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