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