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