Merge branch 'ag/rebase-p'
[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 "commit.h"
13 #include "blob.h"
14 #include "builtin.h"
15 #include "tree-walk.h"
16 #include "diff.h"
17 #include "diffcore.h"
18 #include "tag.h"
19 #include "alloc.h"
20 #include "unpack-trees.h"
21 #include "string-list.h"
22 #include "xdiff-interface.h"
23 #include "ll-merge.h"
24 #include "attr.h"
25 #include "merge-recursive.h"
26 #include "dir.h"
27 #include "submodule.h"
28 #include "revision.h"
29
30 struct path_hashmap_entry {
31         struct hashmap_entry e;
32         char path[FLEX_ARRAY];
33 };
34
35 static int path_hashmap_cmp(const void *cmp_data,
36                             const void *entry,
37                             const void *entry_or_key,
38                             const void *keydata)
39 {
40         const struct path_hashmap_entry *a = entry;
41         const struct path_hashmap_entry *b = entry_or_key;
42         const char *key = keydata;
43
44         if (ignore_case)
45                 return strcasecmp(a->path, key ? key : b->path);
46         else
47                 return strcmp(a->path, key ? key : b->path);
48 }
49
50 static unsigned int path_hash(const char *path)
51 {
52         return ignore_case ? strihash(path) : strhash(path);
53 }
54
55 static struct dir_rename_entry *dir_rename_find_entry(struct hashmap *hashmap,
56                                                       char *dir)
57 {
58         struct dir_rename_entry key;
59
60         if (dir == NULL)
61                 return NULL;
62         hashmap_entry_init(&key, strhash(dir));
63         key.dir = dir;
64         return hashmap_get(hashmap, &key, NULL);
65 }
66
67 static int dir_rename_cmp(const void *unused_cmp_data,
68                           const void *entry,
69                           const void *entry_or_key,
70                           const void *unused_keydata)
71 {
72         const struct dir_rename_entry *e1 = entry;
73         const struct dir_rename_entry *e2 = entry_or_key;
74
75         return strcmp(e1->dir, e2->dir);
76 }
77
78 static void dir_rename_init(struct hashmap *map)
79 {
80         hashmap_init(map, dir_rename_cmp, NULL, 0);
81 }
82
83 static void dir_rename_entry_init(struct dir_rename_entry *entry,
84                                   char *directory)
85 {
86         hashmap_entry_init(entry, strhash(directory));
87         entry->dir = directory;
88         entry->non_unique_new_dir = 0;
89         strbuf_init(&entry->new_dir, 0);
90         string_list_init(&entry->possible_new_dirs, 0);
91 }
92
93 static struct collision_entry *collision_find_entry(struct hashmap *hashmap,
94                                                     char *target_file)
95 {
96         struct collision_entry key;
97
98         hashmap_entry_init(&key, strhash(target_file));
99         key.target_file = target_file;
100         return hashmap_get(hashmap, &key, NULL);
101 }
102
103 static int collision_cmp(void *unused_cmp_data,
104                          const struct collision_entry *e1,
105                          const struct collision_entry *e2,
106                          const void *unused_keydata)
107 {
108         return strcmp(e1->target_file, e2->target_file);
109 }
110
111 static void collision_init(struct hashmap *map)
112 {
113         hashmap_init(map, (hashmap_cmp_fn) collision_cmp, NULL, 0);
114 }
115
116 static void flush_output(struct merge_options *o)
117 {
118         if (o->buffer_output < 2 && o->obuf.len) {
119                 fputs(o->obuf.buf, stdout);
120                 strbuf_reset(&o->obuf);
121         }
122 }
123
124 static int err(struct merge_options *o, const char *err, ...)
125 {
126         va_list params;
127
128         if (o->buffer_output < 2)
129                 flush_output(o);
130         else {
131                 strbuf_complete(&o->obuf, '\n');
132                 strbuf_addstr(&o->obuf, "error: ");
133         }
134         va_start(params, err);
135         strbuf_vaddf(&o->obuf, err, params);
136         va_end(params);
137         if (o->buffer_output > 1)
138                 strbuf_addch(&o->obuf, '\n');
139         else {
140                 error("%s", o->obuf.buf);
141                 strbuf_reset(&o->obuf);
142         }
143
144         return -1;
145 }
146
147 static struct tree *shift_tree_object(struct tree *one, struct tree *two,
148                                       const char *subtree_shift)
149 {
150         struct object_id shifted;
151
152         if (!*subtree_shift) {
153                 shift_tree(&one->object.oid, &two->object.oid, &shifted, 0);
154         } else {
155                 shift_tree_by(&one->object.oid, &two->object.oid, &shifted,
156                               subtree_shift);
157         }
158         if (!oidcmp(&two->object.oid, &shifted))
159                 return two;
160         return lookup_tree(&shifted);
161 }
162
163 static struct commit *make_virtual_commit(struct tree *tree, const char *comment)
164 {
165         struct commit *commit = alloc_commit_node(the_repository);
166
167         set_merge_remote_desc(commit, comment, (struct object *)commit);
168         commit->maybe_tree = tree;
169         commit->object.parsed = 1;
170         return commit;
171 }
172
173 /*
174  * Since we use get_tree_entry(), which does not put the read object into
175  * the object pool, we cannot rely on a == b.
176  */
177 static int oid_eq(const struct object_id *a, const struct object_id *b)
178 {
179         if (!a && !b)
180                 return 2;
181         return a && b && oidcmp(a, b) == 0;
182 }
183
184 enum rename_type {
185         RENAME_NORMAL = 0,
186         RENAME_VIA_DIR,
187         RENAME_DELETE,
188         RENAME_ONE_FILE_TO_ONE,
189         RENAME_ONE_FILE_TO_TWO,
190         RENAME_TWO_FILES_TO_ONE
191 };
192
193 struct rename_conflict_info {
194         enum rename_type rename_type;
195         struct diff_filepair *pair1;
196         struct diff_filepair *pair2;
197         const char *branch1;
198         const char *branch2;
199         struct stage_data *dst_entry1;
200         struct stage_data *dst_entry2;
201         struct diff_filespec ren1_other;
202         struct diff_filespec ren2_other;
203 };
204
205 /*
206  * Since we want to write the index eventually, we cannot reuse the index
207  * for these (temporary) data.
208  */
209 struct stage_data {
210         struct {
211                 unsigned mode;
212                 struct object_id oid;
213         } stages[4];
214         struct rename_conflict_info *rename_conflict_info;
215         unsigned processed:1;
216 };
217
218 static inline void setup_rename_conflict_info(enum rename_type rename_type,
219                                               struct diff_filepair *pair1,
220                                               struct diff_filepair *pair2,
221                                               const char *branch1,
222                                               const char *branch2,
223                                               struct stage_data *dst_entry1,
224                                               struct stage_data *dst_entry2,
225                                               struct merge_options *o,
226                                               struct stage_data *src_entry1,
227                                               struct stage_data *src_entry2)
228 {
229         struct rename_conflict_info *ci = xcalloc(1, sizeof(struct rename_conflict_info));
230         ci->rename_type = rename_type;
231         ci->pair1 = pair1;
232         ci->branch1 = branch1;
233         ci->branch2 = branch2;
234
235         ci->dst_entry1 = dst_entry1;
236         dst_entry1->rename_conflict_info = ci;
237         dst_entry1->processed = 0;
238
239         assert(!pair2 == !dst_entry2);
240         if (dst_entry2) {
241                 ci->dst_entry2 = dst_entry2;
242                 ci->pair2 = pair2;
243                 dst_entry2->rename_conflict_info = ci;
244         }
245
246         if (rename_type == RENAME_TWO_FILES_TO_ONE) {
247                 /*
248                  * For each rename, there could have been
249                  * modifications on the side of history where that
250                  * file was not renamed.
251                  */
252                 int ostage1 = o->branch1 == branch1 ? 3 : 2;
253                 int ostage2 = ostage1 ^ 1;
254
255                 ci->ren1_other.path = pair1->one->path;
256                 oidcpy(&ci->ren1_other.oid, &src_entry1->stages[ostage1].oid);
257                 ci->ren1_other.mode = src_entry1->stages[ostage1].mode;
258
259                 ci->ren2_other.path = pair2->one->path;
260                 oidcpy(&ci->ren2_other.oid, &src_entry2->stages[ostage2].oid);
261                 ci->ren2_other.mode = src_entry2->stages[ostage2].mode;
262         }
263 }
264
265 static int show(struct merge_options *o, int v)
266 {
267         return (!o->call_depth && o->verbosity >= v) || o->verbosity >= 5;
268 }
269
270 __attribute__((format (printf, 3, 4)))
271 static void output(struct merge_options *o, int v, const char *fmt, ...)
272 {
273         va_list ap;
274
275         if (!show(o, v))
276                 return;
277
278         strbuf_addchars(&o->obuf, ' ', o->call_depth * 2);
279
280         va_start(ap, fmt);
281         strbuf_vaddf(&o->obuf, fmt, ap);
282         va_end(ap);
283
284         strbuf_addch(&o->obuf, '\n');
285         if (!o->buffer_output)
286                 flush_output(o);
287 }
288
289 static void output_commit_title(struct merge_options *o, struct commit *commit)
290 {
291         struct merge_remote_desc *desc;
292
293         strbuf_addchars(&o->obuf, ' ', o->call_depth * 2);
294         desc = merge_remote_util(commit);
295         if (desc)
296                 strbuf_addf(&o->obuf, "virtual %s\n", desc->name);
297         else {
298                 strbuf_add_unique_abbrev(&o->obuf, &commit->object.oid,
299                                          DEFAULT_ABBREV);
300                 strbuf_addch(&o->obuf, ' ');
301                 if (parse_commit(commit) != 0)
302                         strbuf_addstr(&o->obuf, _("(bad commit)\n"));
303                 else {
304                         const char *title;
305                         const char *msg = get_commit_buffer(commit, NULL);
306                         int len = find_commit_subject(msg, &title);
307                         if (len)
308                                 strbuf_addf(&o->obuf, "%.*s\n", len, title);
309                         unuse_commit_buffer(commit, msg);
310                 }
311         }
312         flush_output(o);
313 }
314
315 static int add_cacheinfo(struct merge_options *o,
316                          unsigned int mode, const struct object_id *oid,
317                          const char *path, int stage, int refresh, int options)
318 {
319         struct cache_entry *ce;
320         int ret;
321
322         ce = make_cache_entry(mode, oid ? oid->hash : null_sha1, path, stage, 0);
323         if (!ce)
324                 return err(o, _("add_cacheinfo failed for path '%s'; merge aborting."), path);
325
326         ret = add_cache_entry(ce, options);
327         if (refresh) {
328                 struct cache_entry *nce;
329
330                 nce = refresh_cache_entry(ce, CE_MATCH_REFRESH | CE_MATCH_IGNORE_MISSING);
331                 if (!nce)
332                         return err(o, _("add_cacheinfo failed to refresh for path '%s'; merge aborting."), path);
333                 if (nce != ce)
334                         ret = add_cache_entry(nce, options);
335         }
336         return ret;
337 }
338
339 static void init_tree_desc_from_tree(struct tree_desc *desc, struct tree *tree)
340 {
341         parse_tree(tree);
342         init_tree_desc(desc, tree->buffer, tree->size);
343 }
344
345 static int unpack_trees_start(struct merge_options *o,
346                               struct tree *common,
347                               struct tree *head,
348                               struct tree *merge)
349 {
350         int rc;
351         struct tree_desc t[3];
352         struct index_state tmp_index = { NULL };
353
354         memset(&o->unpack_opts, 0, sizeof(o->unpack_opts));
355         if (o->call_depth)
356                 o->unpack_opts.index_only = 1;
357         else
358                 o->unpack_opts.update = 1;
359         o->unpack_opts.merge = 1;
360         o->unpack_opts.head_idx = 2;
361         o->unpack_opts.fn = threeway_merge;
362         o->unpack_opts.src_index = &the_index;
363         o->unpack_opts.dst_index = &tmp_index;
364         o->unpack_opts.aggressive = !merge_detect_rename(o);
365         setup_unpack_trees_porcelain(&o->unpack_opts, "merge");
366
367         init_tree_desc_from_tree(t+0, common);
368         init_tree_desc_from_tree(t+1, head);
369         init_tree_desc_from_tree(t+2, merge);
370
371         rc = unpack_trees(3, t, &o->unpack_opts);
372         cache_tree_free(&active_cache_tree);
373
374         /*
375          * Update the_index to match the new results, AFTER saving a copy
376          * in o->orig_index.  Update src_index to point to the saved copy.
377          * (verify_uptodate() checks src_index, and the original index is
378          * the one that had the necessary modification timestamps.)
379          */
380         o->orig_index = the_index;
381         the_index = tmp_index;
382         o->unpack_opts.src_index = &o->orig_index;
383
384         return rc;
385 }
386
387 static void unpack_trees_finish(struct merge_options *o)
388 {
389         discard_index(&o->orig_index);
390         clear_unpack_trees_porcelain(&o->unpack_opts);
391 }
392
393 struct tree *write_tree_from_memory(struct merge_options *o)
394 {
395         struct tree *result = NULL;
396
397         if (unmerged_cache()) {
398                 int i;
399                 fprintf(stderr, "BUG: There are unmerged index entries:\n");
400                 for (i = 0; i < active_nr; i++) {
401                         const struct cache_entry *ce = active_cache[i];
402                         if (ce_stage(ce))
403                                 fprintf(stderr, "BUG: %d %.*s\n", ce_stage(ce),
404                                         (int)ce_namelen(ce), ce->name);
405                 }
406                 BUG("unmerged index entries in merge-recursive.c");
407         }
408
409         if (!active_cache_tree)
410                 active_cache_tree = cache_tree();
411
412         if (!cache_tree_fully_valid(active_cache_tree) &&
413             cache_tree_update(&the_index, 0) < 0) {
414                 err(o, _("error building trees"));
415                 return NULL;
416         }
417
418         result = lookup_tree(&active_cache_tree->oid);
419
420         return result;
421 }
422
423 static int save_files_dirs(const struct object_id *oid,
424                            struct strbuf *base, const char *path,
425                            unsigned int mode, int stage, void *context)
426 {
427         struct path_hashmap_entry *entry;
428         int baselen = base->len;
429         struct merge_options *o = context;
430
431         strbuf_addstr(base, path);
432
433         FLEX_ALLOC_MEM(entry, path, base->buf, base->len);
434         hashmap_entry_init(entry, path_hash(entry->path));
435         hashmap_add(&o->current_file_dir_set, entry);
436
437         strbuf_setlen(base, baselen);
438         return (S_ISDIR(mode) ? READ_TREE_RECURSIVE : 0);
439 }
440
441 static void get_files_dirs(struct merge_options *o, struct tree *tree)
442 {
443         struct pathspec match_all;
444         memset(&match_all, 0, sizeof(match_all));
445         read_tree_recursive(tree, "", 0, 0, &match_all, save_files_dirs, o);
446 }
447
448 static int get_tree_entry_if_blob(const struct object_id *tree,
449                                   const char *path,
450                                   struct object_id *hashy,
451                                   unsigned int *mode_o)
452 {
453         int ret;
454
455         ret = get_tree_entry(tree, path, hashy, mode_o);
456         if (S_ISDIR(*mode_o)) {
457                 oidcpy(hashy, &null_oid);
458                 *mode_o = 0;
459         }
460         return ret;
461 }
462
463 /*
464  * Returns an index_entry instance which doesn't have to correspond to
465  * a real cache entry in Git's index.
466  */
467 static struct stage_data *insert_stage_data(const char *path,
468                 struct tree *o, struct tree *a, struct tree *b,
469                 struct string_list *entries)
470 {
471         struct string_list_item *item;
472         struct stage_data *e = xcalloc(1, sizeof(struct stage_data));
473         get_tree_entry_if_blob(&o->object.oid, path,
474                                &e->stages[1].oid, &e->stages[1].mode);
475         get_tree_entry_if_blob(&a->object.oid, path,
476                                &e->stages[2].oid, &e->stages[2].mode);
477         get_tree_entry_if_blob(&b->object.oid, path,
478                                &e->stages[3].oid, &e->stages[3].mode);
479         item = string_list_insert(entries, path);
480         item->util = e;
481         return e;
482 }
483
484 /*
485  * Create a dictionary mapping file names to stage_data objects. The
486  * dictionary contains one entry for every path with a non-zero stage entry.
487  */
488 static struct string_list *get_unmerged(void)
489 {
490         struct string_list *unmerged = xcalloc(1, sizeof(struct string_list));
491         int i;
492
493         unmerged->strdup_strings = 1;
494
495         for (i = 0; i < active_nr; i++) {
496                 struct string_list_item *item;
497                 struct stage_data *e;
498                 const struct cache_entry *ce = active_cache[i];
499                 if (!ce_stage(ce))
500                         continue;
501
502                 item = string_list_lookup(unmerged, ce->name);
503                 if (!item) {
504                         item = string_list_insert(unmerged, ce->name);
505                         item->util = xcalloc(1, sizeof(struct stage_data));
506                 }
507                 e = item->util;
508                 e->stages[ce_stage(ce)].mode = ce->ce_mode;
509                 oidcpy(&e->stages[ce_stage(ce)].oid, &ce->oid);
510         }
511
512         return unmerged;
513 }
514
515 static int string_list_df_name_compare(const char *one, const char *two)
516 {
517         int onelen = strlen(one);
518         int twolen = strlen(two);
519         /*
520          * Here we only care that entries for D/F conflicts are
521          * adjacent, in particular with the file of the D/F conflict
522          * appearing before files below the corresponding directory.
523          * The order of the rest of the list is irrelevant for us.
524          *
525          * To achieve this, we sort with df_name_compare and provide
526          * the mode S_IFDIR so that D/F conflicts will sort correctly.
527          * We use the mode S_IFDIR for everything else for simplicity,
528          * since in other cases any changes in their order due to
529          * sorting cause no problems for us.
530          */
531         int cmp = df_name_compare(one, onelen, S_IFDIR,
532                                   two, twolen, S_IFDIR);
533         /*
534          * Now that 'foo' and 'foo/bar' compare equal, we have to make sure
535          * that 'foo' comes before 'foo/bar'.
536          */
537         if (cmp)
538                 return cmp;
539         return onelen - twolen;
540 }
541
542 static void record_df_conflict_files(struct merge_options *o,
543                                      struct string_list *entries)
544 {
545         /* If there is a D/F conflict and the file for such a conflict
546          * currently exists in the working tree, we want to allow it to be
547          * removed to make room for the corresponding directory if needed.
548          * The files underneath the directories of such D/F conflicts will
549          * be processed before the corresponding file involved in the D/F
550          * conflict.  If the D/F directory ends up being removed by the
551          * merge, then we won't have to touch the D/F file.  If the D/F
552          * directory needs to be written to the working copy, then the D/F
553          * file will simply be removed (in make_room_for_path()) to make
554          * room for the necessary paths.  Note that if both the directory
555          * and the file need to be present, then the D/F file will be
556          * reinstated with a new unique name at the time it is processed.
557          */
558         struct string_list df_sorted_entries = STRING_LIST_INIT_NODUP;
559         const char *last_file = NULL;
560         int last_len = 0;
561         int i;
562
563         /*
564          * If we're merging merge-bases, we don't want to bother with
565          * any working directory changes.
566          */
567         if (o->call_depth)
568                 return;
569
570         /* Ensure D/F conflicts are adjacent in the entries list. */
571         for (i = 0; i < entries->nr; i++) {
572                 struct string_list_item *next = &entries->items[i];
573                 string_list_append(&df_sorted_entries, next->string)->util =
574                                    next->util;
575         }
576         df_sorted_entries.cmp = string_list_df_name_compare;
577         string_list_sort(&df_sorted_entries);
578
579         string_list_clear(&o->df_conflict_file_set, 1);
580         for (i = 0; i < df_sorted_entries.nr; i++) {
581                 const char *path = df_sorted_entries.items[i].string;
582                 int len = strlen(path);
583                 struct stage_data *e = df_sorted_entries.items[i].util;
584
585                 /*
586                  * Check if last_file & path correspond to a D/F conflict;
587                  * i.e. whether path is last_file+'/'+<something>.
588                  * If so, record that it's okay to remove last_file to make
589                  * room for path and friends if needed.
590                  */
591                 if (last_file &&
592                     len > last_len &&
593                     memcmp(path, last_file, last_len) == 0 &&
594                     path[last_len] == '/') {
595                         string_list_insert(&o->df_conflict_file_set, last_file);
596                 }
597
598                 /*
599                  * Determine whether path could exist as a file in the
600                  * working directory as a possible D/F conflict.  This
601                  * will only occur when it exists in stage 2 as a
602                  * file.
603                  */
604                 if (S_ISREG(e->stages[2].mode) || S_ISLNK(e->stages[2].mode)) {
605                         last_file = path;
606                         last_len = len;
607                 } else {
608                         last_file = NULL;
609                 }
610         }
611         string_list_clear(&df_sorted_entries, 0);
612 }
613
614 struct rename {
615         struct diff_filepair *pair;
616         /*
617          * Purpose of src_entry and dst_entry:
618          *
619          * If 'before' is renamed to 'after' then src_entry will contain
620          * the versions of 'before' from the merge_base, HEAD, and MERGE in
621          * stages 1, 2, and 3; dst_entry will contain the respective
622          * versions of 'after' in corresponding locations.  Thus, we have a
623          * total of six modes and oids, though some will be null.  (Stage 0
624          * is ignored; we're interested in handling conflicts.)
625          *
626          * Since we don't turn on break-rewrites by default, neither
627          * src_entry nor dst_entry can have all three of their stages have
628          * non-null oids, meaning at most four of the six will be non-null.
629          * Also, since this is a rename, both src_entry and dst_entry will
630          * have at least one non-null oid, meaning at least two will be
631          * non-null.  Of the six oids, a typical rename will have three be
632          * non-null.  Only two implies a rename/delete, and four implies a
633          * rename/add.
634          */
635         struct stage_data *src_entry;
636         struct stage_data *dst_entry;
637         unsigned add_turned_into_rename:1;
638         unsigned processed:1;
639 };
640
641 static int update_stages(struct merge_options *opt, const char *path,
642                          const struct diff_filespec *o,
643                          const struct diff_filespec *a,
644                          const struct diff_filespec *b)
645 {
646
647         /*
648          * NOTE: It is usually a bad idea to call update_stages on a path
649          * before calling update_file on that same path, since it can
650          * sometimes lead to spurious "refusing to lose untracked file..."
651          * messages from update_file (via make_room_for path via
652          * would_lose_untracked).  Instead, reverse the order of the calls
653          * (executing update_file first and then update_stages).
654          */
655         int clear = 1;
656         int options = ADD_CACHE_OK_TO_ADD | ADD_CACHE_SKIP_DFCHECK;
657         if (clear)
658                 if (remove_file_from_cache(path))
659                         return -1;
660         if (o)
661                 if (add_cacheinfo(opt, o->mode, &o->oid, path, 1, 0, options))
662                         return -1;
663         if (a)
664                 if (add_cacheinfo(opt, a->mode, &a->oid, path, 2, 0, options))
665                         return -1;
666         if (b)
667                 if (add_cacheinfo(opt, b->mode, &b->oid, path, 3, 0, options))
668                         return -1;
669         return 0;
670 }
671
672 static int update_stages_for_stage_data(struct merge_options *opt,
673                                         const char *path,
674                                         const struct stage_data *stage_data)
675 {
676         struct diff_filespec o, a, b;
677
678         o.mode = stage_data->stages[1].mode;
679         oidcpy(&o.oid, &stage_data->stages[1].oid);
680
681         a.mode = stage_data->stages[2].mode;
682         oidcpy(&a.oid, &stage_data->stages[2].oid);
683
684         b.mode = stage_data->stages[3].mode;
685         oidcpy(&b.oid, &stage_data->stages[3].oid);
686
687         return update_stages(opt, path,
688                              is_null_oid(&o.oid) ? NULL : &o,
689                              is_null_oid(&a.oid) ? NULL : &a,
690                              is_null_oid(&b.oid) ? NULL : &b);
691 }
692
693 static void update_entry(struct stage_data *entry,
694                          struct diff_filespec *o,
695                          struct diff_filespec *a,
696                          struct diff_filespec *b)
697 {
698         entry->processed = 0;
699         entry->stages[1].mode = o->mode;
700         entry->stages[2].mode = a->mode;
701         entry->stages[3].mode = b->mode;
702         oidcpy(&entry->stages[1].oid, &o->oid);
703         oidcpy(&entry->stages[2].oid, &a->oid);
704         oidcpy(&entry->stages[3].oid, &b->oid);
705 }
706
707 static int remove_file(struct merge_options *o, int clean,
708                        const char *path, int no_wd)
709 {
710         int update_cache = o->call_depth || clean;
711         int update_working_directory = !o->call_depth && !no_wd;
712
713         if (update_cache) {
714                 if (remove_file_from_cache(path))
715                         return -1;
716         }
717         if (update_working_directory) {
718                 if (ignore_case) {
719                         struct cache_entry *ce;
720                         ce = cache_file_exists(path, strlen(path), ignore_case);
721                         if (ce && ce_stage(ce) == 0 && strcmp(path, ce->name))
722                                 return 0;
723                 }
724                 if (remove_path(path))
725                         return -1;
726         }
727         return 0;
728 }
729
730 /* add a string to a strbuf, but converting "/" to "_" */
731 static void add_flattened_path(struct strbuf *out, const char *s)
732 {
733         size_t i = out->len;
734         strbuf_addstr(out, s);
735         for (; i < out->len; i++)
736                 if (out->buf[i] == '/')
737                         out->buf[i] = '_';
738 }
739
740 static char *unique_path(struct merge_options *o, const char *path, const char *branch)
741 {
742         struct path_hashmap_entry *entry;
743         struct strbuf newpath = STRBUF_INIT;
744         int suffix = 0;
745         size_t base_len;
746
747         strbuf_addf(&newpath, "%s~", path);
748         add_flattened_path(&newpath, branch);
749
750         base_len = newpath.len;
751         while (hashmap_get_from_hash(&o->current_file_dir_set,
752                                      path_hash(newpath.buf), newpath.buf) ||
753                (!o->call_depth && file_exists(newpath.buf))) {
754                 strbuf_setlen(&newpath, base_len);
755                 strbuf_addf(&newpath, "_%d", suffix++);
756         }
757
758         FLEX_ALLOC_MEM(entry, path, newpath.buf, newpath.len);
759         hashmap_entry_init(entry, path_hash(entry->path));
760         hashmap_add(&o->current_file_dir_set, entry);
761         return strbuf_detach(&newpath, NULL);
762 }
763
764 /**
765  * Check whether a directory in the index is in the way of an incoming
766  * file.  Return 1 if so.  If check_working_copy is non-zero, also
767  * check the working directory.  If empty_ok is non-zero, also return
768  * 0 in the case where the working-tree dir exists but is empty.
769  */
770 static int dir_in_way(const char *path, int check_working_copy, int empty_ok)
771 {
772         int pos;
773         struct strbuf dirpath = STRBUF_INIT;
774         struct stat st;
775
776         strbuf_addstr(&dirpath, path);
777         strbuf_addch(&dirpath, '/');
778
779         pos = cache_name_pos(dirpath.buf, dirpath.len);
780
781         if (pos < 0)
782                 pos = -1 - pos;
783         if (pos < active_nr &&
784             !strncmp(dirpath.buf, active_cache[pos]->name, dirpath.len)) {
785                 strbuf_release(&dirpath);
786                 return 1;
787         }
788
789         strbuf_release(&dirpath);
790         return check_working_copy && !lstat(path, &st) && S_ISDIR(st.st_mode) &&
791                 !(empty_ok && is_empty_dir(path));
792 }
793
794 /*
795  * Returns whether path was tracked in the index before the merge started,
796  * and its oid and mode match the specified values
797  */
798 static int was_tracked_and_matches(struct merge_options *o, const char *path,
799                                    const struct object_id *oid, unsigned mode)
800 {
801         int pos = index_name_pos(&o->orig_index, path, strlen(path));
802         struct cache_entry *ce;
803
804         if (0 > pos)
805                 /* we were not tracking this path before the merge */
806                 return 0;
807
808         /* See if the file we were tracking before matches */
809         ce = o->orig_index.cache[pos];
810         return (oid_eq(&ce->oid, oid) && ce->ce_mode == mode);
811 }
812
813 /*
814  * Returns whether path was tracked in the index before the merge started
815  */
816 static int was_tracked(struct merge_options *o, const char *path)
817 {
818         int pos = index_name_pos(&o->orig_index, path, strlen(path));
819
820         if (0 <= pos)
821                 /* we were tracking this path before the merge */
822                 return 1;
823
824         return 0;
825 }
826
827 static int would_lose_untracked(const char *path)
828 {
829         /*
830          * This may look like it can be simplified to:
831          *   return !was_tracked(o, path) && file_exists(path)
832          * but it can't.  This function needs to know whether path was in
833          * the working tree due to EITHER having been tracked in the index
834          * before the merge OR having been put into the working copy and
835          * index by unpack_trees().  Due to that either-or requirement, we
836          * check the current index instead of the original one.
837          *
838          * Note that we do not need to worry about merge-recursive itself
839          * updating the index after unpack_trees() and before calling this
840          * function, because we strictly require all code paths in
841          * merge-recursive to update the working tree first and the index
842          * second.  Doing otherwise would break
843          * update_file()/would_lose_untracked(); see every comment in this
844          * file which mentions "update_stages".
845          */
846         int pos = cache_name_pos(path, strlen(path));
847
848         if (pos < 0)
849                 pos = -1 - pos;
850         while (pos < active_nr &&
851                !strcmp(path, active_cache[pos]->name)) {
852                 /*
853                  * If stage #0, it is definitely tracked.
854                  * If it has stage #2 then it was tracked
855                  * before this merge started.  All other
856                  * cases the path was not tracked.
857                  */
858                 switch (ce_stage(active_cache[pos])) {
859                 case 0:
860                 case 2:
861                         return 0;
862                 }
863                 pos++;
864         }
865         return file_exists(path);
866 }
867
868 static int was_dirty(struct merge_options *o, const char *path)
869 {
870         struct cache_entry *ce;
871         int dirty = 1;
872
873         if (o->call_depth || !was_tracked(o, path))
874                 return !dirty;
875
876         ce = index_file_exists(o->unpack_opts.src_index,
877                                path, strlen(path), ignore_case);
878         dirty = verify_uptodate(ce, &o->unpack_opts) != 0;
879         return dirty;
880 }
881
882 static int make_room_for_path(struct merge_options *o, const char *path)
883 {
884         int status, i;
885         const char *msg = _("failed to create path '%s'%s");
886
887         /* Unlink any D/F conflict files that are in the way */
888         for (i = 0; i < o->df_conflict_file_set.nr; i++) {
889                 const char *df_path = o->df_conflict_file_set.items[i].string;
890                 size_t pathlen = strlen(path);
891                 size_t df_pathlen = strlen(df_path);
892                 if (df_pathlen < pathlen &&
893                     path[df_pathlen] == '/' &&
894                     strncmp(path, df_path, df_pathlen) == 0) {
895                         output(o, 3,
896                                _("Removing %s to make room for subdirectory\n"),
897                                df_path);
898                         unlink(df_path);
899                         unsorted_string_list_delete_item(&o->df_conflict_file_set,
900                                                          i, 0);
901                         break;
902                 }
903         }
904
905         /* Make sure leading directories are created */
906         status = safe_create_leading_directories_const(path);
907         if (status) {
908                 if (status == SCLD_EXISTS)
909                         /* something else exists */
910                         return err(o, msg, path, _(": perhaps a D/F conflict?"));
911                 return err(o, msg, path, "");
912         }
913
914         /*
915          * Do not unlink a file in the work tree if we are not
916          * tracking it.
917          */
918         if (would_lose_untracked(path))
919                 return err(o, _("refusing to lose untracked file at '%s'"),
920                            path);
921
922         /* Successful unlink is good.. */
923         if (!unlink(path))
924                 return 0;
925         /* .. and so is no existing file */
926         if (errno == ENOENT)
927                 return 0;
928         /* .. but not some other error (who really cares what?) */
929         return err(o, msg, path, _(": perhaps a D/F conflict?"));
930 }
931
932 static int update_file_flags(struct merge_options *o,
933                              const struct object_id *oid,
934                              unsigned mode,
935                              const char *path,
936                              int update_cache,
937                              int update_wd)
938 {
939         int ret = 0;
940
941         if (o->call_depth)
942                 update_wd = 0;
943
944         if (update_wd) {
945                 enum object_type type;
946                 void *buf;
947                 unsigned long size;
948
949                 if (S_ISGITLINK(mode)) {
950                         /*
951                          * We may later decide to recursively descend into
952                          * the submodule directory and update its index
953                          * and/or work tree, but we do not do that now.
954                          */
955                         update_wd = 0;
956                         goto update_index;
957                 }
958
959                 buf = read_object_file(oid, &type, &size);
960                 if (!buf)
961                         return err(o, _("cannot read object %s '%s'"), oid_to_hex(oid), path);
962                 if (type != OBJ_BLOB) {
963                         ret = err(o, _("blob expected for %s '%s'"), oid_to_hex(oid), path);
964                         goto free_buf;
965                 }
966                 if (S_ISREG(mode)) {
967                         struct strbuf strbuf = STRBUF_INIT;
968                         if (convert_to_working_tree(path, buf, size, &strbuf)) {
969                                 free(buf);
970                                 size = strbuf.len;
971                                 buf = strbuf_detach(&strbuf, NULL);
972                         }
973                 }
974
975                 if (make_room_for_path(o, path) < 0) {
976                         update_wd = 0;
977                         goto free_buf;
978                 }
979                 if (S_ISREG(mode) || (!has_symlinks && S_ISLNK(mode))) {
980                         int fd;
981                         if (mode & 0100)
982                                 mode = 0777;
983                         else
984                                 mode = 0666;
985                         fd = open(path, O_WRONLY | O_TRUNC | O_CREAT, mode);
986                         if (fd < 0) {
987                                 ret = err(o, _("failed to open '%s': %s"),
988                                           path, strerror(errno));
989                                 goto free_buf;
990                         }
991                         write_in_full(fd, buf, size);
992                         close(fd);
993                 } else if (S_ISLNK(mode)) {
994                         char *lnk = xmemdupz(buf, size);
995                         safe_create_leading_directories_const(path);
996                         unlink(path);
997                         if (symlink(lnk, path))
998                                 ret = err(o, _("failed to symlink '%s': %s"),
999                                           path, strerror(errno));
1000                         free(lnk);
1001                 } else
1002                         ret = err(o,
1003                                   _("do not know what to do with %06o %s '%s'"),
1004                                   mode, oid_to_hex(oid), path);
1005         free_buf:
1006                 free(buf);
1007         }
1008 update_index:
1009         if (!ret && update_cache)
1010                 if (add_cacheinfo(o, mode, oid, path, 0, update_wd,
1011                                   ADD_CACHE_OK_TO_ADD))
1012                         return -1;
1013         return ret;
1014 }
1015
1016 static int update_file(struct merge_options *o,
1017                        int clean,
1018                        const struct object_id *oid,
1019                        unsigned mode,
1020                        const char *path)
1021 {
1022         return update_file_flags(o, oid, mode, path, o->call_depth || clean, !o->call_depth);
1023 }
1024
1025 /* Low level file merging, update and removal */
1026
1027 struct merge_file_info {
1028         struct object_id oid;
1029         unsigned mode;
1030         unsigned clean:1,
1031                  merge:1;
1032 };
1033
1034 static int merge_3way(struct merge_options *o,
1035                       mmbuffer_t *result_buf,
1036                       const struct diff_filespec *one,
1037                       const struct diff_filespec *a,
1038                       const struct diff_filespec *b,
1039                       const char *branch1,
1040                       const char *branch2)
1041 {
1042         mmfile_t orig, src1, src2;
1043         struct ll_merge_options ll_opts = {0};
1044         char *base_name, *name1, *name2;
1045         int merge_status;
1046
1047         ll_opts.renormalize = o->renormalize;
1048         ll_opts.xdl_opts = o->xdl_opts;
1049
1050         if (o->call_depth) {
1051                 ll_opts.virtual_ancestor = 1;
1052                 ll_opts.variant = 0;
1053         } else {
1054                 switch (o->recursive_variant) {
1055                 case MERGE_RECURSIVE_OURS:
1056                         ll_opts.variant = XDL_MERGE_FAVOR_OURS;
1057                         break;
1058                 case MERGE_RECURSIVE_THEIRS:
1059                         ll_opts.variant = XDL_MERGE_FAVOR_THEIRS;
1060                         break;
1061                 default:
1062                         ll_opts.variant = 0;
1063                         break;
1064                 }
1065         }
1066
1067         if (strcmp(a->path, b->path) ||
1068             (o->ancestor != NULL && strcmp(a->path, one->path) != 0)) {
1069                 base_name = o->ancestor == NULL ? NULL :
1070                         mkpathdup("%s:%s", o->ancestor, one->path);
1071                 name1 = mkpathdup("%s:%s", branch1, a->path);
1072                 name2 = mkpathdup("%s:%s", branch2, b->path);
1073         } else {
1074                 base_name = o->ancestor == NULL ? NULL :
1075                         mkpathdup("%s", o->ancestor);
1076                 name1 = mkpathdup("%s", branch1);
1077                 name2 = mkpathdup("%s", branch2);
1078         }
1079
1080         read_mmblob(&orig, &one->oid);
1081         read_mmblob(&src1, &a->oid);
1082         read_mmblob(&src2, &b->oid);
1083
1084         merge_status = ll_merge(result_buf, a->path, &orig, base_name,
1085                                 &src1, name1, &src2, name2, &ll_opts);
1086
1087         free(base_name);
1088         free(name1);
1089         free(name2);
1090         free(orig.ptr);
1091         free(src1.ptr);
1092         free(src2.ptr);
1093         return merge_status;
1094 }
1095
1096 static int find_first_merges(struct object_array *result, const char *path,
1097                              struct commit *a, struct commit *b)
1098 {
1099         int i, j;
1100         struct object_array merges = OBJECT_ARRAY_INIT;
1101         struct commit *commit;
1102         int contains_another;
1103
1104         char merged_revision[42];
1105         const char *rev_args[] = { "rev-list", "--merges", "--ancestry-path",
1106                                    "--all", merged_revision, NULL };
1107         struct rev_info revs;
1108         struct setup_revision_opt rev_opts;
1109
1110         memset(result, 0, sizeof(struct object_array));
1111         memset(&rev_opts, 0, sizeof(rev_opts));
1112
1113         /* get all revisions that merge commit a */
1114         xsnprintf(merged_revision, sizeof(merged_revision), "^%s",
1115                   oid_to_hex(&a->object.oid));
1116         init_revisions(&revs, NULL);
1117         rev_opts.submodule = path;
1118         /* FIXME: can't handle linked worktrees in submodules yet */
1119         revs.single_worktree = path != NULL;
1120         setup_revisions(ARRAY_SIZE(rev_args)-1, rev_args, &revs, &rev_opts);
1121
1122         /* save all revisions from the above list that contain b */
1123         if (prepare_revision_walk(&revs))
1124                 die("revision walk setup failed");
1125         while ((commit = get_revision(&revs)) != NULL) {
1126                 struct object *o = &(commit->object);
1127                 if (in_merge_bases(b, commit))
1128                         add_object_array(o, NULL, &merges);
1129         }
1130         reset_revision_walk();
1131
1132         /* Now we've got all merges that contain a and b. Prune all
1133          * merges that contain another found merge and save them in
1134          * result.
1135          */
1136         for (i = 0; i < merges.nr; i++) {
1137                 struct commit *m1 = (struct commit *) merges.objects[i].item;
1138
1139                 contains_another = 0;
1140                 for (j = 0; j < merges.nr; j++) {
1141                         struct commit *m2 = (struct commit *) merges.objects[j].item;
1142                         if (i != j && in_merge_bases(m2, m1)) {
1143                                 contains_another = 1;
1144                                 break;
1145                         }
1146                 }
1147
1148                 if (!contains_another)
1149                         add_object_array(merges.objects[i].item, NULL, result);
1150         }
1151
1152         object_array_clear(&merges);
1153         return result->nr;
1154 }
1155
1156 static void print_commit(struct commit *commit)
1157 {
1158         struct strbuf sb = STRBUF_INIT;
1159         struct pretty_print_context ctx = {0};
1160         ctx.date_mode.type = DATE_NORMAL;
1161         format_commit_message(commit, " %h: %m %s", &sb, &ctx);
1162         fprintf(stderr, "%s\n", sb.buf);
1163         strbuf_release(&sb);
1164 }
1165
1166 static int merge_submodule(struct merge_options *o,
1167                            struct object_id *result, const char *path,
1168                            const struct object_id *base, const struct object_id *a,
1169                            const struct object_id *b)
1170 {
1171         struct commit *commit_base, *commit_a, *commit_b;
1172         int parent_count;
1173         struct object_array merges;
1174
1175         int i;
1176         int search = !o->call_depth;
1177
1178         /* store a in result in case we fail */
1179         oidcpy(result, a);
1180
1181         /* we can not handle deletion conflicts */
1182         if (is_null_oid(base))
1183                 return 0;
1184         if (is_null_oid(a))
1185                 return 0;
1186         if (is_null_oid(b))
1187                 return 0;
1188
1189         if (add_submodule_odb(path)) {
1190                 output(o, 1, _("Failed to merge submodule %s (not checked out)"), path);
1191                 return 0;
1192         }
1193
1194         if (!(commit_base = lookup_commit_reference(base)) ||
1195             !(commit_a = lookup_commit_reference(a)) ||
1196             !(commit_b = lookup_commit_reference(b))) {
1197                 output(o, 1, _("Failed to merge submodule %s (commits not present)"), path);
1198                 return 0;
1199         }
1200
1201         /* check whether both changes are forward */
1202         if (!in_merge_bases(commit_base, commit_a) ||
1203             !in_merge_bases(commit_base, commit_b)) {
1204                 output(o, 1, _("Failed to merge submodule %s (commits don't follow merge-base)"), path);
1205                 return 0;
1206         }
1207
1208         /* Case #1: a is contained in b or vice versa */
1209         if (in_merge_bases(commit_a, commit_b)) {
1210                 oidcpy(result, b);
1211                 if (show(o, 3)) {
1212                         output(o, 3, _("Fast-forwarding submodule %s to the following commit:"), path);
1213                         output_commit_title(o, commit_b);
1214                 } else if (show(o, 2))
1215                         output(o, 2, _("Fast-forwarding submodule %s"), path);
1216                 else
1217                         ; /* no output */
1218
1219                 return 1;
1220         }
1221         if (in_merge_bases(commit_b, commit_a)) {
1222                 oidcpy(result, a);
1223                 if (show(o, 3)) {
1224                         output(o, 3, _("Fast-forwarding submodule %s to the following commit:"), path);
1225                         output_commit_title(o, commit_a);
1226                 } else if (show(o, 2))
1227                         output(o, 2, _("Fast-forwarding submodule %s"), path);
1228                 else
1229                         ; /* no output */
1230
1231                 return 1;
1232         }
1233
1234         /*
1235          * Case #2: There are one or more merges that contain a and b in
1236          * the submodule. If there is only one, then present it as a
1237          * suggestion to the user, but leave it marked unmerged so the
1238          * user needs to confirm the resolution.
1239          */
1240
1241         /* Skip the search if makes no sense to the calling context.  */
1242         if (!search)
1243                 return 0;
1244
1245         /* find commit which merges them */
1246         parent_count = find_first_merges(&merges, path, commit_a, commit_b);
1247         switch (parent_count) {
1248         case 0:
1249                 output(o, 1, _("Failed to merge submodule %s (merge following commits not found)"), path);
1250                 break;
1251
1252         case 1:
1253                 output(o, 1, _("Failed to merge submodule %s (not fast-forward)"), path);
1254                 output(o, 2, _("Found a possible merge resolution for the submodule:\n"));
1255                 print_commit((struct commit *) merges.objects[0].item);
1256                 output(o, 2, _(
1257                        "If this is correct simply add it to the index "
1258                        "for example\n"
1259                        "by using:\n\n"
1260                        "  git update-index --cacheinfo 160000 %s \"%s\"\n\n"
1261                        "which will accept this suggestion.\n"),
1262                        oid_to_hex(&merges.objects[0].item->oid), path);
1263                 break;
1264
1265         default:
1266                 output(o, 1, _("Failed to merge submodule %s (multiple merges found)"), path);
1267                 for (i = 0; i < merges.nr; i++)
1268                         print_commit((struct commit *) merges.objects[i].item);
1269         }
1270
1271         object_array_clear(&merges);
1272         return 0;
1273 }
1274
1275 static int merge_file_1(struct merge_options *o,
1276                         const struct diff_filespec *one,
1277                         const struct diff_filespec *a,
1278                         const struct diff_filespec *b,
1279                         const char *filename,
1280                         const char *branch1,
1281                         const char *branch2,
1282                         struct merge_file_info *result)
1283 {
1284         result->merge = 0;
1285         result->clean = 1;
1286
1287         if ((S_IFMT & a->mode) != (S_IFMT & b->mode)) {
1288                 result->clean = 0;
1289                 if (S_ISREG(a->mode)) {
1290                         result->mode = a->mode;
1291                         oidcpy(&result->oid, &a->oid);
1292                 } else {
1293                         result->mode = b->mode;
1294                         oidcpy(&result->oid, &b->oid);
1295                 }
1296         } else {
1297                 if (!oid_eq(&a->oid, &one->oid) && !oid_eq(&b->oid, &one->oid))
1298                         result->merge = 1;
1299
1300                 /*
1301                  * Merge modes
1302                  */
1303                 if (a->mode == b->mode || a->mode == one->mode)
1304                         result->mode = b->mode;
1305                 else {
1306                         result->mode = a->mode;
1307                         if (b->mode != one->mode) {
1308                                 result->clean = 0;
1309                                 result->merge = 1;
1310                         }
1311                 }
1312
1313                 if (oid_eq(&a->oid, &b->oid) || oid_eq(&a->oid, &one->oid))
1314                         oidcpy(&result->oid, &b->oid);
1315                 else if (oid_eq(&b->oid, &one->oid))
1316                         oidcpy(&result->oid, &a->oid);
1317                 else if (S_ISREG(a->mode)) {
1318                         mmbuffer_t result_buf;
1319                         int ret = 0, merge_status;
1320
1321                         merge_status = merge_3way(o, &result_buf, one, a, b,
1322                                                   branch1, branch2);
1323
1324                         if ((merge_status < 0) || !result_buf.ptr)
1325                                 ret = err(o, _("Failed to execute internal merge"));
1326
1327                         if (!ret &&
1328                             write_object_file(result_buf.ptr, result_buf.size,
1329                                               blob_type, &result->oid))
1330                                 ret = err(o, _("Unable to add %s to database"),
1331                                           a->path);
1332
1333                         free(result_buf.ptr);
1334                         if (ret)
1335                                 return ret;
1336                         result->clean = (merge_status == 0);
1337                 } else if (S_ISGITLINK(a->mode)) {
1338                         result->clean = merge_submodule(o, &result->oid,
1339                                                         one->path,
1340                                                         &one->oid,
1341                                                         &a->oid,
1342                                                         &b->oid);
1343                 } else if (S_ISLNK(a->mode)) {
1344                         switch (o->recursive_variant) {
1345                         case MERGE_RECURSIVE_NORMAL:
1346                                 oidcpy(&result->oid, &a->oid);
1347                                 if (!oid_eq(&a->oid, &b->oid))
1348                                         result->clean = 0;
1349                                 break;
1350                         case MERGE_RECURSIVE_OURS:
1351                                 oidcpy(&result->oid, &a->oid);
1352                                 break;
1353                         case MERGE_RECURSIVE_THEIRS:
1354                                 oidcpy(&result->oid, &b->oid);
1355                                 break;
1356                         }
1357                 } else
1358                         BUG("unsupported object type in the tree");
1359         }
1360
1361         if (result->merge)
1362                 output(o, 2, _("Auto-merging %s"), filename);
1363
1364         return 0;
1365 }
1366
1367 static int merge_file_special_markers(struct merge_options *o,
1368                                       const struct diff_filespec *one,
1369                                       const struct diff_filespec *a,
1370                                       const struct diff_filespec *b,
1371                                       const char *target_filename,
1372                                       const char *branch1,
1373                                       const char *filename1,
1374                                       const char *branch2,
1375                                       const char *filename2,
1376                                       struct merge_file_info *mfi)
1377 {
1378         char *side1 = NULL;
1379         char *side2 = NULL;
1380         int ret;
1381
1382         if (filename1)
1383                 side1 = xstrfmt("%s:%s", branch1, filename1);
1384         if (filename2)
1385                 side2 = xstrfmt("%s:%s", branch2, filename2);
1386
1387         ret = merge_file_1(o, one, a, b, target_filename,
1388                            side1 ? side1 : branch1,
1389                            side2 ? side2 : branch2, mfi);
1390
1391         free(side1);
1392         free(side2);
1393         return ret;
1394 }
1395
1396 static int merge_file_one(struct merge_options *o,
1397                           const char *path,
1398                           const struct object_id *o_oid, int o_mode,
1399                           const struct object_id *a_oid, int a_mode,
1400                           const struct object_id *b_oid, int b_mode,
1401                           const char *branch1,
1402                           const char *branch2,
1403                           struct merge_file_info *mfi)
1404 {
1405         struct diff_filespec one, a, b;
1406
1407         one.path = a.path = b.path = (char *)path;
1408         oidcpy(&one.oid, o_oid);
1409         one.mode = o_mode;
1410         oidcpy(&a.oid, a_oid);
1411         a.mode = a_mode;
1412         oidcpy(&b.oid, b_oid);
1413         b.mode = b_mode;
1414         return merge_file_1(o, &one, &a, &b, path, branch1, branch2, mfi);
1415 }
1416
1417 static int handle_rename_via_dir(struct merge_options *o,
1418                                  struct diff_filepair *pair,
1419                                  const char *rename_branch,
1420                                  const char *other_branch)
1421 {
1422         /*
1423          * Handle file adds that need to be renamed due to directory rename
1424          * detection.  This differs from handle_rename_normal, because
1425          * there is no content merge to do; just move the file into the
1426          * desired final location.
1427          */
1428         const struct diff_filespec *dest = pair->two;
1429
1430         if (!o->call_depth && would_lose_untracked(dest->path)) {
1431                 char *alt_path = unique_path(o, dest->path, rename_branch);
1432
1433                 output(o, 1, _("Error: Refusing to lose untracked file at %s; "
1434                                "writing to %s instead."),
1435                        dest->path, alt_path);
1436                 /*
1437                  * Write the file in worktree at alt_path, but not in the
1438                  * index.  Instead, write to dest->path for the index but
1439                  * only at the higher appropriate stage.
1440                  */
1441                 if (update_file(o, 0, &dest->oid, dest->mode, alt_path))
1442                         return -1;
1443                 free(alt_path);
1444                 return update_stages(o, dest->path, NULL,
1445                                      rename_branch == o->branch1 ? dest : NULL,
1446                                      rename_branch == o->branch1 ? NULL : dest);
1447         }
1448
1449         /* Update dest->path both in index and in worktree */
1450         if (update_file(o, 1, &dest->oid, dest->mode, dest->path))
1451                 return -1;
1452         return 0;
1453 }
1454
1455 static int handle_change_delete(struct merge_options *o,
1456                                 const char *path, const char *old_path,
1457                                 const struct object_id *o_oid, int o_mode,
1458                                 const struct object_id *changed_oid,
1459                                 int changed_mode,
1460                                 const char *change_branch,
1461                                 const char *delete_branch,
1462                                 const char *change, const char *change_past)
1463 {
1464         char *alt_path = NULL;
1465         const char *update_path = path;
1466         int ret = 0;
1467
1468         if (dir_in_way(path, !o->call_depth, 0) ||
1469             (!o->call_depth && would_lose_untracked(path))) {
1470                 update_path = alt_path = unique_path(o, path, change_branch);
1471         }
1472
1473         if (o->call_depth) {
1474                 /*
1475                  * We cannot arbitrarily accept either a_sha or b_sha as
1476                  * correct; since there is no true "middle point" between
1477                  * them, simply reuse the base version for virtual merge base.
1478                  */
1479                 ret = remove_file_from_cache(path);
1480                 if (!ret)
1481                         ret = update_file(o, 0, o_oid, o_mode, update_path);
1482         } else {
1483                 /*
1484                  * Despite the four nearly duplicate messages and argument
1485                  * lists below and the ugliness of the nested if-statements,
1486                  * having complete messages makes the job easier for
1487                  * translators.
1488                  *
1489                  * The slight variance among the cases is due to the fact
1490                  * that:
1491                  *   1) directory/file conflicts (in effect if
1492                  *      !alt_path) could cause us to need to write the
1493                  *      file to a different path.
1494                  *   2) renames (in effect if !old_path) could mean that
1495                  *      there are two names for the path that the user
1496                  *      may know the file by.
1497                  */
1498                 if (!alt_path) {
1499                         if (!old_path) {
1500                                 output(o, 1, _("CONFLICT (%s/delete): %s deleted in %s "
1501                                        "and %s in %s. Version %s of %s left in tree."),
1502                                        change, path, delete_branch, change_past,
1503                                        change_branch, change_branch, path);
1504                         } else {
1505                                 output(o, 1, _("CONFLICT (%s/delete): %s deleted in %s "
1506                                        "and %s to %s in %s. Version %s of %s left in tree."),
1507                                        change, old_path, delete_branch, change_past, path,
1508                                        change_branch, change_branch, path);
1509                         }
1510                 } else {
1511                         if (!old_path) {
1512                                 output(o, 1, _("CONFLICT (%s/delete): %s deleted in %s "
1513                                        "and %s in %s. Version %s of %s left in tree at %s."),
1514                                        change, path, delete_branch, change_past,
1515                                        change_branch, change_branch, path, alt_path);
1516                         } else {
1517                                 output(o, 1, _("CONFLICT (%s/delete): %s deleted in %s "
1518                                        "and %s to %s in %s. Version %s of %s left in tree at %s."),
1519                                        change, old_path, delete_branch, change_past, path,
1520                                        change_branch, change_branch, path, alt_path);
1521                         }
1522                 }
1523                 /*
1524                  * No need to call update_file() on path when change_branch ==
1525                  * o->branch1 && !alt_path, since that would needlessly touch
1526                  * path.  We could call update_file_flags() with update_cache=0
1527                  * and update_wd=0, but that's a no-op.
1528                  */
1529                 if (change_branch != o->branch1 || alt_path)
1530                         ret = update_file(o, 0, changed_oid, changed_mode, update_path);
1531         }
1532         free(alt_path);
1533
1534         return ret;
1535 }
1536
1537 static int handle_rename_delete(struct merge_options *o,
1538                                 struct diff_filepair *pair,
1539                                 const char *rename_branch,
1540                                 const char *delete_branch)
1541 {
1542         const struct diff_filespec *orig = pair->one;
1543         const struct diff_filespec *dest = pair->two;
1544
1545         if (handle_change_delete(o,
1546                                  o->call_depth ? orig->path : dest->path,
1547                                  o->call_depth ? NULL : orig->path,
1548                                  &orig->oid, orig->mode,
1549                                  &dest->oid, dest->mode,
1550                                  rename_branch, delete_branch,
1551                                  _("rename"), _("renamed")))
1552                 return -1;
1553
1554         if (o->call_depth)
1555                 return remove_file_from_cache(dest->path);
1556         else
1557                 return update_stages(o, dest->path, NULL,
1558                                      rename_branch == o->branch1 ? dest : NULL,
1559                                      rename_branch == o->branch1 ? NULL : dest);
1560 }
1561
1562 static struct diff_filespec *filespec_from_entry(struct diff_filespec *target,
1563                                                  struct stage_data *entry,
1564                                                  int stage)
1565 {
1566         struct object_id *oid = &entry->stages[stage].oid;
1567         unsigned mode = entry->stages[stage].mode;
1568         if (mode == 0 || is_null_oid(oid))
1569                 return NULL;
1570         oidcpy(&target->oid, oid);
1571         target->mode = mode;
1572         return target;
1573 }
1574
1575 static int handle_file(struct merge_options *o,
1576                         struct diff_filespec *rename,
1577                         int stage,
1578                         struct rename_conflict_info *ci)
1579 {
1580         char *dst_name = rename->path;
1581         struct stage_data *dst_entry;
1582         const char *cur_branch, *other_branch;
1583         struct diff_filespec other;
1584         struct diff_filespec *add;
1585         int ret;
1586
1587         if (stage == 2) {
1588                 dst_entry = ci->dst_entry1;
1589                 cur_branch = ci->branch1;
1590                 other_branch = ci->branch2;
1591         } else {
1592                 dst_entry = ci->dst_entry2;
1593                 cur_branch = ci->branch2;
1594                 other_branch = ci->branch1;
1595         }
1596
1597         add = filespec_from_entry(&other, dst_entry, stage ^ 1);
1598         if (add) {
1599                 int ren_src_was_dirty = was_dirty(o, rename->path);
1600                 char *add_name = unique_path(o, rename->path, other_branch);
1601                 if (update_file(o, 0, &add->oid, add->mode, add_name))
1602                         return -1;
1603
1604                 if (ren_src_was_dirty) {
1605                         output(o, 1, _("Refusing to lose dirty file at %s"),
1606                                rename->path);
1607                 }
1608                 /*
1609                  * Because the double negatives somehow keep confusing me...
1610                  *    1) update_wd iff !ren_src_was_dirty.
1611                  *    2) no_wd iff !update_wd
1612                  *    3) so, no_wd == !!ren_src_was_dirty == ren_src_was_dirty
1613                  */
1614                 remove_file(o, 0, rename->path, ren_src_was_dirty);
1615                 dst_name = unique_path(o, rename->path, cur_branch);
1616         } else {
1617                 if (dir_in_way(rename->path, !o->call_depth, 0)) {
1618                         dst_name = unique_path(o, rename->path, cur_branch);
1619                         output(o, 1, _("%s is a directory in %s adding as %s instead"),
1620                                rename->path, other_branch, dst_name);
1621                 } else if (!o->call_depth &&
1622                            would_lose_untracked(rename->path)) {
1623                         dst_name = unique_path(o, rename->path, cur_branch);
1624                         output(o, 1, _("Refusing to lose untracked file at %s; "
1625                                        "adding as %s instead"),
1626                                rename->path, dst_name);
1627                 }
1628         }
1629         if ((ret = update_file(o, 0, &rename->oid, rename->mode, dst_name)))
1630                 ; /* fall through, do allow dst_name to be released */
1631         else if (stage == 2)
1632                 ret = update_stages(o, rename->path, NULL, rename, add);
1633         else
1634                 ret = update_stages(o, rename->path, NULL, add, rename);
1635
1636         if (dst_name != rename->path)
1637                 free(dst_name);
1638
1639         return ret;
1640 }
1641
1642 static int handle_rename_rename_1to2(struct merge_options *o,
1643                                      struct rename_conflict_info *ci)
1644 {
1645         /* One file was renamed in both branches, but to different names. */
1646         struct diff_filespec *one = ci->pair1->one;
1647         struct diff_filespec *a = ci->pair1->two;
1648         struct diff_filespec *b = ci->pair2->two;
1649
1650         output(o, 1, _("CONFLICT (rename/rename): "
1651                "Rename \"%s\"->\"%s\" in branch \"%s\" "
1652                "rename \"%s\"->\"%s\" in \"%s\"%s"),
1653                one->path, a->path, ci->branch1,
1654                one->path, b->path, ci->branch2,
1655                o->call_depth ? _(" (left unresolved)") : "");
1656         if (o->call_depth) {
1657                 struct merge_file_info mfi;
1658                 struct diff_filespec other;
1659                 struct diff_filespec *add;
1660                 if (merge_file_one(o, one->path,
1661                                  &one->oid, one->mode,
1662                                  &a->oid, a->mode,
1663                                  &b->oid, b->mode,
1664                                  ci->branch1, ci->branch2, &mfi))
1665                         return -1;
1666
1667                 /*
1668                  * FIXME: For rename/add-source conflicts (if we could detect
1669                  * such), this is wrong.  We should instead find a unique
1670                  * pathname and then either rename the add-source file to that
1671                  * unique path, or use that unique path instead of src here.
1672                  */
1673                 if (update_file(o, 0, &mfi.oid, mfi.mode, one->path))
1674                         return -1;
1675
1676                 /*
1677                  * Above, we put the merged content at the merge-base's
1678                  * path.  Now we usually need to delete both a->path and
1679                  * b->path.  However, the rename on each side of the merge
1680                  * could also be involved in a rename/add conflict.  In
1681                  * such cases, we should keep the added file around,
1682                  * resolving the conflict at that path in its favor.
1683                  */
1684                 add = filespec_from_entry(&other, ci->dst_entry1, 2 ^ 1);
1685                 if (add) {
1686                         if (update_file(o, 0, &add->oid, add->mode, a->path))
1687                                 return -1;
1688                 }
1689                 else
1690                         remove_file_from_cache(a->path);
1691                 add = filespec_from_entry(&other, ci->dst_entry2, 3 ^ 1);
1692                 if (add) {
1693                         if (update_file(o, 0, &add->oid, add->mode, b->path))
1694                                 return -1;
1695                 }
1696                 else
1697                         remove_file_from_cache(b->path);
1698         } else if (handle_file(o, a, 2, ci) || handle_file(o, b, 3, ci))
1699                 return -1;
1700
1701         return 0;
1702 }
1703
1704 static int handle_rename_rename_2to1(struct merge_options *o,
1705                                      struct rename_conflict_info *ci)
1706 {
1707         /* Two files, a & b, were renamed to the same thing, c. */
1708         struct diff_filespec *a = ci->pair1->one;
1709         struct diff_filespec *b = ci->pair2->one;
1710         struct diff_filespec *c1 = ci->pair1->two;
1711         struct diff_filespec *c2 = ci->pair2->two;
1712         char *path = c1->path; /* == c2->path */
1713         char *path_side_1_desc;
1714         char *path_side_2_desc;
1715         struct merge_file_info mfi_c1;
1716         struct merge_file_info mfi_c2;
1717         int ret;
1718
1719         output(o, 1, _("CONFLICT (rename/rename): "
1720                "Rename %s->%s in %s. "
1721                "Rename %s->%s in %s"),
1722                a->path, c1->path, ci->branch1,
1723                b->path, c2->path, ci->branch2);
1724
1725         remove_file(o, 1, a->path, o->call_depth || would_lose_untracked(a->path));
1726         remove_file(o, 1, b->path, o->call_depth || would_lose_untracked(b->path));
1727
1728         path_side_1_desc = xstrfmt("%s (was %s)", path, a->path);
1729         path_side_2_desc = xstrfmt("%s (was %s)", path, b->path);
1730         if (merge_file_special_markers(o, a, c1, &ci->ren1_other,
1731                                        path_side_1_desc,
1732                                        o->branch1, c1->path,
1733                                        o->branch2, ci->ren1_other.path, &mfi_c1) ||
1734             merge_file_special_markers(o, b, &ci->ren2_other, c2,
1735                                        path_side_2_desc,
1736                                        o->branch1, ci->ren2_other.path,
1737                                        o->branch2, c2->path, &mfi_c2))
1738                 return -1;
1739         free(path_side_1_desc);
1740         free(path_side_2_desc);
1741
1742         if (o->call_depth) {
1743                 /*
1744                  * If mfi_c1.clean && mfi_c2.clean, then it might make
1745                  * sense to do a two-way merge of those results.  But, I
1746                  * think in all cases, it makes sense to have the virtual
1747                  * merge base just undo the renames; they can be detected
1748                  * again later for the non-recursive merge.
1749                  */
1750                 remove_file(o, 0, path, 0);
1751                 ret = update_file(o, 0, &mfi_c1.oid, mfi_c1.mode, a->path);
1752                 if (!ret)
1753                         ret = update_file(o, 0, &mfi_c2.oid, mfi_c2.mode,
1754                                           b->path);
1755         } else {
1756                 char *new_path1 = unique_path(o, path, ci->branch1);
1757                 char *new_path2 = unique_path(o, path, ci->branch2);
1758                 output(o, 1, _("Renaming %s to %s and %s to %s instead"),
1759                        a->path, new_path1, b->path, new_path2);
1760                 if (was_dirty(o, path))
1761                         output(o, 1, _("Refusing to lose dirty file at %s"),
1762                                path);
1763                 else if (would_lose_untracked(path))
1764                         /*
1765                          * Only way we get here is if both renames were from
1766                          * a directory rename AND user had an untracked file
1767                          * at the location where both files end up after the
1768                          * two directory renames.  See testcase 10d of t6043.
1769                          */
1770                         output(o, 1, _("Refusing to lose untracked file at "
1771                                        "%s, even though it's in the way."),
1772                                path);
1773                 else
1774                         remove_file(o, 0, path, 0);
1775                 ret = update_file(o, 0, &mfi_c1.oid, mfi_c1.mode, new_path1);
1776                 if (!ret)
1777                         ret = update_file(o, 0, &mfi_c2.oid, mfi_c2.mode,
1778                                           new_path2);
1779                 /*
1780                  * unpack_trees() actually populates the index for us for
1781                  * "normal" rename/rename(2to1) situtations so that the
1782                  * correct entries are at the higher stages, which would
1783                  * make the call below to update_stages_for_stage_data
1784                  * unnecessary.  However, if either of the renames came
1785                  * from a directory rename, then unpack_trees() will not
1786                  * have gotten the right data loaded into the index, so we
1787                  * need to do so now.  (While it'd be tempting to move this
1788                  * call to update_stages_for_stage_data() to
1789                  * apply_directory_rename_modifications(), that would break
1790                  * our intermediate calls to would_lose_untracked() since
1791                  * those rely on the current in-memory index.  See also the
1792                  * big "NOTE" in update_stages()).
1793                  */
1794                 if (update_stages_for_stage_data(o, path, ci->dst_entry1))
1795                         ret = -1;
1796
1797                 free(new_path2);
1798                 free(new_path1);
1799         }
1800
1801         return ret;
1802 }
1803
1804 /*
1805  * Get the diff_filepairs changed between o_tree and tree.
1806  */
1807 static struct diff_queue_struct *get_diffpairs(struct merge_options *o,
1808                                                struct tree *o_tree,
1809                                                struct tree *tree)
1810 {
1811         struct diff_queue_struct *ret;
1812         struct diff_options opts;
1813
1814         diff_setup(&opts);
1815         opts.flags.recursive = 1;
1816         opts.flags.rename_empty = 0;
1817         opts.detect_rename = merge_detect_rename(o);
1818         /*
1819          * We do not have logic to handle the detection of copies.  In
1820          * fact, it may not even make sense to add such logic: would we
1821          * really want a change to a base file to be propagated through
1822          * multiple other files by a merge?
1823          */
1824         if (opts.detect_rename > DIFF_DETECT_RENAME)
1825                 opts.detect_rename = DIFF_DETECT_RENAME;
1826         opts.rename_limit = o->merge_rename_limit >= 0 ? o->merge_rename_limit :
1827                             o->diff_rename_limit >= 0 ? o->diff_rename_limit :
1828                             1000;
1829         opts.rename_score = o->rename_score;
1830         opts.show_rename_progress = o->show_rename_progress;
1831         opts.output_format = DIFF_FORMAT_NO_OUTPUT;
1832         diff_setup_done(&opts);
1833         diff_tree_oid(&o_tree->object.oid, &tree->object.oid, "", &opts);
1834         diffcore_std(&opts);
1835         if (opts.needed_rename_limit > o->needed_rename_limit)
1836                 o->needed_rename_limit = opts.needed_rename_limit;
1837
1838         ret = xmalloc(sizeof(*ret));
1839         *ret = diff_queued_diff;
1840
1841         opts.output_format = DIFF_FORMAT_NO_OUTPUT;
1842         diff_queued_diff.nr = 0;
1843         diff_queued_diff.queue = NULL;
1844         diff_flush(&opts);
1845         return ret;
1846 }
1847
1848 static int tree_has_path(struct tree *tree, const char *path)
1849 {
1850         struct object_id hashy;
1851         unsigned int mode_o;
1852
1853         return !get_tree_entry(&tree->object.oid, path,
1854                                &hashy, &mode_o);
1855 }
1856
1857 /*
1858  * Return a new string that replaces the beginning portion (which matches
1859  * entry->dir), with entry->new_dir.  In perl-speak:
1860  *   new_path_name = (old_path =~ s/entry->dir/entry->new_dir/);
1861  * NOTE:
1862  *   Caller must ensure that old_path starts with entry->dir + '/'.
1863  */
1864 static char *apply_dir_rename(struct dir_rename_entry *entry,
1865                               const char *old_path)
1866 {
1867         struct strbuf new_path = STRBUF_INIT;
1868         int oldlen, newlen;
1869
1870         if (entry->non_unique_new_dir)
1871                 return NULL;
1872
1873         oldlen = strlen(entry->dir);
1874         newlen = entry->new_dir.len + (strlen(old_path) - oldlen) + 1;
1875         strbuf_grow(&new_path, newlen);
1876         strbuf_addbuf(&new_path, &entry->new_dir);
1877         strbuf_addstr(&new_path, &old_path[oldlen]);
1878
1879         return strbuf_detach(&new_path, NULL);
1880 }
1881
1882 static void get_renamed_dir_portion(const char *old_path, const char *new_path,
1883                                     char **old_dir, char **new_dir)
1884 {
1885         char *end_of_old, *end_of_new;
1886         int old_len, new_len;
1887
1888         *old_dir = NULL;
1889         *new_dir = NULL;
1890
1891         /*
1892          * For
1893          *    "a/b/c/d/e/foo.c" -> "a/b/some/thing/else/e/foo.c"
1894          * the "e/foo.c" part is the same, we just want to know that
1895          *    "a/b/c/d" was renamed to "a/b/some/thing/else"
1896          * so, for this example, this function returns "a/b/c/d" in
1897          * *old_dir and "a/b/some/thing/else" in *new_dir.
1898          *
1899          * Also, if the basename of the file changed, we don't care.  We
1900          * want to know which portion of the directory, if any, changed.
1901          */
1902         end_of_old = strrchr(old_path, '/');
1903         end_of_new = strrchr(new_path, '/');
1904
1905         if (end_of_old == NULL || end_of_new == NULL)
1906                 return;
1907         while (*--end_of_new == *--end_of_old &&
1908                end_of_old != old_path &&
1909                end_of_new != new_path)
1910                 ; /* Do nothing; all in the while loop */
1911         /*
1912          * We've found the first non-matching character in the directory
1913          * paths.  That means the current directory we were comparing
1914          * represents the rename.  Move end_of_old and end_of_new back
1915          * to the full directory name.
1916          */
1917         if (*end_of_old == '/')
1918                 end_of_old++;
1919         if (*end_of_old != '/')
1920                 end_of_new++;
1921         end_of_old = strchr(end_of_old, '/');
1922         end_of_new = strchr(end_of_new, '/');
1923
1924         /*
1925          * It may have been the case that old_path and new_path were the same
1926          * directory all along.  Don't claim a rename if they're the same.
1927          */
1928         old_len = end_of_old - old_path;
1929         new_len = end_of_new - new_path;
1930
1931         if (old_len != new_len || strncmp(old_path, new_path, old_len)) {
1932                 *old_dir = xstrndup(old_path, old_len);
1933                 *new_dir = xstrndup(new_path, new_len);
1934         }
1935 }
1936
1937 static void remove_hashmap_entries(struct hashmap *dir_renames,
1938                                    struct string_list *items_to_remove)
1939 {
1940         int i;
1941         struct dir_rename_entry *entry;
1942
1943         for (i = 0; i < items_to_remove->nr; i++) {
1944                 entry = items_to_remove->items[i].util;
1945                 hashmap_remove(dir_renames, entry, NULL);
1946         }
1947         string_list_clear(items_to_remove, 0);
1948 }
1949
1950 /*
1951  * See if there is a directory rename for path, and if there are any file
1952  * level conflicts for the renamed location.  If there is a rename and
1953  * there are no conflicts, return the new name.  Otherwise, return NULL.
1954  */
1955 static char *handle_path_level_conflicts(struct merge_options *o,
1956                                          const char *path,
1957                                          struct dir_rename_entry *entry,
1958                                          struct hashmap *collisions,
1959                                          struct tree *tree)
1960 {
1961         char *new_path = NULL;
1962         struct collision_entry *collision_ent;
1963         int clean = 1;
1964         struct strbuf collision_paths = STRBUF_INIT;
1965
1966         /*
1967          * entry has the mapping of old directory name to new directory name
1968          * that we want to apply to path.
1969          */
1970         new_path = apply_dir_rename(entry, path);
1971
1972         if (!new_path) {
1973                 /* This should only happen when entry->non_unique_new_dir set */
1974                 if (!entry->non_unique_new_dir)
1975                         BUG("entry->non_unqiue_dir not set and !new_path");
1976                 output(o, 1, _("CONFLICT (directory rename split): "
1977                                "Unclear where to place %s because directory "
1978                                "%s was renamed to multiple other directories, "
1979                                "with no destination getting a majority of the "
1980                                "files."),
1981                        path, entry->dir);
1982                 clean = 0;
1983                 return NULL;
1984         }
1985
1986         /*
1987          * The caller needs to have ensured that it has pre-populated
1988          * collisions with all paths that map to new_path.  Do a quick check
1989          * to ensure that's the case.
1990          */
1991         collision_ent = collision_find_entry(collisions, new_path);
1992         if (collision_ent == NULL)
1993                 BUG("collision_ent is NULL");
1994
1995         /*
1996          * Check for one-sided add/add/.../add conflicts, i.e.
1997          * where implicit renames from the other side doing
1998          * directory rename(s) can affect this side of history
1999          * to put multiple paths into the same location.  Warn
2000          * and bail on directory renames for such paths.
2001          */
2002         if (collision_ent->reported_already) {
2003                 clean = 0;
2004         } else if (tree_has_path(tree, new_path)) {
2005                 collision_ent->reported_already = 1;
2006                 strbuf_add_separated_string_list(&collision_paths, ", ",
2007                                                  &collision_ent->source_files);
2008                 output(o, 1, _("CONFLICT (implicit dir rename): Existing "
2009                                "file/dir at %s in the way of implicit "
2010                                "directory rename(s) putting the following "
2011                                "path(s) there: %s."),
2012                        new_path, collision_paths.buf);
2013                 clean = 0;
2014         } else if (collision_ent->source_files.nr > 1) {
2015                 collision_ent->reported_already = 1;
2016                 strbuf_add_separated_string_list(&collision_paths, ", ",
2017                                                  &collision_ent->source_files);
2018                 output(o, 1, _("CONFLICT (implicit dir rename): Cannot map "
2019                                "more than one path to %s; implicit directory "
2020                                "renames tried to put these paths there: %s"),
2021                        new_path, collision_paths.buf);
2022                 clean = 0;
2023         }
2024
2025         /* Free memory we no longer need */
2026         strbuf_release(&collision_paths);
2027         if (!clean && new_path) {
2028                 free(new_path);
2029                 return NULL;
2030         }
2031
2032         return new_path;
2033 }
2034
2035 /*
2036  * There are a couple things we want to do at the directory level:
2037  *   1. Check for both sides renaming to the same thing, in order to avoid
2038  *      implicit renaming of files that should be left in place.  (See
2039  *      testcase 6b in t6043 for details.)
2040  *   2. Prune directory renames if there are still files left in the
2041  *      the original directory.  These represent a partial directory rename,
2042  *      i.e. a rename where only some of the files within the directory
2043  *      were renamed elsewhere.  (Technically, this could be done earlier
2044  *      in get_directory_renames(), except that would prevent us from
2045  *      doing the previous check and thus failing testcase 6b.)
2046  *   3. Check for rename/rename(1to2) conflicts (at the directory level).
2047  *      In the future, we could potentially record this info as well and
2048  *      omit reporting rename/rename(1to2) conflicts for each path within
2049  *      the affected directories, thus cleaning up the merge output.
2050  *   NOTE: We do NOT check for rename/rename(2to1) conflicts at the
2051  *         directory level, because merging directories is fine.  If it
2052  *         causes conflicts for files within those merged directories, then
2053  *         that should be detected at the individual path level.
2054  */
2055 static void handle_directory_level_conflicts(struct merge_options *o,
2056                                              struct hashmap *dir_re_head,
2057                                              struct tree *head,
2058                                              struct hashmap *dir_re_merge,
2059                                              struct tree *merge)
2060 {
2061         struct hashmap_iter iter;
2062         struct dir_rename_entry *head_ent;
2063         struct dir_rename_entry *merge_ent;
2064
2065         struct string_list remove_from_head = STRING_LIST_INIT_NODUP;
2066         struct string_list remove_from_merge = STRING_LIST_INIT_NODUP;
2067
2068         hashmap_iter_init(dir_re_head, &iter);
2069         while ((head_ent = hashmap_iter_next(&iter))) {
2070                 merge_ent = dir_rename_find_entry(dir_re_merge, head_ent->dir);
2071                 if (merge_ent &&
2072                     !head_ent->non_unique_new_dir &&
2073                     !merge_ent->non_unique_new_dir &&
2074                     !strbuf_cmp(&head_ent->new_dir, &merge_ent->new_dir)) {
2075                         /* 1. Renamed identically; remove it from both sides */
2076                         string_list_append(&remove_from_head,
2077                                            head_ent->dir)->util = head_ent;
2078                         strbuf_release(&head_ent->new_dir);
2079                         string_list_append(&remove_from_merge,
2080                                            merge_ent->dir)->util = merge_ent;
2081                         strbuf_release(&merge_ent->new_dir);
2082                 } else if (tree_has_path(head, head_ent->dir)) {
2083                         /* 2. This wasn't a directory rename after all */
2084                         string_list_append(&remove_from_head,
2085                                            head_ent->dir)->util = head_ent;
2086                         strbuf_release(&head_ent->new_dir);
2087                 }
2088         }
2089
2090         remove_hashmap_entries(dir_re_head, &remove_from_head);
2091         remove_hashmap_entries(dir_re_merge, &remove_from_merge);
2092
2093         hashmap_iter_init(dir_re_merge, &iter);
2094         while ((merge_ent = hashmap_iter_next(&iter))) {
2095                 head_ent = dir_rename_find_entry(dir_re_head, merge_ent->dir);
2096                 if (tree_has_path(merge, merge_ent->dir)) {
2097                         /* 2. This wasn't a directory rename after all */
2098                         string_list_append(&remove_from_merge,
2099                                            merge_ent->dir)->util = merge_ent;
2100                 } else if (head_ent &&
2101                            !head_ent->non_unique_new_dir &&
2102                            !merge_ent->non_unique_new_dir) {
2103                         /* 3. rename/rename(1to2) */
2104                         /*
2105                          * We can assume it's not rename/rename(1to1) because
2106                          * that was case (1), already checked above.  So we
2107                          * know that head_ent->new_dir and merge_ent->new_dir
2108                          * are different strings.
2109                          */
2110                         output(o, 1, _("CONFLICT (rename/rename): "
2111                                        "Rename directory %s->%s in %s. "
2112                                        "Rename directory %s->%s in %s"),
2113                                head_ent->dir, head_ent->new_dir.buf, o->branch1,
2114                                head_ent->dir, merge_ent->new_dir.buf, o->branch2);
2115                         string_list_append(&remove_from_head,
2116                                            head_ent->dir)->util = head_ent;
2117                         strbuf_release(&head_ent->new_dir);
2118                         string_list_append(&remove_from_merge,
2119                                            merge_ent->dir)->util = merge_ent;
2120                         strbuf_release(&merge_ent->new_dir);
2121                 }
2122         }
2123
2124         remove_hashmap_entries(dir_re_head, &remove_from_head);
2125         remove_hashmap_entries(dir_re_merge, &remove_from_merge);
2126 }
2127
2128 static struct hashmap *get_directory_renames(struct diff_queue_struct *pairs,
2129                                              struct tree *tree)
2130 {
2131         struct hashmap *dir_renames;
2132         struct hashmap_iter iter;
2133         struct dir_rename_entry *entry;
2134         int i;
2135
2136         /*
2137          * Typically, we think of a directory rename as all files from a
2138          * certain directory being moved to a target directory.  However,
2139          * what if someone first moved two files from the original
2140          * directory in one commit, and then renamed the directory
2141          * somewhere else in a later commit?  At merge time, we just know
2142          * that files from the original directory went to two different
2143          * places, and that the bulk of them ended up in the same place.
2144          * We want each directory rename to represent where the bulk of the
2145          * files from that directory end up; this function exists to find
2146          * where the bulk of the files went.
2147          *
2148          * The first loop below simply iterates through the list of file
2149          * renames, finding out how often each directory rename pair
2150          * possibility occurs.
2151          */
2152         dir_renames = xmalloc(sizeof(*dir_renames));
2153         dir_rename_init(dir_renames);
2154         for (i = 0; i < pairs->nr; ++i) {
2155                 struct string_list_item *item;
2156                 int *count;
2157                 struct diff_filepair *pair = pairs->queue[i];
2158                 char *old_dir, *new_dir;
2159
2160                 /* File not part of directory rename if it wasn't renamed */
2161                 if (pair->status != 'R')
2162                         continue;
2163
2164                 get_renamed_dir_portion(pair->one->path, pair->two->path,
2165                                         &old_dir,        &new_dir);
2166                 if (!old_dir)
2167                         /* Directory didn't change at all; ignore this one. */
2168                         continue;
2169
2170                 entry = dir_rename_find_entry(dir_renames, old_dir);
2171                 if (!entry) {
2172                         entry = xmalloc(sizeof(*entry));
2173                         dir_rename_entry_init(entry, old_dir);
2174                         hashmap_put(dir_renames, entry);
2175                 } else {
2176                         free(old_dir);
2177                 }
2178                 item = string_list_lookup(&entry->possible_new_dirs, new_dir);
2179                 if (!item) {
2180                         item = string_list_insert(&entry->possible_new_dirs,
2181                                                   new_dir);
2182                         item->util = xcalloc(1, sizeof(int));
2183                 } else {
2184                         free(new_dir);
2185                 }
2186                 count = item->util;
2187                 *count += 1;
2188         }
2189
2190         /*
2191          * For each directory with files moved out of it, we find out which
2192          * target directory received the most files so we can declare it to
2193          * be the "winning" target location for the directory rename.  This
2194          * winner gets recorded in new_dir.  If there is no winner
2195          * (multiple target directories received the same number of files),
2196          * we set non_unique_new_dir.  Once we've determined the winner (or
2197          * that there is no winner), we no longer need possible_new_dirs.
2198          */
2199         hashmap_iter_init(dir_renames, &iter);
2200         while ((entry = hashmap_iter_next(&iter))) {
2201                 int max = 0;
2202                 int bad_max = 0;
2203                 char *best = NULL;
2204
2205                 for (i = 0; i < entry->possible_new_dirs.nr; i++) {
2206                         int *count = entry->possible_new_dirs.items[i].util;
2207
2208                         if (*count == max)
2209                                 bad_max = max;
2210                         else if (*count > max) {
2211                                 max = *count;
2212                                 best = entry->possible_new_dirs.items[i].string;
2213                         }
2214                 }
2215                 if (bad_max == max)
2216                         entry->non_unique_new_dir = 1;
2217                 else {
2218                         assert(entry->new_dir.len == 0);
2219                         strbuf_addstr(&entry->new_dir, best);
2220                 }
2221                 /*
2222                  * The relevant directory sub-portion of the original full
2223                  * filepaths were xstrndup'ed before inserting into
2224                  * possible_new_dirs, and instead of manually iterating the
2225                  * list and free'ing each, just lie and tell
2226                  * possible_new_dirs that it did the strdup'ing so that it
2227                  * will free them for us.
2228                  */
2229                 entry->possible_new_dirs.strdup_strings = 1;
2230                 string_list_clear(&entry->possible_new_dirs, 1);
2231         }
2232
2233         return dir_renames;
2234 }
2235
2236 static struct dir_rename_entry *check_dir_renamed(const char *path,
2237                                                   struct hashmap *dir_renames)
2238 {
2239         char *temp = xstrdup(path);
2240         char *end;
2241         struct dir_rename_entry *entry = NULL;;
2242
2243         while ((end = strrchr(temp, '/'))) {
2244                 *end = '\0';
2245                 entry = dir_rename_find_entry(dir_renames, temp);
2246                 if (entry)
2247                         break;
2248         }
2249         free(temp);
2250         return entry;
2251 }
2252
2253 static void compute_collisions(struct hashmap *collisions,
2254                                struct hashmap *dir_renames,
2255                                struct diff_queue_struct *pairs)
2256 {
2257         int i;
2258
2259         /*
2260          * Multiple files can be mapped to the same path due to directory
2261          * renames done by the other side of history.  Since that other
2262          * side of history could have merged multiple directories into one,
2263          * if our side of history added the same file basename to each of
2264          * those directories, then all N of them would get implicitly
2265          * renamed by the directory rename detection into the same path,
2266          * and we'd get an add/add/.../add conflict, and all those adds
2267          * from *this* side of history.  This is not representable in the
2268          * index, and users aren't going to easily be able to make sense of
2269          * it.  So we need to provide a good warning about what's
2270          * happening, and fall back to no-directory-rename detection
2271          * behavior for those paths.
2272          *
2273          * See testcases 9e and all of section 5 from t6043 for examples.
2274          */
2275         collision_init(collisions);
2276
2277         for (i = 0; i < pairs->nr; ++i) {
2278                 struct dir_rename_entry *dir_rename_ent;
2279                 struct collision_entry *collision_ent;
2280                 char *new_path;
2281                 struct diff_filepair *pair = pairs->queue[i];
2282
2283                 if (pair->status != 'A' && pair->status != 'R')
2284                         continue;
2285                 dir_rename_ent = check_dir_renamed(pair->two->path,
2286                                                    dir_renames);
2287                 if (!dir_rename_ent)
2288                         continue;
2289
2290                 new_path = apply_dir_rename(dir_rename_ent, pair->two->path);
2291                 if (!new_path)
2292                         /*
2293                          * dir_rename_ent->non_unique_new_path is true, which
2294                          * means there is no directory rename for us to use,
2295                          * which means it won't cause us any additional
2296                          * collisions.
2297                          */
2298                         continue;
2299                 collision_ent = collision_find_entry(collisions, new_path);
2300                 if (!collision_ent) {
2301                         collision_ent = xcalloc(1,
2302                                                 sizeof(struct collision_entry));
2303                         hashmap_entry_init(collision_ent, strhash(new_path));
2304                         hashmap_put(collisions, collision_ent);
2305                         collision_ent->target_file = new_path;
2306                 } else {
2307                         free(new_path);
2308                 }
2309                 string_list_insert(&collision_ent->source_files,
2310                                    pair->two->path);
2311         }
2312 }
2313
2314 static char *check_for_directory_rename(struct merge_options *o,
2315                                         const char *path,
2316                                         struct tree *tree,
2317                                         struct hashmap *dir_renames,
2318                                         struct hashmap *dir_rename_exclusions,
2319                                         struct hashmap *collisions,
2320                                         int *clean_merge)
2321 {
2322         char *new_path = NULL;
2323         struct dir_rename_entry *entry = check_dir_renamed(path, dir_renames);
2324         struct dir_rename_entry *oentry = NULL;
2325
2326         if (!entry)
2327                 return new_path;
2328
2329         /*
2330          * This next part is a little weird.  We do not want to do an
2331          * implicit rename into a directory we renamed on our side, because
2332          * that will result in a spurious rename/rename(1to2) conflict.  An
2333          * example:
2334          *   Base commit: dumbdir/afile, otherdir/bfile
2335          *   Side 1:      smrtdir/afile, otherdir/bfile
2336          *   Side 2:      dumbdir/afile, dumbdir/bfile
2337          * Here, while working on Side 1, we could notice that otherdir was
2338          * renamed/merged to dumbdir, and change the diff_filepair for
2339          * otherdir/bfile into a rename into dumbdir/bfile.  However, Side
2340          * 2 will notice the rename from dumbdir to smrtdir, and do the
2341          * transitive rename to move it from dumbdir/bfile to
2342          * smrtdir/bfile.  That gives us bfile in dumbdir vs being in
2343          * smrtdir, a rename/rename(1to2) conflict.  We really just want
2344          * the file to end up in smrtdir.  And the way to achieve that is
2345          * to not let Side1 do the rename to dumbdir, since we know that is
2346          * the source of one of our directory renames.
2347          *
2348          * That's why oentry and dir_rename_exclusions is here.
2349          *
2350          * As it turns out, this also prevents N-way transient rename
2351          * confusion; See testcases 9c and 9d of t6043.
2352          */
2353         oentry = dir_rename_find_entry(dir_rename_exclusions, entry->new_dir.buf);
2354         if (oentry) {
2355                 output(o, 1, _("WARNING: Avoiding applying %s -> %s rename "
2356                                "to %s, because %s itself was renamed."),
2357                        entry->dir, entry->new_dir.buf, path, entry->new_dir.buf);
2358         } else {
2359                 new_path = handle_path_level_conflicts(o, path, entry,
2360                                                        collisions, tree);
2361                 *clean_merge &= (new_path != NULL);
2362         }
2363
2364         return new_path;
2365 }
2366
2367 static void apply_directory_rename_modifications(struct merge_options *o,
2368                                                  struct diff_filepair *pair,
2369                                                  char *new_path,
2370                                                  struct rename *re,
2371                                                  struct tree *tree,
2372                                                  struct tree *o_tree,
2373                                                  struct tree *a_tree,
2374                                                  struct tree *b_tree,
2375                                                  struct string_list *entries,
2376                                                  int *clean)
2377 {
2378         struct string_list_item *item;
2379         int stage = (tree == a_tree ? 2 : 3);
2380         int update_wd;
2381
2382         /*
2383          * In all cases where we can do directory rename detection,
2384          * unpack_trees() will have read pair->two->path into the
2385          * index and the working copy.  We need to remove it so that
2386          * we can instead place it at new_path.  It is guaranteed to
2387          * not be untracked (unpack_trees() would have errored out
2388          * saying the file would have been overwritten), but it might
2389          * be dirty, though.
2390          */
2391         update_wd = !was_dirty(o, pair->two->path);
2392         if (!update_wd)
2393                 output(o, 1, _("Refusing to lose dirty file at %s"),
2394                        pair->two->path);
2395         remove_file(o, 1, pair->two->path, !update_wd);
2396
2397         /* Find or create a new re->dst_entry */
2398         item = string_list_lookup(entries, new_path);
2399         if (item) {
2400                 /*
2401                  * Since we're renaming on this side of history, and it's
2402                  * due to a directory rename on the other side of history
2403                  * (which we only allow when the directory in question no
2404                  * longer exists on the other side of history), the
2405                  * original entry for re->dst_entry is no longer
2406                  * necessary...
2407                  */
2408                 re->dst_entry->processed = 1;
2409
2410                 /*
2411                  * ...because we'll be using this new one.
2412                  */
2413                 re->dst_entry = item->util;
2414         } else {
2415                 /*
2416                  * re->dst_entry is for the before-dir-rename path, and we
2417                  * need it to hold information for the after-dir-rename
2418                  * path.  Before creating a new entry, we need to mark the
2419                  * old one as unnecessary (...unless it is shared by
2420                  * src_entry, i.e. this didn't use to be a rename, in which
2421                  * case we can just allow the normal processing to happen
2422                  * for it).
2423                  */
2424                 if (pair->status == 'R')
2425                         re->dst_entry->processed = 1;
2426
2427                 re->dst_entry = insert_stage_data(new_path,
2428                                                   o_tree, a_tree, b_tree,
2429                                                   entries);
2430                 item = string_list_insert(entries, new_path);
2431                 item->util = re->dst_entry;
2432         }
2433
2434         /*
2435          * Update the stage_data with the information about the path we are
2436          * moving into place.  That slot will be empty and available for us
2437          * to write to because of the collision checks in
2438          * handle_path_level_conflicts().  In other words,
2439          * re->dst_entry->stages[stage].oid will be the null_oid, so it's
2440          * open for us to write to.
2441          *
2442          * It may be tempting to actually update the index at this point as
2443          * well, using update_stages_for_stage_data(), but as per the big
2444          * "NOTE" in update_stages(), doing so will modify the current
2445          * in-memory index which will break calls to would_lose_untracked()
2446          * that we need to make.  Instead, we need to just make sure that
2447          * the various handle_rename_*() functions update the index
2448          * explicitly rather than relying on unpack_trees() to have done it.
2449          */
2450         get_tree_entry(&tree->object.oid,
2451                        pair->two->path,
2452                        &re->dst_entry->stages[stage].oid,
2453                        &re->dst_entry->stages[stage].mode);
2454
2455         /* Update pair status */
2456         if (pair->status == 'A') {
2457                 /*
2458                  * Recording rename information for this add makes it look
2459                  * like a rename/delete conflict.  Make sure we can
2460                  * correctly handle this as an add that was moved to a new
2461                  * directory instead of reporting a rename/delete conflict.
2462                  */
2463                 re->add_turned_into_rename = 1;
2464         }
2465         /*
2466          * We don't actually look at pair->status again, but it seems
2467          * pedagogically correct to adjust it.
2468          */
2469         pair->status = 'R';
2470
2471         /*
2472          * Finally, record the new location.
2473          */
2474         pair->two->path = new_path;
2475 }
2476
2477 /*
2478  * Get information of all renames which occurred in 'pairs', making use of
2479  * any implicit directory renames inferred from the other side of history.
2480  * We need the three trees in the merge ('o_tree', 'a_tree' and 'b_tree')
2481  * to be able to associate the correct cache entries with the rename
2482  * information; tree is always equal to either a_tree or b_tree.
2483  */
2484 static struct string_list *get_renames(struct merge_options *o,
2485                                        struct diff_queue_struct *pairs,
2486                                        struct hashmap *dir_renames,
2487                                        struct hashmap *dir_rename_exclusions,
2488                                        struct tree *tree,
2489                                        struct tree *o_tree,
2490                                        struct tree *a_tree,
2491                                        struct tree *b_tree,
2492                                        struct string_list *entries,
2493                                        int *clean_merge)
2494 {
2495         int i;
2496         struct hashmap collisions;
2497         struct hashmap_iter iter;
2498         struct collision_entry *e;
2499         struct string_list *renames;
2500
2501         compute_collisions(&collisions, dir_renames, pairs);
2502         renames = xcalloc(1, sizeof(struct string_list));
2503
2504         for (i = 0; i < pairs->nr; ++i) {
2505                 struct string_list_item *item;
2506                 struct rename *re;
2507                 struct diff_filepair *pair = pairs->queue[i];
2508                 char *new_path; /* non-NULL only with directory renames */
2509
2510                 if (pair->status != 'A' && pair->status != 'R') {
2511                         diff_free_filepair(pair);
2512                         continue;
2513                 }
2514                 new_path = check_for_directory_rename(o, pair->two->path, tree,
2515                                                       dir_renames,
2516                                                       dir_rename_exclusions,
2517                                                       &collisions,
2518                                                       clean_merge);
2519                 if (pair->status != 'R' && !new_path) {
2520                         diff_free_filepair(pair);
2521                         continue;
2522                 }
2523
2524                 re = xmalloc(sizeof(*re));
2525                 re->processed = 0;
2526                 re->add_turned_into_rename = 0;
2527                 re->pair = pair;
2528                 item = string_list_lookup(entries, re->pair->one->path);
2529                 if (!item)
2530                         re->src_entry = insert_stage_data(re->pair->one->path,
2531                                         o_tree, a_tree, b_tree, entries);
2532                 else
2533                         re->src_entry = item->util;
2534
2535                 item = string_list_lookup(entries, re->pair->two->path);
2536                 if (!item)
2537                         re->dst_entry = insert_stage_data(re->pair->two->path,
2538                                         o_tree, a_tree, b_tree, entries);
2539                 else
2540                         re->dst_entry = item->util;
2541                 item = string_list_insert(renames, pair->one->path);
2542                 item->util = re;
2543                 if (new_path)
2544                         apply_directory_rename_modifications(o, pair, new_path,
2545                                                              re, tree, o_tree,
2546                                                              a_tree, b_tree,
2547                                                              entries,
2548                                                              clean_merge);
2549         }
2550
2551         hashmap_iter_init(&collisions, &iter);
2552         while ((e = hashmap_iter_next(&iter))) {
2553                 free(e->target_file);
2554                 string_list_clear(&e->source_files, 0);
2555         }
2556         hashmap_free(&collisions, 1);
2557         return renames;
2558 }
2559
2560 static int process_renames(struct merge_options *o,
2561                            struct string_list *a_renames,
2562                            struct string_list *b_renames)
2563 {
2564         int clean_merge = 1, i, j;
2565         struct string_list a_by_dst = STRING_LIST_INIT_NODUP;
2566         struct string_list b_by_dst = STRING_LIST_INIT_NODUP;
2567         const struct rename *sre;
2568
2569         for (i = 0; i < a_renames->nr; i++) {
2570                 sre = a_renames->items[i].util;
2571                 string_list_insert(&a_by_dst, sre->pair->two->path)->util
2572                         = (void *)sre;
2573         }
2574         for (i = 0; i < b_renames->nr; i++) {
2575                 sre = b_renames->items[i].util;
2576                 string_list_insert(&b_by_dst, sre->pair->two->path)->util
2577                         = (void *)sre;
2578         }
2579
2580         for (i = 0, j = 0; i < a_renames->nr || j < b_renames->nr;) {
2581                 struct string_list *renames1, *renames2Dst;
2582                 struct rename *ren1 = NULL, *ren2 = NULL;
2583                 const char *branch1, *branch2;
2584                 const char *ren1_src, *ren1_dst;
2585                 struct string_list_item *lookup;
2586
2587                 if (i >= a_renames->nr) {
2588                         ren2 = b_renames->items[j++].util;
2589                 } else if (j >= b_renames->nr) {
2590                         ren1 = a_renames->items[i++].util;
2591                 } else {
2592                         int compare = strcmp(a_renames->items[i].string,
2593                                              b_renames->items[j].string);
2594                         if (compare <= 0)
2595                                 ren1 = a_renames->items[i++].util;
2596                         if (compare >= 0)
2597                                 ren2 = b_renames->items[j++].util;
2598                 }
2599
2600                 /* TODO: refactor, so that 1/2 are not needed */
2601                 if (ren1) {
2602                         renames1 = a_renames;
2603                         renames2Dst = &b_by_dst;
2604                         branch1 = o->branch1;
2605                         branch2 = o->branch2;
2606                 } else {
2607                         renames1 = b_renames;
2608                         renames2Dst = &a_by_dst;
2609                         branch1 = o->branch2;
2610                         branch2 = o->branch1;
2611                         SWAP(ren2, ren1);
2612                 }
2613
2614                 if (ren1->processed)
2615                         continue;
2616                 ren1->processed = 1;
2617                 ren1->dst_entry->processed = 1;
2618                 /* BUG: We should only mark src_entry as processed if we
2619                  * are not dealing with a rename + add-source case.
2620                  */
2621                 ren1->src_entry->processed = 1;
2622
2623                 ren1_src = ren1->pair->one->path;
2624                 ren1_dst = ren1->pair->two->path;
2625
2626                 if (ren2) {
2627                         /* One file renamed on both sides */
2628                         const char *ren2_src = ren2->pair->one->path;
2629                         const char *ren2_dst = ren2->pair->two->path;
2630                         enum rename_type rename_type;
2631                         if (strcmp(ren1_src, ren2_src) != 0)
2632                                 BUG("ren1_src != ren2_src");
2633                         ren2->dst_entry->processed = 1;
2634                         ren2->processed = 1;
2635                         if (strcmp(ren1_dst, ren2_dst) != 0) {
2636                                 rename_type = RENAME_ONE_FILE_TO_TWO;
2637                                 clean_merge = 0;
2638                         } else {
2639                                 rename_type = RENAME_ONE_FILE_TO_ONE;
2640                                 /* BUG: We should only remove ren1_src in
2641                                  * the base stage (think of rename +
2642                                  * add-source cases).
2643                                  */
2644                                 remove_file(o, 1, ren1_src, 1);
2645                                 update_entry(ren1->dst_entry,
2646                                              ren1->pair->one,
2647                                              ren1->pair->two,
2648                                              ren2->pair->two);
2649                         }
2650                         setup_rename_conflict_info(rename_type,
2651                                                    ren1->pair,
2652                                                    ren2->pair,
2653                                                    branch1,
2654                                                    branch2,
2655                                                    ren1->dst_entry,
2656                                                    ren2->dst_entry,
2657                                                    o,
2658                                                    NULL,
2659                                                    NULL);
2660                 } else if ((lookup = string_list_lookup(renames2Dst, ren1_dst))) {
2661                         /* Two different files renamed to the same thing */
2662                         char *ren2_dst;
2663                         ren2 = lookup->util;
2664                         ren2_dst = ren2->pair->two->path;
2665                         if (strcmp(ren1_dst, ren2_dst) != 0)
2666                                 BUG("ren1_dst != ren2_dst");
2667
2668                         clean_merge = 0;
2669                         ren2->processed = 1;
2670                         /*
2671                          * BUG: We should only mark src_entry as processed
2672                          * if we are not dealing with a rename + add-source
2673                          * case.
2674                          */
2675                         ren2->src_entry->processed = 1;
2676
2677                         setup_rename_conflict_info(RENAME_TWO_FILES_TO_ONE,
2678                                                    ren1->pair,
2679                                                    ren2->pair,
2680                                                    branch1,
2681                                                    branch2,
2682                                                    ren1->dst_entry,
2683                                                    ren2->dst_entry,
2684                                                    o,
2685                                                    ren1->src_entry,
2686                                                    ren2->src_entry);
2687
2688                 } else {
2689                         /* Renamed in 1, maybe changed in 2 */
2690                         /* we only use sha1 and mode of these */
2691                         struct diff_filespec src_other, dst_other;
2692                         int try_merge;
2693
2694                         /*
2695                          * unpack_trees loads entries from common-commit
2696                          * into stage 1, from head-commit into stage 2, and
2697                          * from merge-commit into stage 3.  We keep track
2698                          * of which side corresponds to the rename.
2699                          */
2700                         int renamed_stage = a_renames == renames1 ? 2 : 3;
2701                         int other_stage =   a_renames == renames1 ? 3 : 2;
2702
2703                         /* BUG: We should only remove ren1_src in the base
2704                          * stage and in other_stage (think of rename +
2705                          * add-source case).
2706                          */
2707                         remove_file(o, 1, ren1_src,
2708                                     renamed_stage == 2 || !was_tracked(o, ren1_src));
2709
2710                         oidcpy(&src_other.oid,
2711                                &ren1->src_entry->stages[other_stage].oid);
2712                         src_other.mode = ren1->src_entry->stages[other_stage].mode;
2713                         oidcpy(&dst_other.oid,
2714                                &ren1->dst_entry->stages[other_stage].oid);
2715                         dst_other.mode = ren1->dst_entry->stages[other_stage].mode;
2716                         try_merge = 0;
2717
2718                         if (oid_eq(&src_other.oid, &null_oid) &&
2719                             ren1->add_turned_into_rename) {
2720                                 setup_rename_conflict_info(RENAME_VIA_DIR,
2721                                                            ren1->pair,
2722                                                            NULL,
2723                                                            branch1,
2724                                                            branch2,
2725                                                            ren1->dst_entry,
2726                                                            NULL,
2727                                                            o,
2728                                                            NULL,
2729                                                            NULL);
2730                         } else if (oid_eq(&src_other.oid, &null_oid)) {
2731                                 setup_rename_conflict_info(RENAME_DELETE,
2732                                                            ren1->pair,
2733                                                            NULL,
2734                                                            branch1,
2735                                                            branch2,
2736                                                            ren1->dst_entry,
2737                                                            NULL,
2738                                                            o,
2739                                                            NULL,
2740                                                            NULL);
2741                         } else if ((dst_other.mode == ren1->pair->two->mode) &&
2742                                    oid_eq(&dst_other.oid, &ren1->pair->two->oid)) {
2743                                 /*
2744                                  * Added file on the other side identical to
2745                                  * the file being renamed: clean merge.
2746                                  * Also, there is no need to overwrite the
2747                                  * file already in the working copy, so call
2748                                  * update_file_flags() instead of
2749                                  * update_file().
2750                                  */
2751                                 if (update_file_flags(o,
2752                                                       &ren1->pair->two->oid,
2753                                                       ren1->pair->two->mode,
2754                                                       ren1_dst,
2755                                                       1, /* update_cache */
2756                                                       0  /* update_wd    */))
2757                                         clean_merge = -1;
2758                         } else if (!oid_eq(&dst_other.oid, &null_oid)) {
2759                                 clean_merge = 0;
2760                                 try_merge = 1;
2761                                 output(o, 1, _("CONFLICT (rename/add): Rename %s->%s in %s. "
2762                                        "%s added in %s"),
2763                                        ren1_src, ren1_dst, branch1,
2764                                        ren1_dst, branch2);
2765                                 if (o->call_depth) {
2766                                         struct merge_file_info mfi;
2767                                         if (merge_file_one(o, ren1_dst, &null_oid, 0,
2768                                                            &ren1->pair->two->oid,
2769                                                            ren1->pair->two->mode,
2770                                                            &dst_other.oid,
2771                                                            dst_other.mode,
2772                                                            branch1, branch2, &mfi)) {
2773                                                 clean_merge = -1;
2774                                                 goto cleanup_and_return;
2775                                         }
2776                                         output(o, 1, _("Adding merged %s"), ren1_dst);
2777                                         if (update_file(o, 0, &mfi.oid,
2778                                                         mfi.mode, ren1_dst))
2779                                                 clean_merge = -1;
2780                                         try_merge = 0;
2781                                 } else {
2782                                         char *new_path = unique_path(o, ren1_dst, branch2);
2783                                         output(o, 1, _("Adding as %s instead"), new_path);
2784                                         if (update_file(o, 0, &dst_other.oid,
2785                                                         dst_other.mode, new_path))
2786                                                 clean_merge = -1;
2787                                         free(new_path);
2788                                 }
2789                         } else
2790                                 try_merge = 1;
2791
2792                         if (clean_merge < 0)
2793                                 goto cleanup_and_return;
2794                         if (try_merge) {
2795                                 struct diff_filespec *one, *a, *b;
2796                                 src_other.path = (char *)ren1_src;
2797
2798                                 one = ren1->pair->one;
2799                                 if (a_renames == renames1) {
2800                                         a = ren1->pair->two;
2801                                         b = &src_other;
2802                                 } else {
2803                                         b = ren1->pair->two;
2804                                         a = &src_other;
2805                                 }
2806                                 update_entry(ren1->dst_entry, one, a, b);
2807                                 setup_rename_conflict_info(RENAME_NORMAL,
2808                                                            ren1->pair,
2809                                                            NULL,
2810                                                            branch1,
2811                                                            NULL,
2812                                                            ren1->dst_entry,
2813                                                            NULL,
2814                                                            o,
2815                                                            NULL,
2816                                                            NULL);
2817                         }
2818                 }
2819         }
2820 cleanup_and_return:
2821         string_list_clear(&a_by_dst, 0);
2822         string_list_clear(&b_by_dst, 0);
2823
2824         return clean_merge;
2825 }
2826
2827 struct rename_info {
2828         struct string_list *head_renames;
2829         struct string_list *merge_renames;
2830 };
2831
2832 static void initial_cleanup_rename(struct diff_queue_struct *pairs,
2833                                    struct hashmap *dir_renames)
2834 {
2835         struct hashmap_iter iter;
2836         struct dir_rename_entry *e;
2837
2838         hashmap_iter_init(dir_renames, &iter);
2839         while ((e = hashmap_iter_next(&iter))) {
2840                 free(e->dir);
2841                 strbuf_release(&e->new_dir);
2842                 /* possible_new_dirs already cleared in get_directory_renames */
2843         }
2844         hashmap_free(dir_renames, 1);
2845         free(dir_renames);
2846
2847         free(pairs->queue);
2848         free(pairs);
2849 }
2850
2851 static int detect_and_process_renames(struct merge_options *o,
2852                                       struct tree *common,
2853                                       struct tree *head,
2854                                       struct tree *merge,
2855                                       struct string_list *entries,
2856                                       struct rename_info *ri)
2857 {
2858         struct diff_queue_struct *head_pairs, *merge_pairs;
2859         struct hashmap *dir_re_head, *dir_re_merge;
2860         int clean = 1;
2861
2862         ri->head_renames = NULL;
2863         ri->merge_renames = NULL;
2864
2865         if (!merge_detect_rename(o))
2866                 return 1;
2867
2868         head_pairs = get_diffpairs(o, common, head);
2869         merge_pairs = get_diffpairs(o, common, merge);
2870
2871         dir_re_head = get_directory_renames(head_pairs, head);
2872         dir_re_merge = get_directory_renames(merge_pairs, merge);
2873
2874         handle_directory_level_conflicts(o,
2875                                          dir_re_head, head,
2876                                          dir_re_merge, merge);
2877
2878         ri->head_renames  = get_renames(o, head_pairs,
2879                                         dir_re_merge, dir_re_head, head,
2880                                         common, head, merge, entries,
2881                                         &clean);
2882         if (clean < 0)
2883                 goto cleanup;
2884         ri->merge_renames = get_renames(o, merge_pairs,
2885                                         dir_re_head, dir_re_merge, merge,
2886                                         common, head, merge, entries,
2887                                         &clean);
2888         if (clean < 0)
2889                 goto cleanup;
2890         clean &= process_renames(o, ri->head_renames, ri->merge_renames);
2891
2892 cleanup:
2893         /*
2894          * Some cleanup is deferred until cleanup_renames() because the
2895          * data structures are still needed and referenced in
2896          * process_entry().  But there are a few things we can free now.
2897          */
2898         initial_cleanup_rename(head_pairs, dir_re_head);
2899         initial_cleanup_rename(merge_pairs, dir_re_merge);
2900
2901         return clean;
2902 }
2903
2904 static void final_cleanup_rename(struct string_list *rename)
2905 {
2906         const struct rename *re;
2907         int i;
2908
2909         if (rename == NULL)
2910                 return;
2911
2912         for (i = 0; i < rename->nr; i++) {
2913                 re = rename->items[i].util;
2914                 diff_free_filepair(re->pair);
2915         }
2916         string_list_clear(rename, 1);
2917         free(rename);
2918 }
2919
2920 static void final_cleanup_renames(struct rename_info *re_info)
2921 {
2922         final_cleanup_rename(re_info->head_renames);
2923         final_cleanup_rename(re_info->merge_renames);
2924 }
2925
2926 static struct object_id *stage_oid(const struct object_id *oid, unsigned mode)
2927 {
2928         return (is_null_oid(oid) || mode == 0) ? NULL: (struct object_id *)oid;
2929 }
2930
2931 static int read_oid_strbuf(struct merge_options *o,
2932                            const struct object_id *oid,
2933                            struct strbuf *dst)
2934 {
2935         void *buf;
2936         enum object_type type;
2937         unsigned long size;
2938         buf = read_object_file(oid, &type, &size);
2939         if (!buf)
2940                 return err(o, _("cannot read object %s"), oid_to_hex(oid));
2941         if (type != OBJ_BLOB) {
2942                 free(buf);
2943                 return err(o, _("object %s is not a blob"), oid_to_hex(oid));
2944         }
2945         strbuf_attach(dst, buf, size, size + 1);
2946         return 0;
2947 }
2948
2949 static int blob_unchanged(struct merge_options *opt,
2950                           const struct object_id *o_oid,
2951                           unsigned o_mode,
2952                           const struct object_id *a_oid,
2953                           unsigned a_mode,
2954                           int renormalize, const char *path)
2955 {
2956         struct strbuf o = STRBUF_INIT;
2957         struct strbuf a = STRBUF_INIT;
2958         int ret = 0; /* assume changed for safety */
2959
2960         if (a_mode != o_mode)
2961                 return 0;
2962         if (oid_eq(o_oid, a_oid))
2963                 return 1;
2964         if (!renormalize)
2965                 return 0;
2966
2967         assert(o_oid && a_oid);
2968         if (read_oid_strbuf(opt, o_oid, &o) || read_oid_strbuf(opt, a_oid, &a))
2969                 goto error_return;
2970         /*
2971          * Note: binary | is used so that both renormalizations are
2972          * performed.  Comparison can be skipped if both files are
2973          * unchanged since their sha1s have already been compared.
2974          */
2975         if (renormalize_buffer(&the_index, path, o.buf, o.len, &o) |
2976             renormalize_buffer(&the_index, path, a.buf, a.len, &a))
2977                 ret = (o.len == a.len && !memcmp(o.buf, a.buf, o.len));
2978
2979 error_return:
2980         strbuf_release(&o);
2981         strbuf_release(&a);
2982         return ret;
2983 }
2984
2985 static int handle_modify_delete(struct merge_options *o,
2986                                 const char *path,
2987                                 struct object_id *o_oid, int o_mode,
2988                                 struct object_id *a_oid, int a_mode,
2989                                 struct object_id *b_oid, int b_mode)
2990 {
2991         const char *modify_branch, *delete_branch;
2992         struct object_id *changed_oid;
2993         int changed_mode;
2994
2995         if (a_oid) {
2996                 modify_branch = o->branch1;
2997                 delete_branch = o->branch2;
2998                 changed_oid = a_oid;
2999                 changed_mode = a_mode;
3000         } else {
3001                 modify_branch = o->branch2;
3002                 delete_branch = o->branch1;
3003                 changed_oid = b_oid;
3004                 changed_mode = b_mode;
3005         }
3006
3007         return handle_change_delete(o,
3008                                     path, NULL,
3009                                     o_oid, o_mode,
3010                                     changed_oid, changed_mode,
3011                                     modify_branch, delete_branch,
3012                                     _("modify"), _("modified"));
3013 }
3014
3015 static int merge_content(struct merge_options *o,
3016                          const char *path,
3017                          int is_dirty,
3018                          struct object_id *o_oid, int o_mode,
3019                          struct object_id *a_oid, int a_mode,
3020                          struct object_id *b_oid, int b_mode,
3021                          struct rename_conflict_info *rename_conflict_info)
3022 {
3023         const char *reason = _("content");
3024         const char *path1 = NULL, *path2 = NULL;
3025         struct merge_file_info mfi;
3026         struct diff_filespec one, a, b;
3027         unsigned df_conflict_remains = 0;
3028
3029         if (!o_oid) {
3030                 reason = _("add/add");
3031                 o_oid = (struct object_id *)&null_oid;
3032         }
3033         one.path = a.path = b.path = (char *)path;
3034         oidcpy(&one.oid, o_oid);
3035         one.mode = o_mode;
3036         oidcpy(&a.oid, a_oid);
3037         a.mode = a_mode;
3038         oidcpy(&b.oid, b_oid);
3039         b.mode = b_mode;
3040
3041         if (rename_conflict_info) {
3042                 struct diff_filepair *pair1 = rename_conflict_info->pair1;
3043
3044                 path1 = (o->branch1 == rename_conflict_info->branch1) ?
3045                         pair1->two->path : pair1->one->path;
3046                 /* If rename_conflict_info->pair2 != NULL, we are in
3047                  * RENAME_ONE_FILE_TO_ONE case.  Otherwise, we have a
3048                  * normal rename.
3049                  */
3050                 path2 = (rename_conflict_info->pair2 ||
3051                          o->branch2 == rename_conflict_info->branch1) ?
3052                         pair1->two->path : pair1->one->path;
3053
3054                 if (dir_in_way(path, !o->call_depth,
3055                                S_ISGITLINK(pair1->two->mode)))
3056                         df_conflict_remains = 1;
3057         }
3058         if (merge_file_special_markers(o, &one, &a, &b, path,
3059                                        o->branch1, path1,
3060                                        o->branch2, path2, &mfi))
3061                 return -1;
3062
3063         /*
3064          * We can skip updating the working tree file iff:
3065          *   a) The merge is clean
3066          *   b) The merge matches what was in HEAD (content, mode, pathname)
3067          *   c) The target path is usable (i.e. not involved in D/F conflict)
3068          */
3069         if (mfi.clean &&
3070             was_tracked_and_matches(o, path, &mfi.oid, mfi.mode) &&
3071             !df_conflict_remains) {
3072                 output(o, 3, _("Skipped %s (merged same as existing)"), path);
3073                 if (add_cacheinfo(o, mfi.mode, &mfi.oid, path,
3074                                   0, (!o->call_depth && !is_dirty), 0))
3075                         return -1;
3076                 return mfi.clean;
3077         }
3078
3079         if (!mfi.clean) {
3080                 if (S_ISGITLINK(mfi.mode))
3081                         reason = _("submodule");
3082                 output(o, 1, _("CONFLICT (%s): Merge conflict in %s"),
3083                                 reason, path);
3084                 if (rename_conflict_info && !df_conflict_remains)
3085                         if (update_stages(o, path, &one, &a, &b))
3086                                 return -1;
3087         }
3088
3089         if (df_conflict_remains || is_dirty) {
3090                 char *new_path;
3091                 if (o->call_depth) {
3092                         remove_file_from_cache(path);
3093                 } else {
3094                         if (!mfi.clean) {
3095                                 if (update_stages(o, path, &one, &a, &b))
3096                                         return -1;
3097                         } else {
3098                                 int file_from_stage2 = was_tracked(o, path);
3099                                 struct diff_filespec merged;
3100                                 oidcpy(&merged.oid, &mfi.oid);
3101                                 merged.mode = mfi.mode;
3102
3103                                 if (update_stages(o, path, NULL,
3104                                                   file_from_stage2 ? &merged : NULL,
3105                                                   file_from_stage2 ? NULL : &merged))
3106                                         return -1;
3107                         }
3108
3109                 }
3110                 new_path = unique_path(o, path, rename_conflict_info->branch1);
3111                 if (is_dirty) {
3112                         output(o, 1, _("Refusing to lose dirty file at %s"),
3113                                path);
3114                 }
3115                 output(o, 1, _("Adding as %s instead"), new_path);
3116                 if (update_file(o, 0, &mfi.oid, mfi.mode, new_path)) {
3117                         free(new_path);
3118                         return -1;
3119                 }
3120                 free(new_path);
3121                 mfi.clean = 0;
3122         } else if (update_file(o, mfi.clean, &mfi.oid, mfi.mode, path))
3123                 return -1;
3124         return !is_dirty && mfi.clean;
3125 }
3126
3127 static int handle_rename_normal(struct merge_options *o,
3128                                 const char *path,
3129                                 struct object_id *o_oid, unsigned int o_mode,
3130                                 struct object_id *a_oid, unsigned int a_mode,
3131                                 struct object_id *b_oid, unsigned int b_mode,
3132                                 struct rename_conflict_info *ci)
3133 {
3134         /* Merge the content and write it out */
3135         return merge_content(o, path, was_dirty(o, path),
3136                              o_oid, o_mode, a_oid, a_mode, b_oid, b_mode,
3137                              ci);
3138 }
3139
3140 /* Per entry merge function */
3141 static int process_entry(struct merge_options *o,
3142                          const char *path, struct stage_data *entry)
3143 {
3144         int clean_merge = 1;
3145         int normalize = o->renormalize;
3146         unsigned o_mode = entry->stages[1].mode;
3147         unsigned a_mode = entry->stages[2].mode;
3148         unsigned b_mode = entry->stages[3].mode;
3149         struct object_id *o_oid = stage_oid(&entry->stages[1].oid, o_mode);
3150         struct object_id *a_oid = stage_oid(&entry->stages[2].oid, a_mode);
3151         struct object_id *b_oid = stage_oid(&entry->stages[3].oid, b_mode);
3152
3153         entry->processed = 1;
3154         if (entry->rename_conflict_info) {
3155                 struct rename_conflict_info *conflict_info = entry->rename_conflict_info;
3156                 switch (conflict_info->rename_type) {
3157                 case RENAME_NORMAL:
3158                 case RENAME_ONE_FILE_TO_ONE:
3159                         clean_merge = handle_rename_normal(o,
3160                                                            path,
3161                                                            o_oid, o_mode,
3162                                                            a_oid, a_mode,
3163                                                            b_oid, b_mode,
3164                                                            conflict_info);
3165                         break;
3166                 case RENAME_VIA_DIR:
3167                         clean_merge = 1;
3168                         if (handle_rename_via_dir(o,
3169                                                   conflict_info->pair1,
3170                                                   conflict_info->branch1,
3171                                                   conflict_info->branch2))
3172                                 clean_merge = -1;
3173                         break;
3174                 case RENAME_DELETE:
3175                         clean_merge = 0;
3176                         if (handle_rename_delete(o,
3177                                                  conflict_info->pair1,
3178                                                  conflict_info->branch1,
3179                                                  conflict_info->branch2))
3180                                 clean_merge = -1;
3181                         break;
3182                 case RENAME_ONE_FILE_TO_TWO:
3183                         clean_merge = 0;
3184                         if (handle_rename_rename_1to2(o, conflict_info))
3185                                 clean_merge = -1;
3186                         break;
3187                 case RENAME_TWO_FILES_TO_ONE:
3188                         clean_merge = 0;
3189                         if (handle_rename_rename_2to1(o, conflict_info))
3190                                 clean_merge = -1;
3191                         break;
3192                 default:
3193                         entry->processed = 0;
3194                         break;
3195                 }
3196         } else if (o_oid && (!a_oid || !b_oid)) {
3197                 /* Case A: Deleted in one */
3198                 if ((!a_oid && !b_oid) ||
3199                     (!b_oid && blob_unchanged(o, o_oid, o_mode, a_oid, a_mode, normalize, path)) ||
3200                     (!a_oid && blob_unchanged(o, o_oid, o_mode, b_oid, b_mode, normalize, path))) {
3201                         /* Deleted in both or deleted in one and
3202                          * unchanged in the other */
3203                         if (a_oid)
3204                                 output(o, 2, _("Removing %s"), path);
3205                         /* do not touch working file if it did not exist */
3206                         remove_file(o, 1, path, !a_oid);
3207                 } else {
3208                         /* Modify/delete; deleted side may have put a directory in the way */
3209                         clean_merge = 0;
3210                         if (handle_modify_delete(o, path, o_oid, o_mode,
3211                                                  a_oid, a_mode, b_oid, b_mode))
3212                                 clean_merge = -1;
3213                 }
3214         } else if ((!o_oid && a_oid && !b_oid) ||
3215                    (!o_oid && !a_oid && b_oid)) {
3216                 /* Case B: Added in one. */
3217                 /* [nothing|directory] -> ([nothing|directory], file) */
3218
3219                 const char *add_branch;
3220                 const char *other_branch;
3221                 unsigned mode;
3222                 const struct object_id *oid;
3223                 const char *conf;
3224
3225                 if (a_oid) {
3226                         add_branch = o->branch1;
3227                         other_branch = o->branch2;
3228                         mode = a_mode;
3229                         oid = a_oid;
3230                         conf = _("file/directory");
3231                 } else {
3232                         add_branch = o->branch2;
3233                         other_branch = o->branch1;
3234                         mode = b_mode;
3235                         oid = b_oid;
3236                         conf = _("directory/file");
3237                 }
3238                 if (dir_in_way(path,
3239                                !o->call_depth && !S_ISGITLINK(a_mode),
3240                                0)) {
3241                         char *new_path = unique_path(o, path, add_branch);
3242                         clean_merge = 0;
3243                         output(o, 1, _("CONFLICT (%s): There is a directory with name %s in %s. "
3244                                "Adding %s as %s"),
3245                                conf, path, other_branch, path, new_path);
3246                         if (update_file(o, 0, oid, mode, new_path))
3247                                 clean_merge = -1;
3248                         else if (o->call_depth)
3249                                 remove_file_from_cache(path);
3250                         free(new_path);
3251                 } else {
3252                         output(o, 2, _("Adding %s"), path);
3253                         /* do not overwrite file if already present */
3254                         if (update_file_flags(o, oid, mode, path, 1, !a_oid))
3255                                 clean_merge = -1;
3256                 }
3257         } else if (a_oid && b_oid) {
3258                 /* Case C: Added in both (check for same permissions) and */
3259                 /* case D: Modified in both, but differently. */
3260                 int is_dirty = 0; /* unpack_trees would have bailed if dirty */
3261                 clean_merge = merge_content(o, path, is_dirty,
3262                                             o_oid, o_mode, a_oid, a_mode, b_oid, b_mode,
3263                                             NULL);
3264         } else if (!o_oid && !a_oid && !b_oid) {
3265                 /*
3266                  * this entry was deleted altogether. a_mode == 0 means
3267                  * we had that path and want to actively remove it.
3268                  */
3269                 remove_file(o, 1, path, !a_mode);
3270         } else
3271                 BUG("fatal merge failure, shouldn't happen.");
3272
3273         return clean_merge;
3274 }
3275
3276 int merge_trees(struct merge_options *o,
3277                 struct tree *head,
3278                 struct tree *merge,
3279                 struct tree *common,
3280                 struct tree **result)
3281 {
3282         int code, clean;
3283
3284         if (o->subtree_shift) {
3285                 merge = shift_tree_object(head, merge, o->subtree_shift);
3286                 common = shift_tree_object(head, common, o->subtree_shift);
3287         }
3288
3289         if (oid_eq(&common->object.oid, &merge->object.oid)) {
3290                 struct strbuf sb = STRBUF_INIT;
3291
3292                 if (!o->call_depth && index_has_changes(&sb)) {
3293                         err(o, _("Dirty index: cannot merge (dirty: %s)"),
3294                             sb.buf);
3295                         return 0;
3296                 }
3297                 output(o, 0, _("Already up to date!"));
3298                 *result = head;
3299                 return 1;
3300         }
3301
3302         code = unpack_trees_start(o, common, head, merge);
3303
3304         if (code != 0) {
3305                 if (show(o, 4) || o->call_depth)
3306                         err(o, _("merging of trees %s and %s failed"),
3307                             oid_to_hex(&head->object.oid),
3308                             oid_to_hex(&merge->object.oid));
3309                 unpack_trees_finish(o);
3310                 return -1;
3311         }
3312
3313         if (unmerged_cache()) {
3314                 struct string_list *entries;
3315                 struct rename_info re_info;
3316                 int i;
3317                 /*
3318                  * Only need the hashmap while processing entries, so
3319                  * initialize it here and free it when we are done running
3320                  * through the entries. Keeping it in the merge_options as
3321                  * opposed to decaring a local hashmap is for convenience
3322                  * so that we don't have to pass it to around.
3323                  */
3324                 hashmap_init(&o->current_file_dir_set, path_hashmap_cmp, NULL, 512);
3325                 get_files_dirs(o, head);
3326                 get_files_dirs(o, merge);
3327
3328                 entries = get_unmerged();
3329                 clean = detect_and_process_renames(o, common, head, merge,
3330                                                    entries, &re_info);
3331                 record_df_conflict_files(o, entries);
3332                 if (clean < 0)
3333                         goto cleanup;
3334                 for (i = entries->nr-1; 0 <= i; i--) {
3335                         const char *path = entries->items[i].string;
3336                         struct stage_data *e = entries->items[i].util;
3337                         if (!e->processed) {
3338                                 int ret = process_entry(o, path, e);
3339                                 if (!ret)
3340                                         clean = 0;
3341                                 else if (ret < 0) {
3342                                         clean = ret;
3343                                         goto cleanup;
3344                                 }
3345                         }
3346                 }
3347                 for (i = 0; i < entries->nr; i++) {
3348                         struct stage_data *e = entries->items[i].util;
3349                         if (!e->processed)
3350                                 BUG("unprocessed path??? %s",
3351                                     entries->items[i].string);
3352                 }
3353
3354         cleanup:
3355                 final_cleanup_renames(&re_info);
3356
3357                 string_list_clear(entries, 1);
3358                 free(entries);
3359
3360                 hashmap_free(&o->current_file_dir_set, 1);
3361
3362                 if (clean < 0) {
3363                         unpack_trees_finish(o);
3364                         return clean;
3365                 }
3366         }
3367         else
3368                 clean = 1;
3369
3370         unpack_trees_finish(o);
3371
3372         if (o->call_depth && !(*result = write_tree_from_memory(o)))
3373                 return -1;
3374
3375         return clean;
3376 }
3377
3378 static struct commit_list *reverse_commit_list(struct commit_list *list)
3379 {
3380         struct commit_list *next = NULL, *current, *backup;
3381         for (current = list; current; current = backup) {
3382                 backup = current->next;
3383                 current->next = next;
3384                 next = current;
3385         }
3386         return next;
3387 }
3388
3389 /*
3390  * Merge the commits h1 and h2, return the resulting virtual
3391  * commit object and a flag indicating the cleanness of the merge.
3392  */
3393 int merge_recursive(struct merge_options *o,
3394                     struct commit *h1,
3395                     struct commit *h2,
3396                     struct commit_list *ca,
3397                     struct commit **result)
3398 {
3399         struct commit_list *iter;
3400         struct commit *merged_common_ancestors;
3401         struct tree *mrtree;
3402         int clean;
3403
3404         if (show(o, 4)) {
3405                 output(o, 4, _("Merging:"));
3406                 output_commit_title(o, h1);
3407                 output_commit_title(o, h2);
3408         }
3409
3410         if (!ca) {
3411                 ca = get_merge_bases(h1, h2);
3412                 ca = reverse_commit_list(ca);
3413         }
3414
3415         if (show(o, 5)) {
3416                 unsigned cnt = commit_list_count(ca);
3417
3418                 output(o, 5, Q_("found %u common ancestor:",
3419                                 "found %u common ancestors:", cnt), cnt);
3420                 for (iter = ca; iter; iter = iter->next)
3421                         output_commit_title(o, iter->item);
3422         }
3423
3424         merged_common_ancestors = pop_commit(&ca);
3425         if (merged_common_ancestors == NULL) {
3426                 /* if there is no common ancestor, use an empty tree */
3427                 struct tree *tree;
3428
3429                 tree = lookup_tree(the_hash_algo->empty_tree);
3430                 merged_common_ancestors = make_virtual_commit(tree, "ancestor");
3431         }
3432
3433         for (iter = ca; iter; iter = iter->next) {
3434                 const char *saved_b1, *saved_b2;
3435                 o->call_depth++;
3436                 /*
3437                  * When the merge fails, the result contains files
3438                  * with conflict markers. The cleanness flag is
3439                  * ignored (unless indicating an error), it was never
3440                  * actually used, as result of merge_trees has always
3441                  * overwritten it: the committed "conflicts" were
3442                  * already resolved.
3443                  */
3444                 discard_cache();
3445                 saved_b1 = o->branch1;
3446                 saved_b2 = o->branch2;
3447                 o->branch1 = "Temporary merge branch 1";
3448                 o->branch2 = "Temporary merge branch 2";
3449                 if (merge_recursive(o, merged_common_ancestors, iter->item,
3450                                     NULL, &merged_common_ancestors) < 0)
3451                         return -1;
3452                 o->branch1 = saved_b1;
3453                 o->branch2 = saved_b2;
3454                 o->call_depth--;
3455
3456                 if (!merged_common_ancestors)
3457                         return err(o, _("merge returned no commit"));
3458         }
3459
3460         discard_cache();
3461         if (!o->call_depth)
3462                 read_cache();
3463
3464         o->ancestor = "merged common ancestors";
3465         clean = merge_trees(o, get_commit_tree(h1), get_commit_tree(h2),
3466                             get_commit_tree(merged_common_ancestors),
3467                             &mrtree);
3468         if (clean < 0) {
3469                 flush_output(o);
3470                 return clean;
3471         }
3472
3473         if (o->call_depth) {
3474                 *result = make_virtual_commit(mrtree, "merged tree");
3475                 commit_list_insert(h1, &(*result)->parents);
3476                 commit_list_insert(h2, &(*result)->parents->next);
3477         }
3478         flush_output(o);
3479         if (!o->call_depth && o->buffer_output < 2)
3480                 strbuf_release(&o->obuf);
3481         if (show(o, 2))
3482                 diff_warn_rename_limit("merge.renamelimit",
3483                                        o->needed_rename_limit, 0);
3484         return clean;
3485 }
3486
3487 static struct commit *get_ref(const struct object_id *oid, const char *name)
3488 {
3489         struct object *object;
3490
3491         object = deref_tag(parse_object(oid), name, strlen(name));
3492         if (!object)
3493                 return NULL;
3494         if (object->type == OBJ_TREE)
3495                 return make_virtual_commit((struct tree*)object, name);
3496         if (object->type != OBJ_COMMIT)
3497                 return NULL;
3498         if (parse_commit((struct commit *)object))
3499                 return NULL;
3500         return (struct commit *)object;
3501 }
3502
3503 int merge_recursive_generic(struct merge_options *o,
3504                             const struct object_id *head,
3505                             const struct object_id *merge,
3506                             int num_base_list,
3507                             const struct object_id **base_list,
3508                             struct commit **result)
3509 {
3510         int clean;
3511         struct lock_file lock = LOCK_INIT;
3512         struct commit *head_commit = get_ref(head, o->branch1);
3513         struct commit *next_commit = get_ref(merge, o->branch2);
3514         struct commit_list *ca = NULL;
3515
3516         if (base_list) {
3517                 int i;
3518                 for (i = 0; i < num_base_list; ++i) {
3519                         struct commit *base;
3520                         if (!(base = get_ref(base_list[i], oid_to_hex(base_list[i]))))
3521                                 return err(o, _("Could not parse object '%s'"),
3522                                            oid_to_hex(base_list[i]));
3523                         commit_list_insert(base, &ca);
3524                 }
3525         }
3526
3527         hold_locked_index(&lock, LOCK_DIE_ON_ERROR);
3528         clean = merge_recursive(o, head_commit, next_commit, ca,
3529                                 result);
3530         if (clean < 0) {
3531                 rollback_lock_file(&lock);
3532                 return clean;
3533         }
3534
3535         if (write_locked_index(&the_index, &lock,
3536                                COMMIT_LOCK | SKIP_IF_UNCHANGED))
3537                 return err(o, _("Unable to write index."));
3538
3539         return clean ? 0 : 1;
3540 }
3541
3542 static void merge_recursive_config(struct merge_options *o)
3543 {
3544         char *value = NULL;
3545         git_config_get_int("merge.verbosity", &o->verbosity);
3546         git_config_get_int("diff.renamelimit", &o->diff_rename_limit);
3547         git_config_get_int("merge.renamelimit", &o->merge_rename_limit);
3548         if (!git_config_get_string("diff.renames", &value)) {
3549                 o->diff_detect_rename = git_config_rename("diff.renames", value);
3550                 free(value);
3551         }
3552         if (!git_config_get_string("merge.renames", &value)) {
3553                 o->merge_detect_rename = git_config_rename("merge.renames", value);
3554                 free(value);
3555         }
3556         git_config(git_xmerge_config, NULL);
3557 }
3558
3559 void init_merge_options(struct merge_options *o)
3560 {
3561         const char *merge_verbosity;
3562         memset(o, 0, sizeof(struct merge_options));
3563         o->verbosity = 2;
3564         o->buffer_output = 1;
3565         o->diff_rename_limit = -1;
3566         o->merge_rename_limit = -1;
3567         o->renormalize = 0;
3568         o->diff_detect_rename = -1;
3569         o->merge_detect_rename = -1;
3570         merge_recursive_config(o);
3571         merge_verbosity = getenv("GIT_MERGE_VERBOSITY");
3572         if (merge_verbosity)
3573                 o->verbosity = strtol(merge_verbosity, NULL, 10);
3574         if (o->verbosity >= 5)
3575                 o->buffer_output = 0;
3576         strbuf_init(&o->obuf, 0);
3577         string_list_init(&o->df_conflict_file_set, 1);
3578 }
3579
3580 int parse_merge_opt(struct merge_options *o, const char *s)
3581 {
3582         const char *arg;
3583
3584         if (!s || !*s)
3585                 return -1;
3586         if (!strcmp(s, "ours"))
3587                 o->recursive_variant = MERGE_RECURSIVE_OURS;
3588         else if (!strcmp(s, "theirs"))
3589                 o->recursive_variant = MERGE_RECURSIVE_THEIRS;
3590         else if (!strcmp(s, "subtree"))
3591                 o->subtree_shift = "";
3592         else if (skip_prefix(s, "subtree=", &arg))
3593                 o->subtree_shift = arg;
3594         else if (!strcmp(s, "patience"))
3595                 o->xdl_opts = DIFF_WITH_ALG(o, PATIENCE_DIFF);
3596         else if (!strcmp(s, "histogram"))
3597                 o->xdl_opts = DIFF_WITH_ALG(o, HISTOGRAM_DIFF);
3598         else if (skip_prefix(s, "diff-algorithm=", &arg)) {
3599                 long value = parse_algorithm_value(arg);
3600                 if (value < 0)
3601                         return -1;
3602                 /* clear out previous settings */
3603                 DIFF_XDL_CLR(o, NEED_MINIMAL);
3604                 o->xdl_opts &= ~XDF_DIFF_ALGORITHM_MASK;
3605                 o->xdl_opts |= value;
3606         }
3607         else if (!strcmp(s, "ignore-space-change"))
3608                 DIFF_XDL_SET(o, IGNORE_WHITESPACE_CHANGE);
3609         else if (!strcmp(s, "ignore-all-space"))
3610                 DIFF_XDL_SET(o, IGNORE_WHITESPACE);
3611         else if (!strcmp(s, "ignore-space-at-eol"))
3612                 DIFF_XDL_SET(o, IGNORE_WHITESPACE_AT_EOL);
3613         else if (!strcmp(s, "ignore-cr-at-eol"))
3614                 DIFF_XDL_SET(o, IGNORE_CR_AT_EOL);
3615         else if (!strcmp(s, "renormalize"))
3616                 o->renormalize = 1;
3617         else if (!strcmp(s, "no-renormalize"))
3618                 o->renormalize = 0;
3619         else if (!strcmp(s, "no-renames"))
3620                 o->merge_detect_rename = 0;
3621         else if (!strcmp(s, "find-renames")) {
3622                 o->merge_detect_rename = 1;
3623                 o->rename_score = 0;
3624         }
3625         else if (skip_prefix(s, "find-renames=", &arg) ||
3626                  skip_prefix(s, "rename-threshold=", &arg)) {
3627                 if ((o->rename_score = parse_rename_score(&arg)) == -1 || *arg != 0)
3628                         return -1;
3629                 o->merge_detect_rename = 1;
3630         }
3631         else
3632                 return -1;
3633         return 0;
3634 }