object-store: provide helpers for loose_objects_cache
[git] / merge-recursive.c
1 /*
2  * Recursive Merge algorithm stolen from git-merge-recursive.py by
3  * Fredrik Kuivinen.
4  * The thieves were Alex Riesen and Johannes Schindelin, in June/July 2006
5  */
6 #include "cache.h"
7 #include "config.h"
8 #include "advice.h"
9 #include "lockfile.h"
10 #include "cache-tree.h"
11 #include "object-store.h"
12 #include "repository.h"
13 #include "commit.h"
14 #include "blob.h"
15 #include "builtin.h"
16 #include "tree-walk.h"
17 #include "diff.h"
18 #include "diffcore.h"
19 #include "tag.h"
20 #include "alloc.h"
21 #include "unpack-trees.h"
22 #include "string-list.h"
23 #include "xdiff-interface.h"
24 #include "ll-merge.h"
25 #include "attr.h"
26 #include "merge-recursive.h"
27 #include "dir.h"
28 #include "submodule.h"
29 #include "revision.h"
30 #include "commit-reach.h"
31
32 struct path_hashmap_entry {
33         struct hashmap_entry e;
34         char path[FLEX_ARRAY];
35 };
36
37 static int path_hashmap_cmp(const void *cmp_data,
38                             const void *entry,
39                             const void *entry_or_key,
40                             const void *keydata)
41 {
42         const struct path_hashmap_entry *a = entry;
43         const struct path_hashmap_entry *b = entry_or_key;
44         const char *key = keydata;
45
46         if (ignore_case)
47                 return strcasecmp(a->path, key ? key : b->path);
48         else
49                 return strcmp(a->path, key ? key : b->path);
50 }
51
52 static unsigned int path_hash(const char *path)
53 {
54         return ignore_case ? strihash(path) : strhash(path);
55 }
56
57 static struct dir_rename_entry *dir_rename_find_entry(struct hashmap *hashmap,
58                                                       char *dir)
59 {
60         struct dir_rename_entry key;
61
62         if (dir == NULL)
63                 return NULL;
64         hashmap_entry_init(&key, strhash(dir));
65         key.dir = dir;
66         return hashmap_get(hashmap, &key, NULL);
67 }
68
69 static int dir_rename_cmp(const void *unused_cmp_data,
70                           const void *entry,
71                           const void *entry_or_key,
72                           const void *unused_keydata)
73 {
74         const struct dir_rename_entry *e1 = entry;
75         const struct dir_rename_entry *e2 = entry_or_key;
76
77         return strcmp(e1->dir, e2->dir);
78 }
79
80 static void dir_rename_init(struct hashmap *map)
81 {
82         hashmap_init(map, dir_rename_cmp, NULL, 0);
83 }
84
85 static void dir_rename_entry_init(struct dir_rename_entry *entry,
86                                   char *directory)
87 {
88         hashmap_entry_init(entry, strhash(directory));
89         entry->dir = directory;
90         entry->non_unique_new_dir = 0;
91         strbuf_init(&entry->new_dir, 0);
92         string_list_init(&entry->possible_new_dirs, 0);
93 }
94
95 static struct collision_entry *collision_find_entry(struct hashmap *hashmap,
96                                                     char *target_file)
97 {
98         struct collision_entry key;
99
100         hashmap_entry_init(&key, strhash(target_file));
101         key.target_file = target_file;
102         return hashmap_get(hashmap, &key, NULL);
103 }
104
105 static int collision_cmp(void *unused_cmp_data,
106                          const struct collision_entry *e1,
107                          const struct collision_entry *e2,
108                          const void *unused_keydata)
109 {
110         return strcmp(e1->target_file, e2->target_file);
111 }
112
113 static void collision_init(struct hashmap *map)
114 {
115         hashmap_init(map, (hashmap_cmp_fn) collision_cmp, NULL, 0);
116 }
117
118 static void flush_output(struct merge_options *o)
119 {
120         if (o->buffer_output < 2 && o->obuf.len) {
121                 fputs(o->obuf.buf, stdout);
122                 strbuf_reset(&o->obuf);
123         }
124 }
125
126 static int err(struct merge_options *o, const char *err, ...)
127 {
128         va_list params;
129
130         if (o->buffer_output < 2)
131                 flush_output(o);
132         else {
133                 strbuf_complete(&o->obuf, '\n');
134                 strbuf_addstr(&o->obuf, "error: ");
135         }
136         va_start(params, err);
137         strbuf_vaddf(&o->obuf, err, params);
138         va_end(params);
139         if (o->buffer_output > 1)
140                 strbuf_addch(&o->obuf, '\n');
141         else {
142                 error("%s", o->obuf.buf);
143                 strbuf_reset(&o->obuf);
144         }
145
146         return -1;
147 }
148
149 static struct tree *shift_tree_object(struct tree *one, struct tree *two,
150                                       const char *subtree_shift)
151 {
152         struct object_id shifted;
153
154         if (!*subtree_shift) {
155                 shift_tree(&one->object.oid, &two->object.oid, &shifted, 0);
156         } else {
157                 shift_tree_by(&one->object.oid, &two->object.oid, &shifted,
158                               subtree_shift);
159         }
160         if (oideq(&two->object.oid, &shifted))
161                 return two;
162         return lookup_tree(the_repository, &shifted);
163 }
164
165 static struct commit *make_virtual_commit(struct tree *tree, const char *comment)
166 {
167         struct commit *commit = alloc_commit_node(the_repository);
168
169         set_merge_remote_desc(commit, comment, (struct object *)commit);
170         commit->maybe_tree = tree;
171         commit->object.parsed = 1;
172         return commit;
173 }
174
175 /*
176  * Since we use get_tree_entry(), which does not put the read object into
177  * the object pool, we cannot rely on a == b.
178  */
179 static int oid_eq(const struct object_id *a, const struct object_id *b)
180 {
181         if (!a && !b)
182                 return 2;
183         return a && b && oideq(a, b);
184 }
185
186 enum rename_type {
187         RENAME_NORMAL = 0,
188         RENAME_VIA_DIR,
189         RENAME_DELETE,
190         RENAME_ONE_FILE_TO_ONE,
191         RENAME_ONE_FILE_TO_TWO,
192         RENAME_TWO_FILES_TO_ONE
193 };
194
195 struct rename_conflict_info {
196         enum rename_type rename_type;
197         struct diff_filepair *pair1;
198         struct diff_filepair *pair2;
199         const char *branch1;
200         const char *branch2;
201         struct stage_data *dst_entry1;
202         struct stage_data *dst_entry2;
203         struct diff_filespec ren1_other;
204         struct diff_filespec ren2_other;
205 };
206
207 /*
208  * Since we want to write the index eventually, we cannot reuse the index
209  * for these (temporary) data.
210  */
211 struct stage_data {
212         struct {
213                 unsigned mode;
214                 struct object_id oid;
215         } stages[4];
216         struct rename_conflict_info *rename_conflict_info;
217         unsigned processed:1;
218 };
219
220 static inline void setup_rename_conflict_info(enum rename_type rename_type,
221                                               struct diff_filepair *pair1,
222                                               struct diff_filepair *pair2,
223                                               const char *branch1,
224                                               const char *branch2,
225                                               struct stage_data *dst_entry1,
226                                               struct stage_data *dst_entry2,
227                                               struct merge_options *o,
228                                               struct stage_data *src_entry1,
229                                               struct stage_data *src_entry2)
230 {
231         struct rename_conflict_info *ci;
232
233         /*
234          * When we have two renames involved, it's easiest to get the
235          * correct things into stage 2 and 3, and to make sure that the
236          * content merge puts HEAD before the other branch if we just
237          * ensure that branch1 == o->branch1.  So, simply flip arguments
238          * around if we don't have that.
239          */
240         if (dst_entry2 && branch1 != o->branch1) {
241                 setup_rename_conflict_info(rename_type,
242                                            pair2,      pair1,
243                                            branch2,    branch1,
244                                            dst_entry2, dst_entry1,
245                                            o,
246                                            src_entry2, src_entry1);
247                 return;
248         }
249
250         ci = xcalloc(1, sizeof(struct rename_conflict_info));
251         ci->rename_type = rename_type;
252         ci->pair1 = pair1;
253         ci->branch1 = branch1;
254         ci->branch2 = branch2;
255
256         ci->dst_entry1 = dst_entry1;
257         dst_entry1->rename_conflict_info = ci;
258         dst_entry1->processed = 0;
259
260         assert(!pair2 == !dst_entry2);
261         if (dst_entry2) {
262                 ci->dst_entry2 = dst_entry2;
263                 ci->pair2 = pair2;
264                 dst_entry2->rename_conflict_info = ci;
265         }
266
267         if (rename_type == RENAME_TWO_FILES_TO_ONE) {
268                 /*
269                  * For each rename, there could have been
270                  * modifications on the side of history where that
271                  * file was not renamed.
272                  */
273                 int ostage1 = o->branch1 == branch1 ? 3 : 2;
274                 int ostage2 = ostage1 ^ 1;
275
276                 ci->ren1_other.path = pair1->one->path;
277                 oidcpy(&ci->ren1_other.oid, &src_entry1->stages[ostage1].oid);
278                 ci->ren1_other.mode = src_entry1->stages[ostage1].mode;
279
280                 ci->ren2_other.path = pair2->one->path;
281                 oidcpy(&ci->ren2_other.oid, &src_entry2->stages[ostage2].oid);
282                 ci->ren2_other.mode = src_entry2->stages[ostage2].mode;
283         }
284 }
285
286 static int show(struct merge_options *o, int v)
287 {
288         return (!o->call_depth && o->verbosity >= v) || o->verbosity >= 5;
289 }
290
291 __attribute__((format (printf, 3, 4)))
292 static void output(struct merge_options *o, int v, const char *fmt, ...)
293 {
294         va_list ap;
295
296         if (!show(o, v))
297                 return;
298
299         strbuf_addchars(&o->obuf, ' ', o->call_depth * 2);
300
301         va_start(ap, fmt);
302         strbuf_vaddf(&o->obuf, fmt, ap);
303         va_end(ap);
304
305         strbuf_addch(&o->obuf, '\n');
306         if (!o->buffer_output)
307                 flush_output(o);
308 }
309
310 static void output_commit_title(struct merge_options *o, struct commit *commit)
311 {
312         struct merge_remote_desc *desc;
313
314         strbuf_addchars(&o->obuf, ' ', o->call_depth * 2);
315         desc = merge_remote_util(commit);
316         if (desc)
317                 strbuf_addf(&o->obuf, "virtual %s\n", desc->name);
318         else {
319                 strbuf_add_unique_abbrev(&o->obuf, &commit->object.oid,
320                                          DEFAULT_ABBREV);
321                 strbuf_addch(&o->obuf, ' ');
322                 if (parse_commit(commit) != 0)
323                         strbuf_addstr(&o->obuf, _("(bad commit)\n"));
324                 else {
325                         const char *title;
326                         const char *msg = get_commit_buffer(commit, NULL);
327                         int len = find_commit_subject(msg, &title);
328                         if (len)
329                                 strbuf_addf(&o->obuf, "%.*s\n", len, title);
330                         unuse_commit_buffer(commit, msg);
331                 }
332         }
333         flush_output(o);
334 }
335
336 static int add_cacheinfo(struct merge_options *o,
337                          unsigned int mode, const struct object_id *oid,
338                          const char *path, int stage, int refresh, int options)
339 {
340         struct cache_entry *ce;
341         int ret;
342
343         ce = make_cache_entry(&the_index, mode, oid ? oid : &null_oid, path, stage, 0);
344         if (!ce)
345                 return err(o, _("add_cacheinfo failed for path '%s'; merge aborting."), path);
346
347         ret = add_cache_entry(ce, options);
348         if (refresh) {
349                 struct cache_entry *nce;
350
351                 nce = refresh_cache_entry(&the_index, ce, CE_MATCH_REFRESH | CE_MATCH_IGNORE_MISSING);
352                 if (!nce)
353                         return err(o, _("add_cacheinfo failed to refresh for path '%s'; merge aborting."), path);
354                 if (nce != ce)
355                         ret = add_cache_entry(nce, options);
356         }
357         return ret;
358 }
359
360 static void init_tree_desc_from_tree(struct tree_desc *desc, struct tree *tree)
361 {
362         parse_tree(tree);
363         init_tree_desc(desc, tree->buffer, tree->size);
364 }
365
366 static int unpack_trees_start(struct merge_options *o,
367                               struct tree *common,
368                               struct tree *head,
369                               struct tree *merge)
370 {
371         int rc;
372         struct tree_desc t[3];
373         struct index_state tmp_index = { NULL };
374
375         memset(&o->unpack_opts, 0, sizeof(o->unpack_opts));
376         if (o->call_depth)
377                 o->unpack_opts.index_only = 1;
378         else
379                 o->unpack_opts.update = 1;
380         o->unpack_opts.merge = 1;
381         o->unpack_opts.head_idx = 2;
382         o->unpack_opts.fn = threeway_merge;
383         o->unpack_opts.src_index = &the_index;
384         o->unpack_opts.dst_index = &tmp_index;
385         o->unpack_opts.aggressive = !merge_detect_rename(o);
386         setup_unpack_trees_porcelain(&o->unpack_opts, "merge");
387
388         init_tree_desc_from_tree(t+0, common);
389         init_tree_desc_from_tree(t+1, head);
390         init_tree_desc_from_tree(t+2, merge);
391
392         rc = unpack_trees(3, t, &o->unpack_opts);
393         cache_tree_free(&active_cache_tree);
394
395         /*
396          * Update the_index to match the new results, AFTER saving a copy
397          * in o->orig_index.  Update src_index to point to the saved copy.
398          * (verify_uptodate() checks src_index, and the original index is
399          * the one that had the necessary modification timestamps.)
400          */
401         o->orig_index = the_index;
402         the_index = tmp_index;
403         o->unpack_opts.src_index = &o->orig_index;
404
405         return rc;
406 }
407
408 static void unpack_trees_finish(struct merge_options *o)
409 {
410         discard_index(&o->orig_index);
411         clear_unpack_trees_porcelain(&o->unpack_opts);
412 }
413
414 struct tree *write_tree_from_memory(struct merge_options *o)
415 {
416         struct tree *result = NULL;
417
418         if (unmerged_cache()) {
419                 int i;
420                 fprintf(stderr, "BUG: There are unmerged index entries:\n");
421                 for (i = 0; i < active_nr; i++) {
422                         const struct cache_entry *ce = active_cache[i];
423                         if (ce_stage(ce))
424                                 fprintf(stderr, "BUG: %d %.*s\n", ce_stage(ce),
425                                         (int)ce_namelen(ce), ce->name);
426                 }
427                 BUG("unmerged index entries in merge-recursive.c");
428         }
429
430         if (!active_cache_tree)
431                 active_cache_tree = cache_tree();
432
433         if (!cache_tree_fully_valid(active_cache_tree) &&
434             cache_tree_update(&the_index, 0) < 0) {
435                 err(o, _("error building trees"));
436                 return NULL;
437         }
438
439         result = lookup_tree(the_repository, &active_cache_tree->oid);
440
441         return result;
442 }
443
444 static int save_files_dirs(const struct object_id *oid,
445                            struct strbuf *base, const char *path,
446                            unsigned int mode, int stage, void *context)
447 {
448         struct path_hashmap_entry *entry;
449         int baselen = base->len;
450         struct merge_options *o = context;
451
452         strbuf_addstr(base, path);
453
454         FLEX_ALLOC_MEM(entry, path, base->buf, base->len);
455         hashmap_entry_init(entry, path_hash(entry->path));
456         hashmap_add(&o->current_file_dir_set, entry);
457
458         strbuf_setlen(base, baselen);
459         return (S_ISDIR(mode) ? READ_TREE_RECURSIVE : 0);
460 }
461
462 static void get_files_dirs(struct merge_options *o, struct tree *tree)
463 {
464         struct pathspec match_all;
465         memset(&match_all, 0, sizeof(match_all));
466         read_tree_recursive(tree, "", 0, 0, &match_all, save_files_dirs, o);
467 }
468
469 static int get_tree_entry_if_blob(const struct object_id *tree,
470                                   const char *path,
471                                   struct object_id *hashy,
472                                   unsigned int *mode_o)
473 {
474         int ret;
475
476         ret = get_tree_entry(tree, path, hashy, mode_o);
477         if (S_ISDIR(*mode_o)) {
478                 oidcpy(hashy, &null_oid);
479                 *mode_o = 0;
480         }
481         return ret;
482 }
483
484 /*
485  * Returns an index_entry instance which doesn't have to correspond to
486  * a real cache entry in Git's index.
487  */
488 static struct stage_data *insert_stage_data(const char *path,
489                 struct tree *o, struct tree *a, struct tree *b,
490                 struct string_list *entries)
491 {
492         struct string_list_item *item;
493         struct stage_data *e = xcalloc(1, sizeof(struct stage_data));
494         get_tree_entry_if_blob(&o->object.oid, path,
495                                &e->stages[1].oid, &e->stages[1].mode);
496         get_tree_entry_if_blob(&a->object.oid, path,
497                                &e->stages[2].oid, &e->stages[2].mode);
498         get_tree_entry_if_blob(&b->object.oid, path,
499                                &e->stages[3].oid, &e->stages[3].mode);
500         item = string_list_insert(entries, path);
501         item->util = e;
502         return e;
503 }
504
505 /*
506  * Create a dictionary mapping file names to stage_data objects. The
507  * dictionary contains one entry for every path with a non-zero stage entry.
508  */
509 static struct string_list *get_unmerged(void)
510 {
511         struct string_list *unmerged = xcalloc(1, sizeof(struct string_list));
512         int i;
513
514         unmerged->strdup_strings = 1;
515
516         for (i = 0; i < active_nr; i++) {
517                 struct string_list_item *item;
518                 struct stage_data *e;
519                 const struct cache_entry *ce = active_cache[i];
520                 if (!ce_stage(ce))
521                         continue;
522
523                 item = string_list_lookup(unmerged, ce->name);
524                 if (!item) {
525                         item = string_list_insert(unmerged, ce->name);
526                         item->util = xcalloc(1, sizeof(struct stage_data));
527                 }
528                 e = item->util;
529                 e->stages[ce_stage(ce)].mode = ce->ce_mode;
530                 oidcpy(&e->stages[ce_stage(ce)].oid, &ce->oid);
531         }
532
533         return unmerged;
534 }
535
536 static int string_list_df_name_compare(const char *one, const char *two)
537 {
538         int onelen = strlen(one);
539         int twolen = strlen(two);
540         /*
541          * Here we only care that entries for D/F conflicts are
542          * adjacent, in particular with the file of the D/F conflict
543          * appearing before files below the corresponding directory.
544          * The order of the rest of the list is irrelevant for us.
545          *
546          * To achieve this, we sort with df_name_compare and provide
547          * the mode S_IFDIR so that D/F conflicts will sort correctly.
548          * We use the mode S_IFDIR for everything else for simplicity,
549          * since in other cases any changes in their order due to
550          * sorting cause no problems for us.
551          */
552         int cmp = df_name_compare(one, onelen, S_IFDIR,
553                                   two, twolen, S_IFDIR);
554         /*
555          * Now that 'foo' and 'foo/bar' compare equal, we have to make sure
556          * that 'foo' comes before 'foo/bar'.
557          */
558         if (cmp)
559                 return cmp;
560         return onelen - twolen;
561 }
562
563 static void record_df_conflict_files(struct merge_options *o,
564                                      struct string_list *entries)
565 {
566         /* If there is a D/F conflict and the file for such a conflict
567          * currently exists in the working tree, we want to allow it to be
568          * removed to make room for the corresponding directory if needed.
569          * The files underneath the directories of such D/F conflicts will
570          * be processed before the corresponding file involved in the D/F
571          * conflict.  If the D/F directory ends up being removed by the
572          * merge, then we won't have to touch the D/F file.  If the D/F
573          * directory needs to be written to the working copy, then the D/F
574          * file will simply be removed (in make_room_for_path()) to make
575          * room for the necessary paths.  Note that if both the directory
576          * and the file need to be present, then the D/F file will be
577          * reinstated with a new unique name at the time it is processed.
578          */
579         struct string_list df_sorted_entries = STRING_LIST_INIT_NODUP;
580         const char *last_file = NULL;
581         int last_len = 0;
582         int i;
583
584         /*
585          * If we're merging merge-bases, we don't want to bother with
586          * any working directory changes.
587          */
588         if (o->call_depth)
589                 return;
590
591         /* Ensure D/F conflicts are adjacent in the entries list. */
592         for (i = 0; i < entries->nr; i++) {
593                 struct string_list_item *next = &entries->items[i];
594                 string_list_append(&df_sorted_entries, next->string)->util =
595                                    next->util;
596         }
597         df_sorted_entries.cmp = string_list_df_name_compare;
598         string_list_sort(&df_sorted_entries);
599
600         string_list_clear(&o->df_conflict_file_set, 1);
601         for (i = 0; i < df_sorted_entries.nr; i++) {
602                 const char *path = df_sorted_entries.items[i].string;
603                 int len = strlen(path);
604                 struct stage_data *e = df_sorted_entries.items[i].util;
605
606                 /*
607                  * Check if last_file & path correspond to a D/F conflict;
608                  * i.e. whether path is last_file+'/'+<something>.
609                  * If so, record that it's okay to remove last_file to make
610                  * room for path and friends if needed.
611                  */
612                 if (last_file &&
613                     len > last_len &&
614                     memcmp(path, last_file, last_len) == 0 &&
615                     path[last_len] == '/') {
616                         string_list_insert(&o->df_conflict_file_set, last_file);
617                 }
618
619                 /*
620                  * Determine whether path could exist as a file in the
621                  * working directory as a possible D/F conflict.  This
622                  * will only occur when it exists in stage 2 as a
623                  * file.
624                  */
625                 if (S_ISREG(e->stages[2].mode) || S_ISLNK(e->stages[2].mode)) {
626                         last_file = path;
627                         last_len = len;
628                 } else {
629                         last_file = NULL;
630                 }
631         }
632         string_list_clear(&df_sorted_entries, 0);
633 }
634
635 struct rename {
636         struct diff_filepair *pair;
637         /*
638          * Purpose of src_entry and dst_entry:
639          *
640          * If 'before' is renamed to 'after' then src_entry will contain
641          * the versions of 'before' from the merge_base, HEAD, and MERGE in
642          * stages 1, 2, and 3; dst_entry will contain the respective
643          * versions of 'after' in corresponding locations.  Thus, we have a
644          * total of six modes and oids, though some will be null.  (Stage 0
645          * is ignored; we're interested in handling conflicts.)
646          *
647          * Since we don't turn on break-rewrites by default, neither
648          * src_entry nor dst_entry can have all three of their stages have
649          * non-null oids, meaning at most four of the six will be non-null.
650          * Also, since this is a rename, both src_entry and dst_entry will
651          * have at least one non-null oid, meaning at least two will be
652          * non-null.  Of the six oids, a typical rename will have three be
653          * non-null.  Only two implies a rename/delete, and four implies a
654          * rename/add.
655          */
656         struct stage_data *src_entry;
657         struct stage_data *dst_entry;
658         unsigned add_turned_into_rename:1;
659         unsigned processed:1;
660 };
661
662 static int update_stages(struct merge_options *opt, const char *path,
663                          const struct diff_filespec *o,
664                          const struct diff_filespec *a,
665                          const struct diff_filespec *b)
666 {
667
668         /*
669          * NOTE: It is usually a bad idea to call update_stages on a path
670          * before calling update_file on that same path, since it can
671          * sometimes lead to spurious "refusing to lose untracked file..."
672          * messages from update_file (via make_room_for path via
673          * would_lose_untracked).  Instead, reverse the order of the calls
674          * (executing update_file first and then update_stages).
675          */
676         int clear = 1;
677         int options = ADD_CACHE_OK_TO_ADD | ADD_CACHE_SKIP_DFCHECK;
678         if (clear)
679                 if (remove_file_from_cache(path))
680                         return -1;
681         if (o)
682                 if (add_cacheinfo(opt, o->mode, &o->oid, path, 1, 0, options))
683                         return -1;
684         if (a)
685                 if (add_cacheinfo(opt, a->mode, &a->oid, path, 2, 0, options))
686                         return -1;
687         if (b)
688                 if (add_cacheinfo(opt, b->mode, &b->oid, path, 3, 0, options))
689                         return -1;
690         return 0;
691 }
692
693 static int update_stages_for_stage_data(struct merge_options *opt,
694                                         const char *path,
695                                         const struct stage_data *stage_data)
696 {
697         struct diff_filespec o, a, b;
698
699         o.mode = stage_data->stages[1].mode;
700         oidcpy(&o.oid, &stage_data->stages[1].oid);
701
702         a.mode = stage_data->stages[2].mode;
703         oidcpy(&a.oid, &stage_data->stages[2].oid);
704
705         b.mode = stage_data->stages[3].mode;
706         oidcpy(&b.oid, &stage_data->stages[3].oid);
707
708         return update_stages(opt, path,
709                              is_null_oid(&o.oid) ? NULL : &o,
710                              is_null_oid(&a.oid) ? NULL : &a,
711                              is_null_oid(&b.oid) ? NULL : &b);
712 }
713
714 static void update_entry(struct stage_data *entry,
715                          struct diff_filespec *o,
716                          struct diff_filespec *a,
717                          struct diff_filespec *b)
718 {
719         entry->processed = 0;
720         entry->stages[1].mode = o->mode;
721         entry->stages[2].mode = a->mode;
722         entry->stages[3].mode = b->mode;
723         oidcpy(&entry->stages[1].oid, &o->oid);
724         oidcpy(&entry->stages[2].oid, &a->oid);
725         oidcpy(&entry->stages[3].oid, &b->oid);
726 }
727
728 static int remove_file(struct merge_options *o, int clean,
729                        const char *path, int no_wd)
730 {
731         int update_cache = o->call_depth || clean;
732         int update_working_directory = !o->call_depth && !no_wd;
733
734         if (update_cache) {
735                 if (remove_file_from_cache(path))
736                         return -1;
737         }
738         if (update_working_directory) {
739                 if (ignore_case) {
740                         struct cache_entry *ce;
741                         ce = cache_file_exists(path, strlen(path), ignore_case);
742                         if (ce && ce_stage(ce) == 0 && strcmp(path, ce->name))
743                                 return 0;
744                 }
745                 if (remove_path(path))
746                         return -1;
747         }
748         return 0;
749 }
750
751 /* add a string to a strbuf, but converting "/" to "_" */
752 static void add_flattened_path(struct strbuf *out, const char *s)
753 {
754         size_t i = out->len;
755         strbuf_addstr(out, s);
756         for (; i < out->len; i++)
757                 if (out->buf[i] == '/')
758                         out->buf[i] = '_';
759 }
760
761 static char *unique_path(struct merge_options *o, const char *path, const char *branch)
762 {
763         struct path_hashmap_entry *entry;
764         struct strbuf newpath = STRBUF_INIT;
765         int suffix = 0;
766         size_t base_len;
767
768         strbuf_addf(&newpath, "%s~", path);
769         add_flattened_path(&newpath, branch);
770
771         base_len = newpath.len;
772         while (hashmap_get_from_hash(&o->current_file_dir_set,
773                                      path_hash(newpath.buf), newpath.buf) ||
774                (!o->call_depth && file_exists(newpath.buf))) {
775                 strbuf_setlen(&newpath, base_len);
776                 strbuf_addf(&newpath, "_%d", suffix++);
777         }
778
779         FLEX_ALLOC_MEM(entry, path, newpath.buf, newpath.len);
780         hashmap_entry_init(entry, path_hash(entry->path));
781         hashmap_add(&o->current_file_dir_set, entry);
782         return strbuf_detach(&newpath, NULL);
783 }
784
785 /**
786  * Check whether a directory in the index is in the way of an incoming
787  * file.  Return 1 if so.  If check_working_copy is non-zero, also
788  * check the working directory.  If empty_ok is non-zero, also return
789  * 0 in the case where the working-tree dir exists but is empty.
790  */
791 static int dir_in_way(const char *path, int check_working_copy, int empty_ok)
792 {
793         int pos;
794         struct strbuf dirpath = STRBUF_INIT;
795         struct stat st;
796
797         strbuf_addstr(&dirpath, path);
798         strbuf_addch(&dirpath, '/');
799
800         pos = cache_name_pos(dirpath.buf, dirpath.len);
801
802         if (pos < 0)
803                 pos = -1 - pos;
804         if (pos < active_nr &&
805             !strncmp(dirpath.buf, active_cache[pos]->name, dirpath.len)) {
806                 strbuf_release(&dirpath);
807                 return 1;
808         }
809
810         strbuf_release(&dirpath);
811         return check_working_copy && !lstat(path, &st) && S_ISDIR(st.st_mode) &&
812                 !(empty_ok && is_empty_dir(path));
813 }
814
815 /*
816  * Returns whether path was tracked in the index before the merge started,
817  * and its oid and mode match the specified values
818  */
819 static int was_tracked_and_matches(struct merge_options *o, const char *path,
820                                    const struct object_id *oid, unsigned mode)
821 {
822         int pos = index_name_pos(&o->orig_index, path, strlen(path));
823         struct cache_entry *ce;
824
825         if (0 > pos)
826                 /* we were not tracking this path before the merge */
827                 return 0;
828
829         /* See if the file we were tracking before matches */
830         ce = o->orig_index.cache[pos];
831         return (oid_eq(&ce->oid, oid) && ce->ce_mode == mode);
832 }
833
834 /*
835  * Returns whether path was tracked in the index before the merge started
836  */
837 static int was_tracked(struct merge_options *o, const char *path)
838 {
839         int pos = index_name_pos(&o->orig_index, path, strlen(path));
840
841         if (0 <= pos)
842                 /* we were tracking this path before the merge */
843                 return 1;
844
845         return 0;
846 }
847
848 static int would_lose_untracked(const char *path)
849 {
850         /*
851          * This may look like it can be simplified to:
852          *   return !was_tracked(o, path) && file_exists(path)
853          * but it can't.  This function needs to know whether path was in
854          * the working tree due to EITHER having been tracked in the index
855          * before the merge OR having been put into the working copy and
856          * index by unpack_trees().  Due to that either-or requirement, we
857          * check the current index instead of the original one.
858          *
859          * Note that we do not need to worry about merge-recursive itself
860          * updating the index after unpack_trees() and before calling this
861          * function, because we strictly require all code paths in
862          * merge-recursive to update the working tree first and the index
863          * second.  Doing otherwise would break
864          * update_file()/would_lose_untracked(); see every comment in this
865          * file which mentions "update_stages".
866          */
867         int pos = cache_name_pos(path, strlen(path));
868
869         if (pos < 0)
870                 pos = -1 - pos;
871         while (pos < active_nr &&
872                !strcmp(path, active_cache[pos]->name)) {
873                 /*
874                  * If stage #0, it is definitely tracked.
875                  * If it has stage #2 then it was tracked
876                  * before this merge started.  All other
877                  * cases the path was not tracked.
878                  */
879                 switch (ce_stage(active_cache[pos])) {
880                 case 0:
881                 case 2:
882                         return 0;
883                 }
884                 pos++;
885         }
886         return file_exists(path);
887 }
888
889 static int was_dirty(struct merge_options *o, const char *path)
890 {
891         struct cache_entry *ce;
892         int dirty = 1;
893
894         if (o->call_depth || !was_tracked(o, path))
895                 return !dirty;
896
897         ce = index_file_exists(o->unpack_opts.src_index,
898                                path, strlen(path), ignore_case);
899         dirty = verify_uptodate(ce, &o->unpack_opts) != 0;
900         return dirty;
901 }
902
903 static int make_room_for_path(struct merge_options *o, const char *path)
904 {
905         int status, i;
906         const char *msg = _("failed to create path '%s'%s");
907
908         /* Unlink any D/F conflict files that are in the way */
909         for (i = 0; i < o->df_conflict_file_set.nr; i++) {
910                 const char *df_path = o->df_conflict_file_set.items[i].string;
911                 size_t pathlen = strlen(path);
912                 size_t df_pathlen = strlen(df_path);
913                 if (df_pathlen < pathlen &&
914                     path[df_pathlen] == '/' &&
915                     strncmp(path, df_path, df_pathlen) == 0) {
916                         output(o, 3,
917                                _("Removing %s to make room for subdirectory\n"),
918                                df_path);
919                         unlink(df_path);
920                         unsorted_string_list_delete_item(&o->df_conflict_file_set,
921                                                          i, 0);
922                         break;
923                 }
924         }
925
926         /* Make sure leading directories are created */
927         status = safe_create_leading_directories_const(path);
928         if (status) {
929                 if (status == SCLD_EXISTS)
930                         /* something else exists */
931                         return err(o, msg, path, _(": perhaps a D/F conflict?"));
932                 return err(o, msg, path, "");
933         }
934
935         /*
936          * Do not unlink a file in the work tree if we are not
937          * tracking it.
938          */
939         if (would_lose_untracked(path))
940                 return err(o, _("refusing to lose untracked file at '%s'"),
941                            path);
942
943         /* Successful unlink is good.. */
944         if (!unlink(path))
945                 return 0;
946         /* .. and so is no existing file */
947         if (errno == ENOENT)
948                 return 0;
949         /* .. but not some other error (who really cares what?) */
950         return err(o, msg, path, _(": perhaps a D/F conflict?"));
951 }
952
953 static int update_file_flags(struct merge_options *o,
954                              const struct object_id *oid,
955                              unsigned mode,
956                              const char *path,
957                              int update_cache,
958                              int update_wd)
959 {
960         int ret = 0;
961
962         if (o->call_depth)
963                 update_wd = 0;
964
965         if (update_wd) {
966                 enum object_type type;
967                 void *buf;
968                 unsigned long size;
969
970                 if (S_ISGITLINK(mode)) {
971                         /*
972                          * We may later decide to recursively descend into
973                          * the submodule directory and update its index
974                          * and/or work tree, but we do not do that now.
975                          */
976                         update_wd = 0;
977                         goto update_index;
978                 }
979
980                 buf = read_object_file(oid, &type, &size);
981                 if (!buf)
982                         return err(o, _("cannot read object %s '%s'"), oid_to_hex(oid), path);
983                 if (type != OBJ_BLOB) {
984                         ret = err(o, _("blob expected for %s '%s'"), oid_to_hex(oid), path);
985                         goto free_buf;
986                 }
987                 if (S_ISREG(mode)) {
988                         struct strbuf strbuf = STRBUF_INIT;
989                         if (convert_to_working_tree(&the_index, path, buf, size, &strbuf)) {
990                                 free(buf);
991                                 size = strbuf.len;
992                                 buf = strbuf_detach(&strbuf, NULL);
993                         }
994                 }
995
996                 if (make_room_for_path(o, path) < 0) {
997                         update_wd = 0;
998                         goto free_buf;
999                 }
1000                 if (S_ISREG(mode) || (!has_symlinks && S_ISLNK(mode))) {
1001                         int fd;
1002                         if (mode & 0100)
1003                                 mode = 0777;
1004                         else
1005                                 mode = 0666;
1006                         fd = open(path, O_WRONLY | O_TRUNC | O_CREAT, mode);
1007                         if (fd < 0) {
1008                                 ret = err(o, _("failed to open '%s': %s"),
1009                                           path, strerror(errno));
1010                                 goto free_buf;
1011                         }
1012                         write_in_full(fd, buf, size);
1013                         close(fd);
1014                 } else if (S_ISLNK(mode)) {
1015                         char *lnk = xmemdupz(buf, size);
1016                         safe_create_leading_directories_const(path);
1017                         unlink(path);
1018                         if (symlink(lnk, path))
1019                                 ret = err(o, _("failed to symlink '%s': %s"),
1020                                           path, strerror(errno));
1021                         free(lnk);
1022                 } else
1023                         ret = err(o,
1024                                   _("do not know what to do with %06o %s '%s'"),
1025                                   mode, oid_to_hex(oid), path);
1026         free_buf:
1027                 free(buf);
1028         }
1029 update_index:
1030         if (!ret && update_cache)
1031                 if (add_cacheinfo(o, mode, oid, path, 0, update_wd,
1032                                   ADD_CACHE_OK_TO_ADD))
1033                         return -1;
1034         return ret;
1035 }
1036
1037 static int update_file(struct merge_options *o,
1038                        int clean,
1039                        const struct object_id *oid,
1040                        unsigned mode,
1041                        const char *path)
1042 {
1043         return update_file_flags(o, oid, mode, path, o->call_depth || clean, !o->call_depth);
1044 }
1045
1046 /* Low level file merging, update and removal */
1047
1048 struct merge_file_info {
1049         struct object_id oid;
1050         unsigned mode;
1051         unsigned clean:1,
1052                  merge:1;
1053 };
1054
1055 static int merge_3way(struct merge_options *o,
1056                       mmbuffer_t *result_buf,
1057                       const struct diff_filespec *one,
1058                       const struct diff_filespec *a,
1059                       const struct diff_filespec *b,
1060                       const char *branch1,
1061                       const char *branch2)
1062 {
1063         mmfile_t orig, src1, src2;
1064         struct ll_merge_options ll_opts = {0};
1065         char *base_name, *name1, *name2;
1066         int merge_status;
1067
1068         ll_opts.renormalize = o->renormalize;
1069         ll_opts.xdl_opts = o->xdl_opts;
1070
1071         if (o->call_depth) {
1072                 ll_opts.virtual_ancestor = 1;
1073                 ll_opts.variant = 0;
1074         } else {
1075                 switch (o->recursive_variant) {
1076                 case MERGE_RECURSIVE_OURS:
1077                         ll_opts.variant = XDL_MERGE_FAVOR_OURS;
1078                         break;
1079                 case MERGE_RECURSIVE_THEIRS:
1080                         ll_opts.variant = XDL_MERGE_FAVOR_THEIRS;
1081                         break;
1082                 default:
1083                         ll_opts.variant = 0;
1084                         break;
1085                 }
1086         }
1087
1088         if (strcmp(a->path, b->path) ||
1089             (o->ancestor != NULL && strcmp(a->path, one->path) != 0)) {
1090                 base_name = o->ancestor == NULL ? NULL :
1091                         mkpathdup("%s:%s", o->ancestor, one->path);
1092                 name1 = mkpathdup("%s:%s", branch1, a->path);
1093                 name2 = mkpathdup("%s:%s", branch2, b->path);
1094         } else {
1095                 base_name = o->ancestor == NULL ? NULL :
1096                         mkpathdup("%s", o->ancestor);
1097                 name1 = mkpathdup("%s", branch1);
1098                 name2 = mkpathdup("%s", branch2);
1099         }
1100
1101         read_mmblob(&orig, &one->oid);
1102         read_mmblob(&src1, &a->oid);
1103         read_mmblob(&src2, &b->oid);
1104
1105         merge_status = ll_merge(result_buf, a->path, &orig, base_name,
1106                                 &src1, name1, &src2, name2,
1107                                 &the_index, &ll_opts);
1108
1109         free(base_name);
1110         free(name1);
1111         free(name2);
1112         free(orig.ptr);
1113         free(src1.ptr);
1114         free(src2.ptr);
1115         return merge_status;
1116 }
1117
1118 static int find_first_merges(struct object_array *result, const char *path,
1119                              struct commit *a, struct commit *b)
1120 {
1121         int i, j;
1122         struct object_array merges = OBJECT_ARRAY_INIT;
1123         struct commit *commit;
1124         int contains_another;
1125
1126         char merged_revision[42];
1127         const char *rev_args[] = { "rev-list", "--merges", "--ancestry-path",
1128                                    "--all", merged_revision, NULL };
1129         struct rev_info revs;
1130         struct setup_revision_opt rev_opts;
1131
1132         memset(result, 0, sizeof(struct object_array));
1133         memset(&rev_opts, 0, sizeof(rev_opts));
1134
1135         /* get all revisions that merge commit a */
1136         xsnprintf(merged_revision, sizeof(merged_revision), "^%s",
1137                   oid_to_hex(&a->object.oid));
1138         repo_init_revisions(the_repository, &revs, NULL);
1139         rev_opts.submodule = path;
1140         /* FIXME: can't handle linked worktrees in submodules yet */
1141         revs.single_worktree = path != NULL;
1142         setup_revisions(ARRAY_SIZE(rev_args)-1, rev_args, &revs, &rev_opts);
1143
1144         /* save all revisions from the above list that contain b */
1145         if (prepare_revision_walk(&revs))
1146                 die("revision walk setup failed");
1147         while ((commit = get_revision(&revs)) != NULL) {
1148                 struct object *o = &(commit->object);
1149                 if (in_merge_bases(b, commit))
1150                         add_object_array(o, NULL, &merges);
1151         }
1152         reset_revision_walk();
1153
1154         /* Now we've got all merges that contain a and b. Prune all
1155          * merges that contain another found merge and save them in
1156          * result.
1157          */
1158         for (i = 0; i < merges.nr; i++) {
1159                 struct commit *m1 = (struct commit *) merges.objects[i].item;
1160
1161                 contains_another = 0;
1162                 for (j = 0; j < merges.nr; j++) {
1163                         struct commit *m2 = (struct commit *) merges.objects[j].item;
1164                         if (i != j && in_merge_bases(m2, m1)) {
1165                                 contains_another = 1;
1166                                 break;
1167                         }
1168                 }
1169
1170                 if (!contains_another)
1171                         add_object_array(merges.objects[i].item, NULL, result);
1172         }
1173
1174         object_array_clear(&merges);
1175         return result->nr;
1176 }
1177
1178 static void print_commit(struct commit *commit)
1179 {
1180         struct strbuf sb = STRBUF_INIT;
1181         struct pretty_print_context ctx = {0};
1182         ctx.date_mode.type = DATE_NORMAL;
1183         format_commit_message(commit, " %h: %m %s", &sb, &ctx);
1184         fprintf(stderr, "%s\n", sb.buf);
1185         strbuf_release(&sb);
1186 }
1187
1188 static int merge_submodule(struct merge_options *o,
1189                            struct object_id *result, const char *path,
1190                            const struct object_id *base, const struct object_id *a,
1191                            const struct object_id *b)
1192 {
1193         struct commit *commit_base, *commit_a, *commit_b;
1194         int parent_count;
1195         struct object_array merges;
1196
1197         int i;
1198         int search = !o->call_depth;
1199
1200         /* store a in result in case we fail */
1201         oidcpy(result, a);
1202
1203         /* we can not handle deletion conflicts */
1204         if (is_null_oid(base))
1205                 return 0;
1206         if (is_null_oid(a))
1207                 return 0;
1208         if (is_null_oid(b))
1209                 return 0;
1210
1211         if (add_submodule_odb(path)) {
1212                 output(o, 1, _("Failed to merge submodule %s (not checked out)"), path);
1213                 return 0;
1214         }
1215
1216         if (!(commit_base = lookup_commit_reference(the_repository, base)) ||
1217             !(commit_a = lookup_commit_reference(the_repository, a)) ||
1218             !(commit_b = lookup_commit_reference(the_repository, b))) {
1219                 output(o, 1, _("Failed to merge submodule %s (commits not present)"), path);
1220                 return 0;
1221         }
1222
1223         /* check whether both changes are forward */
1224         if (!in_merge_bases(commit_base, commit_a) ||
1225             !in_merge_bases(commit_base, commit_b)) {
1226                 output(o, 1, _("Failed to merge submodule %s (commits don't follow merge-base)"), path);
1227                 return 0;
1228         }
1229
1230         /* Case #1: a is contained in b or vice versa */
1231         if (in_merge_bases(commit_a, commit_b)) {
1232                 oidcpy(result, b);
1233                 if (show(o, 3)) {
1234                         output(o, 3, _("Fast-forwarding submodule %s to the following commit:"), path);
1235                         output_commit_title(o, commit_b);
1236                 } else if (show(o, 2))
1237                         output(o, 2, _("Fast-forwarding submodule %s"), path);
1238                 else
1239                         ; /* no output */
1240
1241                 return 1;
1242         }
1243         if (in_merge_bases(commit_b, commit_a)) {
1244                 oidcpy(result, a);
1245                 if (show(o, 3)) {
1246                         output(o, 3, _("Fast-forwarding submodule %s to the following commit:"), path);
1247                         output_commit_title(o, commit_a);
1248                 } else if (show(o, 2))
1249                         output(o, 2, _("Fast-forwarding submodule %s"), path);
1250                 else
1251                         ; /* no output */
1252
1253                 return 1;
1254         }
1255
1256         /*
1257          * Case #2: There are one or more merges that contain a and b in
1258          * the submodule. If there is only one, then present it as a
1259          * suggestion to the user, but leave it marked unmerged so the
1260          * user needs to confirm the resolution.
1261          */
1262
1263         /* Skip the search if makes no sense to the calling context.  */
1264         if (!search)
1265                 return 0;
1266
1267         /* find commit which merges them */
1268         parent_count = find_first_merges(&merges, path, commit_a, commit_b);
1269         switch (parent_count) {
1270         case 0:
1271                 output(o, 1, _("Failed to merge submodule %s (merge following commits not found)"), path);
1272                 break;
1273
1274         case 1:
1275                 output(o, 1, _("Failed to merge submodule %s (not fast-forward)"), path);
1276                 output(o, 2, _("Found a possible merge resolution for the submodule:\n"));
1277                 print_commit((struct commit *) merges.objects[0].item);
1278                 output(o, 2, _(
1279                        "If this is correct simply add it to the index "
1280                        "for example\n"
1281                        "by using:\n\n"
1282                        "  git update-index --cacheinfo 160000 %s \"%s\"\n\n"
1283                        "which will accept this suggestion.\n"),
1284                        oid_to_hex(&merges.objects[0].item->oid), path);
1285                 break;
1286
1287         default:
1288                 output(o, 1, _("Failed to merge submodule %s (multiple merges found)"), path);
1289                 for (i = 0; i < merges.nr; i++)
1290                         print_commit((struct commit *) merges.objects[i].item);
1291         }
1292
1293         object_array_clear(&merges);
1294         return 0;
1295 }
1296
1297 static int merge_mode_and_contents(struct merge_options *o,
1298                                    const struct diff_filespec *one,
1299                                    const struct diff_filespec *a,
1300                                    const struct diff_filespec *b,
1301                                    const char *filename,
1302                                    const char *branch1,
1303                                    const char *branch2,
1304                                    struct merge_file_info *result)
1305 {
1306         if (o->branch1 != branch1) {
1307                 /*
1308                  * It's weird getting a reverse merge with HEAD on the bottom
1309                  * side of the conflict markers and the other branch on the
1310                  * top.  Fix that.
1311                  */
1312                 return merge_mode_and_contents(o, one, b, a,
1313                                                filename,
1314                                                branch2, branch1, result);
1315         }
1316
1317         result->merge = 0;
1318         result->clean = 1;
1319
1320         if ((S_IFMT & a->mode) != (S_IFMT & b->mode)) {
1321                 result->clean = 0;
1322                 if (S_ISREG(a->mode)) {
1323                         result->mode = a->mode;
1324                         oidcpy(&result->oid, &a->oid);
1325                 } else {
1326                         result->mode = b->mode;
1327                         oidcpy(&result->oid, &b->oid);
1328                 }
1329         } else {
1330                 if (!oid_eq(&a->oid, &one->oid) && !oid_eq(&b->oid, &one->oid))
1331                         result->merge = 1;
1332
1333                 /*
1334                  * Merge modes
1335                  */
1336                 if (a->mode == b->mode || a->mode == one->mode)
1337                         result->mode = b->mode;
1338                 else {
1339                         result->mode = a->mode;
1340                         if (b->mode != one->mode) {
1341                                 result->clean = 0;
1342                                 result->merge = 1;
1343                         }
1344                 }
1345
1346                 if (oid_eq(&a->oid, &b->oid) || oid_eq(&a->oid, &one->oid))
1347                         oidcpy(&result->oid, &b->oid);
1348                 else if (oid_eq(&b->oid, &one->oid))
1349                         oidcpy(&result->oid, &a->oid);
1350                 else if (S_ISREG(a->mode)) {
1351                         mmbuffer_t result_buf;
1352                         int ret = 0, merge_status;
1353
1354                         merge_status = merge_3way(o, &result_buf, one, a, b,
1355                                                   branch1, branch2);
1356
1357                         if ((merge_status < 0) || !result_buf.ptr)
1358                                 ret = err(o, _("Failed to execute internal merge"));
1359
1360                         if (!ret &&
1361                             write_object_file(result_buf.ptr, result_buf.size,
1362                                               blob_type, &result->oid))
1363                                 ret = err(o, _("Unable to add %s to database"),
1364                                           a->path);
1365
1366                         free(result_buf.ptr);
1367                         if (ret)
1368                                 return ret;
1369                         result->clean = (merge_status == 0);
1370                 } else if (S_ISGITLINK(a->mode)) {
1371                         result->clean = merge_submodule(o, &result->oid,
1372                                                         one->path,
1373                                                         &one->oid,
1374                                                         &a->oid,
1375                                                         &b->oid);
1376                 } else if (S_ISLNK(a->mode)) {
1377                         switch (o->recursive_variant) {
1378                         case MERGE_RECURSIVE_NORMAL:
1379                                 oidcpy(&result->oid, &a->oid);
1380                                 if (!oid_eq(&a->oid, &b->oid))
1381                                         result->clean = 0;
1382                                 break;
1383                         case MERGE_RECURSIVE_OURS:
1384                                 oidcpy(&result->oid, &a->oid);
1385                                 break;
1386                         case MERGE_RECURSIVE_THEIRS:
1387                                 oidcpy(&result->oid, &b->oid);
1388                                 break;
1389                         }
1390                 } else
1391                         BUG("unsupported object type in the tree");
1392         }
1393
1394         if (result->merge)
1395                 output(o, 2, _("Auto-merging %s"), filename);
1396
1397         return 0;
1398 }
1399
1400 static int handle_rename_via_dir(struct merge_options *o,
1401                                  struct diff_filepair *pair,
1402                                  const char *rename_branch,
1403                                  const char *other_branch)
1404 {
1405         /*
1406          * Handle file adds that need to be renamed due to directory rename
1407          * detection.  This differs from handle_rename_normal, because
1408          * there is no content merge to do; just move the file into the
1409          * desired final location.
1410          */
1411         const struct diff_filespec *dest = pair->two;
1412
1413         if (!o->call_depth && would_lose_untracked(dest->path)) {
1414                 char *alt_path = unique_path(o, dest->path, rename_branch);
1415
1416                 output(o, 1, _("Error: Refusing to lose untracked file at %s; "
1417                                "writing to %s instead."),
1418                        dest->path, alt_path);
1419                 /*
1420                  * Write the file in worktree at alt_path, but not in the
1421                  * index.  Instead, write to dest->path for the index but
1422                  * only at the higher appropriate stage.
1423                  */
1424                 if (update_file(o, 0, &dest->oid, dest->mode, alt_path))
1425                         return -1;
1426                 free(alt_path);
1427                 return update_stages(o, dest->path, NULL,
1428                                      rename_branch == o->branch1 ? dest : NULL,
1429                                      rename_branch == o->branch1 ? NULL : dest);
1430         }
1431
1432         /* Update dest->path both in index and in worktree */
1433         if (update_file(o, 1, &dest->oid, dest->mode, dest->path))
1434                 return -1;
1435         return 0;
1436 }
1437
1438 static int handle_change_delete(struct merge_options *o,
1439                                 const char *path, const char *old_path,
1440                                 const struct object_id *o_oid, int o_mode,
1441                                 const struct object_id *changed_oid,
1442                                 int changed_mode,
1443                                 const char *change_branch,
1444                                 const char *delete_branch,
1445                                 const char *change, const char *change_past)
1446 {
1447         char *alt_path = NULL;
1448         const char *update_path = path;
1449         int ret = 0;
1450
1451         if (dir_in_way(path, !o->call_depth, 0) ||
1452             (!o->call_depth && would_lose_untracked(path))) {
1453                 update_path = alt_path = unique_path(o, path, change_branch);
1454         }
1455
1456         if (o->call_depth) {
1457                 /*
1458                  * We cannot arbitrarily accept either a_sha or b_sha as
1459                  * correct; since there is no true "middle point" between
1460                  * them, simply reuse the base version for virtual merge base.
1461                  */
1462                 ret = remove_file_from_cache(path);
1463                 if (!ret)
1464                         ret = update_file(o, 0, o_oid, o_mode, update_path);
1465         } else {
1466                 /*
1467                  * Despite the four nearly duplicate messages and argument
1468                  * lists below and the ugliness of the nested if-statements,
1469                  * having complete messages makes the job easier for
1470                  * translators.
1471                  *
1472                  * The slight variance among the cases is due to the fact
1473                  * that:
1474                  *   1) directory/file conflicts (in effect if
1475                  *      !alt_path) could cause us to need to write the
1476                  *      file to a different path.
1477                  *   2) renames (in effect if !old_path) could mean that
1478                  *      there are two names for the path that the user
1479                  *      may know the file by.
1480                  */
1481                 if (!alt_path) {
1482                         if (!old_path) {
1483                                 output(o, 1, _("CONFLICT (%s/delete): %s deleted in %s "
1484                                        "and %s in %s. Version %s of %s left in tree."),
1485                                        change, path, delete_branch, change_past,
1486                                        change_branch, change_branch, path);
1487                         } else {
1488                                 output(o, 1, _("CONFLICT (%s/delete): %s deleted in %s "
1489                                        "and %s to %s in %s. Version %s of %s left in tree."),
1490                                        change, old_path, delete_branch, change_past, path,
1491                                        change_branch, change_branch, path);
1492                         }
1493                 } else {
1494                         if (!old_path) {
1495                                 output(o, 1, _("CONFLICT (%s/delete): %s deleted in %s "
1496                                        "and %s in %s. Version %s of %s left in tree at %s."),
1497                                        change, path, delete_branch, change_past,
1498                                        change_branch, change_branch, path, alt_path);
1499                         } else {
1500                                 output(o, 1, _("CONFLICT (%s/delete): %s deleted in %s "
1501                                        "and %s to %s in %s. Version %s of %s left in tree at %s."),
1502                                        change, old_path, delete_branch, change_past, path,
1503                                        change_branch, change_branch, path, alt_path);
1504                         }
1505                 }
1506                 /*
1507                  * No need to call update_file() on path when change_branch ==
1508                  * o->branch1 && !alt_path, since that would needlessly touch
1509                  * path.  We could call update_file_flags() with update_cache=0
1510                  * and update_wd=0, but that's a no-op.
1511                  */
1512                 if (change_branch != o->branch1 || alt_path)
1513                         ret = update_file(o, 0, changed_oid, changed_mode, update_path);
1514         }
1515         free(alt_path);
1516
1517         return ret;
1518 }
1519
1520 static int handle_rename_delete(struct merge_options *o,
1521                                 struct diff_filepair *pair,
1522                                 const char *rename_branch,
1523                                 const char *delete_branch)
1524 {
1525         const struct diff_filespec *orig = pair->one;
1526         const struct diff_filespec *dest = pair->two;
1527
1528         if (handle_change_delete(o,
1529                                  o->call_depth ? orig->path : dest->path,
1530                                  o->call_depth ? NULL : orig->path,
1531                                  &orig->oid, orig->mode,
1532                                  &dest->oid, dest->mode,
1533                                  rename_branch, delete_branch,
1534                                  _("rename"), _("renamed")))
1535                 return -1;
1536
1537         if (o->call_depth)
1538                 return remove_file_from_cache(dest->path);
1539         else
1540                 return update_stages(o, dest->path, NULL,
1541                                      rename_branch == o->branch1 ? dest : NULL,
1542                                      rename_branch == o->branch1 ? NULL : dest);
1543 }
1544
1545 static struct diff_filespec *filespec_from_entry(struct diff_filespec *target,
1546                                                  struct stage_data *entry,
1547                                                  int stage)
1548 {
1549         struct object_id *oid = &entry->stages[stage].oid;
1550         unsigned mode = entry->stages[stage].mode;
1551         if (mode == 0 || is_null_oid(oid))
1552                 return NULL;
1553         oidcpy(&target->oid, oid);
1554         target->mode = mode;
1555         return target;
1556 }
1557
1558 static int handle_file(struct merge_options *o,
1559                         struct diff_filespec *rename,
1560                         int stage,
1561                         struct rename_conflict_info *ci)
1562 {
1563         char *dst_name = rename->path;
1564         struct stage_data *dst_entry;
1565         const char *cur_branch, *other_branch;
1566         struct diff_filespec other;
1567         struct diff_filespec *add;
1568         int ret;
1569
1570         if (stage == 2) {
1571                 dst_entry = ci->dst_entry1;
1572                 cur_branch = ci->branch1;
1573                 other_branch = ci->branch2;
1574         } else {
1575                 dst_entry = ci->dst_entry2;
1576                 cur_branch = ci->branch2;
1577                 other_branch = ci->branch1;
1578         }
1579
1580         add = filespec_from_entry(&other, dst_entry, stage ^ 1);
1581         if (add) {
1582                 int ren_src_was_dirty = was_dirty(o, rename->path);
1583                 char *add_name = unique_path(o, rename->path, other_branch);
1584                 if (update_file(o, 0, &add->oid, add->mode, add_name))
1585                         return -1;
1586
1587                 if (ren_src_was_dirty) {
1588                         output(o, 1, _("Refusing to lose dirty file at %s"),
1589                                rename->path);
1590                 }
1591                 /*
1592                  * Because the double negatives somehow keep confusing me...
1593                  *    1) update_wd iff !ren_src_was_dirty.
1594                  *    2) no_wd iff !update_wd
1595                  *    3) so, no_wd == !!ren_src_was_dirty == ren_src_was_dirty
1596                  */
1597                 remove_file(o, 0, rename->path, ren_src_was_dirty);
1598                 dst_name = unique_path(o, rename->path, cur_branch);
1599         } else {
1600                 if (dir_in_way(rename->path, !o->call_depth, 0)) {
1601                         dst_name = unique_path(o, rename->path, cur_branch);
1602                         output(o, 1, _("%s is a directory in %s adding as %s instead"),
1603                                rename->path, other_branch, dst_name);
1604                 } else if (!o->call_depth &&
1605                            would_lose_untracked(rename->path)) {
1606                         dst_name = unique_path(o, rename->path, cur_branch);
1607                         output(o, 1, _("Refusing to lose untracked file at %s; "
1608                                        "adding as %s instead"),
1609                                rename->path, dst_name);
1610                 }
1611         }
1612         if ((ret = update_file(o, 0, &rename->oid, rename->mode, dst_name)))
1613                 ; /* fall through, do allow dst_name to be released */
1614         else if (stage == 2)
1615                 ret = update_stages(o, rename->path, NULL, rename, add);
1616         else
1617                 ret = update_stages(o, rename->path, NULL, add, rename);
1618
1619         if (dst_name != rename->path)
1620                 free(dst_name);
1621
1622         return ret;
1623 }
1624
1625 static int handle_rename_rename_1to2(struct merge_options *o,
1626                                      struct rename_conflict_info *ci)
1627 {
1628         /* One file was renamed in both branches, but to different names. */
1629         struct diff_filespec *one = ci->pair1->one;
1630         struct diff_filespec *a = ci->pair1->two;
1631         struct diff_filespec *b = ci->pair2->two;
1632
1633         output(o, 1, _("CONFLICT (rename/rename): "
1634                "Rename \"%s\"->\"%s\" in branch \"%s\" "
1635                "rename \"%s\"->\"%s\" in \"%s\"%s"),
1636                one->path, a->path, ci->branch1,
1637                one->path, b->path, ci->branch2,
1638                o->call_depth ? _(" (left unresolved)") : "");
1639         if (o->call_depth) {
1640                 struct merge_file_info mfi;
1641                 struct diff_filespec other;
1642                 struct diff_filespec *add;
1643                 if (merge_mode_and_contents(o, one, a, b, one->path,
1644                                             ci->branch1, ci->branch2, &mfi))
1645                         return -1;
1646
1647                 /*
1648                  * FIXME: For rename/add-source conflicts (if we could detect
1649                  * such), this is wrong.  We should instead find a unique
1650                  * pathname and then either rename the add-source file to that
1651                  * unique path, or use that unique path instead of src here.
1652                  */
1653                 if (update_file(o, 0, &mfi.oid, mfi.mode, one->path))
1654                         return -1;
1655
1656                 /*
1657                  * Above, we put the merged content at the merge-base's
1658                  * path.  Now we usually need to delete both a->path and
1659                  * b->path.  However, the rename on each side of the merge
1660                  * could also be involved in a rename/add conflict.  In
1661                  * such cases, we should keep the added file around,
1662                  * resolving the conflict at that path in its favor.
1663                  */
1664                 add = filespec_from_entry(&other, ci->dst_entry1, 2 ^ 1);
1665                 if (add) {
1666                         if (update_file(o, 0, &add->oid, add->mode, a->path))
1667                                 return -1;
1668                 }
1669                 else
1670                         remove_file_from_cache(a->path);
1671                 add = filespec_from_entry(&other, ci->dst_entry2, 3 ^ 1);
1672                 if (add) {
1673                         if (update_file(o, 0, &add->oid, add->mode, b->path))
1674                                 return -1;
1675                 }
1676                 else
1677                         remove_file_from_cache(b->path);
1678         } else if (handle_file(o, a, 2, ci) || handle_file(o, b, 3, ci))
1679                 return -1;
1680
1681         return 0;
1682 }
1683
1684 static int handle_rename_rename_2to1(struct merge_options *o,
1685                                      struct rename_conflict_info *ci)
1686 {
1687         /* Two files, a & b, were renamed to the same thing, c. */
1688         struct diff_filespec *a = ci->pair1->one;
1689         struct diff_filespec *b = ci->pair2->one;
1690         struct diff_filespec *c1 = ci->pair1->two;
1691         struct diff_filespec *c2 = ci->pair2->two;
1692         char *path = c1->path; /* == c2->path */
1693         char *path_side_1_desc;
1694         char *path_side_2_desc;
1695         struct merge_file_info mfi_c1;
1696         struct merge_file_info mfi_c2;
1697         int ret;
1698
1699         output(o, 1, _("CONFLICT (rename/rename): "
1700                "Rename %s->%s in %s. "
1701                "Rename %s->%s in %s"),
1702                a->path, c1->path, ci->branch1,
1703                b->path, c2->path, ci->branch2);
1704
1705         remove_file(o, 1, a->path, o->call_depth || would_lose_untracked(a->path));
1706         remove_file(o, 1, b->path, o->call_depth || would_lose_untracked(b->path));
1707
1708         path_side_1_desc = xstrfmt("version of %s from %s", path, a->path);
1709         path_side_2_desc = xstrfmt("version of %s from %s", path, b->path);
1710         if (merge_mode_and_contents(o, a, c1, &ci->ren1_other, path_side_1_desc,
1711                                     o->branch1, o->branch2, &mfi_c1) ||
1712             merge_mode_and_contents(o, b, &ci->ren2_other, c2, path_side_2_desc,
1713                                     o->branch1, o->branch2, &mfi_c2))
1714                 return -1;
1715         free(path_side_1_desc);
1716         free(path_side_2_desc);
1717
1718         if (o->call_depth) {
1719                 /*
1720                  * If mfi_c1.clean && mfi_c2.clean, then it might make
1721                  * sense to do a two-way merge of those results.  But, I
1722                  * think in all cases, it makes sense to have the virtual
1723                  * merge base just undo the renames; they can be detected
1724                  * again later for the non-recursive merge.
1725                  */
1726                 remove_file(o, 0, path, 0);
1727                 ret = update_file(o, 0, &mfi_c1.oid, mfi_c1.mode, a->path);
1728                 if (!ret)
1729                         ret = update_file(o, 0, &mfi_c2.oid, mfi_c2.mode,
1730                                           b->path);
1731         } else {
1732                 char *new_path1 = unique_path(o, path, ci->branch1);
1733                 char *new_path2 = unique_path(o, path, ci->branch2);
1734                 output(o, 1, _("Renaming %s to %s and %s to %s instead"),
1735                        a->path, new_path1, b->path, new_path2);
1736                 if (was_dirty(o, path))
1737                         output(o, 1, _("Refusing to lose dirty file at %s"),
1738                                path);
1739                 else if (would_lose_untracked(path))
1740                         /*
1741                          * Only way we get here is if both renames were from
1742                          * a directory rename AND user had an untracked file
1743                          * at the location where both files end up after the
1744                          * two directory renames.  See testcase 10d of t6043.
1745                          */
1746                         output(o, 1, _("Refusing to lose untracked file at "
1747                                        "%s, even though it's in the way."),
1748                                path);
1749                 else
1750                         remove_file(o, 0, path, 0);
1751                 ret = update_file(o, 0, &mfi_c1.oid, mfi_c1.mode, new_path1);
1752                 if (!ret)
1753                         ret = update_file(o, 0, &mfi_c2.oid, mfi_c2.mode,
1754                                           new_path2);
1755                 /*
1756                  * unpack_trees() actually populates the index for us for
1757                  * "normal" rename/rename(2to1) situtations so that the
1758                  * correct entries are at the higher stages, which would
1759                  * make the call below to update_stages_for_stage_data
1760                  * unnecessary.  However, if either of the renames came
1761                  * from a directory rename, then unpack_trees() will not
1762                  * have gotten the right data loaded into the index, so we
1763                  * need to do so now.  (While it'd be tempting to move this
1764                  * call to update_stages_for_stage_data() to
1765                  * apply_directory_rename_modifications(), that would break
1766                  * our intermediate calls to would_lose_untracked() since
1767                  * those rely on the current in-memory index.  See also the
1768                  * big "NOTE" in update_stages()).
1769                  */
1770                 if (update_stages_for_stage_data(o, path, ci->dst_entry1))
1771                         ret = -1;
1772
1773                 free(new_path2);
1774                 free(new_path1);
1775         }
1776
1777         return ret;
1778 }
1779
1780 /*
1781  * Get the diff_filepairs changed between o_tree and tree.
1782  */
1783 static struct diff_queue_struct *get_diffpairs(struct merge_options *o,
1784                                                struct tree *o_tree,
1785                                                struct tree *tree)
1786 {
1787         struct diff_queue_struct *ret;
1788         struct diff_options opts;
1789
1790         repo_diff_setup(the_repository, &opts);
1791         opts.flags.recursive = 1;
1792         opts.flags.rename_empty = 0;
1793         opts.detect_rename = merge_detect_rename(o);
1794         /*
1795          * We do not have logic to handle the detection of copies.  In
1796          * fact, it may not even make sense to add such logic: would we
1797          * really want a change to a base file to be propagated through
1798          * multiple other files by a merge?
1799          */
1800         if (opts.detect_rename > DIFF_DETECT_RENAME)
1801                 opts.detect_rename = DIFF_DETECT_RENAME;
1802         opts.rename_limit = o->merge_rename_limit >= 0 ? o->merge_rename_limit :
1803                             o->diff_rename_limit >= 0 ? o->diff_rename_limit :
1804                             1000;
1805         opts.rename_score = o->rename_score;
1806         opts.show_rename_progress = o->show_rename_progress;
1807         opts.output_format = DIFF_FORMAT_NO_OUTPUT;
1808         diff_setup_done(&opts);
1809         diff_tree_oid(&o_tree->object.oid, &tree->object.oid, "", &opts);
1810         diffcore_std(&opts);
1811         if (opts.needed_rename_limit > o->needed_rename_limit)
1812                 o->needed_rename_limit = opts.needed_rename_limit;
1813
1814         ret = xmalloc(sizeof(*ret));
1815         *ret = diff_queued_diff;
1816
1817         opts.output_format = DIFF_FORMAT_NO_OUTPUT;
1818         diff_queued_diff.nr = 0;
1819         diff_queued_diff.queue = NULL;
1820         diff_flush(&opts);
1821         return ret;
1822 }
1823
1824 static int tree_has_path(struct tree *tree, const char *path)
1825 {
1826         struct object_id hashy;
1827         unsigned int mode_o;
1828
1829         return !get_tree_entry(&tree->object.oid, path,
1830                                &hashy, &mode_o);
1831 }
1832
1833 /*
1834  * Return a new string that replaces the beginning portion (which matches
1835  * entry->dir), with entry->new_dir.  In perl-speak:
1836  *   new_path_name = (old_path =~ s/entry->dir/entry->new_dir/);
1837  * NOTE:
1838  *   Caller must ensure that old_path starts with entry->dir + '/'.
1839  */
1840 static char *apply_dir_rename(struct dir_rename_entry *entry,
1841                               const char *old_path)
1842 {
1843         struct strbuf new_path = STRBUF_INIT;
1844         int oldlen, newlen;
1845
1846         if (entry->non_unique_new_dir)
1847                 return NULL;
1848
1849         oldlen = strlen(entry->dir);
1850         newlen = entry->new_dir.len + (strlen(old_path) - oldlen) + 1;
1851         strbuf_grow(&new_path, newlen);
1852         strbuf_addbuf(&new_path, &entry->new_dir);
1853         strbuf_addstr(&new_path, &old_path[oldlen]);
1854
1855         return strbuf_detach(&new_path, NULL);
1856 }
1857
1858 static void get_renamed_dir_portion(const char *old_path, const char *new_path,
1859                                     char **old_dir, char **new_dir)
1860 {
1861         char *end_of_old, *end_of_new;
1862         int old_len, new_len;
1863
1864         *old_dir = NULL;
1865         *new_dir = NULL;
1866
1867         /*
1868          * For
1869          *    "a/b/c/d/e/foo.c" -> "a/b/some/thing/else/e/foo.c"
1870          * the "e/foo.c" part is the same, we just want to know that
1871          *    "a/b/c/d" was renamed to "a/b/some/thing/else"
1872          * so, for this example, this function returns "a/b/c/d" in
1873          * *old_dir and "a/b/some/thing/else" in *new_dir.
1874          *
1875          * Also, if the basename of the file changed, we don't care.  We
1876          * want to know which portion of the directory, if any, changed.
1877          */
1878         end_of_old = strrchr(old_path, '/');
1879         end_of_new = strrchr(new_path, '/');
1880
1881         if (end_of_old == NULL || end_of_new == NULL)
1882                 return;
1883         while (*--end_of_new == *--end_of_old &&
1884                end_of_old != old_path &&
1885                end_of_new != new_path)
1886                 ; /* Do nothing; all in the while loop */
1887         /*
1888          * We've found the first non-matching character in the directory
1889          * paths.  That means the current directory we were comparing
1890          * represents the rename.  Move end_of_old and end_of_new back
1891          * to the full directory name.
1892          */
1893         if (*end_of_old == '/')
1894                 end_of_old++;
1895         if (*end_of_old != '/')
1896                 end_of_new++;
1897         end_of_old = strchr(end_of_old, '/');
1898         end_of_new = strchr(end_of_new, '/');
1899
1900         /*
1901          * It may have been the case that old_path and new_path were the same
1902          * directory all along.  Don't claim a rename if they're the same.
1903          */
1904         old_len = end_of_old - old_path;
1905         new_len = end_of_new - new_path;
1906
1907         if (old_len != new_len || strncmp(old_path, new_path, old_len)) {
1908                 *old_dir = xstrndup(old_path, old_len);
1909                 *new_dir = xstrndup(new_path, new_len);
1910         }
1911 }
1912
1913 static void remove_hashmap_entries(struct hashmap *dir_renames,
1914                                    struct string_list *items_to_remove)
1915 {
1916         int i;
1917         struct dir_rename_entry *entry;
1918
1919         for (i = 0; i < items_to_remove->nr; i++) {
1920                 entry = items_to_remove->items[i].util;
1921                 hashmap_remove(dir_renames, entry, NULL);
1922         }
1923         string_list_clear(items_to_remove, 0);
1924 }
1925
1926 /*
1927  * See if there is a directory rename for path, and if there are any file
1928  * level conflicts for the renamed location.  If there is a rename and
1929  * there are no conflicts, return the new name.  Otherwise, return NULL.
1930  */
1931 static char *handle_path_level_conflicts(struct merge_options *o,
1932                                          const char *path,
1933                                          struct dir_rename_entry *entry,
1934                                          struct hashmap *collisions,
1935                                          struct tree *tree)
1936 {
1937         char *new_path = NULL;
1938         struct collision_entry *collision_ent;
1939         int clean = 1;
1940         struct strbuf collision_paths = STRBUF_INIT;
1941
1942         /*
1943          * entry has the mapping of old directory name to new directory name
1944          * that we want to apply to path.
1945          */
1946         new_path = apply_dir_rename(entry, path);
1947
1948         if (!new_path) {
1949                 /* This should only happen when entry->non_unique_new_dir set */
1950                 if (!entry->non_unique_new_dir)
1951                         BUG("entry->non_unqiue_dir not set and !new_path");
1952                 output(o, 1, _("CONFLICT (directory rename split): "
1953                                "Unclear where to place %s because directory "
1954                                "%s was renamed to multiple other directories, "
1955                                "with no destination getting a majority of the "
1956                                "files."),
1957                        path, entry->dir);
1958                 clean = 0;
1959                 return NULL;
1960         }
1961
1962         /*
1963          * The caller needs to have ensured that it has pre-populated
1964          * collisions with all paths that map to new_path.  Do a quick check
1965          * to ensure that's the case.
1966          */
1967         collision_ent = collision_find_entry(collisions, new_path);
1968         if (collision_ent == NULL)
1969                 BUG("collision_ent is NULL");
1970
1971         /*
1972          * Check for one-sided add/add/.../add conflicts, i.e.
1973          * where implicit renames from the other side doing
1974          * directory rename(s) can affect this side of history
1975          * to put multiple paths into the same location.  Warn
1976          * and bail on directory renames for such paths.
1977          */
1978         if (collision_ent->reported_already) {
1979                 clean = 0;
1980         } else if (tree_has_path(tree, new_path)) {
1981                 collision_ent->reported_already = 1;
1982                 strbuf_add_separated_string_list(&collision_paths, ", ",
1983                                                  &collision_ent->source_files);
1984                 output(o, 1, _("CONFLICT (implicit dir rename): Existing "
1985                                "file/dir at %s in the way of implicit "
1986                                "directory rename(s) putting the following "
1987                                "path(s) there: %s."),
1988                        new_path, collision_paths.buf);
1989                 clean = 0;
1990         } else if (collision_ent->source_files.nr > 1) {
1991                 collision_ent->reported_already = 1;
1992                 strbuf_add_separated_string_list(&collision_paths, ", ",
1993                                                  &collision_ent->source_files);
1994                 output(o, 1, _("CONFLICT (implicit dir rename): Cannot map "
1995                                "more than one path to %s; implicit directory "
1996                                "renames tried to put these paths there: %s"),
1997                        new_path, collision_paths.buf);
1998                 clean = 0;
1999         }
2000
2001         /* Free memory we no longer need */
2002         strbuf_release(&collision_paths);
2003         if (!clean && new_path) {
2004                 free(new_path);
2005                 return NULL;
2006         }
2007
2008         return new_path;
2009 }
2010
2011 /*
2012  * There are a couple things we want to do at the directory level:
2013  *   1. Check for both sides renaming to the same thing, in order to avoid
2014  *      implicit renaming of files that should be left in place.  (See
2015  *      testcase 6b in t6043 for details.)
2016  *   2. Prune directory renames if there are still files left in the
2017  *      the original directory.  These represent a partial directory rename,
2018  *      i.e. a rename where only some of the files within the directory
2019  *      were renamed elsewhere.  (Technically, this could be done earlier
2020  *      in get_directory_renames(), except that would prevent us from
2021  *      doing the previous check and thus failing testcase 6b.)
2022  *   3. Check for rename/rename(1to2) conflicts (at the directory level).
2023  *      In the future, we could potentially record this info as well and
2024  *      omit reporting rename/rename(1to2) conflicts for each path within
2025  *      the affected directories, thus cleaning up the merge output.
2026  *   NOTE: We do NOT check for rename/rename(2to1) conflicts at the
2027  *         directory level, because merging directories is fine.  If it
2028  *         causes conflicts for files within those merged directories, then
2029  *         that should be detected at the individual path level.
2030  */
2031 static void handle_directory_level_conflicts(struct merge_options *o,
2032                                              struct hashmap *dir_re_head,
2033                                              struct tree *head,
2034                                              struct hashmap *dir_re_merge,
2035                                              struct tree *merge)
2036 {
2037         struct hashmap_iter iter;
2038         struct dir_rename_entry *head_ent;
2039         struct dir_rename_entry *merge_ent;
2040
2041         struct string_list remove_from_head = STRING_LIST_INIT_NODUP;
2042         struct string_list remove_from_merge = STRING_LIST_INIT_NODUP;
2043
2044         hashmap_iter_init(dir_re_head, &iter);
2045         while ((head_ent = hashmap_iter_next(&iter))) {
2046                 merge_ent = dir_rename_find_entry(dir_re_merge, head_ent->dir);
2047                 if (merge_ent &&
2048                     !head_ent->non_unique_new_dir &&
2049                     !merge_ent->non_unique_new_dir &&
2050                     !strbuf_cmp(&head_ent->new_dir, &merge_ent->new_dir)) {
2051                         /* 1. Renamed identically; remove it from both sides */
2052                         string_list_append(&remove_from_head,
2053                                            head_ent->dir)->util = head_ent;
2054                         strbuf_release(&head_ent->new_dir);
2055                         string_list_append(&remove_from_merge,
2056                                            merge_ent->dir)->util = merge_ent;
2057                         strbuf_release(&merge_ent->new_dir);
2058                 } else if (tree_has_path(head, head_ent->dir)) {
2059                         /* 2. This wasn't a directory rename after all */
2060                         string_list_append(&remove_from_head,
2061                                            head_ent->dir)->util = head_ent;
2062                         strbuf_release(&head_ent->new_dir);
2063                 }
2064         }
2065
2066         remove_hashmap_entries(dir_re_head, &remove_from_head);
2067         remove_hashmap_entries(dir_re_merge, &remove_from_merge);
2068
2069         hashmap_iter_init(dir_re_merge, &iter);
2070         while ((merge_ent = hashmap_iter_next(&iter))) {
2071                 head_ent = dir_rename_find_entry(dir_re_head, merge_ent->dir);
2072                 if (tree_has_path(merge, merge_ent->dir)) {
2073                         /* 2. This wasn't a directory rename after all */
2074                         string_list_append(&remove_from_merge,
2075                                            merge_ent->dir)->util = merge_ent;
2076                 } else if (head_ent &&
2077                            !head_ent->non_unique_new_dir &&
2078                            !merge_ent->non_unique_new_dir) {
2079                         /* 3. rename/rename(1to2) */
2080                         /*
2081                          * We can assume it's not rename/rename(1to1) because
2082                          * that was case (1), already checked above.  So we
2083                          * know that head_ent->new_dir and merge_ent->new_dir
2084                          * are different strings.
2085                          */
2086                         output(o, 1, _("CONFLICT (rename/rename): "
2087                                        "Rename directory %s->%s in %s. "
2088                                        "Rename directory %s->%s in %s"),
2089                                head_ent->dir, head_ent->new_dir.buf, o->branch1,
2090                                head_ent->dir, merge_ent->new_dir.buf, o->branch2);
2091                         string_list_append(&remove_from_head,
2092                                            head_ent->dir)->util = head_ent;
2093                         strbuf_release(&head_ent->new_dir);
2094                         string_list_append(&remove_from_merge,
2095                                            merge_ent->dir)->util = merge_ent;
2096                         strbuf_release(&merge_ent->new_dir);
2097                 }
2098         }
2099
2100         remove_hashmap_entries(dir_re_head, &remove_from_head);
2101         remove_hashmap_entries(dir_re_merge, &remove_from_merge);
2102 }
2103
2104 static struct hashmap *get_directory_renames(struct diff_queue_struct *pairs,
2105                                              struct tree *tree)
2106 {
2107         struct hashmap *dir_renames;
2108         struct hashmap_iter iter;
2109         struct dir_rename_entry *entry;
2110         int i;
2111
2112         /*
2113          * Typically, we think of a directory rename as all files from a
2114          * certain directory being moved to a target directory.  However,
2115          * what if someone first moved two files from the original
2116          * directory in one commit, and then renamed the directory
2117          * somewhere else in a later commit?  At merge time, we just know
2118          * that files from the original directory went to two different
2119          * places, and that the bulk of them ended up in the same place.
2120          * We want each directory rename to represent where the bulk of the
2121          * files from that directory end up; this function exists to find
2122          * where the bulk of the files went.
2123          *
2124          * The first loop below simply iterates through the list of file
2125          * renames, finding out how often each directory rename pair
2126          * possibility occurs.
2127          */
2128         dir_renames = xmalloc(sizeof(*dir_renames));
2129         dir_rename_init(dir_renames);
2130         for (i = 0; i < pairs->nr; ++i) {
2131                 struct string_list_item *item;
2132                 int *count;
2133                 struct diff_filepair *pair = pairs->queue[i];
2134                 char *old_dir, *new_dir;
2135
2136                 /* File not part of directory rename if it wasn't renamed */
2137                 if (pair->status != 'R')
2138                         continue;
2139
2140                 get_renamed_dir_portion(pair->one->path, pair->two->path,
2141                                         &old_dir,        &new_dir);
2142                 if (!old_dir)
2143                         /* Directory didn't change at all; ignore this one. */
2144                         continue;
2145
2146                 entry = dir_rename_find_entry(dir_renames, old_dir);
2147                 if (!entry) {
2148                         entry = xmalloc(sizeof(*entry));
2149                         dir_rename_entry_init(entry, old_dir);
2150                         hashmap_put(dir_renames, entry);
2151                 } else {
2152                         free(old_dir);
2153                 }
2154                 item = string_list_lookup(&entry->possible_new_dirs, new_dir);
2155                 if (!item) {
2156                         item = string_list_insert(&entry->possible_new_dirs,
2157                                                   new_dir);
2158                         item->util = xcalloc(1, sizeof(int));
2159                 } else {
2160                         free(new_dir);
2161                 }
2162                 count = item->util;
2163                 *count += 1;
2164         }
2165
2166         /*
2167          * For each directory with files moved out of it, we find out which
2168          * target directory received the most files so we can declare it to
2169          * be the "winning" target location for the directory rename.  This
2170          * winner gets recorded in new_dir.  If there is no winner
2171          * (multiple target directories received the same number of files),
2172          * we set non_unique_new_dir.  Once we've determined the winner (or
2173          * that there is no winner), we no longer need possible_new_dirs.
2174          */
2175         hashmap_iter_init(dir_renames, &iter);
2176         while ((entry = hashmap_iter_next(&iter))) {
2177                 int max = 0;
2178                 int bad_max = 0;
2179                 char *best = NULL;
2180
2181                 for (i = 0; i < entry->possible_new_dirs.nr; i++) {
2182                         int *count = entry->possible_new_dirs.items[i].util;
2183
2184                         if (*count == max)
2185                                 bad_max = max;
2186                         else if (*count > max) {
2187                                 max = *count;
2188                                 best = entry->possible_new_dirs.items[i].string;
2189                         }
2190                 }
2191                 if (bad_max == max)
2192                         entry->non_unique_new_dir = 1;
2193                 else {
2194                         assert(entry->new_dir.len == 0);
2195                         strbuf_addstr(&entry->new_dir, best);
2196                 }
2197                 /*
2198                  * The relevant directory sub-portion of the original full
2199                  * filepaths were xstrndup'ed before inserting into
2200                  * possible_new_dirs, and instead of manually iterating the
2201                  * list and free'ing each, just lie and tell
2202                  * possible_new_dirs that it did the strdup'ing so that it
2203                  * will free them for us.
2204                  */
2205                 entry->possible_new_dirs.strdup_strings = 1;
2206                 string_list_clear(&entry->possible_new_dirs, 1);
2207         }
2208
2209         return dir_renames;
2210 }
2211
2212 static struct dir_rename_entry *check_dir_renamed(const char *path,
2213                                                   struct hashmap *dir_renames)
2214 {
2215         char *temp = xstrdup(path);
2216         char *end;
2217         struct dir_rename_entry *entry = NULL;
2218
2219         while ((end = strrchr(temp, '/'))) {
2220                 *end = '\0';
2221                 entry = dir_rename_find_entry(dir_renames, temp);
2222                 if (entry)
2223                         break;
2224         }
2225         free(temp);
2226         return entry;
2227 }
2228
2229 static void compute_collisions(struct hashmap *collisions,
2230                                struct hashmap *dir_renames,
2231                                struct diff_queue_struct *pairs)
2232 {
2233         int i;
2234
2235         /*
2236          * Multiple files can be mapped to the same path due to directory
2237          * renames done by the other side of history.  Since that other
2238          * side of history could have merged multiple directories into one,
2239          * if our side of history added the same file basename to each of
2240          * those directories, then all N of them would get implicitly
2241          * renamed by the directory rename detection into the same path,
2242          * and we'd get an add/add/.../add conflict, and all those adds
2243          * from *this* side of history.  This is not representable in the
2244          * index, and users aren't going to easily be able to make sense of
2245          * it.  So we need to provide a good warning about what's
2246          * happening, and fall back to no-directory-rename detection
2247          * behavior for those paths.
2248          *
2249          * See testcases 9e and all of section 5 from t6043 for examples.
2250          */
2251         collision_init(collisions);
2252
2253         for (i = 0; i < pairs->nr; ++i) {
2254                 struct dir_rename_entry *dir_rename_ent;
2255                 struct collision_entry *collision_ent;
2256                 char *new_path;
2257                 struct diff_filepair *pair = pairs->queue[i];
2258
2259                 if (pair->status != 'A' && pair->status != 'R')
2260                         continue;
2261                 dir_rename_ent = check_dir_renamed(pair->two->path,
2262                                                    dir_renames);
2263                 if (!dir_rename_ent)
2264                         continue;
2265
2266                 new_path = apply_dir_rename(dir_rename_ent, pair->two->path);
2267                 if (!new_path)
2268                         /*
2269                          * dir_rename_ent->non_unique_new_path is true, which
2270                          * means there is no directory rename for us to use,
2271                          * which means it won't cause us any additional
2272                          * collisions.
2273                          */
2274                         continue;
2275                 collision_ent = collision_find_entry(collisions, new_path);
2276                 if (!collision_ent) {
2277                         collision_ent = xcalloc(1,
2278                                                 sizeof(struct collision_entry));
2279                         hashmap_entry_init(collision_ent, strhash(new_path));
2280                         hashmap_put(collisions, collision_ent);
2281                         collision_ent->target_file = new_path;
2282                 } else {
2283                         free(new_path);
2284                 }
2285                 string_list_insert(&collision_ent->source_files,
2286                                    pair->two->path);
2287         }
2288 }
2289
2290 static char *check_for_directory_rename(struct merge_options *o,
2291                                         const char *path,
2292                                         struct tree *tree,
2293                                         struct hashmap *dir_renames,
2294                                         struct hashmap *dir_rename_exclusions,
2295                                         struct hashmap *collisions,
2296                                         int *clean_merge)
2297 {
2298         char *new_path = NULL;
2299         struct dir_rename_entry *entry = check_dir_renamed(path, dir_renames);
2300         struct dir_rename_entry *oentry = NULL;
2301
2302         if (!entry)
2303                 return new_path;
2304
2305         /*
2306          * This next part is a little weird.  We do not want to do an
2307          * implicit rename into a directory we renamed on our side, because
2308          * that will result in a spurious rename/rename(1to2) conflict.  An
2309          * example:
2310          *   Base commit: dumbdir/afile, otherdir/bfile
2311          *   Side 1:      smrtdir/afile, otherdir/bfile
2312          *   Side 2:      dumbdir/afile, dumbdir/bfile
2313          * Here, while working on Side 1, we could notice that otherdir was
2314          * renamed/merged to dumbdir, and change the diff_filepair for
2315          * otherdir/bfile into a rename into dumbdir/bfile.  However, Side
2316          * 2 will notice the rename from dumbdir to smrtdir, and do the
2317          * transitive rename to move it from dumbdir/bfile to
2318          * smrtdir/bfile.  That gives us bfile in dumbdir vs being in
2319          * smrtdir, a rename/rename(1to2) conflict.  We really just want
2320          * the file to end up in smrtdir.  And the way to achieve that is
2321          * to not let Side1 do the rename to dumbdir, since we know that is
2322          * the source of one of our directory renames.
2323          *
2324          * That's why oentry and dir_rename_exclusions is here.
2325          *
2326          * As it turns out, this also prevents N-way transient rename
2327          * confusion; See testcases 9c and 9d of t6043.
2328          */
2329         oentry = dir_rename_find_entry(dir_rename_exclusions, entry->new_dir.buf);
2330         if (oentry) {
2331                 output(o, 1, _("WARNING: Avoiding applying %s -> %s rename "
2332                                "to %s, because %s itself was renamed."),
2333                        entry->dir, entry->new_dir.buf, path, entry->new_dir.buf);
2334         } else {
2335                 new_path = handle_path_level_conflicts(o, path, entry,
2336                                                        collisions, tree);
2337                 *clean_merge &= (new_path != NULL);
2338         }
2339
2340         return new_path;
2341 }
2342
2343 static void apply_directory_rename_modifications(struct merge_options *o,
2344                                                  struct diff_filepair *pair,
2345                                                  char *new_path,
2346                                                  struct rename *re,
2347                                                  struct tree *tree,
2348                                                  struct tree *o_tree,
2349                                                  struct tree *a_tree,
2350                                                  struct tree *b_tree,
2351                                                  struct string_list *entries,
2352                                                  int *clean)
2353 {
2354         struct string_list_item *item;
2355         int stage = (tree == a_tree ? 2 : 3);
2356         int update_wd;
2357
2358         /*
2359          * In all cases where we can do directory rename detection,
2360          * unpack_trees() will have read pair->two->path into the
2361          * index and the working copy.  We need to remove it so that
2362          * we can instead place it at new_path.  It is guaranteed to
2363          * not be untracked (unpack_trees() would have errored out
2364          * saying the file would have been overwritten), but it might
2365          * be dirty, though.
2366          */
2367         update_wd = !was_dirty(o, pair->two->path);
2368         if (!update_wd)
2369                 output(o, 1, _("Refusing to lose dirty file at %s"),
2370                        pair->two->path);
2371         remove_file(o, 1, pair->two->path, !update_wd);
2372
2373         /* Find or create a new re->dst_entry */
2374         item = string_list_lookup(entries, new_path);
2375         if (item) {
2376                 /*
2377                  * Since we're renaming on this side of history, and it's
2378                  * due to a directory rename on the other side of history
2379                  * (which we only allow when the directory in question no
2380                  * longer exists on the other side of history), the
2381                  * original entry for re->dst_entry is no longer
2382                  * necessary...
2383                  */
2384                 re->dst_entry->processed = 1;
2385
2386                 /*
2387                  * ...because we'll be using this new one.
2388                  */
2389                 re->dst_entry = item->util;
2390         } else {
2391                 /*
2392                  * re->dst_entry is for the before-dir-rename path, and we
2393                  * need it to hold information for the after-dir-rename
2394                  * path.  Before creating a new entry, we need to mark the
2395                  * old one as unnecessary (...unless it is shared by
2396                  * src_entry, i.e. this didn't use to be a rename, in which
2397                  * case we can just allow the normal processing to happen
2398                  * for it).
2399                  */
2400                 if (pair->status == 'R')
2401                         re->dst_entry->processed = 1;
2402
2403                 re->dst_entry = insert_stage_data(new_path,
2404                                                   o_tree, a_tree, b_tree,
2405                                                   entries);
2406                 item = string_list_insert(entries, new_path);
2407                 item->util = re->dst_entry;
2408         }
2409
2410         /*
2411          * Update the stage_data with the information about the path we are
2412          * moving into place.  That slot will be empty and available for us
2413          * to write to because of the collision checks in
2414          * handle_path_level_conflicts().  In other words,
2415          * re->dst_entry->stages[stage].oid will be the null_oid, so it's
2416          * open for us to write to.
2417          *
2418          * It may be tempting to actually update the index at this point as
2419          * well, using update_stages_for_stage_data(), but as per the big
2420          * "NOTE" in update_stages(), doing so will modify the current
2421          * in-memory index which will break calls to would_lose_untracked()
2422          * that we need to make.  Instead, we need to just make sure that
2423          * the various handle_rename_*() functions update the index
2424          * explicitly rather than relying on unpack_trees() to have done it.
2425          */
2426         get_tree_entry(&tree->object.oid,
2427                        pair->two->path,
2428                        &re->dst_entry->stages[stage].oid,
2429                        &re->dst_entry->stages[stage].mode);
2430
2431         /* Update pair status */
2432         if (pair->status == 'A') {
2433                 /*
2434                  * Recording rename information for this add makes it look
2435                  * like a rename/delete conflict.  Make sure we can
2436                  * correctly handle this as an add that was moved to a new
2437                  * directory instead of reporting a rename/delete conflict.
2438                  */
2439                 re->add_turned_into_rename = 1;
2440         }
2441         /*
2442          * We don't actually look at pair->status again, but it seems
2443          * pedagogically correct to adjust it.
2444          */
2445         pair->status = 'R';
2446
2447         /*
2448          * Finally, record the new location.
2449          */
2450         pair->two->path = new_path;
2451 }
2452
2453 /*
2454  * Get information of all renames which occurred in 'pairs', making use of
2455  * any implicit directory renames inferred from the other side of history.
2456  * We need the three trees in the merge ('o_tree', 'a_tree' and 'b_tree')
2457  * to be able to associate the correct cache entries with the rename
2458  * information; tree is always equal to either a_tree or b_tree.
2459  */
2460 static struct string_list *get_renames(struct merge_options *o,
2461                                        struct diff_queue_struct *pairs,
2462                                        struct hashmap *dir_renames,
2463                                        struct hashmap *dir_rename_exclusions,
2464                                        struct tree *tree,
2465                                        struct tree *o_tree,
2466                                        struct tree *a_tree,
2467                                        struct tree *b_tree,
2468                                        struct string_list *entries,
2469                                        int *clean_merge)
2470 {
2471         int i;
2472         struct hashmap collisions;
2473         struct hashmap_iter iter;
2474         struct collision_entry *e;
2475         struct string_list *renames;
2476
2477         compute_collisions(&collisions, dir_renames, pairs);
2478         renames = xcalloc(1, sizeof(struct string_list));
2479
2480         for (i = 0; i < pairs->nr; ++i) {
2481                 struct string_list_item *item;
2482                 struct rename *re;
2483                 struct diff_filepair *pair = pairs->queue[i];
2484                 char *new_path; /* non-NULL only with directory renames */
2485
2486                 if (pair->status != 'A' && pair->status != 'R') {
2487                         diff_free_filepair(pair);
2488                         continue;
2489                 }
2490                 new_path = check_for_directory_rename(o, pair->two->path, tree,
2491                                                       dir_renames,
2492                                                       dir_rename_exclusions,
2493                                                       &collisions,
2494                                                       clean_merge);
2495                 if (pair->status != 'R' && !new_path) {
2496                         diff_free_filepair(pair);
2497                         continue;
2498                 }
2499
2500                 re = xmalloc(sizeof(*re));
2501                 re->processed = 0;
2502                 re->add_turned_into_rename = 0;
2503                 re->pair = pair;
2504                 item = string_list_lookup(entries, re->pair->one->path);
2505                 if (!item)
2506                         re->src_entry = insert_stage_data(re->pair->one->path,
2507                                         o_tree, a_tree, b_tree, entries);
2508                 else
2509                         re->src_entry = item->util;
2510
2511                 item = string_list_lookup(entries, re->pair->two->path);
2512                 if (!item)
2513                         re->dst_entry = insert_stage_data(re->pair->two->path,
2514                                         o_tree, a_tree, b_tree, entries);
2515                 else
2516                         re->dst_entry = item->util;
2517                 item = string_list_insert(renames, pair->one->path);
2518                 item->util = re;
2519                 if (new_path)
2520                         apply_directory_rename_modifications(o, pair, new_path,
2521                                                              re, tree, o_tree,
2522                                                              a_tree, b_tree,
2523                                                              entries,
2524                                                              clean_merge);
2525         }
2526
2527         hashmap_iter_init(&collisions, &iter);
2528         while ((e = hashmap_iter_next(&iter))) {
2529                 free(e->target_file);
2530                 string_list_clear(&e->source_files, 0);
2531         }
2532         hashmap_free(&collisions, 1);
2533         return renames;
2534 }
2535
2536 static int process_renames(struct merge_options *o,
2537                            struct string_list *a_renames,
2538                            struct string_list *b_renames)
2539 {
2540         int clean_merge = 1, i, j;
2541         struct string_list a_by_dst = STRING_LIST_INIT_NODUP;
2542         struct string_list b_by_dst = STRING_LIST_INIT_NODUP;
2543         const struct rename *sre;
2544
2545         for (i = 0; i < a_renames->nr; i++) {
2546                 sre = a_renames->items[i].util;
2547                 string_list_insert(&a_by_dst, sre->pair->two->path)->util
2548                         = (void *)sre;
2549         }
2550         for (i = 0; i < b_renames->nr; i++) {
2551                 sre = b_renames->items[i].util;
2552                 string_list_insert(&b_by_dst, sre->pair->two->path)->util
2553                         = (void *)sre;
2554         }
2555
2556         for (i = 0, j = 0; i < a_renames->nr || j < b_renames->nr;) {
2557                 struct string_list *renames1, *renames2Dst;
2558                 struct rename *ren1 = NULL, *ren2 = NULL;
2559                 const char *branch1, *branch2;
2560                 const char *ren1_src, *ren1_dst;
2561                 struct string_list_item *lookup;
2562
2563                 if (i >= a_renames->nr) {
2564                         ren2 = b_renames->items[j++].util;
2565                 } else if (j >= b_renames->nr) {
2566                         ren1 = a_renames->items[i++].util;
2567                 } else {
2568                         int compare = strcmp(a_renames->items[i].string,
2569                                              b_renames->items[j].string);
2570                         if (compare <= 0)
2571                                 ren1 = a_renames->items[i++].util;
2572                         if (compare >= 0)
2573                                 ren2 = b_renames->items[j++].util;
2574                 }
2575
2576                 /* TODO: refactor, so that 1/2 are not needed */
2577                 if (ren1) {
2578                         renames1 = a_renames;
2579                         renames2Dst = &b_by_dst;
2580                         branch1 = o->branch1;
2581                         branch2 = o->branch2;
2582                 } else {
2583                         renames1 = b_renames;
2584                         renames2Dst = &a_by_dst;
2585                         branch1 = o->branch2;
2586                         branch2 = o->branch1;
2587                         SWAP(ren2, ren1);
2588                 }
2589
2590                 if (ren1->processed)
2591                         continue;
2592                 ren1->processed = 1;
2593                 ren1->dst_entry->processed = 1;
2594                 /* BUG: We should only mark src_entry as processed if we
2595                  * are not dealing with a rename + add-source case.
2596                  */
2597                 ren1->src_entry->processed = 1;
2598
2599                 ren1_src = ren1->pair->one->path;
2600                 ren1_dst = ren1->pair->two->path;
2601
2602                 if (ren2) {
2603                         /* One file renamed on both sides */
2604                         const char *ren2_src = ren2->pair->one->path;
2605                         const char *ren2_dst = ren2->pair->two->path;
2606                         enum rename_type rename_type;
2607                         if (strcmp(ren1_src, ren2_src) != 0)
2608                                 BUG("ren1_src != ren2_src");
2609                         ren2->dst_entry->processed = 1;
2610                         ren2->processed = 1;
2611                         if (strcmp(ren1_dst, ren2_dst) != 0) {
2612                                 rename_type = RENAME_ONE_FILE_TO_TWO;
2613                                 clean_merge = 0;
2614                         } else {
2615                                 rename_type = RENAME_ONE_FILE_TO_ONE;
2616                                 /* BUG: We should only remove ren1_src in
2617                                  * the base stage (think of rename +
2618                                  * add-source cases).
2619                                  */
2620                                 remove_file(o, 1, ren1_src, 1);
2621                                 update_entry(ren1->dst_entry,
2622                                              ren1->pair->one,
2623                                              ren1->pair->two,
2624                                              ren2->pair->two);
2625                         }
2626                         setup_rename_conflict_info(rename_type,
2627                                                    ren1->pair,
2628                                                    ren2->pair,
2629                                                    branch1,
2630                                                    branch2,
2631                                                    ren1->dst_entry,
2632                                                    ren2->dst_entry,
2633                                                    o,
2634                                                    NULL,
2635                                                    NULL);
2636                 } else if ((lookup = string_list_lookup(renames2Dst, ren1_dst))) {
2637                         /* Two different files renamed to the same thing */
2638                         char *ren2_dst;
2639                         ren2 = lookup->util;
2640                         ren2_dst = ren2->pair->two->path;
2641                         if (strcmp(ren1_dst, ren2_dst) != 0)
2642                                 BUG("ren1_dst != ren2_dst");
2643
2644                         clean_merge = 0;
2645                         ren2->processed = 1;
2646                         /*
2647                          * BUG: We should only mark src_entry as processed
2648                          * if we are not dealing with a rename + add-source
2649                          * case.
2650                          */
2651                         ren2->src_entry->processed = 1;
2652
2653                         setup_rename_conflict_info(RENAME_TWO_FILES_TO_ONE,
2654                                                    ren1->pair,
2655                                                    ren2->pair,
2656                                                    branch1,
2657                                                    branch2,
2658                                                    ren1->dst_entry,
2659                                                    ren2->dst_entry,
2660                                                    o,
2661                                                    ren1->src_entry,
2662                                                    ren2->src_entry);
2663
2664                 } else {
2665                         /* Renamed in 1, maybe changed in 2 */
2666                         /* we only use sha1 and mode of these */
2667                         struct diff_filespec src_other, dst_other;
2668                         int try_merge;
2669
2670                         /*
2671                          * unpack_trees loads entries from common-commit
2672                          * into stage 1, from head-commit into stage 2, and
2673                          * from merge-commit into stage 3.  We keep track
2674                          * of which side corresponds to the rename.
2675                          */
2676                         int renamed_stage = a_renames == renames1 ? 2 : 3;
2677                         int other_stage =   a_renames == renames1 ? 3 : 2;
2678
2679                         /* BUG: We should only remove ren1_src in the base
2680                          * stage and in other_stage (think of rename +
2681                          * add-source case).
2682                          */
2683                         remove_file(o, 1, ren1_src,
2684                                     renamed_stage == 2 || !was_tracked(o, ren1_src));
2685
2686                         oidcpy(&src_other.oid,
2687                                &ren1->src_entry->stages[other_stage].oid);
2688                         src_other.mode = ren1->src_entry->stages[other_stage].mode;
2689                         oidcpy(&dst_other.oid,
2690                                &ren1->dst_entry->stages[other_stage].oid);
2691                         dst_other.mode = ren1->dst_entry->stages[other_stage].mode;
2692                         try_merge = 0;
2693
2694                         if (oid_eq(&src_other.oid, &null_oid) &&
2695                             ren1->add_turned_into_rename) {
2696                                 setup_rename_conflict_info(RENAME_VIA_DIR,
2697                                                            ren1->pair,
2698                                                            NULL,
2699                                                            branch1,
2700                                                            branch2,
2701                                                            ren1->dst_entry,
2702                                                            NULL,
2703                                                            o,
2704                                                            NULL,
2705                                                            NULL);
2706                         } else if (oid_eq(&src_other.oid, &null_oid)) {
2707                                 setup_rename_conflict_info(RENAME_DELETE,
2708                                                            ren1->pair,
2709                                                            NULL,
2710                                                            branch1,
2711                                                            branch2,
2712                                                            ren1->dst_entry,
2713                                                            NULL,
2714                                                            o,
2715                                                            NULL,
2716                                                            NULL);
2717                         } else if ((dst_other.mode == ren1->pair->two->mode) &&
2718                                    oid_eq(&dst_other.oid, &ren1->pair->two->oid)) {
2719                                 /*
2720                                  * Added file on the other side identical to
2721                                  * the file being renamed: clean merge.
2722                                  * Also, there is no need to overwrite the
2723                                  * file already in the working copy, so call
2724                                  * update_file_flags() instead of
2725                                  * update_file().
2726                                  */
2727                                 if (update_file_flags(o,
2728                                                       &ren1->pair->two->oid,
2729                                                       ren1->pair->two->mode,
2730                                                       ren1_dst,
2731                                                       1, /* update_cache */
2732                                                       0  /* update_wd    */))
2733                                         clean_merge = -1;
2734                         } else if (!oid_eq(&dst_other.oid, &null_oid)) {
2735                                 clean_merge = 0;
2736                                 try_merge = 1;
2737                                 output(o, 1, _("CONFLICT (rename/add): Rename %s->%s in %s. "
2738                                        "%s added in %s"),
2739                                        ren1_src, ren1_dst, branch1,
2740                                        ren1_dst, branch2);
2741                                 if (o->call_depth) {
2742                                         struct merge_file_info mfi;
2743                                         struct diff_filespec one, a, b;
2744
2745                                         oidcpy(&one.oid, &null_oid);
2746                                         one.mode = 0;
2747                                         one.path = ren1->pair->two->path;
2748
2749                                         oidcpy(&a.oid, &ren1->pair->two->oid);
2750                                         a.mode = ren1->pair->two->mode;
2751                                         a.path = one.path;
2752
2753                                         oidcpy(&b.oid, &dst_other.oid);
2754                                         b.mode = dst_other.mode;
2755                                         b.path = one.path;
2756
2757                                         if (merge_mode_and_contents(o, &one, &a, &b, ren1_dst,
2758                                                                     branch1, branch2,
2759                                                                     &mfi)) {
2760                                                 clean_merge = -1;
2761                                                 goto cleanup_and_return;
2762                                         }
2763                                         output(o, 1, _("Adding merged %s"), ren1_dst);
2764                                         if (update_file(o, 0, &mfi.oid,
2765                                                         mfi.mode, ren1_dst))
2766                                                 clean_merge = -1;
2767                                         try_merge = 0;
2768                                 } else {
2769                                         char *new_path = unique_path(o, ren1_dst, branch2);
2770                                         output(o, 1, _("Adding as %s instead"), new_path);
2771                                         if (update_file(o, 0, &dst_other.oid,
2772                                                         dst_other.mode, new_path))
2773                                                 clean_merge = -1;
2774                                         free(new_path);
2775                                 }
2776                         } else
2777                                 try_merge = 1;
2778
2779                         if (clean_merge < 0)
2780                                 goto cleanup_and_return;
2781                         if (try_merge) {
2782                                 struct diff_filespec *one, *a, *b;
2783                                 src_other.path = (char *)ren1_src;
2784
2785                                 one = ren1->pair->one;
2786                                 if (a_renames == renames1) {
2787                                         a = ren1->pair->two;
2788                                         b = &src_other;
2789                                 } else {
2790                                         b = ren1->pair->two;
2791                                         a = &src_other;
2792                                 }
2793                                 update_entry(ren1->dst_entry, one, a, b);
2794                                 setup_rename_conflict_info(RENAME_NORMAL,
2795                                                            ren1->pair,
2796                                                            NULL,
2797                                                            branch1,
2798                                                            NULL,
2799                                                            ren1->dst_entry,
2800                                                            NULL,
2801                                                            o,
2802                                                            NULL,
2803                                                            NULL);
2804                         }
2805                 }
2806         }
2807 cleanup_and_return:
2808         string_list_clear(&a_by_dst, 0);
2809         string_list_clear(&b_by_dst, 0);
2810
2811         return clean_merge;
2812 }
2813
2814 struct rename_info {
2815         struct string_list *head_renames;
2816         struct string_list *merge_renames;
2817 };
2818
2819 static void initial_cleanup_rename(struct diff_queue_struct *pairs,
2820                                    struct hashmap *dir_renames)
2821 {
2822         struct hashmap_iter iter;
2823         struct dir_rename_entry *e;
2824
2825         hashmap_iter_init(dir_renames, &iter);
2826         while ((e = hashmap_iter_next(&iter))) {
2827                 free(e->dir);
2828                 strbuf_release(&e->new_dir);
2829                 /* possible_new_dirs already cleared in get_directory_renames */
2830         }
2831         hashmap_free(dir_renames, 1);
2832         free(dir_renames);
2833
2834         free(pairs->queue);
2835         free(pairs);
2836 }
2837
2838 static int detect_and_process_renames(struct merge_options *o,
2839                                       struct tree *common,
2840                                       struct tree *head,
2841                                       struct tree *merge,
2842                                       struct string_list *entries,
2843                                       struct rename_info *ri)
2844 {
2845         struct diff_queue_struct *head_pairs, *merge_pairs;
2846         struct hashmap *dir_re_head, *dir_re_merge;
2847         int clean = 1;
2848
2849         ri->head_renames = NULL;
2850         ri->merge_renames = NULL;
2851
2852         if (!merge_detect_rename(o))
2853                 return 1;
2854
2855         head_pairs = get_diffpairs(o, common, head);
2856         merge_pairs = get_diffpairs(o, common, merge);
2857
2858         if (o->detect_directory_renames) {
2859                 dir_re_head = get_directory_renames(head_pairs, head);
2860                 dir_re_merge = get_directory_renames(merge_pairs, merge);
2861
2862                 handle_directory_level_conflicts(o,
2863                                                  dir_re_head, head,
2864                                                  dir_re_merge, merge);
2865         } else {
2866                 dir_re_head  = xmalloc(sizeof(*dir_re_head));
2867                 dir_re_merge = xmalloc(sizeof(*dir_re_merge));
2868                 dir_rename_init(dir_re_head);
2869                 dir_rename_init(dir_re_merge);
2870         }
2871
2872         ri->head_renames  = get_renames(o, head_pairs,
2873                                         dir_re_merge, dir_re_head, head,
2874                                         common, head, merge, entries,
2875                                         &clean);
2876         if (clean < 0)
2877                 goto cleanup;
2878         ri->merge_renames = get_renames(o, merge_pairs,
2879                                         dir_re_head, dir_re_merge, merge,
2880                                         common, head, merge, entries,
2881                                         &clean);
2882         if (clean < 0)
2883                 goto cleanup;
2884         clean &= process_renames(o, ri->head_renames, ri->merge_renames);
2885
2886 cleanup:
2887         /*
2888          * Some cleanup is deferred until cleanup_renames() because the
2889          * data structures are still needed and referenced in
2890          * process_entry().  But there are a few things we can free now.
2891          */
2892         initial_cleanup_rename(head_pairs, dir_re_head);
2893         initial_cleanup_rename(merge_pairs, dir_re_merge);
2894
2895         return clean;
2896 }
2897
2898 static void final_cleanup_rename(struct string_list *rename)
2899 {
2900         const struct rename *re;
2901         int i;
2902
2903         if (rename == NULL)
2904                 return;
2905
2906         for (i = 0; i < rename->nr; i++) {
2907                 re = rename->items[i].util;
2908                 diff_free_filepair(re->pair);
2909         }
2910         string_list_clear(rename, 1);
2911         free(rename);
2912 }
2913
2914 static void final_cleanup_renames(struct rename_info *re_info)
2915 {
2916         final_cleanup_rename(re_info->head_renames);
2917         final_cleanup_rename(re_info->merge_renames);
2918 }
2919
2920 static struct object_id *stage_oid(const struct object_id *oid, unsigned mode)
2921 {
2922         return (is_null_oid(oid) || mode == 0) ? NULL: (struct object_id *)oid;
2923 }
2924
2925 static int read_oid_strbuf(struct merge_options *o,
2926                            const struct object_id *oid,
2927                            struct strbuf *dst)
2928 {
2929         void *buf;
2930         enum object_type type;
2931         unsigned long size;
2932         buf = read_object_file(oid, &type, &size);
2933         if (!buf)
2934                 return err(o, _("cannot read object %s"), oid_to_hex(oid));
2935         if (type != OBJ_BLOB) {
2936                 free(buf);
2937                 return err(o, _("object %s is not a blob"), oid_to_hex(oid));
2938         }
2939         strbuf_attach(dst, buf, size, size + 1);
2940         return 0;
2941 }
2942
2943 static int blob_unchanged(struct merge_options *opt,
2944                           const struct object_id *o_oid,
2945                           unsigned o_mode,
2946                           const struct object_id *a_oid,
2947                           unsigned a_mode,
2948                           int renormalize, const char *path)
2949 {
2950         struct strbuf o = STRBUF_INIT;
2951         struct strbuf a = STRBUF_INIT;
2952         int ret = 0; /* assume changed for safety */
2953
2954         if (a_mode != o_mode)
2955                 return 0;
2956         if (oid_eq(o_oid, a_oid))
2957                 return 1;
2958         if (!renormalize)
2959                 return 0;
2960
2961         assert(o_oid && a_oid);
2962         if (read_oid_strbuf(opt, o_oid, &o) || read_oid_strbuf(opt, a_oid, &a))
2963                 goto error_return;
2964         /*
2965          * Note: binary | is used so that both renormalizations are
2966          * performed.  Comparison can be skipped if both files are
2967          * unchanged since their sha1s have already been compared.
2968          */
2969         if (renormalize_buffer(&the_index, path, o.buf, o.len, &o) |
2970             renormalize_buffer(&the_index, path, a.buf, a.len, &a))
2971                 ret = (o.len == a.len && !memcmp(o.buf, a.buf, o.len));
2972
2973 error_return:
2974         strbuf_release(&o);
2975         strbuf_release(&a);
2976         return ret;
2977 }
2978
2979 static int handle_modify_delete(struct merge_options *o,
2980                                 const char *path,
2981                                 struct object_id *o_oid, int o_mode,
2982                                 struct object_id *a_oid, int a_mode,
2983                                 struct object_id *b_oid, int b_mode)
2984 {
2985         const char *modify_branch, *delete_branch;
2986         struct object_id *changed_oid;
2987         int changed_mode;
2988
2989         if (a_oid) {
2990                 modify_branch = o->branch1;
2991                 delete_branch = o->branch2;
2992                 changed_oid = a_oid;
2993                 changed_mode = a_mode;
2994         } else {
2995                 modify_branch = o->branch2;
2996                 delete_branch = o->branch1;
2997                 changed_oid = b_oid;
2998                 changed_mode = b_mode;
2999         }
3000
3001         return handle_change_delete(o,
3002                                     path, NULL,
3003                                     o_oid, o_mode,
3004                                     changed_oid, changed_mode,
3005                                     modify_branch, delete_branch,
3006                                     _("modify"), _("modified"));
3007 }
3008
3009 static int handle_content_merge(struct merge_options *o,
3010                                 const char *path,
3011                                 int is_dirty,
3012                                 struct object_id *o_oid, int o_mode,
3013                                 struct object_id *a_oid, int a_mode,
3014                                 struct object_id *b_oid, int b_mode,
3015                                 struct rename_conflict_info *rename_conflict_info)
3016 {
3017         const char *reason = _("content");
3018         const char *path1 = NULL, *path2 = NULL;
3019         struct merge_file_info mfi;
3020         struct diff_filespec one, a, b;
3021         unsigned df_conflict_remains = 0;
3022
3023         if (!o_oid) {
3024                 reason = _("add/add");
3025                 o_oid = (struct object_id *)&null_oid;
3026         }
3027         one.path = a.path = b.path = (char *)path;
3028         oidcpy(&one.oid, o_oid);
3029         one.mode = o_mode;
3030         oidcpy(&a.oid, a_oid);
3031         a.mode = a_mode;
3032         oidcpy(&b.oid, b_oid);
3033         b.mode = b_mode;
3034
3035         if (rename_conflict_info) {
3036                 struct diff_filepair *pair1 = rename_conflict_info->pair1;
3037
3038                 path1 = (o->branch1 == rename_conflict_info->branch1) ?
3039                         pair1->two->path : pair1->one->path;
3040                 /* If rename_conflict_info->pair2 != NULL, we are in
3041                  * RENAME_ONE_FILE_TO_ONE case.  Otherwise, we have a
3042                  * normal rename.
3043                  */
3044                 path2 = (rename_conflict_info->pair2 ||
3045                          o->branch2 == rename_conflict_info->branch1) ?
3046                         pair1->two->path : pair1->one->path;
3047                 one.path = pair1->one->path;
3048                 a.path = (char *)path1;
3049                 b.path = (char *)path2;
3050
3051                 if (dir_in_way(path, !o->call_depth,
3052                                S_ISGITLINK(pair1->two->mode)))
3053                         df_conflict_remains = 1;
3054         }
3055         if (merge_mode_and_contents(o, &one, &a, &b, path,
3056                                     o->branch1, o->branch2, &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                 int pos;
3069                 struct cache_entry *ce;
3070
3071                 output(o, 3, _("Skipped %s (merged same as existing)"), path);
3072                 if (add_cacheinfo(o, mfi.mode, &mfi.oid, path,
3073                                   0, (!o->call_depth && !is_dirty), 0))
3074                         return -1;
3075                 /*
3076                  * However, add_cacheinfo() will delete the old cache entry
3077                  * and add a new one.  We need to copy over any skip_worktree
3078                  * flag to avoid making the file appear as if it were
3079                  * deleted by the user.
3080                  */
3081                 pos = index_name_pos(&o->orig_index, path, strlen(path));
3082                 ce = o->orig_index.cache[pos];
3083                 if (ce_skip_worktree(ce)) {
3084                         pos = index_name_pos(&the_index, path, strlen(path));
3085                         ce = the_index.cache[pos];
3086                         ce->ce_flags |= CE_SKIP_WORKTREE;
3087                 }
3088                 return mfi.clean;
3089         }
3090
3091         if (!mfi.clean) {
3092                 if (S_ISGITLINK(mfi.mode))
3093                         reason = _("submodule");
3094                 output(o, 1, _("CONFLICT (%s): Merge conflict in %s"),
3095                                 reason, path);
3096                 if (rename_conflict_info && !df_conflict_remains)
3097                         if (update_stages(o, path, &one, &a, &b))
3098                                 return -1;
3099         }
3100
3101         if (df_conflict_remains || is_dirty) {
3102                 char *new_path;
3103                 if (o->call_depth) {
3104                         remove_file_from_cache(path);
3105                 } else {
3106                         if (!mfi.clean) {
3107                                 if (update_stages(o, path, &one, &a, &b))
3108                                         return -1;
3109                         } else {
3110                                 int file_from_stage2 = was_tracked(o, path);
3111                                 struct diff_filespec merged;
3112                                 oidcpy(&merged.oid, &mfi.oid);
3113                                 merged.mode = mfi.mode;
3114
3115                                 if (update_stages(o, path, NULL,
3116                                                   file_from_stage2 ? &merged : NULL,
3117                                                   file_from_stage2 ? NULL : &merged))
3118                                         return -1;
3119                         }
3120
3121                 }
3122                 new_path = unique_path(o, path, rename_conflict_info->branch1);
3123                 if (is_dirty) {
3124                         output(o, 1, _("Refusing to lose dirty file at %s"),
3125                                path);
3126                 }
3127                 output(o, 1, _("Adding as %s instead"), new_path);
3128                 if (update_file(o, 0, &mfi.oid, mfi.mode, new_path)) {
3129                         free(new_path);
3130                         return -1;
3131                 }
3132                 free(new_path);
3133                 mfi.clean = 0;
3134         } else if (update_file(o, mfi.clean, &mfi.oid, mfi.mode, path))
3135                 return -1;
3136         return !is_dirty && mfi.clean;
3137 }
3138
3139 static int handle_rename_normal(struct merge_options *o,
3140                                 const char *path,
3141                                 struct object_id *o_oid, unsigned int o_mode,
3142                                 struct object_id *a_oid, unsigned int a_mode,
3143                                 struct object_id *b_oid, unsigned int b_mode,
3144                                 struct rename_conflict_info *ci)
3145 {
3146         /* Merge the content and write it out */
3147         return handle_content_merge(o, path, was_dirty(o, path),
3148                                     o_oid, o_mode, a_oid, a_mode, b_oid, b_mode,
3149                                     ci);
3150 }
3151
3152 /* Per entry merge function */
3153 static int process_entry(struct merge_options *o,
3154                          const char *path, struct stage_data *entry)
3155 {
3156         int clean_merge = 1;
3157         int normalize = o->renormalize;
3158         unsigned o_mode = entry->stages[1].mode;
3159         unsigned a_mode = entry->stages[2].mode;
3160         unsigned b_mode = entry->stages[3].mode;
3161         struct object_id *o_oid = stage_oid(&entry->stages[1].oid, o_mode);
3162         struct object_id *a_oid = stage_oid(&entry->stages[2].oid, a_mode);
3163         struct object_id *b_oid = stage_oid(&entry->stages[3].oid, b_mode);
3164
3165         entry->processed = 1;
3166         if (entry->rename_conflict_info) {
3167                 struct rename_conflict_info *conflict_info = entry->rename_conflict_info;
3168                 switch (conflict_info->rename_type) {
3169                 case RENAME_NORMAL:
3170                 case RENAME_ONE_FILE_TO_ONE:
3171                         clean_merge = handle_rename_normal(o,
3172                                                            path,
3173                                                            o_oid, o_mode,
3174                                                            a_oid, a_mode,
3175                                                            b_oid, b_mode,
3176                                                            conflict_info);
3177                         break;
3178                 case RENAME_VIA_DIR:
3179                         clean_merge = 1;
3180                         if (handle_rename_via_dir(o,
3181                                                   conflict_info->pair1,
3182                                                   conflict_info->branch1,
3183                                                   conflict_info->branch2))
3184                                 clean_merge = -1;
3185                         break;
3186                 case RENAME_DELETE:
3187                         clean_merge = 0;
3188                         if (handle_rename_delete(o,
3189                                                  conflict_info->pair1,
3190                                                  conflict_info->branch1,
3191                                                  conflict_info->branch2))
3192                                 clean_merge = -1;
3193                         break;
3194                 case RENAME_ONE_FILE_TO_TWO:
3195                         clean_merge = 0;
3196                         if (handle_rename_rename_1to2(o, conflict_info))
3197                                 clean_merge = -1;
3198                         break;
3199                 case RENAME_TWO_FILES_TO_ONE:
3200                         clean_merge = 0;
3201                         if (handle_rename_rename_2to1(o, conflict_info))
3202                                 clean_merge = -1;
3203                         break;
3204                 default:
3205                         entry->processed = 0;
3206                         break;
3207                 }
3208         } else if (o_oid && (!a_oid || !b_oid)) {
3209                 /* Case A: Deleted in one */
3210                 if ((!a_oid && !b_oid) ||
3211                     (!b_oid && blob_unchanged(o, o_oid, o_mode, a_oid, a_mode, normalize, path)) ||
3212                     (!a_oid && blob_unchanged(o, o_oid, o_mode, b_oid, b_mode, normalize, path))) {
3213                         /* Deleted in both or deleted in one and
3214                          * unchanged in the other */
3215                         if (a_oid)
3216                                 output(o, 2, _("Removing %s"), path);
3217                         /* do not touch working file if it did not exist */
3218                         remove_file(o, 1, path, !a_oid);
3219                 } else {
3220                         /* Modify/delete; deleted side may have put a directory in the way */
3221                         clean_merge = 0;
3222                         if (handle_modify_delete(o, path, o_oid, o_mode,
3223                                                  a_oid, a_mode, b_oid, b_mode))
3224                                 clean_merge = -1;
3225                 }
3226         } else if ((!o_oid && a_oid && !b_oid) ||
3227                    (!o_oid && !a_oid && b_oid)) {
3228                 /* Case B: Added in one. */
3229                 /* [nothing|directory] -> ([nothing|directory], file) */
3230
3231                 const char *add_branch;
3232                 const char *other_branch;
3233                 unsigned mode;
3234                 const struct object_id *oid;
3235                 const char *conf;
3236
3237                 if (a_oid) {
3238                         add_branch = o->branch1;
3239                         other_branch = o->branch2;
3240                         mode = a_mode;
3241                         oid = a_oid;
3242                         conf = _("file/directory");
3243                 } else {
3244                         add_branch = o->branch2;
3245                         other_branch = o->branch1;
3246                         mode = b_mode;
3247                         oid = b_oid;
3248                         conf = _("directory/file");
3249                 }
3250                 if (dir_in_way(path,
3251                                !o->call_depth && !S_ISGITLINK(a_mode),
3252                                0)) {
3253                         char *new_path = unique_path(o, path, add_branch);
3254                         clean_merge = 0;
3255                         output(o, 1, _("CONFLICT (%s): There is a directory with name %s in %s. "
3256                                "Adding %s as %s"),
3257                                conf, path, other_branch, path, new_path);
3258                         if (update_file(o, 0, oid, mode, new_path))
3259                                 clean_merge = -1;
3260                         else if (o->call_depth)
3261                                 remove_file_from_cache(path);
3262                         free(new_path);
3263                 } else {
3264                         output(o, 2, _("Adding %s"), path);
3265                         /* do not overwrite file if already present */
3266                         if (update_file_flags(o, oid, mode, path, 1, !a_oid))
3267                                 clean_merge = -1;
3268                 }
3269         } else if (a_oid && b_oid) {
3270                 /* Case C: Added in both (check for same permissions) and */
3271                 /* case D: Modified in both, but differently. */
3272                 int is_dirty = 0; /* unpack_trees would have bailed if dirty */
3273                 clean_merge = handle_content_merge(o, path, is_dirty,
3274                                                    o_oid, o_mode,
3275                                                    a_oid, a_mode,
3276                                                    b_oid, b_mode,
3277                                                    NULL);
3278         } else if (!o_oid && !a_oid && !b_oid) {
3279                 /*
3280                  * this entry was deleted altogether. a_mode == 0 means
3281                  * we had that path and want to actively remove it.
3282                  */
3283                 remove_file(o, 1, path, !a_mode);
3284         } else
3285                 BUG("fatal merge failure, shouldn't happen.");
3286
3287         return clean_merge;
3288 }
3289
3290 int merge_trees(struct merge_options *o,
3291                 struct tree *head,
3292                 struct tree *merge,
3293                 struct tree *common,
3294                 struct tree **result)
3295 {
3296         int code, clean;
3297         struct strbuf sb = STRBUF_INIT;
3298
3299         if (!o->call_depth && index_has_changes(&the_index, head, &sb)) {
3300                 err(o, _("Your local changes to the following files would be overwritten by merge:\n  %s"),
3301                     sb.buf);
3302                 return -1;
3303         }
3304
3305         if (o->subtree_shift) {
3306                 merge = shift_tree_object(head, merge, o->subtree_shift);
3307                 common = shift_tree_object(head, common, o->subtree_shift);
3308         }
3309
3310         if (oid_eq(&common->object.oid, &merge->object.oid)) {
3311                 output(o, 0, _("Already up to date!"));
3312                 *result = head;
3313                 return 1;
3314         }
3315
3316         code = unpack_trees_start(o, common, head, merge);
3317
3318         if (code != 0) {
3319                 if (show(o, 4) || o->call_depth)
3320                         err(o, _("merging of trees %s and %s failed"),
3321                             oid_to_hex(&head->object.oid),
3322                             oid_to_hex(&merge->object.oid));
3323                 unpack_trees_finish(o);
3324                 return -1;
3325         }
3326
3327         if (unmerged_cache()) {
3328                 struct string_list *entries;
3329                 struct rename_info re_info;
3330                 int i;
3331                 /*
3332                  * Only need the hashmap while processing entries, so
3333                  * initialize it here and free it when we are done running
3334                  * through the entries. Keeping it in the merge_options as
3335                  * opposed to decaring a local hashmap is for convenience
3336                  * so that we don't have to pass it to around.
3337                  */
3338                 hashmap_init(&o->current_file_dir_set, path_hashmap_cmp, NULL, 512);
3339                 get_files_dirs(o, head);
3340                 get_files_dirs(o, merge);
3341
3342                 entries = get_unmerged();
3343                 clean = detect_and_process_renames(o, common, head, merge,
3344                                                    entries, &re_info);
3345                 record_df_conflict_files(o, entries);
3346                 if (clean < 0)
3347                         goto cleanup;
3348                 for (i = entries->nr-1; 0 <= i; i--) {
3349                         const char *path = entries->items[i].string;
3350                         struct stage_data *e = entries->items[i].util;
3351                         if (!e->processed) {
3352                                 int ret = process_entry(o, path, e);
3353                                 if (!ret)
3354                                         clean = 0;
3355                                 else if (ret < 0) {
3356                                         clean = ret;
3357                                         goto cleanup;
3358                                 }
3359                         }
3360                 }
3361                 for (i = 0; i < entries->nr; i++) {
3362                         struct stage_data *e = entries->items[i].util;
3363                         if (!e->processed)
3364                                 BUG("unprocessed path??? %s",
3365                                     entries->items[i].string);
3366                 }
3367
3368         cleanup:
3369                 final_cleanup_renames(&re_info);
3370
3371                 string_list_clear(entries, 1);
3372                 free(entries);
3373
3374                 hashmap_free(&o->current_file_dir_set, 1);
3375
3376                 if (clean < 0) {
3377                         unpack_trees_finish(o);
3378                         return clean;
3379                 }
3380         }
3381         else
3382                 clean = 1;
3383
3384         unpack_trees_finish(o);
3385
3386         if (o->call_depth && !(*result = write_tree_from_memory(o)))
3387                 return -1;
3388
3389         return clean;
3390 }
3391
3392 static struct commit_list *reverse_commit_list(struct commit_list *list)
3393 {
3394         struct commit_list *next = NULL, *current, *backup;
3395         for (current = list; current; current = backup) {
3396                 backup = current->next;
3397                 current->next = next;
3398                 next = current;
3399         }
3400         return next;
3401 }
3402
3403 /*
3404  * Merge the commits h1 and h2, return the resulting virtual
3405  * commit object and a flag indicating the cleanness of the merge.
3406  */
3407 int merge_recursive(struct merge_options *o,
3408                     struct commit *h1,
3409                     struct commit *h2,
3410                     struct commit_list *ca,
3411                     struct commit **result)
3412 {
3413         struct commit_list *iter;
3414         struct commit *merged_common_ancestors;
3415         struct tree *mrtree;
3416         int clean;
3417
3418         if (show(o, 4)) {
3419                 output(o, 4, _("Merging:"));
3420                 output_commit_title(o, h1);
3421                 output_commit_title(o, h2);
3422         }
3423
3424         if (!ca) {
3425                 ca = get_merge_bases(h1, h2);
3426                 ca = reverse_commit_list(ca);
3427         }
3428
3429         if (show(o, 5)) {
3430                 unsigned cnt = commit_list_count(ca);
3431
3432                 output(o, 5, Q_("found %u common ancestor:",
3433                                 "found %u common ancestors:", cnt), cnt);
3434                 for (iter = ca; iter; iter = iter->next)
3435                         output_commit_title(o, iter->item);
3436         }
3437
3438         merged_common_ancestors = pop_commit(&ca);
3439         if (merged_common_ancestors == NULL) {
3440                 /* if there is no common ancestor, use an empty tree */
3441                 struct tree *tree;
3442
3443                 tree = lookup_tree(the_repository, the_repository->hash_algo->empty_tree);
3444                 merged_common_ancestors = make_virtual_commit(tree, "ancestor");
3445         }
3446
3447         for (iter = ca; iter; iter = iter->next) {
3448                 const char *saved_b1, *saved_b2;
3449                 o->call_depth++;
3450                 /*
3451                  * When the merge fails, the result contains files
3452                  * with conflict markers. The cleanness flag is
3453                  * ignored (unless indicating an error), it was never
3454                  * actually used, as result of merge_trees has always
3455                  * overwritten it: the committed "conflicts" were
3456                  * already resolved.
3457                  */
3458                 discard_cache();
3459                 saved_b1 = o->branch1;
3460                 saved_b2 = o->branch2;
3461                 o->branch1 = "Temporary merge branch 1";
3462                 o->branch2 = "Temporary merge branch 2";
3463                 if (merge_recursive(o, merged_common_ancestors, iter->item,
3464                                     NULL, &merged_common_ancestors) < 0)
3465                         return -1;
3466                 o->branch1 = saved_b1;
3467                 o->branch2 = saved_b2;
3468                 o->call_depth--;
3469
3470                 if (!merged_common_ancestors)
3471                         return err(o, _("merge returned no commit"));
3472         }
3473
3474         discard_cache();
3475         if (!o->call_depth)
3476                 read_cache();
3477
3478         o->ancestor = "merged common ancestors";
3479         clean = merge_trees(o, get_commit_tree(h1), get_commit_tree(h2),
3480                             get_commit_tree(merged_common_ancestors),
3481                             &mrtree);
3482         if (clean < 0) {
3483                 flush_output(o);
3484                 return clean;
3485         }
3486
3487         if (o->call_depth) {
3488                 *result = make_virtual_commit(mrtree, "merged tree");
3489                 commit_list_insert(h1, &(*result)->parents);
3490                 commit_list_insert(h2, &(*result)->parents->next);
3491         }
3492         flush_output(o);
3493         if (!o->call_depth && o->buffer_output < 2)
3494                 strbuf_release(&o->obuf);
3495         if (show(o, 2))
3496                 diff_warn_rename_limit("merge.renamelimit",
3497                                        o->needed_rename_limit, 0);
3498         return clean;
3499 }
3500
3501 static struct commit *get_ref(const struct object_id *oid, const char *name)
3502 {
3503         struct object *object;
3504
3505         object = deref_tag(the_repository, parse_object(the_repository, oid),
3506                            name,
3507                            strlen(name));
3508         if (!object)
3509                 return NULL;
3510         if (object->type == OBJ_TREE)
3511                 return make_virtual_commit((struct tree*)object, name);
3512         if (object->type != OBJ_COMMIT)
3513                 return NULL;
3514         if (parse_commit((struct commit *)object))
3515                 return NULL;
3516         return (struct commit *)object;
3517 }
3518
3519 int merge_recursive_generic(struct merge_options *o,
3520                             const struct object_id *head,
3521                             const struct object_id *merge,
3522                             int num_base_list,
3523                             const struct object_id **base_list,
3524                             struct commit **result)
3525 {
3526         int clean;
3527         struct lock_file lock = LOCK_INIT;
3528         struct commit *head_commit = get_ref(head, o->branch1);
3529         struct commit *next_commit = get_ref(merge, o->branch2);
3530         struct commit_list *ca = NULL;
3531
3532         if (base_list) {
3533                 int i;
3534                 for (i = 0; i < num_base_list; ++i) {
3535                         struct commit *base;
3536                         if (!(base = get_ref(base_list[i], oid_to_hex(base_list[i]))))
3537                                 return err(o, _("Could not parse object '%s'"),
3538                                            oid_to_hex(base_list[i]));
3539                         commit_list_insert(base, &ca);
3540                 }
3541         }
3542
3543         hold_locked_index(&lock, LOCK_DIE_ON_ERROR);
3544         clean = merge_recursive(o, head_commit, next_commit, ca,
3545                                 result);
3546         if (clean < 0) {
3547                 rollback_lock_file(&lock);
3548                 return clean;
3549         }
3550
3551         if (write_locked_index(&the_index, &lock,
3552                                COMMIT_LOCK | SKIP_IF_UNCHANGED))
3553                 return err(o, _("Unable to write index."));
3554
3555         return clean ? 0 : 1;
3556 }
3557
3558 static void merge_recursive_config(struct merge_options *o)
3559 {
3560         char *value = NULL;
3561         git_config_get_int("merge.verbosity", &o->verbosity);
3562         git_config_get_int("diff.renamelimit", &o->diff_rename_limit);
3563         git_config_get_int("merge.renamelimit", &o->merge_rename_limit);
3564         if (!git_config_get_string("diff.renames", &value)) {
3565                 o->diff_detect_rename = git_config_rename("diff.renames", value);
3566                 free(value);
3567         }
3568         if (!git_config_get_string("merge.renames", &value)) {
3569                 o->merge_detect_rename = git_config_rename("merge.renames", value);
3570                 free(value);
3571         }
3572         git_config(git_xmerge_config, NULL);
3573 }
3574
3575 void init_merge_options(struct merge_options *o)
3576 {
3577         const char *merge_verbosity;
3578         memset(o, 0, sizeof(struct merge_options));
3579         o->verbosity = 2;
3580         o->buffer_output = 1;
3581         o->diff_rename_limit = -1;
3582         o->merge_rename_limit = -1;
3583         o->renormalize = 0;
3584         o->diff_detect_rename = -1;
3585         o->merge_detect_rename = -1;
3586         o->detect_directory_renames = 1;
3587         merge_recursive_config(o);
3588         merge_verbosity = getenv("GIT_MERGE_VERBOSITY");
3589         if (merge_verbosity)
3590                 o->verbosity = strtol(merge_verbosity, NULL, 10);
3591         if (o->verbosity >= 5)
3592                 o->buffer_output = 0;
3593         strbuf_init(&o->obuf, 0);
3594         string_list_init(&o->df_conflict_file_set, 1);
3595 }
3596
3597 int parse_merge_opt(struct merge_options *o, const char *s)
3598 {
3599         const char *arg;
3600
3601         if (!s || !*s)
3602                 return -1;
3603         if (!strcmp(s, "ours"))
3604                 o->recursive_variant = MERGE_RECURSIVE_OURS;
3605         else if (!strcmp(s, "theirs"))
3606                 o->recursive_variant = MERGE_RECURSIVE_THEIRS;
3607         else if (!strcmp(s, "subtree"))
3608                 o->subtree_shift = "";
3609         else if (skip_prefix(s, "subtree=", &arg))
3610                 o->subtree_shift = arg;
3611         else if (!strcmp(s, "patience"))
3612                 o->xdl_opts = DIFF_WITH_ALG(o, PATIENCE_DIFF);
3613         else if (!strcmp(s, "histogram"))
3614                 o->xdl_opts = DIFF_WITH_ALG(o, HISTOGRAM_DIFF);
3615         else if (skip_prefix(s, "diff-algorithm=", &arg)) {
3616                 long value = parse_algorithm_value(arg);
3617                 if (value < 0)
3618                         return -1;
3619                 /* clear out previous settings */
3620                 DIFF_XDL_CLR(o, NEED_MINIMAL);
3621                 o->xdl_opts &= ~XDF_DIFF_ALGORITHM_MASK;
3622                 o->xdl_opts |= value;
3623         }
3624         else if (!strcmp(s, "ignore-space-change"))
3625                 DIFF_XDL_SET(o, IGNORE_WHITESPACE_CHANGE);
3626         else if (!strcmp(s, "ignore-all-space"))
3627                 DIFF_XDL_SET(o, IGNORE_WHITESPACE);
3628         else if (!strcmp(s, "ignore-space-at-eol"))
3629                 DIFF_XDL_SET(o, IGNORE_WHITESPACE_AT_EOL);
3630         else if (!strcmp(s, "ignore-cr-at-eol"))
3631                 DIFF_XDL_SET(o, IGNORE_CR_AT_EOL);
3632         else if (!strcmp(s, "renormalize"))
3633                 o->renormalize = 1;
3634         else if (!strcmp(s, "no-renormalize"))
3635                 o->renormalize = 0;
3636         else if (!strcmp(s, "no-renames"))
3637                 o->merge_detect_rename = 0;
3638         else if (!strcmp(s, "find-renames")) {
3639                 o->merge_detect_rename = 1;
3640                 o->rename_score = 0;
3641         }
3642         else if (skip_prefix(s, "find-renames=", &arg) ||
3643                  skip_prefix(s, "rename-threshold=", &arg)) {
3644                 if ((o->rename_score = parse_rename_score(&arg)) == -1 || *arg != 0)
3645                         return -1;
3646                 o->merge_detect_rename = 1;
3647         }
3648         else
3649                 return -1;
3650         return 0;
3651 }