merge-recursive: fix rename/add conflict handling
[git] / merge-recursive.c
1 /*
2  * Recursive Merge algorithm stolen from git-merge-recursive.py by
3  * Fredrik Kuivinen.
4  * The thieves were Alex Riesen and Johannes Schindelin, in June/July 2006
5  */
6 #include "cache.h"
7 #include "config.h"
8 #include "advice.h"
9 #include "lockfile.h"
10 #include "cache-tree.h"
11 #include "object-store.h"
12 #include "repository.h"
13 #include "commit.h"
14 #include "blob.h"
15 #include "builtin.h"
16 #include "tree-walk.h"
17 #include "diff.h"
18 #include "diffcore.h"
19 #include "tag.h"
20 #include "alloc.h"
21 #include "unpack-trees.h"
22 #include "string-list.h"
23 #include "xdiff-interface.h"
24 #include "ll-merge.h"
25 #include "attr.h"
26 #include "merge-recursive.h"
27 #include "dir.h"
28 #include "submodule.h"
29 #include "revision.h"
30 #include "commit-reach.h"
31
32 struct path_hashmap_entry {
33         struct hashmap_entry e;
34         char path[FLEX_ARRAY];
35 };
36
37 static int path_hashmap_cmp(const void *cmp_data,
38                             const void *entry,
39                             const void *entry_or_key,
40                             const void *keydata)
41 {
42         const struct path_hashmap_entry *a = entry;
43         const struct path_hashmap_entry *b = entry_or_key;
44         const char *key = keydata;
45
46         if (ignore_case)
47                 return strcasecmp(a->path, key ? key : b->path);
48         else
49                 return strcmp(a->path, key ? key : b->path);
50 }
51
52 static unsigned int path_hash(const char *path)
53 {
54         return ignore_case ? strihash(path) : strhash(path);
55 }
56
57 static struct dir_rename_entry *dir_rename_find_entry(struct hashmap *hashmap,
58                                                       char *dir)
59 {
60         struct dir_rename_entry key;
61
62         if (dir == NULL)
63                 return NULL;
64         hashmap_entry_init(&key, strhash(dir));
65         key.dir = dir;
66         return hashmap_get(hashmap, &key, NULL);
67 }
68
69 static int dir_rename_cmp(const void *unused_cmp_data,
70                           const void *entry,
71                           const void *entry_or_key,
72                           const void *unused_keydata)
73 {
74         const struct dir_rename_entry *e1 = entry;
75         const struct dir_rename_entry *e2 = entry_or_key;
76
77         return strcmp(e1->dir, e2->dir);
78 }
79
80 static void dir_rename_init(struct hashmap *map)
81 {
82         hashmap_init(map, dir_rename_cmp, NULL, 0);
83 }
84
85 static void dir_rename_entry_init(struct dir_rename_entry *entry,
86                                   char *directory)
87 {
88         hashmap_entry_init(entry, strhash(directory));
89         entry->dir = directory;
90         entry->non_unique_new_dir = 0;
91         strbuf_init(&entry->new_dir, 0);
92         string_list_init(&entry->possible_new_dirs, 0);
93 }
94
95 static struct collision_entry *collision_find_entry(struct hashmap *hashmap,
96                                                     char *target_file)
97 {
98         struct collision_entry key;
99
100         hashmap_entry_init(&key, strhash(target_file));
101         key.target_file = target_file;
102         return hashmap_get(hashmap, &key, NULL);
103 }
104
105 static int collision_cmp(void *unused_cmp_data,
106                          const struct collision_entry *e1,
107                          const struct collision_entry *e2,
108                          const void *unused_keydata)
109 {
110         return strcmp(e1->target_file, e2->target_file);
111 }
112
113 static void collision_init(struct hashmap *map)
114 {
115         hashmap_init(map, (hashmap_cmp_fn) collision_cmp, NULL, 0);
116 }
117
118 static void flush_output(struct merge_options *o)
119 {
120         if (o->buffer_output < 2 && o->obuf.len) {
121                 fputs(o->obuf.buf, stdout);
122                 strbuf_reset(&o->obuf);
123         }
124 }
125
126 static int err(struct merge_options *o, const char *err, ...)
127 {
128         va_list params;
129
130         if (o->buffer_output < 2)
131                 flush_output(o);
132         else {
133                 strbuf_complete(&o->obuf, '\n');
134                 strbuf_addstr(&o->obuf, "error: ");
135         }
136         va_start(params, err);
137         strbuf_vaddf(&o->obuf, err, params);
138         va_end(params);
139         if (o->buffer_output > 1)
140                 strbuf_addch(&o->obuf, '\n');
141         else {
142                 error("%s", o->obuf.buf);
143                 strbuf_reset(&o->obuf);
144         }
145
146         return -1;
147 }
148
149 static struct tree *shift_tree_object(struct tree *one, struct tree *two,
150                                       const char *subtree_shift)
151 {
152         struct object_id shifted;
153
154         if (!*subtree_shift) {
155                 shift_tree(&one->object.oid, &two->object.oid, &shifted, 0);
156         } else {
157                 shift_tree_by(&one->object.oid, &two->object.oid, &shifted,
158                               subtree_shift);
159         }
160         if (oideq(&two->object.oid, &shifted))
161                 return two;
162         return lookup_tree(the_repository, &shifted);
163 }
164
165 static struct commit *make_virtual_commit(struct tree *tree, const char *comment)
166 {
167         struct commit *commit = alloc_commit_node(the_repository);
168
169         set_merge_remote_desc(commit, comment, (struct object *)commit);
170         commit->maybe_tree = tree;
171         commit->object.parsed = 1;
172         return commit;
173 }
174
175 /*
176  * Since we use get_tree_entry(), which does not put the read object into
177  * the object pool, we cannot rely on a == b.
178  */
179 static int oid_eq(const struct object_id *a, const struct object_id *b)
180 {
181         if (!a && !b)
182                 return 2;
183         return a && b && oideq(a, b);
184 }
185
186 enum rename_type {
187         RENAME_NORMAL = 0,
188         RENAME_VIA_DIR,
189         RENAME_ADD,
190         RENAME_DELETE,
191         RENAME_ONE_FILE_TO_ONE,
192         RENAME_ONE_FILE_TO_TWO,
193         RENAME_TWO_FILES_TO_ONE
194 };
195
196 struct rename_conflict_info {
197         enum rename_type rename_type;
198         struct diff_filepair *pair1;
199         struct diff_filepair *pair2;
200         const char *branch1;
201         const char *branch2;
202         struct stage_data *dst_entry1;
203         struct stage_data *dst_entry2;
204         struct diff_filespec ren1_other;
205         struct diff_filespec ren2_other;
206 };
207
208 /*
209  * Since we want to write the index eventually, we cannot reuse the index
210  * for these (temporary) data.
211  */
212 struct stage_data {
213         struct {
214                 unsigned mode;
215                 struct object_id oid;
216         } stages[4];
217         struct rename_conflict_info *rename_conflict_info;
218         unsigned processed:1;
219 };
220
221 static inline void setup_rename_conflict_info(enum rename_type rename_type,
222                                               struct diff_filepair *pair1,
223                                               struct diff_filepair *pair2,
224                                               const char *branch1,
225                                               const char *branch2,
226                                               struct stage_data *dst_entry1,
227                                               struct stage_data *dst_entry2,
228                                               struct merge_options *o,
229                                               struct stage_data *src_entry1,
230                                               struct stage_data *src_entry2)
231 {
232         int ostage1 = 0, ostage2;
233         struct rename_conflict_info *ci;
234
235         /*
236          * When we have two renames involved, it's easiest to get the
237          * correct things into stage 2 and 3, and to make sure that the
238          * content merge puts HEAD before the other branch if we just
239          * ensure that branch1 == o->branch1.  So, simply flip arguments
240          * around if we don't have that.
241          */
242         if (dst_entry2 && branch1 != o->branch1) {
243                 setup_rename_conflict_info(rename_type,
244                                            pair2,      pair1,
245                                            branch2,    branch1,
246                                            dst_entry2, dst_entry1,
247                                            o,
248                                            src_entry2, src_entry1);
249                 return;
250         }
251
252         ci = xcalloc(1, sizeof(struct rename_conflict_info));
253         ci->rename_type = rename_type;
254         ci->pair1 = pair1;
255         ci->branch1 = branch1;
256         ci->branch2 = branch2;
257
258         ci->dst_entry1 = dst_entry1;
259         dst_entry1->rename_conflict_info = ci;
260         dst_entry1->processed = 0;
261
262         assert(!pair2 == !dst_entry2);
263         if (dst_entry2) {
264                 ci->dst_entry2 = dst_entry2;
265                 ci->pair2 = pair2;
266                 dst_entry2->rename_conflict_info = ci;
267         }
268
269         /*
270          * For each rename, there could have been
271          * modifications on the side of history where that
272          * file was not renamed.
273          */
274         if (rename_type == RENAME_ADD ||
275             rename_type == RENAME_TWO_FILES_TO_ONE) {
276                 ostage1 = o->branch1 == branch1 ? 3 : 2;
277
278                 ci->ren1_other.path = pair1->one->path;
279                 oidcpy(&ci->ren1_other.oid, &src_entry1->stages[ostage1].oid);
280                 ci->ren1_other.mode = src_entry1->stages[ostage1].mode;
281         }
282
283         if (rename_type == RENAME_TWO_FILES_TO_ONE) {
284                 ostage2 = ostage1 ^ 1;
285
286                 ci->ren2_other.path = pair2->one->path;
287                 oidcpy(&ci->ren2_other.oid, &src_entry2->stages[ostage2].oid);
288                 ci->ren2_other.mode = src_entry2->stages[ostage2].mode;
289         }
290 }
291
292 static int show(struct merge_options *o, int v)
293 {
294         return (!o->call_depth && o->verbosity >= v) || o->verbosity >= 5;
295 }
296
297 __attribute__((format (printf, 3, 4)))
298 static void output(struct merge_options *o, int v, const char *fmt, ...)
299 {
300         va_list ap;
301
302         if (!show(o, v))
303                 return;
304
305         strbuf_addchars(&o->obuf, ' ', o->call_depth * 2);
306
307         va_start(ap, fmt);
308         strbuf_vaddf(&o->obuf, fmt, ap);
309         va_end(ap);
310
311         strbuf_addch(&o->obuf, '\n');
312         if (!o->buffer_output)
313                 flush_output(o);
314 }
315
316 static void output_commit_title(struct merge_options *o, struct commit *commit)
317 {
318         struct merge_remote_desc *desc;
319
320         strbuf_addchars(&o->obuf, ' ', o->call_depth * 2);
321         desc = merge_remote_util(commit);
322         if (desc)
323                 strbuf_addf(&o->obuf, "virtual %s\n", desc->name);
324         else {
325                 strbuf_add_unique_abbrev(&o->obuf, &commit->object.oid,
326                                          DEFAULT_ABBREV);
327                 strbuf_addch(&o->obuf, ' ');
328                 if (parse_commit(commit) != 0)
329                         strbuf_addstr(&o->obuf, _("(bad commit)\n"));
330                 else {
331                         const char *title;
332                         const char *msg = get_commit_buffer(commit, NULL);
333                         int len = find_commit_subject(msg, &title);
334                         if (len)
335                                 strbuf_addf(&o->obuf, "%.*s\n", len, title);
336                         unuse_commit_buffer(commit, msg);
337                 }
338         }
339         flush_output(o);
340 }
341
342 static int add_cacheinfo(struct merge_options *o,
343                          unsigned int mode, const struct object_id *oid,
344                          const char *path, int stage, int refresh, int options)
345 {
346         struct cache_entry *ce;
347         int ret;
348
349         ce = make_cache_entry(&the_index, mode, oid ? oid : &null_oid, path, stage, 0);
350         if (!ce)
351                 return err(o, _("add_cacheinfo failed for path '%s'; merge aborting."), path);
352
353         ret = add_cache_entry(ce, options);
354         if (refresh) {
355                 struct cache_entry *nce;
356
357                 nce = refresh_cache_entry(&the_index, ce, CE_MATCH_REFRESH | CE_MATCH_IGNORE_MISSING);
358                 if (!nce)
359                         return err(o, _("add_cacheinfo failed to refresh for path '%s'; merge aborting."), path);
360                 if (nce != ce)
361                         ret = add_cache_entry(nce, options);
362         }
363         return ret;
364 }
365
366 static void init_tree_desc_from_tree(struct tree_desc *desc, struct tree *tree)
367 {
368         parse_tree(tree);
369         init_tree_desc(desc, tree->buffer, tree->size);
370 }
371
372 static int unpack_trees_start(struct merge_options *o,
373                               struct tree *common,
374                               struct tree *head,
375                               struct tree *merge)
376 {
377         int rc;
378         struct tree_desc t[3];
379         struct index_state tmp_index = { NULL };
380
381         memset(&o->unpack_opts, 0, sizeof(o->unpack_opts));
382         if (o->call_depth)
383                 o->unpack_opts.index_only = 1;
384         else
385                 o->unpack_opts.update = 1;
386         o->unpack_opts.merge = 1;
387         o->unpack_opts.head_idx = 2;
388         o->unpack_opts.fn = threeway_merge;
389         o->unpack_opts.src_index = &the_index;
390         o->unpack_opts.dst_index = &tmp_index;
391         o->unpack_opts.aggressive = !merge_detect_rename(o);
392         setup_unpack_trees_porcelain(&o->unpack_opts, "merge");
393
394         init_tree_desc_from_tree(t+0, common);
395         init_tree_desc_from_tree(t+1, head);
396         init_tree_desc_from_tree(t+2, merge);
397
398         rc = unpack_trees(3, t, &o->unpack_opts);
399         cache_tree_free(&active_cache_tree);
400
401         /*
402          * Update the_index to match the new results, AFTER saving a copy
403          * in o->orig_index.  Update src_index to point to the saved copy.
404          * (verify_uptodate() checks src_index, and the original index is
405          * the one that had the necessary modification timestamps.)
406          */
407         o->orig_index = the_index;
408         the_index = tmp_index;
409         o->unpack_opts.src_index = &o->orig_index;
410
411         return rc;
412 }
413
414 static void unpack_trees_finish(struct merge_options *o)
415 {
416         discard_index(&o->orig_index);
417         clear_unpack_trees_porcelain(&o->unpack_opts);
418 }
419
420 struct tree *write_tree_from_memory(struct merge_options *o)
421 {
422         struct tree *result = NULL;
423
424         if (unmerged_cache()) {
425                 int i;
426                 fprintf(stderr, "BUG: There are unmerged index entries:\n");
427                 for (i = 0; i < active_nr; i++) {
428                         const struct cache_entry *ce = active_cache[i];
429                         if (ce_stage(ce))
430                                 fprintf(stderr, "BUG: %d %.*s\n", ce_stage(ce),
431                                         (int)ce_namelen(ce), ce->name);
432                 }
433                 BUG("unmerged index entries in merge-recursive.c");
434         }
435
436         if (!active_cache_tree)
437                 active_cache_tree = cache_tree();
438
439         if (!cache_tree_fully_valid(active_cache_tree) &&
440             cache_tree_update(&the_index, 0) < 0) {
441                 err(o, _("error building trees"));
442                 return NULL;
443         }
444
445         result = lookup_tree(the_repository, &active_cache_tree->oid);
446
447         return result;
448 }
449
450 static int save_files_dirs(const struct object_id *oid,
451                            struct strbuf *base, const char *path,
452                            unsigned int mode, int stage, void *context)
453 {
454         struct path_hashmap_entry *entry;
455         int baselen = base->len;
456         struct merge_options *o = context;
457
458         strbuf_addstr(base, path);
459
460         FLEX_ALLOC_MEM(entry, path, base->buf, base->len);
461         hashmap_entry_init(entry, path_hash(entry->path));
462         hashmap_add(&o->current_file_dir_set, entry);
463
464         strbuf_setlen(base, baselen);
465         return (S_ISDIR(mode) ? READ_TREE_RECURSIVE : 0);
466 }
467
468 static void get_files_dirs(struct merge_options *o, struct tree *tree)
469 {
470         struct pathspec match_all;
471         memset(&match_all, 0, sizeof(match_all));
472         read_tree_recursive(tree, "", 0, 0, &match_all, save_files_dirs, o);
473 }
474
475 static int get_tree_entry_if_blob(const struct object_id *tree,
476                                   const char *path,
477                                   struct object_id *hashy,
478                                   unsigned int *mode_o)
479 {
480         int ret;
481
482         ret = get_tree_entry(tree, path, hashy, mode_o);
483         if (S_ISDIR(*mode_o)) {
484                 oidcpy(hashy, &null_oid);
485                 *mode_o = 0;
486         }
487         return ret;
488 }
489
490 /*
491  * Returns an index_entry instance which doesn't have to correspond to
492  * a real cache entry in Git's index.
493  */
494 static struct stage_data *insert_stage_data(const char *path,
495                 struct tree *o, struct tree *a, struct tree *b,
496                 struct string_list *entries)
497 {
498         struct string_list_item *item;
499         struct stage_data *e = xcalloc(1, sizeof(struct stage_data));
500         get_tree_entry_if_blob(&o->object.oid, path,
501                                &e->stages[1].oid, &e->stages[1].mode);
502         get_tree_entry_if_blob(&a->object.oid, path,
503                                &e->stages[2].oid, &e->stages[2].mode);
504         get_tree_entry_if_blob(&b->object.oid, path,
505                                &e->stages[3].oid, &e->stages[3].mode);
506         item = string_list_insert(entries, path);
507         item->util = e;
508         return e;
509 }
510
511 /*
512  * Create a dictionary mapping file names to stage_data objects. The
513  * dictionary contains one entry for every path with a non-zero stage entry.
514  */
515 static struct string_list *get_unmerged(void)
516 {
517         struct string_list *unmerged = xcalloc(1, sizeof(struct string_list));
518         int i;
519
520         unmerged->strdup_strings = 1;
521
522         for (i = 0; i < active_nr; i++) {
523                 struct string_list_item *item;
524                 struct stage_data *e;
525                 const struct cache_entry *ce = active_cache[i];
526                 if (!ce_stage(ce))
527                         continue;
528
529                 item = string_list_lookup(unmerged, ce->name);
530                 if (!item) {
531                         item = string_list_insert(unmerged, ce->name);
532                         item->util = xcalloc(1, sizeof(struct stage_data));
533                 }
534                 e = item->util;
535                 e->stages[ce_stage(ce)].mode = ce->ce_mode;
536                 oidcpy(&e->stages[ce_stage(ce)].oid, &ce->oid);
537         }
538
539         return unmerged;
540 }
541
542 static int string_list_df_name_compare(const char *one, const char *two)
543 {
544         int onelen = strlen(one);
545         int twolen = strlen(two);
546         /*
547          * Here we only care that entries for D/F conflicts are
548          * adjacent, in particular with the file of the D/F conflict
549          * appearing before files below the corresponding directory.
550          * The order of the rest of the list is irrelevant for us.
551          *
552          * To achieve this, we sort with df_name_compare and provide
553          * the mode S_IFDIR so that D/F conflicts will sort correctly.
554          * We use the mode S_IFDIR for everything else for simplicity,
555          * since in other cases any changes in their order due to
556          * sorting cause no problems for us.
557          */
558         int cmp = df_name_compare(one, onelen, S_IFDIR,
559                                   two, twolen, S_IFDIR);
560         /*
561          * Now that 'foo' and 'foo/bar' compare equal, we have to make sure
562          * that 'foo' comes before 'foo/bar'.
563          */
564         if (cmp)
565                 return cmp;
566         return onelen - twolen;
567 }
568
569 static void record_df_conflict_files(struct merge_options *o,
570                                      struct string_list *entries)
571 {
572         /* If there is a D/F conflict and the file for such a conflict
573          * currently exists in the working tree, we want to allow it to be
574          * removed to make room for the corresponding directory if needed.
575          * The files underneath the directories of such D/F conflicts will
576          * be processed before the corresponding file involved in the D/F
577          * conflict.  If the D/F directory ends up being removed by the
578          * merge, then we won't have to touch the D/F file.  If the D/F
579          * directory needs to be written to the working copy, then the D/F
580          * file will simply be removed (in make_room_for_path()) to make
581          * room for the necessary paths.  Note that if both the directory
582          * and the file need to be present, then the D/F file will be
583          * reinstated with a new unique name at the time it is processed.
584          */
585         struct string_list df_sorted_entries = STRING_LIST_INIT_NODUP;
586         const char *last_file = NULL;
587         int last_len = 0;
588         int i;
589
590         /*
591          * If we're merging merge-bases, we don't want to bother with
592          * any working directory changes.
593          */
594         if (o->call_depth)
595                 return;
596
597         /* Ensure D/F conflicts are adjacent in the entries list. */
598         for (i = 0; i < entries->nr; i++) {
599                 struct string_list_item *next = &entries->items[i];
600                 string_list_append(&df_sorted_entries, next->string)->util =
601                                    next->util;
602         }
603         df_sorted_entries.cmp = string_list_df_name_compare;
604         string_list_sort(&df_sorted_entries);
605
606         string_list_clear(&o->df_conflict_file_set, 1);
607         for (i = 0; i < df_sorted_entries.nr; i++) {
608                 const char *path = df_sorted_entries.items[i].string;
609                 int len = strlen(path);
610                 struct stage_data *e = df_sorted_entries.items[i].util;
611
612                 /*
613                  * Check if last_file & path correspond to a D/F conflict;
614                  * i.e. whether path is last_file+'/'+<something>.
615                  * If so, record that it's okay to remove last_file to make
616                  * room for path and friends if needed.
617                  */
618                 if (last_file &&
619                     len > last_len &&
620                     memcmp(path, last_file, last_len) == 0 &&
621                     path[last_len] == '/') {
622                         string_list_insert(&o->df_conflict_file_set, last_file);
623                 }
624
625                 /*
626                  * Determine whether path could exist as a file in the
627                  * working directory as a possible D/F conflict.  This
628                  * will only occur when it exists in stage 2 as a
629                  * file.
630                  */
631                 if (S_ISREG(e->stages[2].mode) || S_ISLNK(e->stages[2].mode)) {
632                         last_file = path;
633                         last_len = len;
634                 } else {
635                         last_file = NULL;
636                 }
637         }
638         string_list_clear(&df_sorted_entries, 0);
639 }
640
641 struct rename {
642         struct diff_filepair *pair;
643         /*
644          * Purpose of src_entry and dst_entry:
645          *
646          * If 'before' is renamed to 'after' then src_entry will contain
647          * the versions of 'before' from the merge_base, HEAD, and MERGE in
648          * stages 1, 2, and 3; dst_entry will contain the respective
649          * versions of 'after' in corresponding locations.  Thus, we have a
650          * total of six modes and oids, though some will be null.  (Stage 0
651          * is ignored; we're interested in handling conflicts.)
652          *
653          * Since we don't turn on break-rewrites by default, neither
654          * src_entry nor dst_entry can have all three of their stages have
655          * non-null oids, meaning at most four of the six will be non-null.
656          * Also, since this is a rename, both src_entry and dst_entry will
657          * have at least one non-null oid, meaning at least two will be
658          * non-null.  Of the six oids, a typical rename will have three be
659          * non-null.  Only two implies a rename/delete, and four implies a
660          * rename/add.
661          */
662         struct stage_data *src_entry;
663         struct stage_data *dst_entry;
664         unsigned add_turned_into_rename:1;
665         unsigned processed:1;
666 };
667
668 static int update_stages(struct merge_options *opt, const char *path,
669                          const struct diff_filespec *o,
670                          const struct diff_filespec *a,
671                          const struct diff_filespec *b)
672 {
673
674         /*
675          * NOTE: It is usually a bad idea to call update_stages on a path
676          * before calling update_file on that same path, since it can
677          * sometimes lead to spurious "refusing to lose untracked file..."
678          * messages from update_file (via make_room_for path via
679          * would_lose_untracked).  Instead, reverse the order of the calls
680          * (executing update_file first and then update_stages).
681          */
682         int clear = 1;
683         int options = ADD_CACHE_OK_TO_ADD | ADD_CACHE_SKIP_DFCHECK;
684         if (clear)
685                 if (remove_file_from_cache(path))
686                         return -1;
687         if (o)
688                 if (add_cacheinfo(opt, o->mode, &o->oid, path, 1, 0, options))
689                         return -1;
690         if (a)
691                 if (add_cacheinfo(opt, a->mode, &a->oid, path, 2, 0, options))
692                         return -1;
693         if (b)
694                 if (add_cacheinfo(opt, b->mode, &b->oid, path, 3, 0, options))
695                         return -1;
696         return 0;
697 }
698
699 static int update_stages_for_stage_data(struct merge_options *opt,
700                                         const char *path,
701                                         const struct stage_data *stage_data)
702 {
703         struct diff_filespec o, a, b;
704
705         o.mode = stage_data->stages[1].mode;
706         oidcpy(&o.oid, &stage_data->stages[1].oid);
707
708         a.mode = stage_data->stages[2].mode;
709         oidcpy(&a.oid, &stage_data->stages[2].oid);
710
711         b.mode = stage_data->stages[3].mode;
712         oidcpy(&b.oid, &stage_data->stages[3].oid);
713
714         return update_stages(opt, path,
715                              is_null_oid(&o.oid) ? NULL : &o,
716                              is_null_oid(&a.oid) ? NULL : &a,
717                              is_null_oid(&b.oid) ? NULL : &b);
718 }
719
720 static void update_entry(struct stage_data *entry,
721                          struct diff_filespec *o,
722                          struct diff_filespec *a,
723                          struct diff_filespec *b)
724 {
725         entry->processed = 0;
726         entry->stages[1].mode = o->mode;
727         entry->stages[2].mode = a->mode;
728         entry->stages[3].mode = b->mode;
729         oidcpy(&entry->stages[1].oid, &o->oid);
730         oidcpy(&entry->stages[2].oid, &a->oid);
731         oidcpy(&entry->stages[3].oid, &b->oid);
732 }
733
734 static int remove_file(struct merge_options *o, int clean,
735                        const char *path, int no_wd)
736 {
737         int update_cache = o->call_depth || clean;
738         int update_working_directory = !o->call_depth && !no_wd;
739
740         if (update_cache) {
741                 if (remove_file_from_cache(path))
742                         return -1;
743         }
744         if (update_working_directory) {
745                 if (ignore_case) {
746                         struct cache_entry *ce;
747                         ce = cache_file_exists(path, strlen(path), ignore_case);
748                         if (ce && ce_stage(ce) == 0 && strcmp(path, ce->name))
749                                 return 0;
750                 }
751                 if (remove_path(path))
752                         return -1;
753         }
754         return 0;
755 }
756
757 /* add a string to a strbuf, but converting "/" to "_" */
758 static void add_flattened_path(struct strbuf *out, const char *s)
759 {
760         size_t i = out->len;
761         strbuf_addstr(out, s);
762         for (; i < out->len; i++)
763                 if (out->buf[i] == '/')
764                         out->buf[i] = '_';
765 }
766
767 static char *unique_path(struct merge_options *o, const char *path, const char *branch)
768 {
769         struct path_hashmap_entry *entry;
770         struct strbuf newpath = STRBUF_INIT;
771         int suffix = 0;
772         size_t base_len;
773
774         strbuf_addf(&newpath, "%s~", path);
775         add_flattened_path(&newpath, branch);
776
777         base_len = newpath.len;
778         while (hashmap_get_from_hash(&o->current_file_dir_set,
779                                      path_hash(newpath.buf), newpath.buf) ||
780                (!o->call_depth && file_exists(newpath.buf))) {
781                 strbuf_setlen(&newpath, base_len);
782                 strbuf_addf(&newpath, "_%d", suffix++);
783         }
784
785         FLEX_ALLOC_MEM(entry, path, newpath.buf, newpath.len);
786         hashmap_entry_init(entry, path_hash(entry->path));
787         hashmap_add(&o->current_file_dir_set, entry);
788         return strbuf_detach(&newpath, NULL);
789 }
790
791 /**
792  * Check whether a directory in the index is in the way of an incoming
793  * file.  Return 1 if so.  If check_working_copy is non-zero, also
794  * check the working directory.  If empty_ok is non-zero, also return
795  * 0 in the case where the working-tree dir exists but is empty.
796  */
797 static int dir_in_way(const char *path, int check_working_copy, int empty_ok)
798 {
799         int pos;
800         struct strbuf dirpath = STRBUF_INIT;
801         struct stat st;
802
803         strbuf_addstr(&dirpath, path);
804         strbuf_addch(&dirpath, '/');
805
806         pos = cache_name_pos(dirpath.buf, dirpath.len);
807
808         if (pos < 0)
809                 pos = -1 - pos;
810         if (pos < active_nr &&
811             !strncmp(dirpath.buf, active_cache[pos]->name, dirpath.len)) {
812                 strbuf_release(&dirpath);
813                 return 1;
814         }
815
816         strbuf_release(&dirpath);
817         return check_working_copy && !lstat(path, &st) && S_ISDIR(st.st_mode) &&
818                 !(empty_ok && is_empty_dir(path));
819 }
820
821 /*
822  * Returns whether path was tracked in the index before the merge started,
823  * and its oid and mode match the specified values
824  */
825 static int was_tracked_and_matches(struct merge_options *o, const char *path,
826                                    const struct object_id *oid, unsigned mode)
827 {
828         int pos = index_name_pos(&o->orig_index, path, strlen(path));
829         struct cache_entry *ce;
830
831         if (0 > pos)
832                 /* we were not tracking this path before the merge */
833                 return 0;
834
835         /* See if the file we were tracking before matches */
836         ce = o->orig_index.cache[pos];
837         return (oid_eq(&ce->oid, oid) && ce->ce_mode == mode);
838 }
839
840 /*
841  * Returns whether path was tracked in the index before the merge started
842  */
843 static int was_tracked(struct merge_options *o, const char *path)
844 {
845         int pos = index_name_pos(&o->orig_index, path, strlen(path));
846
847         if (0 <= pos)
848                 /* we were tracking this path before the merge */
849                 return 1;
850
851         return 0;
852 }
853
854 static int would_lose_untracked(const char *path)
855 {
856         /*
857          * This may look like it can be simplified to:
858          *   return !was_tracked(o, path) && file_exists(path)
859          * but it can't.  This function needs to know whether path was in
860          * the working tree due to EITHER having been tracked in the index
861          * before the merge OR having been put into the working copy and
862          * index by unpack_trees().  Due to that either-or requirement, we
863          * check the current index instead of the original one.
864          *
865          * Note that we do not need to worry about merge-recursive itself
866          * updating the index after unpack_trees() and before calling this
867          * function, because we strictly require all code paths in
868          * merge-recursive to update the working tree first and the index
869          * second.  Doing otherwise would break
870          * update_file()/would_lose_untracked(); see every comment in this
871          * file which mentions "update_stages".
872          */
873         int pos = cache_name_pos(path, strlen(path));
874
875         if (pos < 0)
876                 pos = -1 - pos;
877         while (pos < active_nr &&
878                !strcmp(path, active_cache[pos]->name)) {
879                 /*
880                  * If stage #0, it is definitely tracked.
881                  * If it has stage #2 then it was tracked
882                  * before this merge started.  All other
883                  * cases the path was not tracked.
884                  */
885                 switch (ce_stage(active_cache[pos])) {
886                 case 0:
887                 case 2:
888                         return 0;
889                 }
890                 pos++;
891         }
892         return file_exists(path);
893 }
894
895 static int was_dirty(struct merge_options *o, const char *path)
896 {
897         struct cache_entry *ce;
898         int dirty = 1;
899
900         if (o->call_depth || !was_tracked(o, path))
901                 return !dirty;
902
903         ce = index_file_exists(o->unpack_opts.src_index,
904                                path, strlen(path), ignore_case);
905         dirty = verify_uptodate(ce, &o->unpack_opts) != 0;
906         return dirty;
907 }
908
909 static int make_room_for_path(struct merge_options *o, const char *path)
910 {
911         int status, i;
912         const char *msg = _("failed to create path '%s'%s");
913
914         /* Unlink any D/F conflict files that are in the way */
915         for (i = 0; i < o->df_conflict_file_set.nr; i++) {
916                 const char *df_path = o->df_conflict_file_set.items[i].string;
917                 size_t pathlen = strlen(path);
918                 size_t df_pathlen = strlen(df_path);
919                 if (df_pathlen < pathlen &&
920                     path[df_pathlen] == '/' &&
921                     strncmp(path, df_path, df_pathlen) == 0) {
922                         output(o, 3,
923                                _("Removing %s to make room for subdirectory\n"),
924                                df_path);
925                         unlink(df_path);
926                         unsorted_string_list_delete_item(&o->df_conflict_file_set,
927                                                          i, 0);
928                         break;
929                 }
930         }
931
932         /* Make sure leading directories are created */
933         status = safe_create_leading_directories_const(path);
934         if (status) {
935                 if (status == SCLD_EXISTS)
936                         /* something else exists */
937                         return err(o, msg, path, _(": perhaps a D/F conflict?"));
938                 return err(o, msg, path, "");
939         }
940
941         /*
942          * Do not unlink a file in the work tree if we are not
943          * tracking it.
944          */
945         if (would_lose_untracked(path))
946                 return err(o, _("refusing to lose untracked file at '%s'"),
947                            path);
948
949         /* Successful unlink is good.. */
950         if (!unlink(path))
951                 return 0;
952         /* .. and so is no existing file */
953         if (errno == ENOENT)
954                 return 0;
955         /* .. but not some other error (who really cares what?) */
956         return err(o, msg, path, _(": perhaps a D/F conflict?"));
957 }
958
959 static int update_file_flags(struct merge_options *o,
960                              const struct object_id *oid,
961                              unsigned mode,
962                              const char *path,
963                              int update_cache,
964                              int update_wd)
965 {
966         int ret = 0;
967
968         if (o->call_depth)
969                 update_wd = 0;
970
971         if (update_wd) {
972                 enum object_type type;
973                 void *buf;
974                 unsigned long size;
975
976                 if (S_ISGITLINK(mode)) {
977                         /*
978                          * We may later decide to recursively descend into
979                          * the submodule directory and update its index
980                          * and/or work tree, but we do not do that now.
981                          */
982                         update_wd = 0;
983                         goto update_index;
984                 }
985
986                 buf = read_object_file(oid, &type, &size);
987                 if (!buf)
988                         return err(o, _("cannot read object %s '%s'"), oid_to_hex(oid), path);
989                 if (type != OBJ_BLOB) {
990                         ret = err(o, _("blob expected for %s '%s'"), oid_to_hex(oid), path);
991                         goto free_buf;
992                 }
993                 if (S_ISREG(mode)) {
994                         struct strbuf strbuf = STRBUF_INIT;
995                         if (convert_to_working_tree(&the_index, path, buf, size, &strbuf)) {
996                                 free(buf);
997                                 size = strbuf.len;
998                                 buf = strbuf_detach(&strbuf, NULL);
999                         }
1000                 }
1001
1002                 if (make_room_for_path(o, path) < 0) {
1003                         update_wd = 0;
1004                         goto free_buf;
1005                 }
1006                 if (S_ISREG(mode) || (!has_symlinks && S_ISLNK(mode))) {
1007                         int fd;
1008                         if (mode & 0100)
1009                                 mode = 0777;
1010                         else
1011                                 mode = 0666;
1012                         fd = open(path, O_WRONLY | O_TRUNC | O_CREAT, mode);
1013                         if (fd < 0) {
1014                                 ret = err(o, _("failed to open '%s': %s"),
1015                                           path, strerror(errno));
1016                                 goto free_buf;
1017                         }
1018                         write_in_full(fd, buf, size);
1019                         close(fd);
1020                 } else if (S_ISLNK(mode)) {
1021                         char *lnk = xmemdupz(buf, size);
1022                         safe_create_leading_directories_const(path);
1023                         unlink(path);
1024                         if (symlink(lnk, path))
1025                                 ret = err(o, _("failed to symlink '%s': %s"),
1026                                           path, strerror(errno));
1027                         free(lnk);
1028                 } else
1029                         ret = err(o,
1030                                   _("do not know what to do with %06o %s '%s'"),
1031                                   mode, oid_to_hex(oid), path);
1032         free_buf:
1033                 free(buf);
1034         }
1035 update_index:
1036         if (!ret && update_cache)
1037                 if (add_cacheinfo(o, mode, oid, path, 0, update_wd,
1038                                   ADD_CACHE_OK_TO_ADD))
1039                         return -1;
1040         return ret;
1041 }
1042
1043 static int update_file(struct merge_options *o,
1044                        int clean,
1045                        const struct object_id *oid,
1046                        unsigned mode,
1047                        const char *path)
1048 {
1049         return update_file_flags(o, oid, mode, path, o->call_depth || clean, !o->call_depth);
1050 }
1051
1052 /* Low level file merging, update and removal */
1053
1054 struct merge_file_info {
1055         struct object_id oid;
1056         unsigned mode;
1057         unsigned clean:1,
1058                  merge:1;
1059 };
1060
1061 static int merge_3way(struct merge_options *o,
1062                       mmbuffer_t *result_buf,
1063                       const struct diff_filespec *one,
1064                       const struct diff_filespec *a,
1065                       const struct diff_filespec *b,
1066                       const char *branch1,
1067                       const char *branch2,
1068                       const int extra_marker_size)
1069 {
1070         mmfile_t orig, src1, src2;
1071         struct ll_merge_options ll_opts = {0};
1072         char *base_name, *name1, *name2;
1073         int merge_status;
1074
1075         ll_opts.renormalize = o->renormalize;
1076         ll_opts.extra_marker_size = extra_marker_size;
1077         ll_opts.xdl_opts = o->xdl_opts;
1078
1079         if (o->call_depth) {
1080                 ll_opts.virtual_ancestor = 1;
1081                 ll_opts.variant = 0;
1082         } else {
1083                 switch (o->recursive_variant) {
1084                 case MERGE_RECURSIVE_OURS:
1085                         ll_opts.variant = XDL_MERGE_FAVOR_OURS;
1086                         break;
1087                 case MERGE_RECURSIVE_THEIRS:
1088                         ll_opts.variant = XDL_MERGE_FAVOR_THEIRS;
1089                         break;
1090                 default:
1091                         ll_opts.variant = 0;
1092                         break;
1093                 }
1094         }
1095
1096         if (strcmp(a->path, b->path) ||
1097             (o->ancestor != NULL && strcmp(a->path, one->path) != 0)) {
1098                 base_name = o->ancestor == NULL ? NULL :
1099                         mkpathdup("%s:%s", o->ancestor, one->path);
1100                 name1 = mkpathdup("%s:%s", branch1, a->path);
1101                 name2 = mkpathdup("%s:%s", branch2, b->path);
1102         } else {
1103                 base_name = o->ancestor == NULL ? NULL :
1104                         mkpathdup("%s", o->ancestor);
1105                 name1 = mkpathdup("%s", branch1);
1106                 name2 = mkpathdup("%s", branch2);
1107         }
1108
1109         read_mmblob(&orig, &one->oid);
1110         read_mmblob(&src1, &a->oid);
1111         read_mmblob(&src2, &b->oid);
1112
1113         merge_status = ll_merge(result_buf, a->path, &orig, base_name,
1114                                 &src1, name1, &src2, name2, &ll_opts);
1115
1116         free(base_name);
1117         free(name1);
1118         free(name2);
1119         free(orig.ptr);
1120         free(src1.ptr);
1121         free(src2.ptr);
1122         return merge_status;
1123 }
1124
1125 static int find_first_merges(struct object_array *result, const char *path,
1126                              struct commit *a, struct commit *b)
1127 {
1128         int i, j;
1129         struct object_array merges = OBJECT_ARRAY_INIT;
1130         struct commit *commit;
1131         int contains_another;
1132
1133         char merged_revision[42];
1134         const char *rev_args[] = { "rev-list", "--merges", "--ancestry-path",
1135                                    "--all", merged_revision, NULL };
1136         struct rev_info revs;
1137         struct setup_revision_opt rev_opts;
1138
1139         memset(result, 0, sizeof(struct object_array));
1140         memset(&rev_opts, 0, sizeof(rev_opts));
1141
1142         /* get all revisions that merge commit a */
1143         xsnprintf(merged_revision, sizeof(merged_revision), "^%s",
1144                   oid_to_hex(&a->object.oid));
1145         init_revisions(&revs, NULL);
1146         rev_opts.submodule = path;
1147         /* FIXME: can't handle linked worktrees in submodules yet */
1148         revs.single_worktree = path != NULL;
1149         setup_revisions(ARRAY_SIZE(rev_args)-1, rev_args, &revs, &rev_opts);
1150
1151         /* save all revisions from the above list that contain b */
1152         if (prepare_revision_walk(&revs))
1153                 die("revision walk setup failed");
1154         while ((commit = get_revision(&revs)) != NULL) {
1155                 struct object *o = &(commit->object);
1156                 if (in_merge_bases(b, commit))
1157                         add_object_array(o, NULL, &merges);
1158         }
1159         reset_revision_walk();
1160
1161         /* Now we've got all merges that contain a and b. Prune all
1162          * merges that contain another found merge and save them in
1163          * result.
1164          */
1165         for (i = 0; i < merges.nr; i++) {
1166                 struct commit *m1 = (struct commit *) merges.objects[i].item;
1167
1168                 contains_another = 0;
1169                 for (j = 0; j < merges.nr; j++) {
1170                         struct commit *m2 = (struct commit *) merges.objects[j].item;
1171                         if (i != j && in_merge_bases(m2, m1)) {
1172                                 contains_another = 1;
1173                                 break;
1174                         }
1175                 }
1176
1177                 if (!contains_another)
1178                         add_object_array(merges.objects[i].item, NULL, result);
1179         }
1180
1181         object_array_clear(&merges);
1182         return result->nr;
1183 }
1184
1185 static void print_commit(struct commit *commit)
1186 {
1187         struct strbuf sb = STRBUF_INIT;
1188         struct pretty_print_context ctx = {0};
1189         ctx.date_mode.type = DATE_NORMAL;
1190         format_commit_message(commit, " %h: %m %s", &sb, &ctx);
1191         fprintf(stderr, "%s\n", sb.buf);
1192         strbuf_release(&sb);
1193 }
1194
1195 static int merge_submodule(struct merge_options *o,
1196                            struct object_id *result, const char *path,
1197                            const struct object_id *base, const struct object_id *a,
1198                            const struct object_id *b)
1199 {
1200         struct commit *commit_base, *commit_a, *commit_b;
1201         int parent_count;
1202         struct object_array merges;
1203
1204         int i;
1205         int search = !o->call_depth;
1206
1207         /* store a in result in case we fail */
1208         oidcpy(result, a);
1209
1210         /* we can not handle deletion conflicts */
1211         if (is_null_oid(base))
1212                 return 0;
1213         if (is_null_oid(a))
1214                 return 0;
1215         if (is_null_oid(b))
1216                 return 0;
1217
1218         if (add_submodule_odb(path)) {
1219                 output(o, 1, _("Failed to merge submodule %s (not checked out)"), path);
1220                 return 0;
1221         }
1222
1223         if (!(commit_base = lookup_commit_reference(the_repository, base)) ||
1224             !(commit_a = lookup_commit_reference(the_repository, a)) ||
1225             !(commit_b = lookup_commit_reference(the_repository, b))) {
1226                 output(o, 1, _("Failed to merge submodule %s (commits not present)"), path);
1227                 return 0;
1228         }
1229
1230         /* check whether both changes are forward */
1231         if (!in_merge_bases(commit_base, commit_a) ||
1232             !in_merge_bases(commit_base, commit_b)) {
1233                 output(o, 1, _("Failed to merge submodule %s (commits don't follow merge-base)"), path);
1234                 return 0;
1235         }
1236
1237         /* Case #1: a is contained in b or vice versa */
1238         if (in_merge_bases(commit_a, commit_b)) {
1239                 oidcpy(result, b);
1240                 if (show(o, 3)) {
1241                         output(o, 3, _("Fast-forwarding submodule %s to the following commit:"), path);
1242                         output_commit_title(o, commit_b);
1243                 } else if (show(o, 2))
1244                         output(o, 2, _("Fast-forwarding submodule %s"), path);
1245                 else
1246                         ; /* no output */
1247
1248                 return 1;
1249         }
1250         if (in_merge_bases(commit_b, commit_a)) {
1251                 oidcpy(result, a);
1252                 if (show(o, 3)) {
1253                         output(o, 3, _("Fast-forwarding submodule %s to the following commit:"), path);
1254                         output_commit_title(o, commit_a);
1255                 } else if (show(o, 2))
1256                         output(o, 2, _("Fast-forwarding submodule %s"), path);
1257                 else
1258                         ; /* no output */
1259
1260                 return 1;
1261         }
1262
1263         /*
1264          * Case #2: There are one or more merges that contain a and b in
1265          * the submodule. If there is only one, then present it as a
1266          * suggestion to the user, but leave it marked unmerged so the
1267          * user needs to confirm the resolution.
1268          */
1269
1270         /* Skip the search if makes no sense to the calling context.  */
1271         if (!search)
1272                 return 0;
1273
1274         /* find commit which merges them */
1275         parent_count = find_first_merges(&merges, path, commit_a, commit_b);
1276         switch (parent_count) {
1277         case 0:
1278                 output(o, 1, _("Failed to merge submodule %s (merge following commits not found)"), path);
1279                 break;
1280
1281         case 1:
1282                 output(o, 1, _("Failed to merge submodule %s (not fast-forward)"), path);
1283                 output(o, 2, _("Found a possible merge resolution for the submodule:\n"));
1284                 print_commit((struct commit *) merges.objects[0].item);
1285                 output(o, 2, _(
1286                        "If this is correct simply add it to the index "
1287                        "for example\n"
1288                        "by using:\n\n"
1289                        "  git update-index --cacheinfo 160000 %s \"%s\"\n\n"
1290                        "which will accept this suggestion.\n"),
1291                        oid_to_hex(&merges.objects[0].item->oid), path);
1292                 break;
1293
1294         default:
1295                 output(o, 1, _("Failed to merge submodule %s (multiple merges found)"), path);
1296                 for (i = 0; i < merges.nr; i++)
1297                         print_commit((struct commit *) merges.objects[i].item);
1298         }
1299
1300         object_array_clear(&merges);
1301         return 0;
1302 }
1303
1304 static int merge_mode_and_contents(struct merge_options *o,
1305                                    const struct diff_filespec *one,
1306                                    const struct diff_filespec *a,
1307                                    const struct diff_filespec *b,
1308                                    const char *filename,
1309                                    const char *branch1,
1310                                    const char *branch2,
1311                                    const int extra_marker_size,
1312                                    struct merge_file_info *result)
1313 {
1314         if (o->branch1 != branch1) {
1315                 /*
1316                  * It's weird getting a reverse merge with HEAD on the bottom
1317                  * side of the conflict markers and the other branch on the
1318                  * top.  Fix that.
1319                  */
1320                 return merge_mode_and_contents(o, one, b, a,
1321                                                filename,
1322                                                branch2, branch1,
1323                                                extra_marker_size, result);
1324         }
1325
1326         result->merge = 0;
1327         result->clean = 1;
1328
1329         if ((S_IFMT & a->mode) != (S_IFMT & b->mode)) {
1330                 result->clean = 0;
1331                 if (S_ISREG(a->mode)) {
1332                         result->mode = a->mode;
1333                         oidcpy(&result->oid, &a->oid);
1334                 } else {
1335                         result->mode = b->mode;
1336                         oidcpy(&result->oid, &b->oid);
1337                 }
1338         } else {
1339                 if (!oid_eq(&a->oid, &one->oid) && !oid_eq(&b->oid, &one->oid))
1340                         result->merge = 1;
1341
1342                 /*
1343                  * Merge modes
1344                  */
1345                 if (a->mode == b->mode || a->mode == one->mode)
1346                         result->mode = b->mode;
1347                 else {
1348                         result->mode = a->mode;
1349                         if (b->mode != one->mode) {
1350                                 result->clean = 0;
1351                                 result->merge = 1;
1352                         }
1353                 }
1354
1355                 if (oid_eq(&a->oid, &b->oid) || oid_eq(&a->oid, &one->oid))
1356                         oidcpy(&result->oid, &b->oid);
1357                 else if (oid_eq(&b->oid, &one->oid))
1358                         oidcpy(&result->oid, &a->oid);
1359                 else if (S_ISREG(a->mode)) {
1360                         mmbuffer_t result_buf;
1361                         int ret = 0, merge_status;
1362
1363                         merge_status = merge_3way(o, &result_buf, one, a, b,
1364                                                   branch1, branch2,
1365                                                   extra_marker_size);
1366
1367                         if ((merge_status < 0) || !result_buf.ptr)
1368                                 ret = err(o, _("Failed to execute internal merge"));
1369
1370                         if (!ret &&
1371                             write_object_file(result_buf.ptr, result_buf.size,
1372                                               blob_type, &result->oid))
1373                                 ret = err(o, _("Unable to add %s to database"),
1374                                           a->path);
1375
1376                         free(result_buf.ptr);
1377                         if (ret)
1378                                 return ret;
1379                         result->clean = (merge_status == 0);
1380                 } else if (S_ISGITLINK(a->mode)) {
1381                         result->clean = merge_submodule(o, &result->oid,
1382                                                         one->path,
1383                                                         &one->oid,
1384                                                         &a->oid,
1385                                                         &b->oid);
1386                 } else if (S_ISLNK(a->mode)) {
1387                         switch (o->recursive_variant) {
1388                         case MERGE_RECURSIVE_NORMAL:
1389                                 oidcpy(&result->oid, &a->oid);
1390                                 if (!oid_eq(&a->oid, &b->oid))
1391                                         result->clean = 0;
1392                                 break;
1393                         case MERGE_RECURSIVE_OURS:
1394                                 oidcpy(&result->oid, &a->oid);
1395                                 break;
1396                         case MERGE_RECURSIVE_THEIRS:
1397                                 oidcpy(&result->oid, &b->oid);
1398                                 break;
1399                         }
1400                 } else
1401                         BUG("unsupported object type in the tree");
1402         }
1403
1404         if (result->merge)
1405                 output(o, 2, _("Auto-merging %s"), filename);
1406
1407         return 0;
1408 }
1409
1410 static int handle_rename_via_dir(struct merge_options *o,
1411                                  struct diff_filepair *pair,
1412                                  const char *rename_branch,
1413                                  const char *other_branch)
1414 {
1415         /*
1416          * Handle file adds that need to be renamed due to directory rename
1417          * detection.  This differs from handle_rename_normal, because
1418          * there is no content merge to do; just move the file into the
1419          * desired final location.
1420          */
1421         const struct diff_filespec *dest = pair->two;
1422
1423         if (!o->call_depth && would_lose_untracked(dest->path)) {
1424                 char *alt_path = unique_path(o, dest->path, rename_branch);
1425
1426                 output(o, 1, _("Error: Refusing to lose untracked file at %s; "
1427                                "writing to %s instead."),
1428                        dest->path, alt_path);
1429                 /*
1430                  * Write the file in worktree at alt_path, but not in the
1431                  * index.  Instead, write to dest->path for the index but
1432                  * only at the higher appropriate stage.
1433                  */
1434                 if (update_file(o, 0, &dest->oid, dest->mode, alt_path))
1435                         return -1;
1436                 free(alt_path);
1437                 return update_stages(o, dest->path, NULL,
1438                                      rename_branch == o->branch1 ? dest : NULL,
1439                                      rename_branch == o->branch1 ? NULL : dest);
1440         }
1441
1442         /* Update dest->path both in index and in worktree */
1443         if (update_file(o, 1, &dest->oid, dest->mode, dest->path))
1444                 return -1;
1445         return 0;
1446 }
1447
1448 static int handle_change_delete(struct merge_options *o,
1449                                 const char *path, const char *old_path,
1450                                 const struct object_id *o_oid, int o_mode,
1451                                 const struct object_id *changed_oid,
1452                                 int changed_mode,
1453                                 const char *change_branch,
1454                                 const char *delete_branch,
1455                                 const char *change, const char *change_past)
1456 {
1457         char *alt_path = NULL;
1458         const char *update_path = path;
1459         int ret = 0;
1460
1461         if (dir_in_way(path, !o->call_depth, 0) ||
1462             (!o->call_depth && would_lose_untracked(path))) {
1463                 update_path = alt_path = unique_path(o, path, change_branch);
1464         }
1465
1466         if (o->call_depth) {
1467                 /*
1468                  * We cannot arbitrarily accept either a_sha or b_sha as
1469                  * correct; since there is no true "middle point" between
1470                  * them, simply reuse the base version for virtual merge base.
1471                  */
1472                 ret = remove_file_from_cache(path);
1473                 if (!ret)
1474                         ret = update_file(o, 0, o_oid, o_mode, update_path);
1475         } else {
1476                 /*
1477                  * Despite the four nearly duplicate messages and argument
1478                  * lists below and the ugliness of the nested if-statements,
1479                  * having complete messages makes the job easier for
1480                  * translators.
1481                  *
1482                  * The slight variance among the cases is due to the fact
1483                  * that:
1484                  *   1) directory/file conflicts (in effect if
1485                  *      !alt_path) could cause us to need to write the
1486                  *      file to a different path.
1487                  *   2) renames (in effect if !old_path) could mean that
1488                  *      there are two names for the path that the user
1489                  *      may know the file by.
1490                  */
1491                 if (!alt_path) {
1492                         if (!old_path) {
1493                                 output(o, 1, _("CONFLICT (%s/delete): %s deleted in %s "
1494                                        "and %s in %s. Version %s of %s left in tree."),
1495                                        change, path, delete_branch, change_past,
1496                                        change_branch, change_branch, path);
1497                         } else {
1498                                 output(o, 1, _("CONFLICT (%s/delete): %s deleted in %s "
1499                                        "and %s to %s in %s. Version %s of %s left in tree."),
1500                                        change, old_path, delete_branch, change_past, path,
1501                                        change_branch, change_branch, path);
1502                         }
1503                 } else {
1504                         if (!old_path) {
1505                                 output(o, 1, _("CONFLICT (%s/delete): %s deleted in %s "
1506                                        "and %s in %s. Version %s of %s left in tree at %s."),
1507                                        change, path, delete_branch, change_past,
1508                                        change_branch, change_branch, path, alt_path);
1509                         } else {
1510                                 output(o, 1, _("CONFLICT (%s/delete): %s deleted in %s "
1511                                        "and %s to %s in %s. Version %s of %s left in tree at %s."),
1512                                        change, old_path, delete_branch, change_past, path,
1513                                        change_branch, change_branch, path, alt_path);
1514                         }
1515                 }
1516                 /*
1517                  * No need to call update_file() on path when change_branch ==
1518                  * o->branch1 && !alt_path, since that would needlessly touch
1519                  * path.  We could call update_file_flags() with update_cache=0
1520                  * and update_wd=0, but that's a no-op.
1521                  */
1522                 if (change_branch != o->branch1 || alt_path)
1523                         ret = update_file(o, 0, changed_oid, changed_mode, update_path);
1524         }
1525         free(alt_path);
1526
1527         return ret;
1528 }
1529
1530 static int handle_rename_delete(struct merge_options *o,
1531                                 struct diff_filepair *pair,
1532                                 const char *rename_branch,
1533                                 const char *delete_branch)
1534 {
1535         const struct diff_filespec *orig = pair->one;
1536         const struct diff_filespec *dest = pair->two;
1537
1538         if (handle_change_delete(o,
1539                                  o->call_depth ? orig->path : dest->path,
1540                                  o->call_depth ? NULL : orig->path,
1541                                  &orig->oid, orig->mode,
1542                                  &dest->oid, dest->mode,
1543                                  rename_branch, delete_branch,
1544                                  _("rename"), _("renamed")))
1545                 return -1;
1546
1547         if (o->call_depth)
1548                 return remove_file_from_cache(dest->path);
1549         else
1550                 return update_stages(o, dest->path, NULL,
1551                                      rename_branch == o->branch1 ? dest : NULL,
1552                                      rename_branch == o->branch1 ? NULL : dest);
1553 }
1554
1555 static struct diff_filespec *filespec_from_entry(struct diff_filespec *target,
1556                                                  struct stage_data *entry,
1557                                                  int stage)
1558 {
1559         struct object_id *oid = &entry->stages[stage].oid;
1560         unsigned mode = entry->stages[stage].mode;
1561         if (mode == 0 || is_null_oid(oid))
1562                 return NULL;
1563         oidcpy(&target->oid, oid);
1564         target->mode = mode;
1565         return target;
1566 }
1567
1568 static int handle_file_collision(struct merge_options *o,
1569                                  const char *collide_path,
1570                                  const char *prev_path1,
1571                                  const char *prev_path2,
1572                                  const char *branch1, const char *branch2,
1573                                  const struct object_id *a_oid,
1574                                  unsigned int a_mode,
1575                                  const struct object_id *b_oid,
1576                                  unsigned int b_mode)
1577 {
1578         struct merge_file_info mfi;
1579         struct diff_filespec null, a, b;
1580         char *alt_path = NULL;
1581         const char *update_path = collide_path;
1582
1583         /*
1584          * It's easiest to get the correct things into stage 2 and 3, and
1585          * to make sure that the content merge puts HEAD before the other
1586          * branch if we just ensure that branch1 == o->branch1.  So, simply
1587          * flip arguments around if we don't have that.
1588          */
1589         if (branch1 != o->branch1) {
1590                 return handle_file_collision(o, collide_path,
1591                                              prev_path2, prev_path1,
1592                                              branch2, branch1,
1593                                              b_oid, b_mode,
1594                                              a_oid, a_mode);
1595         }
1596
1597         /*
1598          * In the recursive case, we just opt to undo renames
1599          */
1600         if (o->call_depth && (prev_path1 || prev_path2)) {
1601                 /* Put first file (a_oid, a_mode) in its original spot */
1602                 if (prev_path1) {
1603                         if (update_file(o, 1, a_oid, a_mode, prev_path1))
1604                                 return -1;
1605                 } else {
1606                         if (update_file(o, 1, a_oid, a_mode, collide_path))
1607                                 return -1;
1608                 }
1609
1610                 /* Put second file (b_oid, b_mode) in its original spot */
1611                 if (prev_path2) {
1612                         if (update_file(o, 1, b_oid, b_mode, prev_path2))
1613                                 return -1;
1614                 } else {
1615                         if (update_file(o, 1, b_oid, b_mode, collide_path))
1616                                 return -1;
1617                 }
1618
1619                 /* Don't leave something at collision path if unrenaming both */
1620                 if (prev_path1 && prev_path2)
1621                         remove_file(o, 1, collide_path, 0);
1622
1623                 return 0;
1624         }
1625
1626         /* Remove rename sources if rename/add or rename/rename(2to1) */
1627         if (prev_path1)
1628                 remove_file(o, 1, prev_path1,
1629                             o->call_depth || would_lose_untracked(prev_path1));
1630         if (prev_path2)
1631                 remove_file(o, 1, prev_path2,
1632                             o->call_depth || would_lose_untracked(prev_path2));
1633
1634         /*
1635          * Remove the collision path, if it wouldn't cause dirty contents
1636          * or an untracked file to get lost.  We'll either overwrite with
1637          * merged contents, or just write out to differently named files.
1638          */
1639         if (was_dirty(o, collide_path)) {
1640                 output(o, 1, _("Refusing to lose dirty file at %s"),
1641                        collide_path);
1642                 update_path = alt_path = unique_path(o, collide_path, "merged");
1643         } else if (would_lose_untracked(collide_path)) {
1644                 /*
1645                  * Only way we get here is if both renames were from
1646                  * a directory rename AND user had an untracked file
1647                  * at the location where both files end up after the
1648                  * two directory renames.  See testcase 10d of t6043.
1649                  */
1650                 output(o, 1, _("Refusing to lose untracked file at "
1651                                "%s, even though it's in the way."),
1652                        collide_path);
1653                 update_path = alt_path = unique_path(o, collide_path, "merged");
1654         } else {
1655                 /*
1656                  * FIXME: It's possible that the two files are identical
1657                  * and that the current working copy happens to match, in
1658                  * which case we are unnecessarily touching the working
1659                  * tree file.  It's not a likely enough scenario that I
1660                  * want to code up the checks for it and a better fix is
1661                  * available if we restructure how unpack_trees() and
1662                  * merge-recursive interoperate anyway, so punting for
1663                  * now...
1664                  */
1665                 remove_file(o, 0, collide_path, 0);
1666         }
1667
1668         /* Store things in diff_filespecs for functions that need it */
1669         memset(&a, 0, sizeof(struct diff_filespec));
1670         memset(&b, 0, sizeof(struct diff_filespec));
1671         null.path = a.path = b.path = (char *)collide_path;
1672         oidcpy(&null.oid, &null_oid);
1673         null.mode = 0;
1674         oidcpy(&a.oid, a_oid);
1675         a.mode = a_mode;
1676         a.oid_valid = 1;
1677         oidcpy(&b.oid, b_oid);
1678         b.mode = b_mode;
1679         b.oid_valid = 1;
1680
1681         if (merge_mode_and_contents(o, &null, &a, &b, collide_path,
1682                                     branch1, branch2, o->call_depth * 2, &mfi))
1683                 return -1;
1684         mfi.clean &= !alt_path;
1685         if (update_file(o, mfi.clean, &mfi.oid, mfi.mode, update_path))
1686                 return -1;
1687         if (!mfi.clean && !o->call_depth &&
1688             update_stages(o, collide_path, NULL, &a, &b))
1689                 return -1;
1690         free(alt_path);
1691         /*
1692          * FIXME: If both a & b both started with conflicts (only possible
1693          * if they came from a rename/rename(2to1)), but had IDENTICAL
1694          * contents including those conflicts, then in the next line we claim
1695          * it was clean.  If someone cares about this case, we should have the
1696          * caller notify us if we started with conflicts.
1697          */
1698         return mfi.clean;
1699 }
1700
1701 static int handle_rename_add(struct merge_options *o,
1702                              struct rename_conflict_info *ci)
1703 {
1704         /* a was renamed to c, and a separate c was added. */
1705         struct diff_filespec *a = ci->pair1->one;
1706         struct diff_filespec *c = ci->pair1->two;
1707         char *path = c->path;
1708         char *prev_path_desc;
1709         struct merge_file_info mfi;
1710
1711         int other_stage = (ci->branch1 == o->branch1 ? 3 : 2);
1712
1713         output(o, 1, _("CONFLICT (rename/add): "
1714                "Rename %s->%s in %s.  Added %s in %s"),
1715                a->path, c->path, ci->branch1,
1716                c->path, ci->branch2);
1717
1718         prev_path_desc = xstrfmt("version of %s from %s", path, a->path);
1719         if (merge_mode_and_contents(o, a, c, &ci->ren1_other, prev_path_desc,
1720                                     o->branch1, o->branch2,
1721                                     1 + o->call_depth * 2, &mfi))
1722                 return -1;
1723         free(prev_path_desc);
1724
1725         return handle_file_collision(o,
1726                                      c->path, a->path, NULL,
1727                                      ci->branch1, ci->branch2,
1728                                      &mfi.oid, mfi.mode,
1729                                      &ci->dst_entry1->stages[other_stage].oid,
1730                                      ci->dst_entry1->stages[other_stage].mode);
1731 }
1732
1733 static int handle_file(struct merge_options *o,
1734                         struct diff_filespec *rename,
1735                         int stage,
1736                         struct rename_conflict_info *ci)
1737 {
1738         char *dst_name = rename->path;
1739         struct stage_data *dst_entry;
1740         const char *cur_branch, *other_branch;
1741         struct diff_filespec other;
1742         struct diff_filespec *add;
1743         int ret;
1744
1745         if (stage == 2) {
1746                 dst_entry = ci->dst_entry1;
1747                 cur_branch = ci->branch1;
1748                 other_branch = ci->branch2;
1749         } else {
1750                 dst_entry = ci->dst_entry2;
1751                 cur_branch = ci->branch2;
1752                 other_branch = ci->branch1;
1753         }
1754
1755         add = filespec_from_entry(&other, dst_entry, stage ^ 1);
1756         if (add) {
1757                 int ren_src_was_dirty = was_dirty(o, rename->path);
1758                 char *add_name = unique_path(o, rename->path, other_branch);
1759                 if (update_file(o, 0, &add->oid, add->mode, add_name))
1760                         return -1;
1761
1762                 if (ren_src_was_dirty) {
1763                         output(o, 1, _("Refusing to lose dirty file at %s"),
1764                                rename->path);
1765                 }
1766                 /*
1767                  * Because the double negatives somehow keep confusing me...
1768                  *    1) update_wd iff !ren_src_was_dirty.
1769                  *    2) no_wd iff !update_wd
1770                  *    3) so, no_wd == !!ren_src_was_dirty == ren_src_was_dirty
1771                  */
1772                 remove_file(o, 0, rename->path, ren_src_was_dirty);
1773                 dst_name = unique_path(o, rename->path, cur_branch);
1774         } else {
1775                 if (dir_in_way(rename->path, !o->call_depth, 0)) {
1776                         dst_name = unique_path(o, rename->path, cur_branch);
1777                         output(o, 1, _("%s is a directory in %s adding as %s instead"),
1778                                rename->path, other_branch, dst_name);
1779                 } else if (!o->call_depth &&
1780                            would_lose_untracked(rename->path)) {
1781                         dst_name = unique_path(o, rename->path, cur_branch);
1782                         output(o, 1, _("Refusing to lose untracked file at %s; "
1783                                        "adding as %s instead"),
1784                                rename->path, dst_name);
1785                 }
1786         }
1787         if ((ret = update_file(o, 0, &rename->oid, rename->mode, dst_name)))
1788                 ; /* fall through, do allow dst_name to be released */
1789         else if (stage == 2)
1790                 ret = update_stages(o, rename->path, NULL, rename, add);
1791         else
1792                 ret = update_stages(o, rename->path, NULL, add, rename);
1793
1794         if (dst_name != rename->path)
1795                 free(dst_name);
1796
1797         return ret;
1798 }
1799
1800 static int handle_rename_rename_1to2(struct merge_options *o,
1801                                      struct rename_conflict_info *ci)
1802 {
1803         /* One file was renamed in both branches, but to different names. */
1804         struct diff_filespec *one = ci->pair1->one;
1805         struct diff_filespec *a = ci->pair1->two;
1806         struct diff_filespec *b = ci->pair2->two;
1807
1808         output(o, 1, _("CONFLICT (rename/rename): "
1809                "Rename \"%s\"->\"%s\" in branch \"%s\" "
1810                "rename \"%s\"->\"%s\" in \"%s\"%s"),
1811                one->path, a->path, ci->branch1,
1812                one->path, b->path, ci->branch2,
1813                o->call_depth ? _(" (left unresolved)") : "");
1814         if (o->call_depth) {
1815                 struct merge_file_info mfi;
1816                 struct diff_filespec other;
1817                 struct diff_filespec *add;
1818                 if (merge_mode_and_contents(o, one, a, b, one->path,
1819                                             ci->branch1, ci->branch2,
1820                                             o->call_depth * 2, &mfi))
1821                         return -1;
1822
1823                 /*
1824                  * FIXME: For rename/add-source conflicts (if we could detect
1825                  * such), this is wrong.  We should instead find a unique
1826                  * pathname and then either rename the add-source file to that
1827                  * unique path, or use that unique path instead of src here.
1828                  */
1829                 if (update_file(o, 0, &mfi.oid, mfi.mode, one->path))
1830                         return -1;
1831
1832                 /*
1833                  * Above, we put the merged content at the merge-base's
1834                  * path.  Now we usually need to delete both a->path and
1835                  * b->path.  However, the rename on each side of the merge
1836                  * could also be involved in a rename/add conflict.  In
1837                  * such cases, we should keep the added file around,
1838                  * resolving the conflict at that path in its favor.
1839                  */
1840                 add = filespec_from_entry(&other, ci->dst_entry1, 2 ^ 1);
1841                 if (add) {
1842                         if (update_file(o, 0, &add->oid, add->mode, a->path))
1843                                 return -1;
1844                 }
1845                 else
1846                         remove_file_from_cache(a->path);
1847                 add = filespec_from_entry(&other, ci->dst_entry2, 3 ^ 1);
1848                 if (add) {
1849                         if (update_file(o, 0, &add->oid, add->mode, b->path))
1850                                 return -1;
1851                 }
1852                 else
1853                         remove_file_from_cache(b->path);
1854         } else if (handle_file(o, a, 2, ci) || handle_file(o, b, 3, ci))
1855                 return -1;
1856
1857         return 0;
1858 }
1859
1860 static int handle_rename_rename_2to1(struct merge_options *o,
1861                                      struct rename_conflict_info *ci)
1862 {
1863         /* Two files, a & b, were renamed to the same thing, c. */
1864         struct diff_filespec *a = ci->pair1->one;
1865         struct diff_filespec *b = ci->pair2->one;
1866         struct diff_filespec *c1 = ci->pair1->two;
1867         struct diff_filespec *c2 = ci->pair2->two;
1868         char *path = c1->path; /* == c2->path */
1869         char *path_side_1_desc;
1870         char *path_side_2_desc;
1871         struct merge_file_info mfi_c1;
1872         struct merge_file_info mfi_c2;
1873         int ret;
1874
1875         output(o, 1, _("CONFLICT (rename/rename): "
1876                "Rename %s->%s in %s. "
1877                "Rename %s->%s in %s"),
1878                a->path, c1->path, ci->branch1,
1879                b->path, c2->path, ci->branch2);
1880
1881         remove_file(o, 1, a->path, o->call_depth || would_lose_untracked(a->path));
1882         remove_file(o, 1, b->path, o->call_depth || would_lose_untracked(b->path));
1883
1884         path_side_1_desc = xstrfmt("version of %s from %s", path, a->path);
1885         path_side_2_desc = xstrfmt("version of %s from %s", path, b->path);
1886         if (merge_mode_and_contents(o, a, c1, &ci->ren1_other, path_side_1_desc,
1887                                     o->branch1, o->branch2,
1888                                     o->call_depth * 2, &mfi_c1) ||
1889             merge_mode_and_contents(o, b, &ci->ren2_other, c2, path_side_2_desc,
1890                                     o->branch1, o->branch2,
1891                                     o->call_depth * 2, &mfi_c2))
1892                 return -1;
1893         free(path_side_1_desc);
1894         free(path_side_2_desc);
1895
1896         if (o->call_depth) {
1897                 /*
1898                  * If mfi_c1.clean && mfi_c2.clean, then it might make
1899                  * sense to do a two-way merge of those results.  But, I
1900                  * think in all cases, it makes sense to have the virtual
1901                  * merge base just undo the renames; they can be detected
1902                  * again later for the non-recursive merge.
1903                  */
1904                 remove_file(o, 0, path, 0);
1905                 ret = update_file(o, 0, &mfi_c1.oid, mfi_c1.mode, a->path);
1906                 if (!ret)
1907                         ret = update_file(o, 0, &mfi_c2.oid, mfi_c2.mode,
1908                                           b->path);
1909         } else {
1910                 char *new_path1 = unique_path(o, path, ci->branch1);
1911                 char *new_path2 = unique_path(o, path, ci->branch2);
1912                 output(o, 1, _("Renaming %s to %s and %s to %s instead"),
1913                        a->path, new_path1, b->path, new_path2);
1914                 if (was_dirty(o, path))
1915                         output(o, 1, _("Refusing to lose dirty file at %s"),
1916                                path);
1917                 else if (would_lose_untracked(path))
1918                         /*
1919                          * Only way we get here is if both renames were from
1920                          * a directory rename AND user had an untracked file
1921                          * at the location where both files end up after the
1922                          * two directory renames.  See testcase 10d of t6043.
1923                          */
1924                         output(o, 1, _("Refusing to lose untracked file at "
1925                                        "%s, even though it's in the way."),
1926                                path);
1927                 else
1928                         remove_file(o, 0, path, 0);
1929                 ret = update_file(o, 0, &mfi_c1.oid, mfi_c1.mode, new_path1);
1930                 if (!ret)
1931                         ret = update_file(o, 0, &mfi_c2.oid, mfi_c2.mode,
1932                                           new_path2);
1933                 /*
1934                  * unpack_trees() actually populates the index for us for
1935                  * "normal" rename/rename(2to1) situtations so that the
1936                  * correct entries are at the higher stages, which would
1937                  * make the call below to update_stages_for_stage_data
1938                  * unnecessary.  However, if either of the renames came
1939                  * from a directory rename, then unpack_trees() will not
1940                  * have gotten the right data loaded into the index, so we
1941                  * need to do so now.  (While it'd be tempting to move this
1942                  * call to update_stages_for_stage_data() to
1943                  * apply_directory_rename_modifications(), that would break
1944                  * our intermediate calls to would_lose_untracked() since
1945                  * those rely on the current in-memory index.  See also the
1946                  * big "NOTE" in update_stages()).
1947                  */
1948                 if (update_stages_for_stage_data(o, path, ci->dst_entry1))
1949                         ret = -1;
1950
1951                 free(new_path2);
1952                 free(new_path1);
1953         }
1954
1955         return ret;
1956 }
1957
1958 /*
1959  * Get the diff_filepairs changed between o_tree and tree.
1960  */
1961 static struct diff_queue_struct *get_diffpairs(struct merge_options *o,
1962                                                struct tree *o_tree,
1963                                                struct tree *tree)
1964 {
1965         struct diff_queue_struct *ret;
1966         struct diff_options opts;
1967
1968         diff_setup(&opts);
1969         opts.flags.recursive = 1;
1970         opts.flags.rename_empty = 0;
1971         opts.detect_rename = merge_detect_rename(o);
1972         /*
1973          * We do not have logic to handle the detection of copies.  In
1974          * fact, it may not even make sense to add such logic: would we
1975          * really want a change to a base file to be propagated through
1976          * multiple other files by a merge?
1977          */
1978         if (opts.detect_rename > DIFF_DETECT_RENAME)
1979                 opts.detect_rename = DIFF_DETECT_RENAME;
1980         opts.rename_limit = o->merge_rename_limit >= 0 ? o->merge_rename_limit :
1981                             o->diff_rename_limit >= 0 ? o->diff_rename_limit :
1982                             1000;
1983         opts.rename_score = o->rename_score;
1984         opts.show_rename_progress = o->show_rename_progress;
1985         opts.output_format = DIFF_FORMAT_NO_OUTPUT;
1986         diff_setup_done(&opts);
1987         diff_tree_oid(&o_tree->object.oid, &tree->object.oid, "", &opts);
1988         diffcore_std(&opts);
1989         if (opts.needed_rename_limit > o->needed_rename_limit)
1990                 o->needed_rename_limit = opts.needed_rename_limit;
1991
1992         ret = xmalloc(sizeof(*ret));
1993         *ret = diff_queued_diff;
1994
1995         opts.output_format = DIFF_FORMAT_NO_OUTPUT;
1996         diff_queued_diff.nr = 0;
1997         diff_queued_diff.queue = NULL;
1998         diff_flush(&opts);
1999         return ret;
2000 }
2001
2002 static int tree_has_path(struct tree *tree, const char *path)
2003 {
2004         struct object_id hashy;
2005         unsigned int mode_o;
2006
2007         return !get_tree_entry(&tree->object.oid, path,
2008                                &hashy, &mode_o);
2009 }
2010
2011 /*
2012  * Return a new string that replaces the beginning portion (which matches
2013  * entry->dir), with entry->new_dir.  In perl-speak:
2014  *   new_path_name = (old_path =~ s/entry->dir/entry->new_dir/);
2015  * NOTE:
2016  *   Caller must ensure that old_path starts with entry->dir + '/'.
2017  */
2018 static char *apply_dir_rename(struct dir_rename_entry *entry,
2019                               const char *old_path)
2020 {
2021         struct strbuf new_path = STRBUF_INIT;
2022         int oldlen, newlen;
2023
2024         if (entry->non_unique_new_dir)
2025                 return NULL;
2026
2027         oldlen = strlen(entry->dir);
2028         newlen = entry->new_dir.len + (strlen(old_path) - oldlen) + 1;
2029         strbuf_grow(&new_path, newlen);
2030         strbuf_addbuf(&new_path, &entry->new_dir);
2031         strbuf_addstr(&new_path, &old_path[oldlen]);
2032
2033         return strbuf_detach(&new_path, NULL);
2034 }
2035
2036 static void get_renamed_dir_portion(const char *old_path, const char *new_path,
2037                                     char **old_dir, char **new_dir)
2038 {
2039         char *end_of_old, *end_of_new;
2040         int old_len, new_len;
2041
2042         *old_dir = NULL;
2043         *new_dir = NULL;
2044
2045         /*
2046          * For
2047          *    "a/b/c/d/e/foo.c" -> "a/b/some/thing/else/e/foo.c"
2048          * the "e/foo.c" part is the same, we just want to know that
2049          *    "a/b/c/d" was renamed to "a/b/some/thing/else"
2050          * so, for this example, this function returns "a/b/c/d" in
2051          * *old_dir and "a/b/some/thing/else" in *new_dir.
2052          *
2053          * Also, if the basename of the file changed, we don't care.  We
2054          * want to know which portion of the directory, if any, changed.
2055          */
2056         end_of_old = strrchr(old_path, '/');
2057         end_of_new = strrchr(new_path, '/');
2058
2059         if (end_of_old == NULL || end_of_new == NULL)
2060                 return;
2061         while (*--end_of_new == *--end_of_old &&
2062                end_of_old != old_path &&
2063                end_of_new != new_path)
2064                 ; /* Do nothing; all in the while loop */
2065         /*
2066          * We've found the first non-matching character in the directory
2067          * paths.  That means the current directory we were comparing
2068          * represents the rename.  Move end_of_old and end_of_new back
2069          * to the full directory name.
2070          */
2071         if (*end_of_old == '/')
2072                 end_of_old++;
2073         if (*end_of_old != '/')
2074                 end_of_new++;
2075         end_of_old = strchr(end_of_old, '/');
2076         end_of_new = strchr(end_of_new, '/');
2077
2078         /*
2079          * It may have been the case that old_path and new_path were the same
2080          * directory all along.  Don't claim a rename if they're the same.
2081          */
2082         old_len = end_of_old - old_path;
2083         new_len = end_of_new - new_path;
2084
2085         if (old_len != new_len || strncmp(old_path, new_path, old_len)) {
2086                 *old_dir = xstrndup(old_path, old_len);
2087                 *new_dir = xstrndup(new_path, new_len);
2088         }
2089 }
2090
2091 static void remove_hashmap_entries(struct hashmap *dir_renames,
2092                                    struct string_list *items_to_remove)
2093 {
2094         int i;
2095         struct dir_rename_entry *entry;
2096
2097         for (i = 0; i < items_to_remove->nr; i++) {
2098                 entry = items_to_remove->items[i].util;
2099                 hashmap_remove(dir_renames, entry, NULL);
2100         }
2101         string_list_clear(items_to_remove, 0);
2102 }
2103
2104 /*
2105  * See if there is a directory rename for path, and if there are any file
2106  * level conflicts for the renamed location.  If there is a rename and
2107  * there are no conflicts, return the new name.  Otherwise, return NULL.
2108  */
2109 static char *handle_path_level_conflicts(struct merge_options *o,
2110                                          const char *path,
2111                                          struct dir_rename_entry *entry,
2112                                          struct hashmap *collisions,
2113                                          struct tree *tree)
2114 {
2115         char *new_path = NULL;
2116         struct collision_entry *collision_ent;
2117         int clean = 1;
2118         struct strbuf collision_paths = STRBUF_INIT;
2119
2120         /*
2121          * entry has the mapping of old directory name to new directory name
2122          * that we want to apply to path.
2123          */
2124         new_path = apply_dir_rename(entry, path);
2125
2126         if (!new_path) {
2127                 /* This should only happen when entry->non_unique_new_dir set */
2128                 if (!entry->non_unique_new_dir)
2129                         BUG("entry->non_unqiue_dir not set and !new_path");
2130                 output(o, 1, _("CONFLICT (directory rename split): "
2131                                "Unclear where to place %s because directory "
2132                                "%s was renamed to multiple other directories, "
2133                                "with no destination getting a majority of the "
2134                                "files."),
2135                        path, entry->dir);
2136                 clean = 0;
2137                 return NULL;
2138         }
2139
2140         /*
2141          * The caller needs to have ensured that it has pre-populated
2142          * collisions with all paths that map to new_path.  Do a quick check
2143          * to ensure that's the case.
2144          */
2145         collision_ent = collision_find_entry(collisions, new_path);
2146         if (collision_ent == NULL)
2147                 BUG("collision_ent is NULL");
2148
2149         /*
2150          * Check for one-sided add/add/.../add conflicts, i.e.
2151          * where implicit renames from the other side doing
2152          * directory rename(s) can affect this side of history
2153          * to put multiple paths into the same location.  Warn
2154          * and bail on directory renames for such paths.
2155          */
2156         if (collision_ent->reported_already) {
2157                 clean = 0;
2158         } else if (tree_has_path(tree, new_path)) {
2159                 collision_ent->reported_already = 1;
2160                 strbuf_add_separated_string_list(&collision_paths, ", ",
2161                                                  &collision_ent->source_files);
2162                 output(o, 1, _("CONFLICT (implicit dir rename): Existing "
2163                                "file/dir at %s in the way of implicit "
2164                                "directory rename(s) putting the following "
2165                                "path(s) there: %s."),
2166                        new_path, collision_paths.buf);
2167                 clean = 0;
2168         } else if (collision_ent->source_files.nr > 1) {
2169                 collision_ent->reported_already = 1;
2170                 strbuf_add_separated_string_list(&collision_paths, ", ",
2171                                                  &collision_ent->source_files);
2172                 output(o, 1, _("CONFLICT (implicit dir rename): Cannot map "
2173                                "more than one path to %s; implicit directory "
2174                                "renames tried to put these paths there: %s"),
2175                        new_path, collision_paths.buf);
2176                 clean = 0;
2177         }
2178
2179         /* Free memory we no longer need */
2180         strbuf_release(&collision_paths);
2181         if (!clean && new_path) {
2182                 free(new_path);
2183                 return NULL;
2184         }
2185
2186         return new_path;
2187 }
2188
2189 /*
2190  * There are a couple things we want to do at the directory level:
2191  *   1. Check for both sides renaming to the same thing, in order to avoid
2192  *      implicit renaming of files that should be left in place.  (See
2193  *      testcase 6b in t6043 for details.)
2194  *   2. Prune directory renames if there are still files left in the
2195  *      the original directory.  These represent a partial directory rename,
2196  *      i.e. a rename where only some of the files within the directory
2197  *      were renamed elsewhere.  (Technically, this could be done earlier
2198  *      in get_directory_renames(), except that would prevent us from
2199  *      doing the previous check and thus failing testcase 6b.)
2200  *   3. Check for rename/rename(1to2) conflicts (at the directory level).
2201  *      In the future, we could potentially record this info as well and
2202  *      omit reporting rename/rename(1to2) conflicts for each path within
2203  *      the affected directories, thus cleaning up the merge output.
2204  *   NOTE: We do NOT check for rename/rename(2to1) conflicts at the
2205  *         directory level, because merging directories is fine.  If it
2206  *         causes conflicts for files within those merged directories, then
2207  *         that should be detected at the individual path level.
2208  */
2209 static void handle_directory_level_conflicts(struct merge_options *o,
2210                                              struct hashmap *dir_re_head,
2211                                              struct tree *head,
2212                                              struct hashmap *dir_re_merge,
2213                                              struct tree *merge)
2214 {
2215         struct hashmap_iter iter;
2216         struct dir_rename_entry *head_ent;
2217         struct dir_rename_entry *merge_ent;
2218
2219         struct string_list remove_from_head = STRING_LIST_INIT_NODUP;
2220         struct string_list remove_from_merge = STRING_LIST_INIT_NODUP;
2221
2222         hashmap_iter_init(dir_re_head, &iter);
2223         while ((head_ent = hashmap_iter_next(&iter))) {
2224                 merge_ent = dir_rename_find_entry(dir_re_merge, head_ent->dir);
2225                 if (merge_ent &&
2226                     !head_ent->non_unique_new_dir &&
2227                     !merge_ent->non_unique_new_dir &&
2228                     !strbuf_cmp(&head_ent->new_dir, &merge_ent->new_dir)) {
2229                         /* 1. Renamed identically; remove it from both sides */
2230                         string_list_append(&remove_from_head,
2231                                            head_ent->dir)->util = head_ent;
2232                         strbuf_release(&head_ent->new_dir);
2233                         string_list_append(&remove_from_merge,
2234                                            merge_ent->dir)->util = merge_ent;
2235                         strbuf_release(&merge_ent->new_dir);
2236                 } else if (tree_has_path(head, head_ent->dir)) {
2237                         /* 2. This wasn't a directory rename after all */
2238                         string_list_append(&remove_from_head,
2239                                            head_ent->dir)->util = head_ent;
2240                         strbuf_release(&head_ent->new_dir);
2241                 }
2242         }
2243
2244         remove_hashmap_entries(dir_re_head, &remove_from_head);
2245         remove_hashmap_entries(dir_re_merge, &remove_from_merge);
2246
2247         hashmap_iter_init(dir_re_merge, &iter);
2248         while ((merge_ent = hashmap_iter_next(&iter))) {
2249                 head_ent = dir_rename_find_entry(dir_re_head, merge_ent->dir);
2250                 if (tree_has_path(merge, merge_ent->dir)) {
2251                         /* 2. This wasn't a directory rename after all */
2252                         string_list_append(&remove_from_merge,
2253                                            merge_ent->dir)->util = merge_ent;
2254                 } else if (head_ent &&
2255                            !head_ent->non_unique_new_dir &&
2256                            !merge_ent->non_unique_new_dir) {
2257                         /* 3. rename/rename(1to2) */
2258                         /*
2259                          * We can assume it's not rename/rename(1to1) because
2260                          * that was case (1), already checked above.  So we
2261                          * know that head_ent->new_dir and merge_ent->new_dir
2262                          * are different strings.
2263                          */
2264                         output(o, 1, _("CONFLICT (rename/rename): "
2265                                        "Rename directory %s->%s in %s. "
2266                                        "Rename directory %s->%s in %s"),
2267                                head_ent->dir, head_ent->new_dir.buf, o->branch1,
2268                                head_ent->dir, merge_ent->new_dir.buf, o->branch2);
2269                         string_list_append(&remove_from_head,
2270                                            head_ent->dir)->util = head_ent;
2271                         strbuf_release(&head_ent->new_dir);
2272                         string_list_append(&remove_from_merge,
2273                                            merge_ent->dir)->util = merge_ent;
2274                         strbuf_release(&merge_ent->new_dir);
2275                 }
2276         }
2277
2278         remove_hashmap_entries(dir_re_head, &remove_from_head);
2279         remove_hashmap_entries(dir_re_merge, &remove_from_merge);
2280 }
2281
2282 static struct hashmap *get_directory_renames(struct diff_queue_struct *pairs,
2283                                              struct tree *tree)
2284 {
2285         struct hashmap *dir_renames;
2286         struct hashmap_iter iter;
2287         struct dir_rename_entry *entry;
2288         int i;
2289
2290         /*
2291          * Typically, we think of a directory rename as all files from a
2292          * certain directory being moved to a target directory.  However,
2293          * what if someone first moved two files from the original
2294          * directory in one commit, and then renamed the directory
2295          * somewhere else in a later commit?  At merge time, we just know
2296          * that files from the original directory went to two different
2297          * places, and that the bulk of them ended up in the same place.
2298          * We want each directory rename to represent where the bulk of the
2299          * files from that directory end up; this function exists to find
2300          * where the bulk of the files went.
2301          *
2302          * The first loop below simply iterates through the list of file
2303          * renames, finding out how often each directory rename pair
2304          * possibility occurs.
2305          */
2306         dir_renames = xmalloc(sizeof(*dir_renames));
2307         dir_rename_init(dir_renames);
2308         for (i = 0; i < pairs->nr; ++i) {
2309                 struct string_list_item *item;
2310                 int *count;
2311                 struct diff_filepair *pair = pairs->queue[i];
2312                 char *old_dir, *new_dir;
2313
2314                 /* File not part of directory rename if it wasn't renamed */
2315                 if (pair->status != 'R')
2316                         continue;
2317
2318                 get_renamed_dir_portion(pair->one->path, pair->two->path,
2319                                         &old_dir,        &new_dir);
2320                 if (!old_dir)
2321                         /* Directory didn't change at all; ignore this one. */
2322                         continue;
2323
2324                 entry = dir_rename_find_entry(dir_renames, old_dir);
2325                 if (!entry) {
2326                         entry = xmalloc(sizeof(*entry));
2327                         dir_rename_entry_init(entry, old_dir);
2328                         hashmap_put(dir_renames, entry);
2329                 } else {
2330                         free(old_dir);
2331                 }
2332                 item = string_list_lookup(&entry->possible_new_dirs, new_dir);
2333                 if (!item) {
2334                         item = string_list_insert(&entry->possible_new_dirs,
2335                                                   new_dir);
2336                         item->util = xcalloc(1, sizeof(int));
2337                 } else {
2338                         free(new_dir);
2339                 }
2340                 count = item->util;
2341                 *count += 1;
2342         }
2343
2344         /*
2345          * For each directory with files moved out of it, we find out which
2346          * target directory received the most files so we can declare it to
2347          * be the "winning" target location for the directory rename.  This
2348          * winner gets recorded in new_dir.  If there is no winner
2349          * (multiple target directories received the same number of files),
2350          * we set non_unique_new_dir.  Once we've determined the winner (or
2351          * that there is no winner), we no longer need possible_new_dirs.
2352          */
2353         hashmap_iter_init(dir_renames, &iter);
2354         while ((entry = hashmap_iter_next(&iter))) {
2355                 int max = 0;
2356                 int bad_max = 0;
2357                 char *best = NULL;
2358
2359                 for (i = 0; i < entry->possible_new_dirs.nr; i++) {
2360                         int *count = entry->possible_new_dirs.items[i].util;
2361
2362                         if (*count == max)
2363                                 bad_max = max;
2364                         else if (*count > max) {
2365                                 max = *count;
2366                                 best = entry->possible_new_dirs.items[i].string;
2367                         }
2368                 }
2369                 if (bad_max == max)
2370                         entry->non_unique_new_dir = 1;
2371                 else {
2372                         assert(entry->new_dir.len == 0);
2373                         strbuf_addstr(&entry->new_dir, best);
2374                 }
2375                 /*
2376                  * The relevant directory sub-portion of the original full
2377                  * filepaths were xstrndup'ed before inserting into
2378                  * possible_new_dirs, and instead of manually iterating the
2379                  * list and free'ing each, just lie and tell
2380                  * possible_new_dirs that it did the strdup'ing so that it
2381                  * will free them for us.
2382                  */
2383                 entry->possible_new_dirs.strdup_strings = 1;
2384                 string_list_clear(&entry->possible_new_dirs, 1);
2385         }
2386
2387         return dir_renames;
2388 }
2389
2390 static struct dir_rename_entry *check_dir_renamed(const char *path,
2391                                                   struct hashmap *dir_renames)
2392 {
2393         char *temp = xstrdup(path);
2394         char *end;
2395         struct dir_rename_entry *entry = NULL;;
2396
2397         while ((end = strrchr(temp, '/'))) {
2398                 *end = '\0';
2399                 entry = dir_rename_find_entry(dir_renames, temp);
2400                 if (entry)
2401                         break;
2402         }
2403         free(temp);
2404         return entry;
2405 }
2406
2407 static void compute_collisions(struct hashmap *collisions,
2408                                struct hashmap *dir_renames,
2409                                struct diff_queue_struct *pairs)
2410 {
2411         int i;
2412
2413         /*
2414          * Multiple files can be mapped to the same path due to directory
2415          * renames done by the other side of history.  Since that other
2416          * side of history could have merged multiple directories into one,
2417          * if our side of history added the same file basename to each of
2418          * those directories, then all N of them would get implicitly
2419          * renamed by the directory rename detection into the same path,
2420          * and we'd get an add/add/.../add conflict, and all those adds
2421          * from *this* side of history.  This is not representable in the
2422          * index, and users aren't going to easily be able to make sense of
2423          * it.  So we need to provide a good warning about what's
2424          * happening, and fall back to no-directory-rename detection
2425          * behavior for those paths.
2426          *
2427          * See testcases 9e and all of section 5 from t6043 for examples.
2428          */
2429         collision_init(collisions);
2430
2431         for (i = 0; i < pairs->nr; ++i) {
2432                 struct dir_rename_entry *dir_rename_ent;
2433                 struct collision_entry *collision_ent;
2434                 char *new_path;
2435                 struct diff_filepair *pair = pairs->queue[i];
2436
2437                 if (pair->status != 'A' && pair->status != 'R')
2438                         continue;
2439                 dir_rename_ent = check_dir_renamed(pair->two->path,
2440                                                    dir_renames);
2441                 if (!dir_rename_ent)
2442                         continue;
2443
2444                 new_path = apply_dir_rename(dir_rename_ent, pair->two->path);
2445                 if (!new_path)
2446                         /*
2447                          * dir_rename_ent->non_unique_new_path is true, which
2448                          * means there is no directory rename for us to use,
2449                          * which means it won't cause us any additional
2450                          * collisions.
2451                          */
2452                         continue;
2453                 collision_ent = collision_find_entry(collisions, new_path);
2454                 if (!collision_ent) {
2455                         collision_ent = xcalloc(1,
2456                                                 sizeof(struct collision_entry));
2457                         hashmap_entry_init(collision_ent, strhash(new_path));
2458                         hashmap_put(collisions, collision_ent);
2459                         collision_ent->target_file = new_path;
2460                 } else {
2461                         free(new_path);
2462                 }
2463                 string_list_insert(&collision_ent->source_files,
2464                                    pair->two->path);
2465         }
2466 }
2467
2468 static char *check_for_directory_rename(struct merge_options *o,
2469                                         const char *path,
2470                                         struct tree *tree,
2471                                         struct hashmap *dir_renames,
2472                                         struct hashmap *dir_rename_exclusions,
2473                                         struct hashmap *collisions,
2474                                         int *clean_merge)
2475 {
2476         char *new_path = NULL;
2477         struct dir_rename_entry *entry = check_dir_renamed(path, dir_renames);
2478         struct dir_rename_entry *oentry = NULL;
2479
2480         if (!entry)
2481                 return new_path;
2482
2483         /*
2484          * This next part is a little weird.  We do not want to do an
2485          * implicit rename into a directory we renamed on our side, because
2486          * that will result in a spurious rename/rename(1to2) conflict.  An
2487          * example:
2488          *   Base commit: dumbdir/afile, otherdir/bfile
2489          *   Side 1:      smrtdir/afile, otherdir/bfile
2490          *   Side 2:      dumbdir/afile, dumbdir/bfile
2491          * Here, while working on Side 1, we could notice that otherdir was
2492          * renamed/merged to dumbdir, and change the diff_filepair for
2493          * otherdir/bfile into a rename into dumbdir/bfile.  However, Side
2494          * 2 will notice the rename from dumbdir to smrtdir, and do the
2495          * transitive rename to move it from dumbdir/bfile to
2496          * smrtdir/bfile.  That gives us bfile in dumbdir vs being in
2497          * smrtdir, a rename/rename(1to2) conflict.  We really just want
2498          * the file to end up in smrtdir.  And the way to achieve that is
2499          * to not let Side1 do the rename to dumbdir, since we know that is
2500          * the source of one of our directory renames.
2501          *
2502          * That's why oentry and dir_rename_exclusions is here.
2503          *
2504          * As it turns out, this also prevents N-way transient rename
2505          * confusion; See testcases 9c and 9d of t6043.
2506          */
2507         oentry = dir_rename_find_entry(dir_rename_exclusions, entry->new_dir.buf);
2508         if (oentry) {
2509                 output(o, 1, _("WARNING: Avoiding applying %s -> %s rename "
2510                                "to %s, because %s itself was renamed."),
2511                        entry->dir, entry->new_dir.buf, path, entry->new_dir.buf);
2512         } else {
2513                 new_path = handle_path_level_conflicts(o, path, entry,
2514                                                        collisions, tree);
2515                 *clean_merge &= (new_path != NULL);
2516         }
2517
2518         return new_path;
2519 }
2520
2521 static void apply_directory_rename_modifications(struct merge_options *o,
2522                                                  struct diff_filepair *pair,
2523                                                  char *new_path,
2524                                                  struct rename *re,
2525                                                  struct tree *tree,
2526                                                  struct tree *o_tree,
2527                                                  struct tree *a_tree,
2528                                                  struct tree *b_tree,
2529                                                  struct string_list *entries,
2530                                                  int *clean)
2531 {
2532         struct string_list_item *item;
2533         int stage = (tree == a_tree ? 2 : 3);
2534         int update_wd;
2535
2536         /*
2537          * In all cases where we can do directory rename detection,
2538          * unpack_trees() will have read pair->two->path into the
2539          * index and the working copy.  We need to remove it so that
2540          * we can instead place it at new_path.  It is guaranteed to
2541          * not be untracked (unpack_trees() would have errored out
2542          * saying the file would have been overwritten), but it might
2543          * be dirty, though.
2544          */
2545         update_wd = !was_dirty(o, pair->two->path);
2546         if (!update_wd)
2547                 output(o, 1, _("Refusing to lose dirty file at %s"),
2548                        pair->two->path);
2549         remove_file(o, 1, pair->two->path, !update_wd);
2550
2551         /* Find or create a new re->dst_entry */
2552         item = string_list_lookup(entries, new_path);
2553         if (item) {
2554                 /*
2555                  * Since we're renaming on this side of history, and it's
2556                  * due to a directory rename on the other side of history
2557                  * (which we only allow when the directory in question no
2558                  * longer exists on the other side of history), the
2559                  * original entry for re->dst_entry is no longer
2560                  * necessary...
2561                  */
2562                 re->dst_entry->processed = 1;
2563
2564                 /*
2565                  * ...because we'll be using this new one.
2566                  */
2567                 re->dst_entry = item->util;
2568         } else {
2569                 /*
2570                  * re->dst_entry is for the before-dir-rename path, and we
2571                  * need it to hold information for the after-dir-rename
2572                  * path.  Before creating a new entry, we need to mark the
2573                  * old one as unnecessary (...unless it is shared by
2574                  * src_entry, i.e. this didn't use to be a rename, in which
2575                  * case we can just allow the normal processing to happen
2576                  * for it).
2577                  */
2578                 if (pair->status == 'R')
2579                         re->dst_entry->processed = 1;
2580
2581                 re->dst_entry = insert_stage_data(new_path,
2582                                                   o_tree, a_tree, b_tree,
2583                                                   entries);
2584                 item = string_list_insert(entries, new_path);
2585                 item->util = re->dst_entry;
2586         }
2587
2588         /*
2589          * Update the stage_data with the information about the path we are
2590          * moving into place.  That slot will be empty and available for us
2591          * to write to because of the collision checks in
2592          * handle_path_level_conflicts().  In other words,
2593          * re->dst_entry->stages[stage].oid will be the null_oid, so it's
2594          * open for us to write to.
2595          *
2596          * It may be tempting to actually update the index at this point as
2597          * well, using update_stages_for_stage_data(), but as per the big
2598          * "NOTE" in update_stages(), doing so will modify the current
2599          * in-memory index which will break calls to would_lose_untracked()
2600          * that we need to make.  Instead, we need to just make sure that
2601          * the various handle_rename_*() functions update the index
2602          * explicitly rather than relying on unpack_trees() to have done it.
2603          */
2604         get_tree_entry(&tree->object.oid,
2605                        pair->two->path,
2606                        &re->dst_entry->stages[stage].oid,
2607                        &re->dst_entry->stages[stage].mode);
2608
2609         /* Update pair status */
2610         if (pair->status == 'A') {
2611                 /*
2612                  * Recording rename information for this add makes it look
2613                  * like a rename/delete conflict.  Make sure we can
2614                  * correctly handle this as an add that was moved to a new
2615                  * directory instead of reporting a rename/delete conflict.
2616                  */
2617                 re->add_turned_into_rename = 1;
2618         }
2619         /*
2620          * We don't actually look at pair->status again, but it seems
2621          * pedagogically correct to adjust it.
2622          */
2623         pair->status = 'R';
2624
2625         /*
2626          * Finally, record the new location.
2627          */
2628         pair->two->path = new_path;
2629 }
2630
2631 /*
2632  * Get information of all renames which occurred in 'pairs', making use of
2633  * any implicit directory renames inferred from the other side of history.
2634  * We need the three trees in the merge ('o_tree', 'a_tree' and 'b_tree')
2635  * to be able to associate the correct cache entries with the rename
2636  * information; tree is always equal to either a_tree or b_tree.
2637  */
2638 static struct string_list *get_renames(struct merge_options *o,
2639                                        struct diff_queue_struct *pairs,
2640                                        struct hashmap *dir_renames,
2641                                        struct hashmap *dir_rename_exclusions,
2642                                        struct tree *tree,
2643                                        struct tree *o_tree,
2644                                        struct tree *a_tree,
2645                                        struct tree *b_tree,
2646                                        struct string_list *entries,
2647                                        int *clean_merge)
2648 {
2649         int i;
2650         struct hashmap collisions;
2651         struct hashmap_iter iter;
2652         struct collision_entry *e;
2653         struct string_list *renames;
2654
2655         compute_collisions(&collisions, dir_renames, pairs);
2656         renames = xcalloc(1, sizeof(struct string_list));
2657
2658         for (i = 0; i < pairs->nr; ++i) {
2659                 struct string_list_item *item;
2660                 struct rename *re;
2661                 struct diff_filepair *pair = pairs->queue[i];
2662                 char *new_path; /* non-NULL only with directory renames */
2663
2664                 if (pair->status != 'A' && pair->status != 'R') {
2665                         diff_free_filepair(pair);
2666                         continue;
2667                 }
2668                 new_path = check_for_directory_rename(o, pair->two->path, tree,
2669                                                       dir_renames,
2670                                                       dir_rename_exclusions,
2671                                                       &collisions,
2672                                                       clean_merge);
2673                 if (pair->status != 'R' && !new_path) {
2674                         diff_free_filepair(pair);
2675                         continue;
2676                 }
2677
2678                 re = xmalloc(sizeof(*re));
2679                 re->processed = 0;
2680                 re->add_turned_into_rename = 0;
2681                 re->pair = pair;
2682                 item = string_list_lookup(entries, re->pair->one->path);
2683                 if (!item)
2684                         re->src_entry = insert_stage_data(re->pair->one->path,
2685                                         o_tree, a_tree, b_tree, entries);
2686                 else
2687                         re->src_entry = item->util;
2688
2689                 item = string_list_lookup(entries, re->pair->two->path);
2690                 if (!item)
2691                         re->dst_entry = insert_stage_data(re->pair->two->path,
2692                                         o_tree, a_tree, b_tree, entries);
2693                 else
2694                         re->dst_entry = item->util;
2695                 item = string_list_insert(renames, pair->one->path);
2696                 item->util = re;
2697                 if (new_path)
2698                         apply_directory_rename_modifications(o, pair, new_path,
2699                                                              re, tree, o_tree,
2700                                                              a_tree, b_tree,
2701                                                              entries,
2702                                                              clean_merge);
2703         }
2704
2705         hashmap_iter_init(&collisions, &iter);
2706         while ((e = hashmap_iter_next(&iter))) {
2707                 free(e->target_file);
2708                 string_list_clear(&e->source_files, 0);
2709         }
2710         hashmap_free(&collisions, 1);
2711         return renames;
2712 }
2713
2714 static int process_renames(struct merge_options *o,
2715                            struct string_list *a_renames,
2716                            struct string_list *b_renames)
2717 {
2718         int clean_merge = 1, i, j;
2719         struct string_list a_by_dst = STRING_LIST_INIT_NODUP;
2720         struct string_list b_by_dst = STRING_LIST_INIT_NODUP;
2721         const struct rename *sre;
2722
2723         for (i = 0; i < a_renames->nr; i++) {
2724                 sre = a_renames->items[i].util;
2725                 string_list_insert(&a_by_dst, sre->pair->two->path)->util
2726                         = (void *)sre;
2727         }
2728         for (i = 0; i < b_renames->nr; i++) {
2729                 sre = b_renames->items[i].util;
2730                 string_list_insert(&b_by_dst, sre->pair->two->path)->util
2731                         = (void *)sre;
2732         }
2733
2734         for (i = 0, j = 0; i < a_renames->nr || j < b_renames->nr;) {
2735                 struct string_list *renames1, *renames2Dst;
2736                 struct rename *ren1 = NULL, *ren2 = NULL;
2737                 const char *branch1, *branch2;
2738                 const char *ren1_src, *ren1_dst;
2739                 struct string_list_item *lookup;
2740
2741                 if (i >= a_renames->nr) {
2742                         ren2 = b_renames->items[j++].util;
2743                 } else if (j >= b_renames->nr) {
2744                         ren1 = a_renames->items[i++].util;
2745                 } else {
2746                         int compare = strcmp(a_renames->items[i].string,
2747                                              b_renames->items[j].string);
2748                         if (compare <= 0)
2749                                 ren1 = a_renames->items[i++].util;
2750                         if (compare >= 0)
2751                                 ren2 = b_renames->items[j++].util;
2752                 }
2753
2754                 /* TODO: refactor, so that 1/2 are not needed */
2755                 if (ren1) {
2756                         renames1 = a_renames;
2757                         renames2Dst = &b_by_dst;
2758                         branch1 = o->branch1;
2759                         branch2 = o->branch2;
2760                 } else {
2761                         renames1 = b_renames;
2762                         renames2Dst = &a_by_dst;
2763                         branch1 = o->branch2;
2764                         branch2 = o->branch1;
2765                         SWAP(ren2, ren1);
2766                 }
2767
2768                 if (ren1->processed)
2769                         continue;
2770                 ren1->processed = 1;
2771                 ren1->dst_entry->processed = 1;
2772                 /* BUG: We should only mark src_entry as processed if we
2773                  * are not dealing with a rename + add-source case.
2774                  */
2775                 ren1->src_entry->processed = 1;
2776
2777                 ren1_src = ren1->pair->one->path;
2778                 ren1_dst = ren1->pair->two->path;
2779
2780                 if (ren2) {
2781                         /* One file renamed on both sides */
2782                         const char *ren2_src = ren2->pair->one->path;
2783                         const char *ren2_dst = ren2->pair->two->path;
2784                         enum rename_type rename_type;
2785                         if (strcmp(ren1_src, ren2_src) != 0)
2786                                 BUG("ren1_src != ren2_src");
2787                         ren2->dst_entry->processed = 1;
2788                         ren2->processed = 1;
2789                         if (strcmp(ren1_dst, ren2_dst) != 0) {
2790                                 rename_type = RENAME_ONE_FILE_TO_TWO;
2791                                 clean_merge = 0;
2792                         } else {
2793                                 rename_type = RENAME_ONE_FILE_TO_ONE;
2794                                 /* BUG: We should only remove ren1_src in
2795                                  * the base stage (think of rename +
2796                                  * add-source cases).
2797                                  */
2798                                 remove_file(o, 1, ren1_src, 1);
2799                                 update_entry(ren1->dst_entry,
2800                                              ren1->pair->one,
2801                                              ren1->pair->two,
2802                                              ren2->pair->two);
2803                         }
2804                         setup_rename_conflict_info(rename_type,
2805                                                    ren1->pair,
2806                                                    ren2->pair,
2807                                                    branch1,
2808                                                    branch2,
2809                                                    ren1->dst_entry,
2810                                                    ren2->dst_entry,
2811                                                    o,
2812                                                    NULL,
2813                                                    NULL);
2814                 } else if ((lookup = string_list_lookup(renames2Dst, ren1_dst))) {
2815                         /* Two different files renamed to the same thing */
2816                         char *ren2_dst;
2817                         ren2 = lookup->util;
2818                         ren2_dst = ren2->pair->two->path;
2819                         if (strcmp(ren1_dst, ren2_dst) != 0)
2820                                 BUG("ren1_dst != ren2_dst");
2821
2822                         clean_merge = 0;
2823                         ren2->processed = 1;
2824                         /*
2825                          * BUG: We should only mark src_entry as processed
2826                          * if we are not dealing with a rename + add-source
2827                          * case.
2828                          */
2829                         ren2->src_entry->processed = 1;
2830
2831                         setup_rename_conflict_info(RENAME_TWO_FILES_TO_ONE,
2832                                                    ren1->pair,
2833                                                    ren2->pair,
2834                                                    branch1,
2835                                                    branch2,
2836                                                    ren1->dst_entry,
2837                                                    ren2->dst_entry,
2838                                                    o,
2839                                                    ren1->src_entry,
2840                                                    ren2->src_entry);
2841
2842                 } else {
2843                         /* Renamed in 1, maybe changed in 2 */
2844                         /* we only use sha1 and mode of these */
2845                         struct diff_filespec src_other, dst_other;
2846                         int try_merge;
2847
2848                         /*
2849                          * unpack_trees loads entries from common-commit
2850                          * into stage 1, from head-commit into stage 2, and
2851                          * from merge-commit into stage 3.  We keep track
2852                          * of which side corresponds to the rename.
2853                          */
2854                         int renamed_stage = a_renames == renames1 ? 2 : 3;
2855                         int other_stage =   a_renames == renames1 ? 3 : 2;
2856
2857                         /* BUG: We should only remove ren1_src in the base
2858                          * stage and in other_stage (think of rename +
2859                          * add-source case).
2860                          */
2861                         remove_file(o, 1, ren1_src,
2862                                     renamed_stage == 2 || !was_tracked(o, ren1_src));
2863
2864                         oidcpy(&src_other.oid,
2865                                &ren1->src_entry->stages[other_stage].oid);
2866                         src_other.mode = ren1->src_entry->stages[other_stage].mode;
2867                         oidcpy(&dst_other.oid,
2868                                &ren1->dst_entry->stages[other_stage].oid);
2869                         dst_other.mode = ren1->dst_entry->stages[other_stage].mode;
2870                         try_merge = 0;
2871
2872                         if (oid_eq(&src_other.oid, &null_oid) &&
2873                             ren1->add_turned_into_rename) {
2874                                 setup_rename_conflict_info(RENAME_VIA_DIR,
2875                                                            ren1->pair,
2876                                                            NULL,
2877                                                            branch1,
2878                                                            branch2,
2879                                                            ren1->dst_entry,
2880                                                            NULL,
2881                                                            o,
2882                                                            NULL,
2883                                                            NULL);
2884                         } else if (oid_eq(&src_other.oid, &null_oid)) {
2885                                 setup_rename_conflict_info(RENAME_DELETE,
2886                                                            ren1->pair,
2887                                                            NULL,
2888                                                            branch1,
2889                                                            branch2,
2890                                                            ren1->dst_entry,
2891                                                            NULL,
2892                                                            o,
2893                                                            NULL,
2894                                                            NULL);
2895                         } else if ((dst_other.mode == ren1->pair->two->mode) &&
2896                                    oid_eq(&dst_other.oid, &ren1->pair->two->oid)) {
2897                                 /*
2898                                  * Added file on the other side identical to
2899                                  * the file being renamed: clean merge.
2900                                  * Also, there is no need to overwrite the
2901                                  * file already in the working copy, so call
2902                                  * update_file_flags() instead of
2903                                  * update_file().
2904                                  */
2905                                 if (update_file_flags(o,
2906                                                       &ren1->pair->two->oid,
2907                                                       ren1->pair->two->mode,
2908                                                       ren1_dst,
2909                                                       1, /* update_cache */
2910                                                       0  /* update_wd    */))
2911                                         clean_merge = -1;
2912                         } else if (!oid_eq(&dst_other.oid, &null_oid)) {
2913                                 /*
2914                                  * Probably not a clean merge, but it's
2915                                  * premature to set clean_merge to 0 here,
2916                                  * because if the rename merges cleanly and
2917                                  * the merge exactly matches the newly added
2918                                  * file, then the merge will be clean.
2919                                  */
2920                                 setup_rename_conflict_info(RENAME_ADD,
2921                                                            ren1->pair,
2922                                                            NULL,
2923                                                            branch1,
2924                                                            branch2,
2925                                                            ren1->dst_entry,
2926                                                            NULL,
2927                                                            o,
2928                                                            ren1->src_entry,
2929                                                            NULL);
2930                         } else
2931                                 try_merge = 1;
2932
2933                         if (clean_merge < 0)
2934                                 goto cleanup_and_return;
2935                         if (try_merge) {
2936                                 struct diff_filespec *one, *a, *b;
2937                                 src_other.path = (char *)ren1_src;
2938
2939                                 one = ren1->pair->one;
2940                                 if (a_renames == renames1) {
2941                                         a = ren1->pair->two;
2942                                         b = &src_other;
2943                                 } else {
2944                                         b = ren1->pair->two;
2945                                         a = &src_other;
2946                                 }
2947                                 update_entry(ren1->dst_entry, one, a, b);
2948                                 setup_rename_conflict_info(RENAME_NORMAL,
2949                                                            ren1->pair,
2950                                                            NULL,
2951                                                            branch1,
2952                                                            NULL,
2953                                                            ren1->dst_entry,
2954                                                            NULL,
2955                                                            o,
2956                                                            NULL,
2957                                                            NULL);
2958                         }
2959                 }
2960         }
2961 cleanup_and_return:
2962         string_list_clear(&a_by_dst, 0);
2963         string_list_clear(&b_by_dst, 0);
2964
2965         return clean_merge;
2966 }
2967
2968 struct rename_info {
2969         struct string_list *head_renames;
2970         struct string_list *merge_renames;
2971 };
2972
2973 static void initial_cleanup_rename(struct diff_queue_struct *pairs,
2974                                    struct hashmap *dir_renames)
2975 {
2976         struct hashmap_iter iter;
2977         struct dir_rename_entry *e;
2978
2979         hashmap_iter_init(dir_renames, &iter);
2980         while ((e = hashmap_iter_next(&iter))) {
2981                 free(e->dir);
2982                 strbuf_release(&e->new_dir);
2983                 /* possible_new_dirs already cleared in get_directory_renames */
2984         }
2985         hashmap_free(dir_renames, 1);
2986         free(dir_renames);
2987
2988         free(pairs->queue);
2989         free(pairs);
2990 }
2991
2992 static int detect_and_process_renames(struct merge_options *o,
2993                                       struct tree *common,
2994                                       struct tree *head,
2995                                       struct tree *merge,
2996                                       struct string_list *entries,
2997                                       struct rename_info *ri)
2998 {
2999         struct diff_queue_struct *head_pairs, *merge_pairs;
3000         struct hashmap *dir_re_head, *dir_re_merge;
3001         int clean = 1;
3002
3003         ri->head_renames = NULL;
3004         ri->merge_renames = NULL;
3005
3006         if (!merge_detect_rename(o))
3007                 return 1;
3008
3009         head_pairs = get_diffpairs(o, common, head);
3010         merge_pairs = get_diffpairs(o, common, merge);
3011
3012         if (o->detect_directory_renames) {
3013                 dir_re_head = get_directory_renames(head_pairs, head);
3014                 dir_re_merge = get_directory_renames(merge_pairs, merge);
3015
3016                 handle_directory_level_conflicts(o,
3017                                                  dir_re_head, head,
3018                                                  dir_re_merge, merge);
3019         } else {
3020                 dir_re_head  = xmalloc(sizeof(*dir_re_head));
3021                 dir_re_merge = xmalloc(sizeof(*dir_re_merge));
3022                 dir_rename_init(dir_re_head);
3023                 dir_rename_init(dir_re_merge);
3024         }
3025
3026         ri->head_renames  = get_renames(o, head_pairs,
3027                                         dir_re_merge, dir_re_head, head,
3028                                         common, head, merge, entries,
3029                                         &clean);
3030         if (clean < 0)
3031                 goto cleanup;
3032         ri->merge_renames = get_renames(o, merge_pairs,
3033                                         dir_re_head, dir_re_merge, merge,
3034                                         common, head, merge, entries,
3035                                         &clean);
3036         if (clean < 0)
3037                 goto cleanup;
3038         clean &= process_renames(o, ri->head_renames, ri->merge_renames);
3039
3040 cleanup:
3041         /*
3042          * Some cleanup is deferred until cleanup_renames() because the
3043          * data structures are still needed and referenced in
3044          * process_entry().  But there are a few things we can free now.
3045          */
3046         initial_cleanup_rename(head_pairs, dir_re_head);
3047         initial_cleanup_rename(merge_pairs, dir_re_merge);
3048
3049         return clean;
3050 }
3051
3052 static void final_cleanup_rename(struct string_list *rename)
3053 {
3054         const struct rename *re;
3055         int i;
3056
3057         if (rename == NULL)
3058                 return;
3059
3060         for (i = 0; i < rename->nr; i++) {
3061                 re = rename->items[i].util;
3062                 diff_free_filepair(re->pair);
3063         }
3064         string_list_clear(rename, 1);
3065         free(rename);
3066 }
3067
3068 static void final_cleanup_renames(struct rename_info *re_info)
3069 {
3070         final_cleanup_rename(re_info->head_renames);
3071         final_cleanup_rename(re_info->merge_renames);
3072 }
3073
3074 static struct object_id *stage_oid(const struct object_id *oid, unsigned mode)
3075 {
3076         return (is_null_oid(oid) || mode == 0) ? NULL: (struct object_id *)oid;
3077 }
3078
3079 static int read_oid_strbuf(struct merge_options *o,
3080                            const struct object_id *oid,
3081                            struct strbuf *dst)
3082 {
3083         void *buf;
3084         enum object_type type;
3085         unsigned long size;
3086         buf = read_object_file(oid, &type, &size);
3087         if (!buf)
3088                 return err(o, _("cannot read object %s"), oid_to_hex(oid));
3089         if (type != OBJ_BLOB) {
3090                 free(buf);
3091                 return err(o, _("object %s is not a blob"), oid_to_hex(oid));
3092         }
3093         strbuf_attach(dst, buf, size, size + 1);
3094         return 0;
3095 }
3096
3097 static int blob_unchanged(struct merge_options *opt,
3098                           const struct object_id *o_oid,
3099                           unsigned o_mode,
3100                           const struct object_id *a_oid,
3101                           unsigned a_mode,
3102                           int renormalize, const char *path)
3103 {
3104         struct strbuf o = STRBUF_INIT;
3105         struct strbuf a = STRBUF_INIT;
3106         int ret = 0; /* assume changed for safety */
3107
3108         if (a_mode != o_mode)
3109                 return 0;
3110         if (oid_eq(o_oid, a_oid))
3111                 return 1;
3112         if (!renormalize)
3113                 return 0;
3114
3115         assert(o_oid && a_oid);
3116         if (read_oid_strbuf(opt, o_oid, &o) || read_oid_strbuf(opt, a_oid, &a))
3117                 goto error_return;
3118         /*
3119          * Note: binary | is used so that both renormalizations are
3120          * performed.  Comparison can be skipped if both files are
3121          * unchanged since their sha1s have already been compared.
3122          */
3123         if (renormalize_buffer(&the_index, path, o.buf, o.len, &o) |
3124             renormalize_buffer(&the_index, path, a.buf, a.len, &a))
3125                 ret = (o.len == a.len && !memcmp(o.buf, a.buf, o.len));
3126
3127 error_return:
3128         strbuf_release(&o);
3129         strbuf_release(&a);
3130         return ret;
3131 }
3132
3133 static int handle_modify_delete(struct merge_options *o,
3134                                 const char *path,
3135                                 struct object_id *o_oid, int o_mode,
3136                                 struct object_id *a_oid, int a_mode,
3137                                 struct object_id *b_oid, int b_mode)
3138 {
3139         const char *modify_branch, *delete_branch;
3140         struct object_id *changed_oid;
3141         int changed_mode;
3142
3143         if (a_oid) {
3144                 modify_branch = o->branch1;
3145                 delete_branch = o->branch2;
3146                 changed_oid = a_oid;
3147                 changed_mode = a_mode;
3148         } else {
3149                 modify_branch = o->branch2;
3150                 delete_branch = o->branch1;
3151                 changed_oid = b_oid;
3152                 changed_mode = b_mode;
3153         }
3154
3155         return handle_change_delete(o,
3156                                     path, NULL,
3157                                     o_oid, o_mode,
3158                                     changed_oid, changed_mode,
3159                                     modify_branch, delete_branch,
3160                                     _("modify"), _("modified"));
3161 }
3162
3163 static int handle_content_merge(struct merge_options *o,
3164                                 const char *path,
3165                                 int is_dirty,
3166                                 struct object_id *o_oid, int o_mode,
3167                                 struct object_id *a_oid, int a_mode,
3168                                 struct object_id *b_oid, int b_mode,
3169                                 struct rename_conflict_info *rename_conflict_info)
3170 {
3171         const char *reason = _("content");
3172         const char *path1 = NULL, *path2 = NULL;
3173         struct merge_file_info mfi;
3174         struct diff_filespec one, a, b;
3175         unsigned df_conflict_remains = 0;
3176
3177         if (!o_oid) {
3178                 reason = _("add/add");
3179                 o_oid = (struct object_id *)&null_oid;
3180         }
3181         one.path = a.path = b.path = (char *)path;
3182         oidcpy(&one.oid, o_oid);
3183         one.mode = o_mode;
3184         oidcpy(&a.oid, a_oid);
3185         a.mode = a_mode;
3186         oidcpy(&b.oid, b_oid);
3187         b.mode = b_mode;
3188
3189         if (rename_conflict_info) {
3190                 struct diff_filepair *pair1 = rename_conflict_info->pair1;
3191
3192                 path1 = (o->branch1 == rename_conflict_info->branch1) ?
3193                         pair1->two->path : pair1->one->path;
3194                 /* If rename_conflict_info->pair2 != NULL, we are in
3195                  * RENAME_ONE_FILE_TO_ONE case.  Otherwise, we have a
3196                  * normal rename.
3197                  */
3198                 path2 = (rename_conflict_info->pair2 ||
3199                          o->branch2 == rename_conflict_info->branch1) ?
3200                         pair1->two->path : pair1->one->path;
3201                 one.path = pair1->one->path;
3202                 a.path = (char *)path1;
3203                 b.path = (char *)path2;
3204
3205                 if (dir_in_way(path, !o->call_depth,
3206                                S_ISGITLINK(pair1->two->mode)))
3207                         df_conflict_remains = 1;
3208         }
3209         if (merge_mode_and_contents(o, &one, &a, &b, path,
3210                                     o->branch1, o->branch2,
3211                                     o->call_depth * 2, &mfi))
3212                 return -1;
3213
3214         /*
3215          * We can skip updating the working tree file iff:
3216          *   a) The merge is clean
3217          *   b) The merge matches what was in HEAD (content, mode, pathname)
3218          *   c) The target path is usable (i.e. not involved in D/F conflict)
3219          */
3220         if (mfi.clean &&
3221             was_tracked_and_matches(o, path, &mfi.oid, mfi.mode) &&
3222             !df_conflict_remains) {
3223                 int pos;
3224                 struct cache_entry *ce;
3225
3226                 output(o, 3, _("Skipped %s (merged same as existing)"), path);
3227                 if (add_cacheinfo(o, mfi.mode, &mfi.oid, path,
3228                                   0, (!o->call_depth && !is_dirty), 0))
3229                         return -1;
3230                 /*
3231                  * However, add_cacheinfo() will delete the old cache entry
3232                  * and add a new one.  We need to copy over any skip_worktree
3233                  * flag to avoid making the file appear as if it were
3234                  * deleted by the user.
3235                  */
3236                 pos = index_name_pos(&o->orig_index, path, strlen(path));
3237                 ce = o->orig_index.cache[pos];
3238                 if (ce_skip_worktree(ce)) {
3239                         pos = index_name_pos(&the_index, path, strlen(path));
3240                         ce = the_index.cache[pos];
3241                         ce->ce_flags |= CE_SKIP_WORKTREE;
3242                 }
3243                 return mfi.clean;
3244         }
3245
3246         if (!mfi.clean) {
3247                 if (S_ISGITLINK(mfi.mode))
3248                         reason = _("submodule");
3249                 output(o, 1, _("CONFLICT (%s): Merge conflict in %s"),
3250                                 reason, path);
3251                 if (rename_conflict_info && !df_conflict_remains)
3252                         if (update_stages(o, path, &one, &a, &b))
3253                                 return -1;
3254         }
3255
3256         if (df_conflict_remains || is_dirty) {
3257                 char *new_path;
3258                 if (o->call_depth) {
3259                         remove_file_from_cache(path);
3260                 } else {
3261                         if (!mfi.clean) {
3262                                 if (update_stages(o, path, &one, &a, &b))
3263                                         return -1;
3264                         } else {
3265                                 int file_from_stage2 = was_tracked(o, path);
3266                                 struct diff_filespec merged;
3267                                 oidcpy(&merged.oid, &mfi.oid);
3268                                 merged.mode = mfi.mode;
3269
3270                                 if (update_stages(o, path, NULL,
3271                                                   file_from_stage2 ? &merged : NULL,
3272                                                   file_from_stage2 ? NULL : &merged))
3273                                         return -1;
3274                         }
3275
3276                 }
3277                 new_path = unique_path(o, path, rename_conflict_info->branch1);
3278                 if (is_dirty) {
3279                         output(o, 1, _("Refusing to lose dirty file at %s"),
3280                                path);
3281                 }
3282                 output(o, 1, _("Adding as %s instead"), new_path);
3283                 if (update_file(o, 0, &mfi.oid, mfi.mode, new_path)) {
3284                         free(new_path);
3285                         return -1;
3286                 }
3287                 free(new_path);
3288                 mfi.clean = 0;
3289         } else if (update_file(o, mfi.clean, &mfi.oid, mfi.mode, path))
3290                 return -1;
3291         return !is_dirty && mfi.clean;
3292 }
3293
3294 static int handle_rename_normal(struct merge_options *o,
3295                                 const char *path,
3296                                 struct object_id *o_oid, unsigned int o_mode,
3297                                 struct object_id *a_oid, unsigned int a_mode,
3298                                 struct object_id *b_oid, unsigned int b_mode,
3299                                 struct rename_conflict_info *ci)
3300 {
3301         /* Merge the content and write it out */
3302         return handle_content_merge(o, path, was_dirty(o, path),
3303                                     o_oid, o_mode, a_oid, a_mode, b_oid, b_mode,
3304                                     ci);
3305 }
3306
3307 /* Per entry merge function */
3308 static int process_entry(struct merge_options *o,
3309                          const char *path, struct stage_data *entry)
3310 {
3311         int clean_merge = 1;
3312         int normalize = o->renormalize;
3313         unsigned o_mode = entry->stages[1].mode;
3314         unsigned a_mode = entry->stages[2].mode;
3315         unsigned b_mode = entry->stages[3].mode;
3316         struct object_id *o_oid = stage_oid(&entry->stages[1].oid, o_mode);
3317         struct object_id *a_oid = stage_oid(&entry->stages[2].oid, a_mode);
3318         struct object_id *b_oid = stage_oid(&entry->stages[3].oid, b_mode);
3319
3320         entry->processed = 1;
3321         if (entry->rename_conflict_info) {
3322                 struct rename_conflict_info *conflict_info = entry->rename_conflict_info;
3323                 switch (conflict_info->rename_type) {
3324                 case RENAME_NORMAL:
3325                 case RENAME_ONE_FILE_TO_ONE:
3326                         clean_merge = handle_rename_normal(o,
3327                                                            path,
3328                                                            o_oid, o_mode,
3329                                                            a_oid, a_mode,
3330                                                            b_oid, b_mode,
3331                                                            conflict_info);
3332                         break;
3333                 case RENAME_VIA_DIR:
3334                         clean_merge = 1;
3335                         if (handle_rename_via_dir(o,
3336                                                   conflict_info->pair1,
3337                                                   conflict_info->branch1,
3338                                                   conflict_info->branch2))
3339                                 clean_merge = -1;
3340                         break;
3341                 case RENAME_ADD:
3342                         /*
3343                          * Probably unclean merge, but if the renamed file
3344                          * merges cleanly and the result can then be
3345                          * two-way merged cleanly with the added file, I
3346                          * guess it's a clean merge?
3347                          */
3348                         clean_merge = handle_rename_add(o, conflict_info);
3349                         break;
3350                 case RENAME_DELETE:
3351                         clean_merge = 0;
3352                         if (handle_rename_delete(o,
3353                                                  conflict_info->pair1,
3354                                                  conflict_info->branch1,
3355                                                  conflict_info->branch2))
3356                                 clean_merge = -1;
3357                         break;
3358                 case RENAME_ONE_FILE_TO_TWO:
3359                         clean_merge = 0;
3360                         if (handle_rename_rename_1to2(o, conflict_info))
3361                                 clean_merge = -1;
3362                         break;
3363                 case RENAME_TWO_FILES_TO_ONE:
3364                         clean_merge = 0;
3365                         if (handle_rename_rename_2to1(o, conflict_info))
3366                                 clean_merge = -1;
3367                         break;
3368                 default:
3369                         entry->processed = 0;
3370                         break;
3371                 }
3372         } else if (o_oid && (!a_oid || !b_oid)) {
3373                 /* Case A: Deleted in one */
3374                 if ((!a_oid && !b_oid) ||
3375                     (!b_oid && blob_unchanged(o, o_oid, o_mode, a_oid, a_mode, normalize, path)) ||
3376                     (!a_oid && blob_unchanged(o, o_oid, o_mode, b_oid, b_mode, normalize, path))) {
3377                         /* Deleted in both or deleted in one and
3378                          * unchanged in the other */
3379                         if (a_oid)
3380                                 output(o, 2, _("Removing %s"), path);
3381                         /* do not touch working file if it did not exist */
3382                         remove_file(o, 1, path, !a_oid);
3383                 } else {
3384                         /* Modify/delete; deleted side may have put a directory in the way */
3385                         clean_merge = 0;
3386                         if (handle_modify_delete(o, path, o_oid, o_mode,
3387                                                  a_oid, a_mode, b_oid, b_mode))
3388                                 clean_merge = -1;
3389                 }
3390         } else if ((!o_oid && a_oid && !b_oid) ||
3391                    (!o_oid && !a_oid && b_oid)) {
3392                 /* Case B: Added in one. */
3393                 /* [nothing|directory] -> ([nothing|directory], file) */
3394
3395                 const char *add_branch;
3396                 const char *other_branch;
3397                 unsigned mode;
3398                 const struct object_id *oid;
3399                 const char *conf;
3400
3401                 if (a_oid) {
3402                         add_branch = o->branch1;
3403                         other_branch = o->branch2;
3404                         mode = a_mode;
3405                         oid = a_oid;
3406                         conf = _("file/directory");
3407                 } else {
3408                         add_branch = o->branch2;
3409                         other_branch = o->branch1;
3410                         mode = b_mode;
3411                         oid = b_oid;
3412                         conf = _("directory/file");
3413                 }
3414                 if (dir_in_way(path,
3415                                !o->call_depth && !S_ISGITLINK(a_mode),
3416                                0)) {
3417                         char *new_path = unique_path(o, path, add_branch);
3418                         clean_merge = 0;
3419                         output(o, 1, _("CONFLICT (%s): There is a directory with name %s in %s. "
3420                                "Adding %s as %s"),
3421                                conf, path, other_branch, path, new_path);
3422                         if (update_file(o, 0, oid, mode, new_path))
3423                                 clean_merge = -1;
3424                         else if (o->call_depth)
3425                                 remove_file_from_cache(path);
3426                         free(new_path);
3427                 } else {
3428                         output(o, 2, _("Adding %s"), path);
3429                         /* do not overwrite file if already present */
3430                         if (update_file_flags(o, oid, mode, path, 1, !a_oid))
3431                                 clean_merge = -1;
3432                 }
3433         } else if (a_oid && b_oid) {
3434                 /* Case C: Added in both (check for same permissions) and */
3435                 /* case D: Modified in both, but differently. */
3436                 int is_dirty = 0; /* unpack_trees would have bailed if dirty */
3437                 clean_merge = handle_content_merge(o, path, is_dirty,
3438                                                    o_oid, o_mode,
3439                                                    a_oid, a_mode,
3440                                                    b_oid, b_mode,
3441                                                    NULL);
3442         } else if (!o_oid && !a_oid && !b_oid) {
3443                 /*
3444                  * this entry was deleted altogether. a_mode == 0 means
3445                  * we had that path and want to actively remove it.
3446                  */
3447                 remove_file(o, 1, path, !a_mode);
3448         } else
3449                 BUG("fatal merge failure, shouldn't happen.");
3450
3451         return clean_merge;
3452 }
3453
3454 int merge_trees(struct merge_options *o,
3455                 struct tree *head,
3456                 struct tree *merge,
3457                 struct tree *common,
3458                 struct tree **result)
3459 {
3460         int code, clean;
3461         struct strbuf sb = STRBUF_INIT;
3462
3463         if (!o->call_depth && index_has_changes(&the_index, head, &sb)) {
3464                 err(o, _("Your local changes to the following files would be overwritten by merge:\n  %s"),
3465                     sb.buf);
3466                 return -1;
3467         }
3468
3469         if (o->subtree_shift) {
3470                 merge = shift_tree_object(head, merge, o->subtree_shift);
3471                 common = shift_tree_object(head, common, o->subtree_shift);
3472         }
3473
3474         if (oid_eq(&common->object.oid, &merge->object.oid)) {
3475                 output(o, 0, _("Already up to date!"));
3476                 *result = head;
3477                 return 1;
3478         }
3479
3480         code = unpack_trees_start(o, common, head, merge);
3481
3482         if (code != 0) {
3483                 if (show(o, 4) || o->call_depth)
3484                         err(o, _("merging of trees %s and %s failed"),
3485                             oid_to_hex(&head->object.oid),
3486                             oid_to_hex(&merge->object.oid));
3487                 unpack_trees_finish(o);
3488                 return -1;
3489         }
3490
3491         if (unmerged_cache()) {
3492                 struct string_list *entries;
3493                 struct rename_info re_info;
3494                 int i;
3495                 /*
3496                  * Only need the hashmap while processing entries, so
3497                  * initialize it here and free it when we are done running
3498                  * through the entries. Keeping it in the merge_options as
3499                  * opposed to decaring a local hashmap is for convenience
3500                  * so that we don't have to pass it to around.
3501                  */
3502                 hashmap_init(&o->current_file_dir_set, path_hashmap_cmp, NULL, 512);
3503                 get_files_dirs(o, head);
3504                 get_files_dirs(o, merge);
3505
3506                 entries = get_unmerged();
3507                 clean = detect_and_process_renames(o, common, head, merge,
3508                                                    entries, &re_info);
3509                 record_df_conflict_files(o, entries);
3510                 if (clean < 0)
3511                         goto cleanup;
3512                 for (i = entries->nr-1; 0 <= i; i--) {
3513                         const char *path = entries->items[i].string;
3514                         struct stage_data *e = entries->items[i].util;
3515                         if (!e->processed) {
3516                                 int ret = process_entry(o, path, e);
3517                                 if (!ret)
3518                                         clean = 0;
3519                                 else if (ret < 0) {
3520                                         clean = ret;
3521                                         goto cleanup;
3522                                 }
3523                         }
3524                 }
3525                 for (i = 0; i < entries->nr; i++) {
3526                         struct stage_data *e = entries->items[i].util;
3527                         if (!e->processed)
3528                                 BUG("unprocessed path??? %s",
3529                                     entries->items[i].string);
3530                 }
3531
3532         cleanup:
3533                 final_cleanup_renames(&re_info);
3534
3535                 string_list_clear(entries, 1);
3536                 free(entries);
3537
3538                 hashmap_free(&o->current_file_dir_set, 1);
3539
3540                 if (clean < 0) {
3541                         unpack_trees_finish(o);
3542                         return clean;
3543                 }
3544         }
3545         else
3546                 clean = 1;
3547
3548         unpack_trees_finish(o);
3549
3550         if (o->call_depth && !(*result = write_tree_from_memory(o)))
3551                 return -1;
3552
3553         return clean;
3554 }
3555
3556 static struct commit_list *reverse_commit_list(struct commit_list *list)
3557 {
3558         struct commit_list *next = NULL, *current, *backup;
3559         for (current = list; current; current = backup) {
3560                 backup = current->next;
3561                 current->next = next;
3562                 next = current;
3563         }
3564         return next;
3565 }
3566
3567 /*
3568  * Merge the commits h1 and h2, return the resulting virtual
3569  * commit object and a flag indicating the cleanness of the merge.
3570  */
3571 int merge_recursive(struct merge_options *o,
3572                     struct commit *h1,
3573                     struct commit *h2,
3574                     struct commit_list *ca,
3575                     struct commit **result)
3576 {
3577         struct commit_list *iter;
3578         struct commit *merged_common_ancestors;
3579         struct tree *mrtree;
3580         int clean;
3581
3582         if (show(o, 4)) {
3583                 output(o, 4, _("Merging:"));
3584                 output_commit_title(o, h1);
3585                 output_commit_title(o, h2);
3586         }
3587
3588         if (!ca) {
3589                 ca = get_merge_bases(h1, h2);
3590                 ca = reverse_commit_list(ca);
3591         }
3592
3593         if (show(o, 5)) {
3594                 unsigned cnt = commit_list_count(ca);
3595
3596                 output(o, 5, Q_("found %u common ancestor:",
3597                                 "found %u common ancestors:", cnt), cnt);
3598                 for (iter = ca; iter; iter = iter->next)
3599                         output_commit_title(o, iter->item);
3600         }
3601
3602         merged_common_ancestors = pop_commit(&ca);
3603         if (merged_common_ancestors == NULL) {
3604                 /* if there is no common ancestor, use an empty tree */
3605                 struct tree *tree;
3606
3607                 tree = lookup_tree(the_repository, the_repository->hash_algo->empty_tree);
3608                 merged_common_ancestors = make_virtual_commit(tree, "ancestor");
3609         }
3610
3611         for (iter = ca; iter; iter = iter->next) {
3612                 const char *saved_b1, *saved_b2;
3613                 o->call_depth++;
3614                 /*
3615                  * When the merge fails, the result contains files
3616                  * with conflict markers. The cleanness flag is
3617                  * ignored (unless indicating an error), it was never
3618                  * actually used, as result of merge_trees has always
3619                  * overwritten it: the committed "conflicts" were
3620                  * already resolved.
3621                  */
3622                 discard_cache();
3623                 saved_b1 = o->branch1;
3624                 saved_b2 = o->branch2;
3625                 o->branch1 = "Temporary merge branch 1";
3626                 o->branch2 = "Temporary merge branch 2";
3627                 if (merge_recursive(o, merged_common_ancestors, iter->item,
3628                                     NULL, &merged_common_ancestors) < 0)
3629                         return -1;
3630                 o->branch1 = saved_b1;
3631                 o->branch2 = saved_b2;
3632                 o->call_depth--;
3633
3634                 if (!merged_common_ancestors)
3635                         return err(o, _("merge returned no commit"));
3636         }
3637
3638         discard_cache();
3639         if (!o->call_depth)
3640                 read_cache();
3641
3642         o->ancestor = "merged common ancestors";
3643         clean = merge_trees(o, get_commit_tree(h1), get_commit_tree(h2),
3644                             get_commit_tree(merged_common_ancestors),
3645                             &mrtree);
3646         if (clean < 0) {
3647                 flush_output(o);
3648                 return clean;
3649         }
3650
3651         if (o->call_depth) {
3652                 *result = make_virtual_commit(mrtree, "merged tree");
3653                 commit_list_insert(h1, &(*result)->parents);
3654                 commit_list_insert(h2, &(*result)->parents->next);
3655         }
3656         flush_output(o);
3657         if (!o->call_depth && o->buffer_output < 2)
3658                 strbuf_release(&o->obuf);
3659         if (show(o, 2))
3660                 diff_warn_rename_limit("merge.renamelimit",
3661                                        o->needed_rename_limit, 0);
3662         return clean;
3663 }
3664
3665 static struct commit *get_ref(const struct object_id *oid, const char *name)
3666 {
3667         struct object *object;
3668
3669         object = deref_tag(the_repository, parse_object(the_repository, oid),
3670                            name,
3671                            strlen(name));
3672         if (!object)
3673                 return NULL;
3674         if (object->type == OBJ_TREE)
3675                 return make_virtual_commit((struct tree*)object, name);
3676         if (object->type != OBJ_COMMIT)
3677                 return NULL;
3678         if (parse_commit((struct commit *)object))
3679                 return NULL;
3680         return (struct commit *)object;
3681 }
3682
3683 int merge_recursive_generic(struct merge_options *o,
3684                             const struct object_id *head,
3685                             const struct object_id *merge,
3686                             int num_base_list,
3687                             const struct object_id **base_list,
3688                             struct commit **result)
3689 {
3690         int clean;
3691         struct lock_file lock = LOCK_INIT;
3692         struct commit *head_commit = get_ref(head, o->branch1);
3693         struct commit *next_commit = get_ref(merge, o->branch2);
3694         struct commit_list *ca = NULL;
3695
3696         if (base_list) {
3697                 int i;
3698                 for (i = 0; i < num_base_list; ++i) {
3699                         struct commit *base;
3700                         if (!(base = get_ref(base_list[i], oid_to_hex(base_list[i]))))
3701                                 return err(o, _("Could not parse object '%s'"),
3702                                            oid_to_hex(base_list[i]));
3703                         commit_list_insert(base, &ca);
3704                 }
3705         }
3706
3707         hold_locked_index(&lock, LOCK_DIE_ON_ERROR);
3708         clean = merge_recursive(o, head_commit, next_commit, ca,
3709                                 result);
3710         if (clean < 0) {
3711                 rollback_lock_file(&lock);
3712                 return clean;
3713         }
3714
3715         if (write_locked_index(&the_index, &lock,
3716                                COMMIT_LOCK | SKIP_IF_UNCHANGED))
3717                 return err(o, _("Unable to write index."));
3718
3719         return clean ? 0 : 1;
3720 }
3721
3722 static void merge_recursive_config(struct merge_options *o)
3723 {
3724         char *value = NULL;
3725         git_config_get_int("merge.verbosity", &o->verbosity);
3726         git_config_get_int("diff.renamelimit", &o->diff_rename_limit);
3727         git_config_get_int("merge.renamelimit", &o->merge_rename_limit);
3728         if (!git_config_get_string("diff.renames", &value)) {
3729                 o->diff_detect_rename = git_config_rename("diff.renames", value);
3730                 free(value);
3731         }
3732         if (!git_config_get_string("merge.renames", &value)) {
3733                 o->merge_detect_rename = git_config_rename("merge.renames", value);
3734                 free(value);
3735         }
3736         git_config(git_xmerge_config, NULL);
3737 }
3738
3739 void init_merge_options(struct merge_options *o)
3740 {
3741         const char *merge_verbosity;
3742         memset(o, 0, sizeof(struct merge_options));
3743         o->verbosity = 2;
3744         o->buffer_output = 1;
3745         o->diff_rename_limit = -1;
3746         o->merge_rename_limit = -1;
3747         o->renormalize = 0;
3748         o->diff_detect_rename = -1;
3749         o->merge_detect_rename = -1;
3750         o->detect_directory_renames = 1;
3751         merge_recursive_config(o);
3752         merge_verbosity = getenv("GIT_MERGE_VERBOSITY");
3753         if (merge_verbosity)
3754                 o->verbosity = strtol(merge_verbosity, NULL, 10);
3755         if (o->verbosity >= 5)
3756                 o->buffer_output = 0;
3757         strbuf_init(&o->obuf, 0);
3758         string_list_init(&o->df_conflict_file_set, 1);
3759 }
3760
3761 int parse_merge_opt(struct merge_options *o, const char *s)
3762 {
3763         const char *arg;
3764
3765         if (!s || !*s)
3766                 return -1;
3767         if (!strcmp(s, "ours"))
3768                 o->recursive_variant = MERGE_RECURSIVE_OURS;
3769         else if (!strcmp(s, "theirs"))
3770                 o->recursive_variant = MERGE_RECURSIVE_THEIRS;
3771         else if (!strcmp(s, "subtree"))
3772                 o->subtree_shift = "";
3773         else if (skip_prefix(s, "subtree=", &arg))
3774                 o->subtree_shift = arg;
3775         else if (!strcmp(s, "patience"))
3776                 o->xdl_opts = DIFF_WITH_ALG(o, PATIENCE_DIFF);
3777         else if (!strcmp(s, "histogram"))
3778                 o->xdl_opts = DIFF_WITH_ALG(o, HISTOGRAM_DIFF);
3779         else if (skip_prefix(s, "diff-algorithm=", &arg)) {
3780                 long value = parse_algorithm_value(arg);
3781                 if (value < 0)
3782                         return -1;
3783                 /* clear out previous settings */
3784                 DIFF_XDL_CLR(o, NEED_MINIMAL);
3785                 o->xdl_opts &= ~XDF_DIFF_ALGORITHM_MASK;
3786                 o->xdl_opts |= value;
3787         }
3788         else if (!strcmp(s, "ignore-space-change"))
3789                 DIFF_XDL_SET(o, IGNORE_WHITESPACE_CHANGE);
3790         else if (!strcmp(s, "ignore-all-space"))
3791                 DIFF_XDL_SET(o, IGNORE_WHITESPACE);
3792         else if (!strcmp(s, "ignore-space-at-eol"))
3793                 DIFF_XDL_SET(o, IGNORE_WHITESPACE_AT_EOL);
3794         else if (!strcmp(s, "ignore-cr-at-eol"))
3795                 DIFF_XDL_SET(o, IGNORE_CR_AT_EOL);
3796         else if (!strcmp(s, "renormalize"))
3797                 o->renormalize = 1;
3798         else if (!strcmp(s, "no-renormalize"))
3799                 o->renormalize = 0;
3800         else if (!strcmp(s, "no-renames"))
3801                 o->merge_detect_rename = 0;
3802         else if (!strcmp(s, "find-renames")) {
3803                 o->merge_detect_rename = 1;
3804                 o->rename_score = 0;
3805         }
3806         else if (skip_prefix(s, "find-renames=", &arg) ||
3807                  skip_prefix(s, "rename-threshold=", &arg)) {
3808                 if ((o->rename_score = parse_rename_score(&arg)) == -1 || *arg != 0)
3809                         return -1;
3810                 o->merge_detect_rename = 1;
3811         }
3812         else
3813                 return -1;
3814         return 0;
3815 }