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