ent->dir)->util = head_ent;
2165                         strbuf_release(&head_ent->new_dir);
2166                         string_list_append(&remove_from_merge,
2167                                            merge_ent->dir)->util = merge_ent;
2168                         strbuf_release(&merge_ent->new_dir);
2169                 } else if (tree_has_path(head, head_ent->dir)) {
2170                         /* 2. This wasn't a directory rename after all */
2171                         string_list_append(&remove_from_head,
2172                                            head_ent->dir)->util = head_ent;
2173                         strbuf_release(&head_ent->new_dir);
2174                 }
2175         }
2176
2177         remove_hashmap_entries(dir_re_head, &remove_from_head);
2178         remove_hashmap_entries(dir_re_merge, &remove_from_merge);
2179
2180         hashmap_iter_init(dir_re_merge, &iter);
2181         while ((merge_ent = hashmap_iter_next(&iter))) {
2182                 head_ent = dir_rename_find_entry(dir_re_head, merge_ent->dir);
2183                 if (tree_has_path(merge, merge_ent->dir)) {
2184                         /* 2. This wasn't a directory rename after all */
2185                         string_list_append(&remove_from_merge,
2186                                            merge_ent->dir)->util = merge_ent;
2187                 } else if (head_ent &&
2188                            !head_ent->non_unique_new_dir &&
2189                            !merge_ent->non_unique_new_dir) {
2190                         /* 3. rename/rename(1to2) */
2191                         /*
2192                          * We can assume it's not rename/rename(1to1) because
2193                          * that was case (1), already checked above.  So we
2194                          * know that head_ent->new_dir and merge_ent->new_dir
2195                          * are different strings.
2196                          */
2197                         output(o, 1, _("CONFLICT (rename/rename): "
2198                                        "Rename directory %s->%s in %s. "
2199                                        "Rename directory %s->%s in %s"),
2200                                head_ent->dir, head_ent->new_dir.buf, o->branch1,
2201                                head_ent->dir, merge_ent->new_dir.buf, o->branch2);
2202                         string_list_append(&remove_from_head,
2203                                            head_ent->dir)->util = head_ent;
2204                         strbuf_release(&head_ent->new_dir);
2205                         string_list_append(&remove_from_merge,
2206                                            merge_ent->dir)->util = merge_ent;
2207                         strbuf_release(&merge_ent->new_dir);
2208                 }
2209         }
2210
2211         remove_hashmap_entries(dir_re_head, &remove_from_head);
2212         remove_hashmap_entries(dir_re_merge, &remove_from_merge);
2213 }
2214
2215 static struct hashmap *get_directory_renames(struct diff_queue_struct *pairs,
2216                                              struct tree *tree)
2217 {
2218         struct hashmap *dir_renames;
2219         struct hashmap_iter iter;
2220         struct dir_rename_entry *entry;
2221         int i;
2222
2223         /*
2224          * Typically, we think of a directory rename as all files from a
2225          * certain directory being moved to a target directory.  However,
2226          * what if someone first moved two files from the original
2227          * directory in one commit, and then renamed the directory
2228          * somewhere else in a later commit?  At merge time, we just know
2229          * that files from the original directory went to two different
2230          * places, and that the bulk of them ended up in the same place.
2231          * We want each directory rename to represent where the bulk of the
2232          * files from that directory end up; this function exists to find
2233          * where the bulk of the files went.
2234          *
2235          * The first loop below simply iterates through the list of file
2236          * renames, finding out how often each directory rename pair
2237          * possibility occurs.
2238          */
2239         dir_renames = xmalloc(sizeof(*dir_renames));
2240         dir_rename_init(dir_renames);
2241         for (i = 0; i < pairs->nr; ++i) {
2242                 struct string_list_item *item;
2243                 int *count;
2244                 struct diff_filepair *pair = pairs->queue[i];
2245                 char *old_dir, *new_dir;
2246
2247                 /* File not part of directory rename if it wasn't renamed */
2248                 if (pair->status != 'R')
2249                         continue;
2250
2251                 get_renamed_dir_portion(pair->one->path, pair->two->path,
2252                                         &old_dir,        &new_dir);
2253                 if (!old_dir)
2254                         /* Directory didn't change at all; ignore this one. */
2255                         continue;
2256
2257                 entry = dir_rename_find_entry(dir_renames, old_dir);
2258                 if (!entry) {
2259                         entry = xmalloc(sizeof(*entry));
2260                         dir_rename_entry_init(entry, old_dir);
2261                         hashmap_put(dir_renames, entry);
2262                 } else {
2263                         free(old_dir);
2264                 }
2265                 item = string_list_lookup(&entry->possible_new_dirs, new_dir);
2266                 if (!item) {
2267                         item = string_list_insert(&entry->possible_new_dirs,
2268                                                   new_dir);
2269                         item->util = xcalloc(1, sizeof(int));
2270                 } else {
2271                         free(new_dir);
2272                 }
2273                 count = item->util;
2274                 *count += 1;
2275         }
2276
2277         /*
2278          * For each directory with files moved out of it, we find out which
2279          * target directory received the most files so we can declare it to
2280          * be the "winning" target location for the directory rename.  This
2281          * winner gets recorded in new_dir.  If there is no winner
2282          * (multiple target directories received the same number of files),
2283          * we set non_unique_new_dir.  Once we've determined the winner (or
2284          * that there is no winner), we no longer need possible_new_dirs.
2285          */
2286         hashmap_iter_init(dir_renames, &iter);
2287         while ((entry = hashmap_iter_next(&iter))) {
2288                 int max = 0;
2289                 int bad_max = 0;
2290                 char *best = NULL;
2291
2292                 for (i = 0; i < entry->possible_new_dirs.nr; i++) {
2293                         int *count = entry->possible_new_dirs.items[i].util;
2294
2295                         if (*count == max)
2296                                 bad_max = max;
2297                         else if (*count > max) {
2298                                 max = *count;
2299                                 best = entry->possible_new_dirs.items[i].string;
2300                         }
2301                 }
2302                 if (bad_max == max)
2303                         entry->non_unique_new_dir = 1;
2304                 else {
2305                         assert(entry->new_dir.len == 0);
2306                         strbuf_addstr(&entry->new_dir, best);
2307                 }
2308                 /*
2309                  * The relevant directory sub-portion of the original full
2310                  * filepaths were xstrndup'ed before inserting into
2311                  * possible_new_dirs, and instead of manually iterating the
2312                  * list and free'ing each, just lie and tell
2313                  * possible_new_dirs that it did the strdup'ing so that it
2314                  * will free them for us.
2315                  */
2316                 entry->possible_new_dirs.strdup_strings = 1;
2317                 string_list_clear(&entry->possible_new_dirs, 1);
2318         }
2319
2320         return dir_renames;
2321 }
2322
2323 static struct dir_rename_entry *check_dir_renamed(const char *path,
2324                                                   struct hashmap *dir_renames)
2325 {
2326         char *temp = xstrdup(path);
2327         char *end;
2328         struct dir_rename_entry *entry = NULL;
2329
2330         while ((end = strrchr(temp, '/'))) {
2331                 *end = '\0';
2332                 entry = dir_rename_find_entry(dir_renames, temp);
2333                 if (entry)
2334                         break;
2335         }
2336         free(temp);
2337         return entry;
2338 }
2339
2340 static void compute_collisions(struct hashmap *collisions,
2341                                struct hashmap *dir_renames,
2342                                struct diff_queue_struct *pairs)
2343 {
2344         int i;
2345
2346         /*
2347          * Multiple files can be mapped to the same path due to directory
2348          * renames done by the other side of history.  Since that other
2349          * side of history could have merged multiple directories into one,
2350          * if our side of history added the same file basename to each of
2351          * those directories, then all N of them would get implicitly
2352          * renamed by the directory rename detection into the same path,
2353          * and we'd get an add/add/.../add conflict, and all those adds
2354          * from *this* side of history.  This is not representable in the
2355          * index, and users aren't going to easily be able to make sense of
2356          * it.  So we need to provide a good warning about what's
2357          * happening, and fall back to no-directory-rename detection
2358          * behavior for those paths.
2359          *
2360          * See testcases 9e and all of section 5 from t6043 for examples.
2361          */
2362         collision_init(collisions);
2363
2364         for (i = 0; i < pairs->nr; ++i) {
2365                 struct dir_rename_entry *dir_rename_ent;
2366                 struct collision_entry *collision_ent;
2367                 char *new_path;
2368                 struct diff_filepair *pair = pairs->queue[i];
2369
2370                 if (pair->status != 'A' && pair->status != 'R')
2371                         continue;
2372                 dir_rename_ent = check_dir_renamed(pair->two->path,
2373                                                    dir_renames);
2374                 if (!dir_rename_ent)
2375                         continue;
2376
2377                 new_path = apply_dir_rename(dir_rename_ent, pair->two->path);
2378                 if (!new_path)
2379                         /*
2380                          * dir_rename_ent->non_unique_new_path is true, which
2381                          * means there is no directory rename for us to use,
2382                          * which means it won't cause us any additional
2383                          * collisions.
2384                          */
2385                         continue;
2386                 collision_ent = collision_find_entry(collisions, new_path);
2387                 if (!collision_ent) {
2388                         collision_ent = xcalloc(1,
2389                                                 sizeof(struct collision_entry));
2390                         hashmap_entry_init(collision_ent, strhash(new_path));
2391                         hashmap_put(collisions, collision_ent);
2392                         collision_ent->target_file = new_path;
2393                 } else {
2394                         free(new_path);
2395                 }
2396                 string_list_insert(&collision_ent->source_files,
2397                                    pair->two->path);
2398         }
2399 }
2400
2401 static char *check_for_directory_rename(struct merge_options *o,
2402                                         const char *path,
2403                                         struct tree *tree,
2404                                         struct hashmap *dir_renames,
2405                                         struct hashmap *dir_rename_exclusions,
2406                                         struct hashmap *collisions,
2407                                         int *clean_merge)
2408 {
2409         char *new_path = NULL;
2410         struct dir_rename_entry *entry = check_dir_renamed(path, dir_renames);
2411         struct dir_rename_entry *oentry = NULL;
2412
2413         if (!entry)
2414                 return new_path;
2415
2416         /*
2417          * This next part is a little weird.  We do not want to do an
2418          * implicit rename into a directory we renamed on our side, because
2419          * that will result in a spurious rename/rename(1to2) conflict.  An
2420          * example:
2421          *   Base commit: dumbdir/afile, otherdir/bfile
2422          *   Side 1:      smrtdir/afile, otherdir/bfile
2423          *   Side 2:      dumbdir/afile, dumbdir/bfile
2424          * Here, while working on Side 1, we could notice that otherdir was
2425          * renamed/merged to dumbdir, and change the diff_filepair for
2426          * otherdir/bfile into a rename into dumbdir/bfile.  However, Side
2427          * 2 will notice the rename from dumbdir to smrtdir, and do the
2428          * transitive rename to move it from dumbdir/bfile to
2429          * smrtdir/bfile.  That gives us bfile in dumbdir vs being in
2430          * smrtdir, a rename/rename(1to2) conflict.  We really just want
2431          * the file to end up in smrtdir.  And the way to achieve that is
2432          * to not let Side1 do the rename to dumbdir, since we know that is
2433          * the source of one of our directory renames.
2434          *
2435          * That's why oentry and dir_rename_exclusions is here.
2436          *
2437          * As it turns out, this also prevents N-way transient rename
2438          * confusion; See testcases 9c and 9d of t6043.
2439          */
2440         oentry = dir_rename_find_entry(dir_rename_exclusions, entry->new_dir.buf);
2441         if (oentry) {
2442                 output(o, 1, _("WARNING: Avoiding applying %s -> %s rename "
2443                                "to %s, because %s itself was renamed."),
2444                        entry->dir, entry->new_dir.buf, path, entry->new_dir.buf);
2445         } else {
2446                 new_path = handle_path_level_conflicts(o, path, entry,
2447                                                        collisions, tree);
2448                 *clean_merge &= (new_path != NULL);
2449         }
2450
2451         return new_path;
2452 }
2453
2454 static void apply_directory_rename_modifications(struct merge_options *o,
2455                                                  struct diff_filepair *pair,
2456                                                  char *new_path,
2457                                                  struct rename *re,
2458                                                  struct tree *tree,
2459                                                  struct tree *o_tree,
2460                                                  struct tree *a_tree,
2461                                                  struct tree *b_tree,
2462                                                  struct string_list *entries,
2463                                                  int *clean)
2464 {
2465         struct string_list_item *item;
2466         int stage = (tree == a_tree ? 2 : 3);
2467         int update_wd;
2468
2469         /*
2470          * In all cases where we can do directory rename detection,
2471          * unpack_trees() will have read pair->two->path into the
2472          * index and the working copy.  We need to remove it so that
2473          * we can instead place it at new_path.  It is guaranteed to
2474          * not be untracked (unpack_trees() would have errored out
2475          * saying the file would have been overwritten), but it might
2476          * be dirty, though.
2477          */
2478         update_wd = !was_dirty(o, pair->two->path);
2479         if (!update_wd)
2480                 output(o, 1, _("Refusing to lose dirty file at %s"),
2481                        pair->two->path);
2482         remove_file(o, 1, pair->two->path, !update_wd);
2483
2484         /* Find or create a new re->dst_entry */
2485         item = string_list_lookup(entries, new_path);
2486         if (item) {
2487                 /*
2488                  * Since we're renaming on this side of history, and it's
2489                  * due to a directory rename on the other side of history
2490                  * (which we only allow when the directory in question no
2491                  * longer exists on the other side of history), the
2492                  * original entry for re->dst_entry is no longer
2493                  * necessary...
2494                  */
2495                 re->dst_entry->processed = 1;
2496
2497                 /*
2498                  * ...because we'll be using this new one.
2499                  */
2500                 re->dst_entry = item->util;
2501         } else {
2502                 /*
2503                  * re->dst_entry is for the before-dir-rename path, and we
2504                  * need it to hold information for the after-dir-rename
2505                  * path.  Before creating a new entry, we need to mark the
2506                  * old one as unnecessary (...unless it is shared by
2507                  * src_entry, i.e. this didn't use to be a rename, in which
2508                  * case we can just allow the normal processing to happen
2509                  * for it).
2510                  */
2511                 if (pair->status == 'R')
2512                         re->dst_entry->processed = 1;
2513
2514                 re->dst_entry = insert_stage_data(new_path,
2515                                                   o_tree, a_tree, b_tree,
2516                                                   entries);
2517                 item = string_list_insert(entries, new_path);
2518                 item->util = re->dst_entry;
2519         }
2520
2521         /*
2522          * Update the stage_data with the information about the path we are
2523          * moving into place.  That slot will be empty and available for us
2524          * to write to because of the collision checks in
2525          * handle_path_level_conflicts().  In other words,
2526          * re->dst_entry->stages[stage].oid will be the null_oid, so it's
2527          * open for us to write to.
2528          *
2529          * It may be tempting to actually update the index at this point as
2530          * well, using update_stages_for_stage_data(), but as per the big
2531          * "NOTE" in update_stages(), doing so will modify the current
2532          * in-memory index which will break calls to would_lose_untracked()
2533          * that we need to make.  Instead, we need to just make sure that
2534          * the various handle_rename_*() functions update the index
2535          * explicitly rather than relying on unpack_trees() to have done it.
2536          */
2537         get_tree_entry(&tree->object.oid,
2538                        pair->two->path,
2539                        &re->dst_entry->stages[stage].oid,
2540                        &re->dst_entry->stages[stage].mode);
2541
2542         /* Update pair status */
2543         if (pair->status == 'A') {
2544                 /*
2545                  * Recording rename information for this add makes it look
2546                  * like a rename/delete conflict.  Make sure we can
2547                  * correctly handle this as an add that was moved to a new
2548                  * directory instead of reporting a rename/delete conflict.
2549                  */
2550                 re->add_turned_into_rename = 1;
2551         }
2552         /*
2553          * We don't actually look at pair->status again, but it seems
2554          * pedagogically correct to adjust it.
2555          */
2556         pair->status = 'R';
2557
2558         /*
2559          * Finally, record the new location.
2560          */
2561         pair->two->path = new_path;
2562 }
2563
2564 /*
2565  * Get information of all renames which occurred in 'pairs', making use of
2566  * any implicit directory renames inferred from the other side of history.
2567  * We need the three trees in the merge ('o_tree', 'a_tree' and 'b_tree')
2568  * to be able to associate the correct cache entries with the rename
2569  * information; tree is always equal to either a_tree or b_tree.
2570  */
2571 static struct string_list *get_renames(struct merge_options *o,
2572                                        struct diff_queue_struct *pairs,
2573                                        struct hashmap *dir_renames,
2574                                        struct hashmap *dir_rename_exclusions,
2575                                        struct tree *tree,
2576                                        struct tree *o_tree,
2577                                        struct tree *a_tree,
2578                                        struct tree *b_tree,
2579                                        struct string_list *entries,
2580                                        int *clean_merge)
2581 {
2582         int i;
2583         struct hashmap collisions;
2584         struct hashmap_iter iter;
2585         struct collision_entry *e;
2586         struct string_list *renames;
2587
2588         compute_collisions(&collisions, dir_renames, pairs);
2589         renames = xcalloc(1, sizeof(struct string_list));
2590
2591         for (i = 0; i < pairs->nr; ++i) {
2592                 struct string_list_item *item;
2593                 struct rename *re;
2594                 struct diff_filepair *pair = pairs->queue[i];
2595                 char *new_path; /* non-NULL only with directory renames */
2596
2597                 if (pair->status != 'A' && pair->status != 'R') {
2598                         diff_free_filepair(pair);
2599                         continue;
2600                 }
2601                 new_path = check_for_directory_rename(o, pair->two->path, tree,
2602                                                       dir_renames,
2603                                                       dir_rename_exclusions,
2604                                                       &collisions,
2605                                                       clean_merge);
2606                 if (pair->status != 'R' && !new_path) {
2607                         diff_free_filepair(pair);
2608                         continue;
2609                 }
2610
2611                 re = xmalloc(sizeof(*re));
2612                 re->processed = 0;
2613                 re->add_turned_into_rename = 0;
2614                 re->pair = pair;
2615                 item = string_list_lookup(entries, re->pair->one->path);
2616                 if (!item)
2617                         re->src_entry = insert_stage_data(re->pair->one->path,
2618                                         o_tree, a_tree, b_tree, entries);
2619                 else
2620                         re->src_entry = item->util;
2621
2622                 item = string_list_lookup(entries, re->pair->two->path);
2623                 if (!item)
2624                         re->dst_entry = insert_stage_data(re->pair->two->path,
2625                                         o_tree, a_tree, b_tree, entries);
2626                 else
2627                         re->dst_entry = item->util;
2628                 item = string_list_insert(renames, pair->one->path);
2629                 item->util = re;
2630                 if (new_path)
2631                         apply_directory_rename_modifications(o, pair, new_path,
2632                                                              re, tree, o_tree,
2633                                                              a_tree, b_tree,
2634                                                              entries,
2635                                                              clean_merge);
2636         }
2637
2638         hashmap_iter_init(&collisions, &iter);
2639         while ((e = hashmap_iter_next(&iter))) {
2640                 free(e->target_file);
2641                 string_list_clear(&e->source_files, 0);
2642         }
2643         hashmap_free(&collisions, 1);
2644         return renames;
2645 }
2646
2647 static int process_renames(struct merge_options *o,
2648                            struct string_list *a_renames,
2649                            struct string_list *b_renames)
2650 {
2651         int clean_merge = 1, i, j;
2652         struct string_list a_by_dst = STRING_LIST_INIT_NODUP;
2653         struct string_list b_by_dst = STRING_LIST_INIT_NODUP;
2654         const struct rename *sre;
2655
2656         for (i = 0; i < a_renames->nr; i++) {
2657                 sre = a_renames->items[i].util;
2658                 string_list_insert(&a_by_dst, sre->pair->two->path)->util
2659                         = (void *)sre;
2660         }
2661         for (i = 0; i < b_renames->nr; i++) {
2662                 sre = b_renames->items[i].util;
2663                 string_list_insert(&b_by_dst, sre->pair->two->path)->util
2664                         = (void *)sre;
2665         }
2666
2667         for (i = 0, j = 0; i < a_renames->nr || j < b_renames->nr;) {
2668                 struct string_list *renames1, *renames2Dst;
2669                 struct rename *ren1 = NULL, *ren2 = NULL;
2670                 const char *branch1, *branch2;
2671                 const char *ren1_src, *ren1_dst;
2672                 struct string_list_item *lookup;
2673
2674                 if (i >= a_renames->nr) {
2675                         ren2 = b_renames->items[j++].util;
2676                 } else if (j >= b_renames->nr) {
2677                         ren1 = a_renames->items[i++].util;
2678                 } else {
2679                         int compare = strcmp(a_renames->items[i].string,
2680                                              b_renames->items[j].string);
2681                         if (compare <= 0)
2682                                 ren1 = a_renames->items[i++].util;
2683                         if (compare >= 0)
2684                                 ren2 = b_renames->items[j++].util;
2685                 }
2686
2687                 /* TODO: refactor, so that 1/2 are not needed */
2688                 if (ren1) {
2689                         renames1 = a_renames;
2690                         renames2Dst = &b_by_dst;
2691                         branch1 = o->branch1;
2692                         branch2 = o->branch2;
2693                 } else {
2694                         renames1 = b_renames;
2695                         renames2Dst = &a_by_dst;
2696                         branch1 = o->branch2;
2697                         branch2 = o->branch1;
2698                         SWAP(ren2, ren1);
2699                 }
2700
2701                 if (ren1->processed)
2702                         continue;
2703                 ren1->processed = 1;
2704                 ren1->dst_entry->processed = 1;
2705                 /* BUG: We should only mark src_entry as processed if we
2706                  * are not dealing with a rename + add-source case.
2707                  */
2708                 ren1->src_entry->processed = 1;
2709
2710                 ren1_src = ren1->pair->one->path;
2711                 ren1_dst = ren1->pair->two->path;
2712
2713                 if (ren2) {
2714                         /* One file renamed on both sides */
2715                         const char *ren2_src = ren2->pair->one->path;
2716                         const char *ren2_dst = ren2->pair->two->path;
2717                         enum rename_type rename_type;
2718                         if (strcmp(ren1_src, ren2_src) != 0)
2719                                 BUG("ren1_src != ren2_src");
2720                         ren2->dst_entry->processed = 1;
2721                         ren2->processed = 1;
2722                         if (strcmp(ren1_dst, ren2_dst) != 0) {
2723                                 rename_type = RENAME_ONE_FILE_TO_TWO;
2724                                 clean_merge = 0;
2725                         } else {
2726                                 rename_type = RENAME_ONE_FILE_TO_ONE;
2727                                 /* BUG: We should only remove ren1_src in
2728                                  * the base stage (think of rename +
2729                                  * add-source cases).
2730                                  */
2731                                 remove_file(o, 1, ren1_src, 1);
2732                                 update_entry(ren1->dst_entry,
2733                                              ren1->pair->one,
2734                                              ren1->pair->two,
2735                                              ren2->pair->two);
2736                         }
2737                         setup_rename_conflict_info(rename_type,
2738                                                    ren1->pair,
2739                                                    ren2->pair,
2740                                                    branch1,
2741                                                    branch2,
2742                                                    ren1->dst_entry,
2743                                                    ren2->dst_entry,
2744                                                    o,
2745                                                    NULL,
2746                                                    NULL);
2747                 } else if ((lookup = string_list_lookup(renames2Dst, ren1_dst))) {
2748                         /* Two different files renamed to the same thing */
2749                         char *ren2_dst;
2750                         ren2 = lookup->util;
2751                         ren2_dst = ren2->pair->two->path;
2752                         if (strcmp(ren1_dst, ren2_dst) != 0)
2753                                 BUG("ren1_dst != ren2_dst");
2754
2755                         clean_merge = 0;
2756                         ren2->processed = 1;
2757                         /*
2758                          * BUG: We should only mark src_entry as processed
2759                          * if we are not dealing with a rename + add-source
2760                          * case.
2761                          */
2762                         ren2->src_entry->processed = 1;
2763
2764                         setup_rename_conflict_info(RENAME_TWO_FILES_TO_ONE,
2765                                                    ren1->pair,
2766                                                    ren2->pair,
2767                                                    branch1,
2768                                                    branch2,
2769                                                    ren1->dst_entry,
2770                                                    ren2->dst_entry,
2771                                                    o,
2772                                                    ren1->src_entry,
2773                                                    ren2->src_entry);
2774
2775                 } else {
2776                         /* Renamed in 1, maybe changed in 2 */
2777                         /* we only use sha1 and mode of these */
2778                         struct diff_filespec src_other, dst_other;
2779                         int try_merge;
2780
2781                         /*
2782                          * unpack_trees loads entries from common-commit
2783                          * into stage 1, from head-commit into stage 2, and
2784                          * from merge-commit into stage 3.  We keep track
2785                          * of which side corresponds to the rename.
2786                          */
2787                         int renamed_stage = a_renames == renames1 ? 2 : 3;
2788                         int other_stage =   a_renames == renames1 ? 3 : 2;
2789
2790                         /* BUG: We should only remove ren1_src in the base
2791                          * stage and in other_stage (think of rename +
2792                          * add-source case).
2793                          */
2794                         remove_file(o, 1, ren1_src,
2795                                     renamed_stage == 2 || !was_tracked(o, ren1_src));
2796
2797                         oidcpy(&src_other.oid,
2798                                &ren1->src_entry->stages[other_stage].oid);
2799                         src_other.mode = ren1->src_entry->stages[other_stage].mode;
2800                         oidcpy(&dst_other.oid,
2801                                &ren1->dst_entry->stages[other_stage].oid);
2802                         dst_other.mode = ren1->dst_entry->stages[other_stage].mode;
2803                         try_merge = 0;
2804
2805                         if (oid_eq(&src_other.oid, &null_oid) &&
2806                             ren1->add_turned_into_rename) {
2807                                 setup_rename_conflict_info(RENAME_VIA_DIR,
2808                                                            ren1->pair,
2809                                                            NULL,
2810                                                            branch1,
2811                                                            branch2,
2812                                                            ren1->dst_entry,
2813                                                            NULL,
2814                                                            o,
2815                                                            NULL,
2816                                                            NULL);
2817                         } else if (oid_eq(&src_other.oid, &null_oid)) {
2818                                 setup_rename_conflict_info(RENAME_DELETE,
2819                                                            ren1->pair,
2820                                                            NULL,
2821                                                            branch1,
2822                                                            branch2,
2823                                                            ren1->dst_entry,
2824                                                            NULL,
2825                                                            o,
2826                                                            NULL,
2827                                                            NULL);
2828                         } else if ((dst_other.mode == ren1->pair->two->mode) &&
2829                                    oid_eq(&dst_other.oid, &ren1->pair->two->oid)) {
2830                                 /*
2831                                  * Added file on the other side identical to
2832                                  * the file being renamed: clean merge.
2833                                  * Also, there is no need to overwrite the
2834                                  * file already in the working copy, so call
2835                                  * update_file_flags() instead of
2836                                  * update_file().
2837                                  */
2838                                 if (update_file_flags(o,
2839                                                       &ren1->pair->two->oid,
2840                                                       ren1->pair->two->mode,
2841                                                       ren1_dst,
2842                                                       1, /* update_cache */
2843                                                       0  /* update_wd    */))
2844                                         clean_merge = -1;
2845                         } else if (!oid_eq(&dst_other.oid, &null_oid)) {
2846                                 /*
2847                                  * Probably not a clean merge, but it's
2848                                  * premature to set clean_merge to 0 here,
2849                                  * because if the rename merges cleanly and
2850                                  * the merge exactly matches the newly added
2851                                  * file, then the merge will be clean.
2852                                  */
2853                                 setup_rename_conflict_info(RENAME_ADD,
2854                                                            ren1->pair,
2855                                                            NULL,
2856                                                            branch1,
2857                                                            branch2,
2858                                                            ren1->dst_entry,
2859                                                            NULL,
2860                                                            o,
2861                                                            ren1->src_entry,
2862                                                            NULL);
2863                         } else
2864                                 try_merge = 1;
2865
2866                         if (clean_merge < 0)
2867                                 goto cleanup_and_return;
2868                         if (try_merge) {
2869                                 struct diff_filespec *one, *a, *b;
2870                                 src_other.path = (char *)ren1_src;
2871
2872                                 one = ren1->pair->one;
2873                                 if (a_renames == renames1) {
2874                                         a = ren1->pair->two;
2875                                         b = &src_other;
2876                                 } else {
2877                                         b = ren1->pair->two;
2878                                         a = &src_other;
2879                                 }
2880                                 update_entry(ren1->dst_entry, one, a, b);
2881                                 setup_rename_conflict_info(RENAME_NORMAL,
2882                                                            ren1->pair,
2883                                                            NULL,
2884                                                            branch1,
2885                                                            NULL,
2886                                                            ren1->dst_entry,
2887                                                            NULL,
2888                                                            o,
2889                                                            NULL,
2890                                                            NULL);
2891                         }
2892                 }
2893         }
2894 cleanup_and_return:
2895         string_list_clear(&a_by_dst, 0);
2896         string_list_clear(&b_by_dst, 0);
2897
2898         return clean_merge;
2899 }
2900
2901 struct rename_info {
2902         struct string_list *head_renames;
2903         struct string_list *merge_renames;
2904 };
2905
2906 static void initial_cleanup_rename(struct diff_queue_struct *pairs,
2907                                    struct hashmap *dir_renames)
2908 {
2909         struct hashmap_iter iter;
2910         struct dir_rename_entry *e;
2911
2912         hashmap_iter_init(dir_renames, &iter);
2913         while ((e = hashmap_iter_next(&iter))) {
2914                 free(e->dir);
2915                 strbuf_release(&e->new_dir);
2916                 /* possible_new_dirs already cleared in get_directory_renames */
2917         }
2918         hashmap_free(dir_renames, 1);
2919         free(dir_renames);
2920
2921         free(pairs->queue);
2922         free(pairs);
2923 }
2924
2925 static int detect_and_process_renames(struct merge_options *o,
2926                                       struct tree *common,
2927                                       struct tree *head,
2928                                       struct tree *merge,
2929                                       struct string_list *entries,
2930                                       struct rename_info *ri)
2931 {
2932         struct diff_queue_struct *head_pairs, *merge_pairs;
2933         struct hashmap *dir_re_head, *dir_re_merge;
2934         int clean = 1;
2935
2936         ri->head_renames = NULL;
2937         ri->merge_renames = NULL;
2938
2939         if (!merge_detect_rename(o))
2940                 return 1;
2941
2942         head_pairs = get_diffpairs(o, common, head);
2943         merge_pairs = get_diffpairs(o, common, merge);
2944
2945         if (o->detect_directory_renames) {
2946                 dir_re_head = get_directory_renames(head_pairs, head);
2947                 dir_re_merge = get_directory_renames(merge_pairs, merge);
2948
2949                 handle_directory_level_conflicts(o,
2950                                                  dir_re_head, head,
2951                                                  dir_re_merge, merge);
2952         } else {
2953                 dir_re_head  = xmalloc(sizeof(*dir_re_head));
2954                 dir_re_merge = xmalloc(sizeof(*dir_re_merge));
2955                 dir_rename_init(dir_re_head);
2956                 dir_rename_init(dir_re_merge);
2957         }
2958
2959         ri->head_renames  = get_renames(o, head_pairs,
2960                                         dir_re_merge, dir_re_head, head,
2961                                         common, head, merge, entries,
2962                                         &clean);
2963         if (clean < 0)
2964                 goto cleanup;
2965         ri->merge_renames = get_renames(o, merge_pairs,
2966                                         dir_re_head, dir_re_merge, merge,
2967                                         common, head, merge, entries,
2968                                         &clean);
2969         if (clean < 0)
2970                 goto cleanup;
2971         clean &= process_renames(o, ri->head_renames, ri->merge_renames);
2972
2973 cleanup:
2974         /*
2975          * Some cleanup is deferred until cleanup_renames() because the
2976          * data structures are still needed and referenced in
2977          * process_entry().  But there are a few things we can free now.
2978          */
2979         initial_cleanup_rename(head_pairs, dir_re_head);
2980         initial_cleanup_rename(merge_pairs, dir_re_merge);
2981
2982         return clean;
2983 }
2984
2985 static void final_cleanup_rename(struct string_list *rename)
2986 {
2987         const struct rename *re;
2988         int i;
2989
2990         if (rename == NULL)
2991                 return;
2992
2993         for (i = 0; i < rename->nr; i++) {
2994                 re = rename->items[i].util;
2995                 diff_free_filepair(re->pair);
2996         }
2997         string_list_clear(rename, 1);
2998         free(rename);
2999 }
3000
3001 static void final_cleanup_renames(struct rename_info *re_info)
3002 {
3003         final_cleanup_rename(re_info->head_renames);
3004         final_cleanup_rename(re_info->merge_renames);
3005 }
3006
3007 static struct object_id *stage_oid(const struct object_id *oid, unsigned mode)
3008 {
3009         return (is_null_oid(oid) || mode == 0) ? NULL: (struct object_id *)oid;
3010 }
3011
3012 static int read_oid_strbuf(struct merge_options *o,
3013                            const struct object_id *oid,
3014                            struct strbuf *dst)
3015 {
3016         void *buf;
3017         enum object_type type;
3018         unsigned long size;
3019         buf = read_object_file(oid, &type, &size);
3020         if (!buf)
3021                 return err(o, _("cannot read object %s"), oid_to_hex(oid));
3022         if (type != OBJ_BLOB) {
3023                 free(buf);
3024                 return err(o, _("object %s is not a blob"), oid_to_hex(oid));
3025         }
3026         strbuf_attach(dst, buf, size, size + 1);
3027         return 0;
3028 }
3029
3030 static int blob_unchanged(struct merge_options *opt,
3031                           const struct object_id *o_oid,
3032                           unsigned o_mode,
3033                           const struct object_id *a_oid,
3034                           unsigned a_mode,
3035                           int renormalize, const char *path)
3036 {
3037         struct strbuf o = STRBUF_INIT;
3038         struct strbuf a = STRBUF_INIT;
3039         int ret = 0; /* assume changed for safety */
3040
3041         if (a_mode != o_mode)
3042                 return 0;
3043         if (oid_eq(o_oid, a_oid))
3044                 return 1;
3045         if (!renormalize)
3046                 return 0;
3047
3048         assert(o_oid && a_oid);
3049         if (read_oid_strbuf(opt, o_oid, &o) || read_oid_strbuf(opt, a_oid, &a))
3050                 goto error_return;
3051         /*
3052          * Note: binary | is used so that both renormalizations are
3053          * performed.  Comparison can be skipped if both files are
3054          * unchanged since their sha1s have already been compared.
3055          */
3056         if (renormalize_buffer(opt->repo->index, path, o.buf, o.len, &o) |
3057             renormalize_buffer(opt->repo->index, path, a.buf, a.len, &a))
3058                 ret = (o.len == a.len && !memcmp(o.buf, a.buf, o.len));
3059
3060 error_return:
3061         strbuf_release(&o);
3062         strbuf_release(&a);
3063         return ret;
3064 }
3065
3066 static int handle_modify_delete(struct merge_options *o,
3067                                 const char *path,
3068                                 struct object_id *o_oid, int o_mode,
3069                                 struct object_id *a_oid, int a_mode,
3070                                 struct object_id *b_oid, int b_mode)
3071 {
3072         const char *modify_branch, *delete_branch;
3073         struct object_id *changed_oid;
3074         int changed_mode;
3075
3076         if (a_oid) {
3077                 modify_branch = o->branch1;
3078                 delete_branch = o->branch2;
3079                 changed_oid = a_oid;
3080                 changed_mode = a_mode;
3081         } else {
3082                 modify_branch = o->branch2;
3083                 delete_branch = o->branch1;
3084                 changed_oid = b_oid;
3085                 changed_mode = b_mode;
3086         }
3087
3088         return handle_change_delete(o,
3089                                     path, NULL,
3090                                     o_oid, o_mode,
3091                                     changed_oid, changed_mode,
3092                                     modify_branch, delete_branch,
3093                                     _("modify"), _("modified"));
3094 }
3095
3096 static int handle_content_merge(struct merge_options *o,
3097                                 const char *path,
3098                                 int is_dirty,
3099                                 struct object_id *o_oid, int o_mode,
3100                                 struct object_id *a_oid, int a_mode,
3101                                 struct object_id *b_oid, int b_mode,
3102                                 struct rename_conflict_info *rename_conflict_info)
3103 {
3104         const char *reason = _("content");
3105         const char *path1 = NULL, *path2 = NULL;
3106         struct merge_file_info mfi;
3107         struct diff_filespec one, a, b;
3108         unsigned df_conflict_remains = 0;
3109
3110         if (!o_oid) {
3111                 reason = _("add/add");
3112                 o_oid = (struct object_id *)&null_oid;
3113         }
3114         one.path = a.path = b.path = (char *)path;
3115         oidcpy(&one.oid, o_oid);
3116         one.mode = o_mode;
3117         oidcpy(&a.oid, a_oid);
3118         a.mode = a_mode;
3119         oidcpy(&b.oid, b_oid);
3120         b.mode = b_mode;
3121
3122         if (rename_conflict_info) {
3123                 struct diff_filepair *pair1 = rename_conflict_info->pair1;
3124
3125                 path1 = (o->branch1 == rename_conflict_info->branch1) ?
3126                         pair1->two->path : pair1->one->path;
3127                 /* If rename_conflict_info->pair2 != NULL, we are in
3128                  * RENAME_ONE_FILE_TO_ONE case.  Otherwise, we have a
3129                  * normal rename.
3130                  */
3131                 path2 = (rename_conflict_info->pair2 ||
3132                          o->branch2 == rename_conflict_info->branch1) ?
3133                         pair1->two->path : pair1->one->path;
3134                 one.path = pair1->one->path;
3135                 a.path = (char *)path1;
3136                 b.path = (char *)path2;
3137
3138                 if (dir_in_way(o->repo->index, path, !o->call_depth,
3139                                S_ISGITLINK(pair1->two->mode)))
3140                         df_conflict_remains = 1;
3141         }
3142         if (merge_mode_and_contents(o, &one, &a, &b, path,
3143                                     o->branch1, o->branch2,
3144                                     o->call_depth * 2, &mfi))
3145                 return -1;
3146
3147         /*
3148          * We can skip updating the working tree file iff:
3149          *   a) The merge is clean
3150          *   b) The merge matches what was in HEAD (content, mode, pathname)
3151          *   c) The target path is usable (i.e. not involved in D/F conflict)
3152          */
3153         if (mfi.clean &&
3154             was_tracked_and_matches(o, path, &mfi.oid, mfi.mode) &&
3155             !df_conflict_remains) {
3156                 int pos;
3157                 struct cache_entry *ce;
3158
3159                 output(o, 3, _("Skipped %s (merged same as existing)"), path);
3160                 if (add_cacheinfo(o, mfi.mode, &mfi.oid, path,
3161                                   0, (!o->call_depth && !is_dirty), 0))
3162                         return -1;
3163                 /*
3164                  * However, add_cacheinfo() will delete the old cache entry
3165                  * and add a new one.  We need to copy over any skip_worktree
3166                  * flag to avoid making the file appear as if it were
3167                  * deleted by the user.
3168                  */
3169                 pos = index_name_pos(&o->orig_index, path, strlen(path));
3170                 ce = o->orig_index.cache[pos];
3171                 if (ce_skip_worktree(ce)) {
3172                         pos = index_name_pos(o->repo->index, path, strlen(path));
3173                         ce = o->repo->index->cache[pos];
3174                         ce->ce_flags |= CE_SKIP_WORKTREE;
3175                 }
3176                 return mfi.clean;
3177         }
3178
3179         if (!mfi.clean) {
3180                 if (S_ISGITLINK(mfi.mode))
3181                         reason = _("submodule");
3182                 output(o, 1, _("CONFLICT (%s): Merge conflict in %s"),
3183                                 reason, path);
3184                 if (rename_conflict_info && !df_conflict_remains)
3185                         if (update_stages(o, path, &one, &a, &b))
3186                                 return -1;
3187         }
3188
3189         if (df_conflict_remains || is_dirty) {
3190                 char *new_path;
3191                 if (o->call_depth) {
3192                         remove_file_from_index(o->repo->index, path);
3193                 } else {
3194                         if (!mfi.clean) {
3195                                 if (update_stages(o, path, &one, &a, &b))
3196                                         return -1;
3197                         } else {
3198                                 int file_from_stage2 = was_tracked(o, path);
3199                                 struct diff_filespec merged;
3200                                 oidcpy(&merged.oid, &mfi.oid);
3201                                 merged.mode = mfi.mode;
3202
3203                                 if (update_stages(o, path, NULL,
3204                                                   file_from_stage2 ? &merged : NULL,
3205                                                   file_from_stage2 ? NULL : &merged))
3206                                         return -1;
3207                         }
3208
3209                 }
3210                 new_path = unique_path(o, path, rename_conflict_info->branch1);
3211                 if (is_dirty) {
3212                         output(o, 1, _("Refusing to lose dirty file at %s"),
3213                                path);
3214                 }
3215                 output(o, 1, _("Adding as %s instead"), new_path);
3216                 if (update_file(o, 0, &mfi.oid, mfi.mode, new_path)) {
3217                         free(new_path);
3218                         return -1;
3219                 }
3220                 free(new_path);
3221                 mfi.clean = 0;
3222         } else if (update_file(o, mfi.clean, &mfi.oid, mfi.mode, path))
3223                 return -1;
3224         return !is_dirty && mfi.clean;
3225 }
3226
3227 static int handle_rename_normal(struct merge_options *o,
3228                                 const char *path,
3229                                 struct object_id *o_oid, unsigned int o_mode,
3230                                 struct object_id *a_oid, unsigned int a_mode,
3231                                 struct object_id *b_oid, unsigned int b_mode,
3232                                 struct rename_conflict_info *ci)
3233 {
3234         /* Merge the content and write it out */
3235         return handle_content_merge(o, path, was_dirty(o, path),
3236                                     o_oid, o_mode, a_oid, a_mode, b_oid, b_mode,
3237                                     ci);
3238 }
3239
3240 /* Per entry merge function */
3241 static int process_entry(struct merge_options *o,
3242                          const char *path, struct stage_data *entry)
3243 {
3244         int clean_merge = 1;
3245         int normalize = o->renormalize;
3246         unsigned o_mode = entry->stages[1].mode;
3247         unsigned a_mode = entry->stages[2].mode;
3248         unsigned b_mode = entry->stages[3].mode;
3249         struct object_id *o_oid = stage_oid(&entry->stages[1].oid, o_mode);
3250         struct object_id *a_oid = stage_oid(&entry->stages[2].oid, a_mode);
3251         struct object_id *b_oid = stage_oid(&entry->stages[3].oid, b_mode);
3252
3253         entry->processed = 1;
3254         if (entry->rename_conflict_info) {
3255                 struct rename_conflict_info *conflict_info = entry->rename_conflict_info;
3256                 switch (conflict_info->rename_type) {
3257                 case RENAME_NORMAL:
3258                 case RENAME_ONE_FILE_TO_ONE:
3259                         clean_merge = handle_rename_normal(o,
3260                                                            path,
3261                                                            o_oid, o_mode,
3262                                                            a_oid, a_mode,
3263                                                            b_oid, b_mode,
3264                                                            conflict_info);
3265                         break;
3266                 case RENAME_VIA_DIR:
3267                         clean_merge = 1;
3268                         if (handle_rename_via_dir(o,
3269                                                   conflict_info->pair1,
3270                                                   conflict_info->branch1,
3271                                                   conflict_info->branch2))
3272                                 clean_merge = -1;
3273                         break;
3274                 case RENAME_ADD:
3275                         /*
3276                          * Probably unclean merge, but if the renamed file
3277                          * merges cleanly and the result can then be
3278                          * two-way merged cleanly with the added file, I
3279                          * guess it's a clean merge?
3280                          */
3281                         clean_merge = handle_rename_add(o, conflict_info);
3282                         break;
3283                 case RENAME_DELETE:
3284                         clean_merge = 0;
3285                         if (handle_rename_delete(o,
3286                                                  conflict_info->pair1,
3287                                                  conflict_info->branch1,
3288                                                  conflict_info->branch2))
3289                                 clean_merge = -1;
3290                         break;
3291                 case RENAME_ONE_FILE_TO_TWO:
3292                         clean_merge = 0;
3293                         if (handle_rename_rename_1to2(o, conflict_info))
3294                                 clean_merge = -1;
3295                         break;
3296                 case RENAME_TWO_FILES_TO_ONE:
3297                         /*
3298                          * Probably unclean merge, but if the two renamed
3299                          * files merge cleanly and the two resulting files
3300                          * can then be two-way merged cleanly, I guess it's
3301                          * a clean merge?
3302                          */
3303                         clean_merge = handle_rename_rename_2to1(o,
3304                                                                 conflict_info);
3305                         break;
3306                 default:
3307                         entry->processed = 0;
3308                         break;
3309                 }
3310         } else if (o_oid && (!a_oid || !b_oid)) {
3311                 /* Case A: Deleted in one */
3312                 if ((!a_oid && !b_oid) ||
3313                     (!b_oid && blob_unchanged(o, o_oid, o_mode, a_oid, a_mode, normalize, path)) ||
3314                     (!a_oid && blob_unchanged(o, o_oid, o_mode, b_oid, b_mode, normalize, path))) {
3315                         /* Deleted in both or deleted in one and
3316                          * unchanged in the other */
3317                         if (a_oid)
3318                                 output(o, 2, _("Removing %s"), path);
3319                         /* do not touch working file if it did not exist */
3320                         remove_file(o, 1, path, !a_oid);
3321                 } else {
3322                         /* Modify/delete; deleted side may have put a directory in the way */
3323                         clean_merge = 0;
3324                         if (handle_modify_delete(o, path, o_oid, o_mode,
3325                                                  a_oid, a_mode, b_oid, b_mode))
3326                                 clean_merge = -1;
3327                 }
3328         } else if ((!o_oid && a_oid && !b_oid) ||
3329                    (!o_oid && !a_oid && b_oid)) {
3330                 /* Case B: Added in one. */
3331                 /* [nothing|directory] -> ([nothing|directory], file) */
3332
3333                 const char *add_branch;
3334                 const char *other_branch;
3335                 unsigned mode;
3336                 const struct object_id *oid;
3337                 const char *conf;
3338
3339                 if (a_oid) {
3340                         add_branch = o->branch1;
3341                         other_branch = o->branch2;
3342                         mode = a_mode;
3343                         oid = a_oid;
3344                         conf = _("file/directory");
3345                 } else {
3346                         add_branch = o->branch2;
3347                         other_branch = o->branch1;
3348                         mode = b_mode;
3349                         oid = b_oid;
3350                         conf = _("directory/file");
3351                 }
3352                 if (dir_in_way(o->repo->index, path,
3353                                !o->call_depth && !S_ISGITLINK(a_mode),
3354                                0)) {
3355                         char *new_path = unique_path(o, path, add_branch);
3356                         clean_merge = 0;
3357                         output(o, 1, _("CONFLICT (%s): There is a directory with name %s in %s. "
3358                                "Adding %s as %s"),
3359                                conf, path, other_branch, path, new_path);
3360                         if (update_file(o, 0, oid, mode, new_path))
3361                                 clean_merge = -1;
3362                         else if (o->call_depth)
3363                                 remove_file_from_index(o->repo->index, path);
3364                         free(new_path);
3365                 } else {
3366                         output(o, 2, _("Adding %s"), path);
3367                         /* do not overwrite file if already present */
3368                         if (update_file_flags(o, oid, mode, path, 1, !a_oid))
3369                                 clean_merge = -1;
3370                 }
3371         } else if (a_oid && b_oid) {
3372                 if (!o_oid) {
3373                         /* Case C: Added in both (check for same permissions) */
3374                         output(o, 1,
3375                                _("CONFLICT (add/add): Merge conflict in %s"),
3376                                path);
3377                         clean_merge = handle_file_collision(o,
3378                                                             path, NULL, NULL,
3379                                                             o->branch1,
3380                                                             o->branch2,
3381                                                             a_oid, a_mode,
3382                                                             b_oid, b_mode);
3383                 } else {
3384                         /* case D: Modified in both, but differently. */
3385                         int is_dirty = 0; /* unpack_trees would have bailed if dirty */
3386                         clean_merge = handle_content_merge(o, path,
3387                                                            is_dirty,
3388                                                            o_oid, o_mode,
3389                                                            a_oid, a_mode,
3390                                                            b_oid, b_mode,
3391                                                            NULL);
3392                 }
3393         } else if (!o_oid && !a_oid && !b_oid) {
3394                 /*
3395                  * this entry was deleted altogether. a_mode == 0 means
3396                  * we had that path and want to actively remove it.
3397                  */
3398                 remove_file(o, 1, path, !a_mode);
3399         } else
3400                 BUG("fatal merge failure, shouldn't happen.");
3401
3402         return clean_merge;
3403 }
3404
3405 int merge_trees(struct merge_options *o,
3406                 struct tree *head,
3407                 struct tree *merge,
3408                 struct tree *common,
3409                 struct tree **result)
3410 {
3411         struct index_state *istate = o->repo->index;
3412         int code, clean;
3413         struct strbuf sb = STRBUF_INIT;
3414
3415         if (!o->call_depth && repo_index_has_changes(o->repo, head, &sb)) {
3416                 err(o, _("Your local changes to the following files would be overwritten by merge:\n  %s"),
3417                     sb.buf);
3418                 return -1;
3419         }
3420
3421         if (o->subtree_shift) {
3422                 merge = shift_tree_object(o->repo, head, merge, o->subtree_shift);
3423                 common = shift_tree_object(o->repo, head, common, o->subtree_shift);
3424         }
3425
3426         if (oid_eq(&common->object.oid, &merge->object.oid)) {
3427                 output(o, 0, _("Already up to date!"));
3428                 *result = head;
3429                 return 1;
3430         }
3431
3432         code = unpack_trees_start(o, common, head, merge);
3433
3434         if (code != 0) {
3435                 if (show(o, 4) || o->call_depth)
3436                         err(o, _("merging of trees %s and %s failed"),
3437                             oid_to_hex(&head->object.oid),
3438                             oid_to_hex(&merge->object.oid));
3439                 unpack_trees_finish(o);
3440                 return -1;
3441         }
3442
3443         if (unmerged_index(istate)) {
3444                 struct string_list *entries;
3445                 struct rename_info re_info;
3446                 int i;
3447                 /*
3448                  * Only need the hashmap while processing entries, so
3449                  * initialize it here and free it when we are done running
3450                  * through the entries. Keeping it in the merge_options as
3451                  * opposed to decaring a local hashmap is for convenience
3452                  * so that we don't have to pass it to around.
3453                  */
3454                 hashmap_init(&o->current_file_dir_set, path_hashmap_cmp, NULL, 512);
3455                 get_files_dirs(o, head);
3456                 get_files_dirs(o, merge);
3457
3458                 entries = get_unmerged(o->repo->index);
3459                 clean = detect_and_process_renames(o, common, head, merge,
3460                                                    entries, &re_info);
3461                 record_df_conflict_files(o, entries);
3462                 if (clean < 0)
3463                         goto cleanup;
3464                 for (i = entries->nr-1; 0 <= i; i--) {
3465                         const char *path = entries->items[i].string;
3466                         struct stage_data *e = entries->items[i].util;
3467                         if (!e->processed) {
3468                                 int ret = process_entry(o, path, e);
3469                                 if (!ret)
3470                                         clean = 0;
3471                                 else if (ret < 0) {
3472                                         clean = ret;
3473                                         goto cleanup;
3474                                 }
3475                         }
3476                 }
3477                 for (i = 0; i < entries->nr; i++) {
3478                         struct stage_data *e = entries->items[i].util;
3479                         if (!e->processed)
3480                                 BUG("unprocessed path??? %s",
3481                                     entries->items[i].string);
3482                 }
3483
3484         cleanup:
3485                 final_cleanup_renames(&re_info);
3486
3487                 string_list_clear(entries, 1);
3488                 free(entries);
3489
3490                 hashmap_free(&o->current_file_dir_set, 1);
3491
3492                 if (clean < 0) {
3493                         unpack_trees_finish(o);
3494                         return clean;
3495                 }
3496         }
3497         else
3498                 clean = 1;
3499
3500         unpack_trees_finish(o);
3501
3502         if (o->call_depth && !(*result = write_tree_from_memory(o)))
3503                 return -1;
3504
3505         return clean;
3506 }
3507
3508 static struct commit_list *reverse_commit_list(struct commit_list *list)
3509 {
3510         struct commit_list *next = NULL, *current, *backup;
3511         for (current = list; current; current = backup) {
3512                 backup = current->next;
3513                 current->next = next;
3514                 next = current;
3515         }
3516         return next;
3517 }
3518
3519 /*
3520  * Merge the commits h1 and h2, return the resulting virtual
3521  * commit object and a flag indicating the cleanness of the merge.
3522  */
3523 int merge_recursive(struct merge_options *o,
3524                     struct commit *h1,
3525                     struct commit *h2,
3526                     struct commit_list *ca,
3527                     struct commit **result)
3528 {
3529         struct commit_list *iter;
3530         struct commit *merged_common_ancestors;
3531         struct tree *mrtree;
3532         int clean;
3533
3534         if (show(o, 4)) {
3535                 output(o, 4, _("Merging:"));
3536                 output_commit_title(o, h1);
3537                 output_commit_title(o, h2);
3538         }
3539
3540         if (!ca) {
3541                 ca = get_merge_bases(h1, h2);
3542                 ca = reverse_commit_list(ca);
3543         }
3544
3545         if (show(o, 5)) {
3546                 unsigned cnt = commit_list_count(ca);
3547
3548                 output(o, 5, Q_("found %u common ancestor:",
3549                                 "found %u common ancestors:", cnt), cnt);
3550                 for (iter = ca; iter; iter = iter->next)
3551                         output_commit_title(o, iter->item);
3552         }
3553
3554         merged_common_ancestors = pop_commit(&ca);
3555         if (merged_common_ancestors == NULL) {
3556                 /* if there is no common ancestor, use an empty tree */
3557                 struct tree *tree;
3558
3559                 tree = lookup_tree(o->repo, o->repo->hash_algo->empty_tree);
3560                 merged_common_ancestors = make_virtual_commit(o->repo, tree, "ancestor");
3561         }
3562
3563         for (iter = ca; iter; iter = iter->next) {
3564                 const char *saved_b1, *saved_b2;
3565                 o->call_depth++;
3566                 /*
3567                  * When the merge fails, the result contains files
3568                  * with conflict markers. The cleanness flag is
3569                  * ignored (unless indicating an error), it was never
3570                  * actually used, as result of merge_trees has always
3571                  * overwritten it: the committed "conflicts" were
3572                  * already resolved.
3573                  */
3574                 discard_index(o->repo->index);
3575                 saved_b1 = o->branch1;
3576                 saved_b2 = o->branch2;
3577                 o->branch1 = "Temporary merge branch 1";
3578                 o->branch2 = "Temporary merge branch 2";
3579                 if (merge_recursive(o, merged_common_ancestors, iter->item,
3580                                     NULL, &merged_common_ancestors) < 0)
3581                         return -1;
3582                 o->branch1 = saved_b1;
3583                 o->branch2 = saved_b2;
3584                 o->call_depth--;
3585
3586                 if (!merged_common_ancestors)
3587                         return err(o, _("merge returned no commit"));
3588         }
3589
3590         discard_index(o->repo->index);
3591         if (!o->call_depth)
3592                 repo_read_index(o->repo);
3593
3594         o->ancestor = "merged common ancestors";
3595         clean = merge_trees(o, get_commit_tree(h1), get_commit_tree(h2),
3596                             get_commit_tree(merged_common_ancestors),
3597                             &mrtree);
3598         if (clean < 0) {
3599                 flush_output(o);
3600                 return clean;
3601         }
3602
3603         if (o->call_depth) {
3604                 *result = make_virtual_commit(o->repo, mrtree, "merged tree");
3605                 commit_list_insert(h1, &(*result)->parents);
3606                 commit_list_insert(h2, &(*result)->parents->next);
3607         }
3608         flush_output(o);
3609         if (!o->call_depth && o->buffer_output < 2)
3610                 strbuf_release(&o->obuf);
3611         if (show(o, 2))
3612                 diff_warn_rename_limit("merge.renamelimit",
3613                                        o->needed_rename_limit, 0);
3614         return clean;
3615 }
3616
3617 static struct commit *get_ref(struct repository *repo, const struct object_id *oid,
3618                               const char *name)
3619 {
3620         struct object *object;
3621
3622         object = deref_tag(repo, parse_object(repo, oid),
3623                            name, strlen(name));
3624         if (!object)
3625                 return NULL;
3626         if (object->type == OBJ_TREE)
3627                 return make_virtual_commit(repo, (struct tree*)object, name);
3628         if (object->type != OBJ_COMMIT)
3629                 return NULL;
3630         if (parse_commit((struct commit *)object))
3631                 return NULL;
3632         return (struct commit *)object;
3633 }
3634
3635 int merge_recursive_generic(struct merge_options *o,
3636                             const struct object_id *head,
3637                             const struct object_id *merge,
3638                             int num_base_list,
3639                             const struct object_id **base_list,
3640                             struct commit **result)
3641 {
3642         int clean;
3643         struct lock_file lock = LOCK_INIT;
3644         struct commit *head_commit = get_ref(o->repo, head, o->branch1);
3645         struct commit *next_commit = get_ref(o->repo, merge, o->branch2);
3646         struct commit_list *ca = NULL;
3647
3648         if (base_list) {
3649                 int i;
3650                 for (i = 0; i < num_base_list; ++i) {
3651                         struct commit *base;
3652                         if (!(base = get_ref(o->repo, base_list[i], oid_to_hex(base_list[i]))))
3653                                 return err(o, _("Could not parse object '%s'"),
3654                                            oid_to_hex(base_list[i]));
3655                         commit_list_insert(base, &ca);
3656                 }
3657         }
3658
3659         repo_hold_locked_index(o->repo, &lock, LOCK_DIE_ON_ERROR);
3660         clean = merge_recursive(o, head_commit, next_commit, ca,
3661                                 result);
3662         if (clean < 0) {
3663                 rollback_lock_file(&lock);
3664                 return clean;
3665         }
3666
3667         if (write_locked_index(o->repo->index, &lock,
3668                                COMMIT_LOCK | SKIP_IF_UNCHANGED))
3669                 return err(o, _("Unable to write index."));
3670
3671         return clean ? 0 : 1;
3672 }
3673
3674 static void merge_recursive_config(struct merge_options *o)
3675 {
3676         char *value = NULL;
3677         git_config_get_int("merge.verbosity", &o->verbosity);
3678         git_config_get_int("diff.renamelimit", &o->diff_rename_limit);
3679         git_config_get_int("merge.renamelimit", &o->merge_rename_limit);
3680         if (!git_config_get_string("diff.renames", &value)) {
3681                 o->diff_detect_rename = git_config_rename("diff.renames", value);
3682                 free(value);
3683         }
3684         if (!git_config_get_string("merge.renames", &value)) {
3685                 o->merge_detect_rename = git_config_rename("merge.renames", value);
3686                 free(value);
3687         }
3688         git_config(git_xmerge_config, NULL);
3689 }
3690
3691 void init_merge_options(struct merge_options *o,
3692                         struct repository *repo)
3693 {
3694         const char *merge_verbosity;
3695         memset(o, 0, sizeof(struct merge_options));
3696         o->repo = repo;
3697         o->verbosity = 2;
3698         o->buffer_output = 1;
3699         o->diff_rename_limit = -1;
3700         o->merge_rename_limit = -1;
3701         o->renormalize = 0;
3702         o->diff_detect_rename = -1;
3703         o->merge_detect_rename = -1;
3704         o->detect_directory_renames = 1;
3705         merge_recursive_config(o);
3706         merge_verbosity = getenv("GIT_MERGE_VERBOSITY");
3707         if (merge_verbosity)
3708                 o->verbosity = strtol(merge_verbosity, NULL, 10);
3709         if (o->verbosity >= 5)
3710                 o->buffer_output = 0;
3711         strbuf_init(&o->obuf, 0);
3712         string_list_init(&o->df_conflict_file_set, 1);
3713 }
3714
3715 int parse_merge_opt(struct merge_options *o, const char *s)
3716 {
3717         const char *arg;
3718
3719         if (!s || !*s)
3720                 return -1;
3721         if (!strcmp(s, "ours"))
3722                 o->recursive_variant = MERGE_RECURSIVE_OURS;
3723         else if (!strcmp(s, "theirs"))
3724                 o->recursive_variant = MERGE_RECURSIVE_THEIRS;
3725         else if (!strcmp(s, "subtree"))
3726                 o->subtree_shift = "";
3727         else if (skip_prefix(s, "subtree=", &arg))
3728                 o->subtree_shift = arg;
3729         else if (!strcmp(s, "patience"))
3730                 o->xdl_opts = DIFF_WITH_ALG(o, PATIENCE_DIFF);
3731         else if (!strcmp(s, "histogram"))
3732                 o->xdl_opts = DIFF_WITH_ALG(o, HISTOGRAM_DIFF);
3733         else if (skip_prefix(s, "diff-algorithm=", &arg)) {
3734                 long value = parse_algorithm_value(arg);
3735                 if (value < 0)
3736                         return -1;
3737                 /* clear out previous settings */
3738                 DIFF_XDL_CLR(o, NEED_MINIMAL);
3739                 o->xdl_opts &= ~XDF_DIFF_ALGORITHM_MASK;
3740                 o->xdl_opts |= value;
3741         }
3742         else if (!strcmp(s, "ignore-space-change"))
3743                 DIFF_XDL_SET(o, IGNORE_WHITESPACE_CHANGE);
3744         else if (!strcmp(s, "ignore-all-space"))
3745                 DIFF_XDL_SET(o, IGNORE_WHITESPACE);
3746         else if (!strcmp(s, "ignore-space-at-eol"))
3747                 DIFF_XDL_SET(o, IGNORE_WHITESPACE_AT_EOL);
3748         else if (!strcmp(s, "ignore-cr-at-eol"))
3749                 DIFF_XDL_SET(o, IGNORE_CR_AT_EOL);
3750         else if (!strcmp(s, "renormalize"))
3751                 o->renormalize = 1;
3752         else if (!strcmp(s, "no-renormalize"))
3753                 o->renormalize = 0;
3754         else if (!strcmp(s, "no-renames"))
3755                 o->merge_detect_rename = 0;
3756         else if (!strcmp(s, "find-renames")) {
3757                 o->merge_detect_rename = 1;
3758                 o->rename_score = 0;
3759         }
3760         else if (skip_prefix(s, "find-renames=", &arg) ||
3761                  skip_prefix(s, "rename-threshold=", &arg)) {
3762                 if ((o->rename_score = parse_rename_score(&arg)) == -1 || *arg != 0)
3763                         return -1;
3764                 o->merge_detect_rename = 1;
3765         }
3766         else
3767                 return -1;
3768         return 0;
3769 }