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