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