Merge branch 'ds/commit-graph-lockfile-fix'
[git] / merge-recursive.c
1 /*
2  * Recursive Merge algorithm stolen from git-merge-recursive.py by
3  * Fredrik Kuivinen.
4  * The thieves were Alex Riesen and Johannes Schindelin, in June/July 2006
5  */
6 #include "cache.h"
7 #include "config.h"
8 #include "advice.h"
9 #include "lockfile.h"
10 #include "cache-tree.h"
11 #include "commit.h"
12 #include "blob.h"
13 #include "builtin.h"
14 #include "tree-walk.h"
15 #include "diff.h"
16 #include "diffcore.h"
17 #include "tag.h"
18 #include "unpack-trees.h"
19 #include "string-list.h"
20 #include "xdiff-interface.h"
21 #include "ll-merge.h"
22 #include "attr.h"
23 #include "merge-recursive.h"
24 #include "dir.h"
25 #include "submodule.h"
26 #include "revision.h"
27
28 struct path_hashmap_entry {
29         struct hashmap_entry e;
30         char path[FLEX_ARRAY];
31 };
32
33 static int path_hashmap_cmp(const void *cmp_data,
34                             const void *entry,
35                             const void *entry_or_key,
36                             const void *keydata)
37 {
38         const struct path_hashmap_entry *a = entry;
39         const struct path_hashmap_entry *b = entry_or_key;
40         const char *key = keydata;
41
42         if (ignore_case)
43                 return strcasecmp(a->path, key ? key : b->path);
44         else
45                 return strcmp(a->path, key ? key : b->path);
46 }
47
48 static unsigned int path_hash(const char *path)
49 {
50         return ignore_case ? strihash(path) : strhash(path);
51 }
52
53 static struct dir_rename_entry *dir_rename_find_entry(struct hashmap *hashmap,
54                                                       char *dir)
55 {
56         struct dir_rename_entry key;
57
58         if (dir == NULL)
59                 return NULL;
60         hashmap_entry_init(&key, strhash(dir));
61         key.dir = dir;
62         return hashmap_get(hashmap, &key, NULL);
63 }
64
65 static int dir_rename_cmp(const void *unused_cmp_data,
66                           const void *entry,
67                           const void *entry_or_key,
68                           const void *unused_keydata)
69 {
70         const struct dir_rename_entry *e1 = entry;
71         const struct dir_rename_entry *e2 = entry_or_key;
72
73         return strcmp(e1->dir, e2->dir);
74 }
75
76 static void dir_rename_init(struct hashmap *map)
77 {
78         hashmap_init(map, dir_rename_cmp, NULL, 0);
79 }
80
81 static void dir_rename_entry_init(struct dir_rename_entry *entry,
82                                   char *directory)
83 {
84         hashmap_entry_init(entry, strhash(directory));
85         entry->dir = directory;
86         entry->non_unique_new_dir = 0;
87         strbuf_init(&entry->new_dir, 0);
88         string_list_init(&entry->possible_new_dirs, 0);
89 }
90
91 static struct collision_entry *collision_find_entry(struct hashmap *hashmap,
92                                                     char *target_file)
93 {
94         struct collision_entry key;
95
96         hashmap_entry_init(&key, strhash(target_file));
97         key.target_file = target_file;
98         return hashmap_get(hashmap, &key, NULL);
99 }
100
101 static int collision_cmp(void *unused_cmp_data,
102                          const struct collision_entry *e1,
103                          const struct collision_entry *e2,
104                          const void *unused_keydata)
105 {
106         return strcmp(e1->target_file, e2->target_file);
107 }
108
109 static void collision_init(struct hashmap *map)
110 {
111         hashmap_init(map, (hashmap_cmp_fn) collision_cmp, NULL, 0);
112 }
113
114 static void flush_output(struct merge_options *o)
115 {
116         if (o->buffer_output < 2 && o->obuf.len) {
117                 fputs(o->obuf.buf, stdout);
118                 strbuf_reset(&o->obuf);
119         }
120 }
121
122 static int err(struct merge_options *o, const char *err, ...)
123 {
124         va_list params;
125
126         if (o->buffer_output < 2)
127                 flush_output(o);
128         else {
129                 strbuf_complete(&o->obuf, '\n');
130                 strbuf_addstr(&o->obuf, "error: ");
131         }
132         va_start(params, err);
133         strbuf_vaddf(&o->obuf, err, params);
134         va_end(params);
135         if (o->buffer_output > 1)
136                 strbuf_addch(&o->obuf, '\n');
137         else {
138                 error("%s", o->obuf.buf);
139                 strbuf_reset(&o->obuf);
140         }
141
142         return -1;
143 }
144
145 static struct tree *shift_tree_object(struct tree *one, struct tree *two,
146                                       const char *subtree_shift)
147 {
148         struct object_id shifted;
149
150         if (!*subtree_shift) {
151                 shift_tree(&one->object.oid, &two->object.oid, &shifted, 0);
152         } else {
153                 shift_tree_by(&one->object.oid, &two->object.oid, &shifted,
154                               subtree_shift);
155         }
156         if (!oidcmp(&two->object.oid, &shifted))
157                 return two;
158         return lookup_tree(&shifted);
159 }
160
161 static struct commit *make_virtual_commit(struct tree *tree, const char *comment)
162 {
163         struct commit *commit = alloc_commit_node();
164
165         set_merge_remote_desc(commit, comment, (struct object *)commit);
166         commit->maybe_tree = tree;
167         commit->object.parsed = 1;
168         return commit;
169 }
170
171 /*
172  * Since we use get_tree_entry(), which does not put the read object into
173  * the object pool, we cannot rely on a == b.
174  */
175 static int oid_eq(const struct object_id *a, const struct object_id *b)
176 {
177         if (!a && !b)
178                 return 2;
179         return a && b && oidcmp(a, b) == 0;
180 }
181
182 enum rename_type {
183         RENAME_NORMAL = 0,
184         RENAME_DIR,
185         RENAME_DELETE,
186         RENAME_ONE_FILE_TO_ONE,
187         RENAME_ONE_FILE_TO_TWO,
188         RENAME_TWO_FILES_TO_ONE
189 };
190
191 struct rename_conflict_info {
192         enum rename_type rename_type;
193         struct diff_filepair *pair1;
194         struct diff_filepair *pair2;
195         const char *branch1;
196         const char *branch2;
197         struct stage_data *dst_entry1;
198         struct stage_data *dst_entry2;
199         struct diff_filespec ren1_other;
200         struct diff_filespec ren2_other;
201 };
202
203 /*
204  * Since we want to write the index eventually, we cannot reuse the index
205  * for these (temporary) data.
206  */
207 struct stage_data {
208         struct {
209                 unsigned mode;
210                 struct object_id oid;
211         } stages[4];
212         struct rename_conflict_info *rename_conflict_info;
213         unsigned processed:1;
214 };
215
216 static inline void setup_rename_conflict_info(enum rename_type rename_type,
217                                               struct diff_filepair *pair1,
218                                               struct diff_filepair *pair2,
219                                               const char *branch1,
220                                               const char *branch2,
221                                               struct stage_data *dst_entry1,
222                                               struct stage_data *dst_entry2,
223                                               struct merge_options *o,
224                                               struct stage_data *src_entry1,
225                                               struct stage_data *src_entry2)
226 {
227         struct rename_conflict_info *ci = xcalloc(1, sizeof(struct rename_conflict_info));
228         ci->rename_type = rename_type;
229         ci->pair1 = pair1;
230         ci->branch1 = branch1;
231         ci->branch2 = branch2;
232
233         ci->dst_entry1 = dst_entry1;
234         dst_entry1->rename_conflict_info = ci;
235         dst_entry1->processed = 0;
236
237         assert(!pair2 == !dst_entry2);
238         if (dst_entry2) {
239                 ci->dst_entry2 = dst_entry2;
240                 ci->pair2 = pair2;
241                 dst_entry2->rename_conflict_info = ci;
242         }
243
244         if (rename_type == RENAME_TWO_FILES_TO_ONE) {
245                 /*
246                  * For each rename, there could have been
247                  * modifications on the side of history where that
248                  * file was not renamed.
249                  */
250                 int ostage1 = o->branch1 == branch1 ? 3 : 2;
251                 int ostage2 = ostage1 ^ 1;
252
253                 ci->ren1_other.path = pair1->one->path;
254                 oidcpy(&ci->ren1_other.oid, &src_entry1->stages[ostage1].oid);
255                 ci->ren1_other.mode = src_entry1->stages[ostage1].mode;
256
257                 ci->ren2_other.path = pair2->one->path;
258                 oidcpy(&ci->ren2_other.oid, &src_entry2->stages[ostage2].oid);
259                 ci->ren2_other.mode = src_entry2->stages[ostage2].mode;
260         }
261 }
262
263 static int show(struct merge_options *o, int v)
264 {
265         return (!o->call_depth && o->verbosity >= v) || o->verbosity >= 5;
266 }
267
268 __attribute__((format (printf, 3, 4)))
269 static void output(struct merge_options *o, int v, const char *fmt, ...)
270 {
271         va_list ap;
272
273         if (!show(o, v))
274                 return;
275
276         strbuf_addchars(&o->obuf, ' ', o->call_depth * 2);
277
278         va_start(ap, fmt);
279         strbuf_vaddf(&o->obuf, fmt, ap);
280         va_end(ap);
281
282         strbuf_addch(&o->obuf, '\n');
283         if (!o->buffer_output)
284                 flush_output(o);
285 }
286
287 static void output_commit_title(struct merge_options *o, struct commit *commit)
288 {
289         struct merge_remote_desc *desc;
290
291         strbuf_addchars(&o->obuf, ' ', o->call_depth * 2);
292         desc = merge_remote_util(commit);
293         if (desc)
294                 strbuf_addf(&o->obuf, "virtual %s\n", desc->name);
295         else {
296                 strbuf_add_unique_abbrev(&o->obuf, &commit->object.oid,
297                                          DEFAULT_ABBREV);
298                 strbuf_addch(&o->obuf, ' ');
299                 if (parse_commit(commit) != 0)
300                         strbuf_addstr(&o->obuf, _("(bad commit)\n"));
301                 else {
302                         const char *title;
303                         const char *msg = get_commit_buffer(commit, NULL);
304                         int len = find_commit_subject(msg, &title);
305                         if (len)
306                                 strbuf_addf(&o->obuf, "%.*s\n", len, title);
307                         unuse_commit_buffer(commit, msg);
308                 }
309         }
310         flush_output(o);
311 }
312
313 static int add_cacheinfo(struct merge_options *o,
314                 unsigned int mode, const struct object_id *oid,
315                 const char *path, int stage, int refresh, int options)
316 {
317         struct cache_entry *ce;
318         int ret;
319
320         ce = make_cache_entry(mode, oid ? oid->hash : null_sha1, path, stage, 0);
321         if (!ce)
322                 return err(o, _("add_cacheinfo failed for path '%s'; merge aborting."), path);
323
324         ret = add_cache_entry(ce, options);
325         if (refresh) {
326                 struct cache_entry *nce;
327
328                 nce = refresh_cache_entry(ce, CE_MATCH_REFRESH | CE_MATCH_IGNORE_MISSING);
329                 if (!nce)
330                         return err(o, _("add_cacheinfo failed to refresh for path '%s'; merge aborting."), path);
331                 if (nce != ce)
332                         ret = add_cache_entry(nce, options);
333         }
334         return ret;
335 }
336
337 static void init_tree_desc_from_tree(struct tree_desc *desc, struct tree *tree)
338 {
339         parse_tree(tree);
340         init_tree_desc(desc, tree->buffer, tree->size);
341 }
342
343 static int unpack_trees_start(struct merge_options *o,
344                               struct tree *common,
345                               struct tree *head,
346                               struct tree *merge)
347 {
348         int rc;
349         struct tree_desc t[3];
350         struct index_state tmp_index = { NULL };
351
352         memset(&o->unpack_opts, 0, sizeof(o->unpack_opts));
353         if (o->call_depth)
354                 o->unpack_opts.index_only = 1;
355         else
356                 o->unpack_opts.update = 1;
357         o->unpack_opts.merge = 1;
358         o->unpack_opts.head_idx = 2;
359         o->unpack_opts.fn = threeway_merge;
360         o->unpack_opts.src_index = &the_index;
361         o->unpack_opts.dst_index = &tmp_index;
362         o->unpack_opts.aggressive = !merge_detect_rename(o);
363         setup_unpack_trees_porcelain(&o->unpack_opts, "merge");
364
365         init_tree_desc_from_tree(t+0, common);
366         init_tree_desc_from_tree(t+1, head);
367         init_tree_desc_from_tree(t+2, merge);
368
369         rc = unpack_trees(3, t, &o->unpack_opts);
370         cache_tree_free(&active_cache_tree);
371
372         /*
373          * Update the_index to match the new results, AFTER saving a copy
374          * in o->orig_index.  Update src_index to point to the saved copy.
375          * (verify_uptodate() checks src_index, and the original index is
376          * the one that had the necessary modification timestamps.)
377          */
378         o->orig_index = the_index;
379         the_index = tmp_index;
380         o->unpack_opts.src_index = &o->orig_index;
381
382         return rc;
383 }
384
385 static void unpack_trees_finish(struct merge_options *o)
386 {
387         discard_index(&o->orig_index);
388         clear_unpack_trees_porcelain(&o->unpack_opts);
389 }
390
391 struct tree *write_tree_from_memory(struct merge_options *o)
392 {
393         struct tree *result = NULL;
394
395         if (unmerged_cache()) {
396                 int i;
397                 fprintf(stderr, "BUG: There are unmerged index entries:\n");
398                 for (i = 0; i < active_nr; i++) {
399                         const struct cache_entry *ce = active_cache[i];
400                         if (ce_stage(ce))
401                                 fprintf(stderr, "BUG: %d %.*s\n", ce_stage(ce),
402                                         (int)ce_namelen(ce), ce->name);
403                 }
404                 BUG("unmerged index entries in merge-recursive.c");
405         }
406
407         if (!active_cache_tree)
408                 active_cache_tree = cache_tree();
409
410         if (!cache_tree_fully_valid(active_cache_tree) &&
411             cache_tree_update(&the_index, 0) < 0) {
412                 err(o, _("error building trees"));
413                 return NULL;
414         }
415
416         result = lookup_tree(&active_cache_tree->oid);
417
418         return result;
419 }
420
421 static int save_files_dirs(const struct object_id *oid,
422                 struct strbuf *base, const char *path,
423                 unsigned int mode, int stage, void *context)
424 {
425         struct path_hashmap_entry *entry;
426         int baselen = base->len;
427         struct merge_options *o = context;
428
429         strbuf_addstr(base, path);
430
431         FLEX_ALLOC_MEM(entry, path, base->buf, base->len);
432         hashmap_entry_init(entry, path_hash(entry->path));
433         hashmap_add(&o->current_file_dir_set, entry);
434
435         strbuf_setlen(base, baselen);
436         return (S_ISDIR(mode) ? READ_TREE_RECURSIVE : 0);
437 }
438
439 static void get_files_dirs(struct merge_options *o, struct tree *tree)
440 {
441         struct pathspec match_all;
442         memset(&match_all, 0, sizeof(match_all));
443         read_tree_recursive(tree, "", 0, 0, &match_all, save_files_dirs, o);
444 }
445
446 static int get_tree_entry_if_blob(const struct object_id *tree,
447                                   const char *path,
448                                   struct object_id *hashy,
449                                   unsigned int *mode_o)
450 {
451         int ret;
452
453         ret = get_tree_entry(tree, path, hashy, mode_o);
454         if (S_ISDIR(*mode_o)) {
455                 oidcpy(hashy, &null_oid);
456                 *mode_o = 0;
457         }
458         return ret;
459 }
460
461 /*
462  * Returns an index_entry instance which doesn't have to correspond to
463  * a real cache entry in Git's index.
464  */
465 static struct stage_data *insert_stage_data(const char *path,
466                 struct tree *o, struct tree *a, struct tree *b,
467                 struct string_list *entries)
468 {
469         struct string_list_item *item;
470         struct stage_data *e = xcalloc(1, sizeof(struct stage_data));
471         get_tree_entry_if_blob(&o->object.oid, path,
472                                &e->stages[1].oid, &e->stages[1].mode);
473         get_tree_entry_if_blob(&a->object.oid, path,
474                                &e->stages[2].oid, &e->stages[2].mode);
475         get_tree_entry_if_blob(&b->object.oid, path,
476                                &e->stages[3].oid, &e->stages[3].mode);
477         item = string_list_insert(entries, path);
478         item->util = e;
479         return e;
480 }
481
482 /*
483  * Create a dictionary mapping file names to stage_data objects. The
484  * dictionary contains one entry for every path with a non-zero stage entry.
485  */
486 static struct string_list *get_unmerged(void)
487 {
488         struct string_list *unmerged = xcalloc(1, sizeof(struct string_list));
489         int i;
490
491         unmerged->strdup_strings = 1;
492
493         for (i = 0; i < active_nr; i++) {
494                 struct string_list_item *item;
495                 struct stage_data *e;
496                 const struct cache_entry *ce = active_cache[i];
497                 if (!ce_stage(ce))
498                         continue;
499
500                 item = string_list_lookup(unmerged, ce->name);
501                 if (!item) {
502                         item = string_list_insert(unmerged, ce->name);
503                         item->util = xcalloc(1, sizeof(struct stage_data));
504                 }
505                 e = item->util;
506                 e->stages[ce_stage(ce)].mode = ce->ce_mode;
507                 oidcpy(&e->stages[ce_stage(ce)].oid, &ce->oid);
508         }
509
510         return unmerged;
511 }
512
513 static int string_list_df_name_compare(const char *one, const char *two)
514 {
515         int onelen = strlen(one);
516         int twolen = strlen(two);
517         /*
518          * Here we only care that entries for D/F conflicts are
519          * adjacent, in particular with the file of the D/F conflict
520          * appearing before files below the corresponding directory.
521          * The order of the rest of the list is irrelevant for us.
522          *
523          * To achieve this, we sort with df_name_compare and provide
524          * the mode S_IFDIR so that D/F conflicts will sort correctly.
525          * We use the mode S_IFDIR for everything else for simplicity,
526          * since in other cases any changes in their order due to
527          * sorting cause no problems for us.
528          */
529         int cmp = df_name_compare(one, onelen, S_IFDIR,
530                                   two, twolen, S_IFDIR);
531         /*
532          * Now that 'foo' and 'foo/bar' compare equal, we have to make sure
533          * that 'foo' comes before 'foo/bar'.
534          */
535         if (cmp)
536                 return cmp;
537         return onelen - twolen;
538 }
539
540 static void record_df_conflict_files(struct merge_options *o,
541                                      struct string_list *entries)
542 {
543         /* If there is a D/F conflict and the file for such a conflict
544          * currently exist in the working tree, we want to allow it to be
545          * removed to make room for the corresponding directory if needed.
546          * The files underneath the directories of such D/F conflicts will
547          * be processed before the corresponding file involved in the D/F
548          * conflict.  If the D/F directory ends up being removed by the
549          * merge, then we won't have to touch the D/F file.  If the D/F
550          * directory needs to be written to the working copy, then the D/F
551          * file will simply be removed (in make_room_for_path()) to make
552          * room for the necessary paths.  Note that if both the directory
553          * and the file need to be present, then the D/F file will be
554          * reinstated with a new unique name at the time it is processed.
555          */
556         struct string_list df_sorted_entries = STRING_LIST_INIT_NODUP;
557         const char *last_file = NULL;
558         int last_len = 0;
559         int i;
560
561         /*
562          * If we're merging merge-bases, we don't want to bother with
563          * any working directory changes.
564          */
565         if (o->call_depth)
566                 return;
567
568         /* Ensure D/F conflicts are adjacent in the entries list. */
569         for (i = 0; i < entries->nr; i++) {
570                 struct string_list_item *next = &entries->items[i];
571                 string_list_append(&df_sorted_entries, next->string)->util =
572                                    next->util;
573         }
574         df_sorted_entries.cmp = string_list_df_name_compare;
575         string_list_sort(&df_sorted_entries);
576
577         string_list_clear(&o->df_conflict_file_set, 1);
578         for (i = 0; i < df_sorted_entries.nr; i++) {
579                 const char *path = df_sorted_entries.items[i].string;
580                 int len = strlen(path);
581                 struct stage_data *e = df_sorted_entries.items[i].util;
582
583                 /*
584                  * Check if last_file & path correspond to a D/F conflict;
585                  * i.e. whether path is last_file+'/'+<something>.
586                  * If so, record that it's okay to remove last_file to make
587                  * room for path and friends if needed.
588                  */
589                 if (last_file &&
590                     len > last_len &&
591                     memcmp(path, last_file, last_len) == 0 &&
592                     path[last_len] == '/') {
593                         string_list_insert(&o->df_conflict_file_set, last_file);
594                 }
595
596                 /*
597                  * Determine whether path could exist as a file in the
598                  * working directory as a possible D/F conflict.  This
599                  * will only occur when it exists in stage 2 as a
600                  * file.
601                  */
602                 if (S_ISREG(e->stages[2].mode) || S_ISLNK(e->stages[2].mode)) {
603                         last_file = path;
604                         last_len = len;
605                 } else {
606                         last_file = NULL;
607                 }
608         }
609         string_list_clear(&df_sorted_entries, 0);
610 }
611
612 struct rename {
613         struct diff_filepair *pair;
614         /*
615          * Purpose of src_entry and dst_entry:
616          *
617          * If 'before' is renamed to 'after' then src_entry will contain
618          * the versions of 'before' from the merge_base, HEAD, and MERGE in
619          * stages 1, 2, and 3; dst_entry will contain the respective
620          * versions of 'after' in corresponding locations.  Thus, we have a
621          * total of six modes and oids, though some will be null.  (Stage 0
622          * is ignored; we're interested in handling conflicts.)
623          *
624          * Since we don't turn on break-rewrites by default, neither
625          * src_entry nor dst_entry can have all three of their stages have
626          * non-null oids, meaning at most four of the six will be non-null.
627          * Also, since this is a rename, both src_entry and dst_entry will
628          * have at least one non-null oid, meaning at least two will be
629          * non-null.  Of the six oids, a typical rename will have three be
630          * non-null.  Only two implies a rename/delete, and four implies a
631          * rename/add.
632          */
633         struct stage_data *src_entry;
634         struct stage_data *dst_entry;
635         unsigned add_turned_into_rename:1;
636         unsigned processed:1;
637 };
638
639 static int update_stages(struct merge_options *opt, const char *path,
640                          const struct diff_filespec *o,
641                          const struct diff_filespec *a,
642                          const struct diff_filespec *b)
643 {
644
645         /*
646          * NOTE: It is usually a bad idea to call update_stages on a path
647          * before calling update_file on that same path, since it can
648          * sometimes lead to spurious "refusing to lose untracked file..."
649          * messages from update_file (via make_room_for path via
650          * would_lose_untracked).  Instead, reverse the order of the calls
651          * (executing update_file first and then update_stages).
652          */
653         int clear = 1;
654         int options = ADD_CACHE_OK_TO_ADD | ADD_CACHE_SKIP_DFCHECK;
655         if (clear)
656                 if (remove_file_from_cache(path))
657                         return -1;
658         if (o)
659                 if (add_cacheinfo(opt, o->mode, &o->oid, path, 1, 0, options))
660                         return -1;
661         if (a)
662                 if (add_cacheinfo(opt, a->mode, &a->oid, path, 2, 0, options))
663                         return -1;
664         if (b)
665                 if (add_cacheinfo(opt, b->mode, &b->oid, path, 3, 0, options))
666                         return -1;
667         return 0;
668 }
669
670 static int update_stages_for_stage_data(struct merge_options *opt,
671                                         const char *path,
672                                         const struct stage_data *stage_data)
673 {
674         struct diff_filespec o, a, b;
675
676         o.mode = stage_data->stages[1].mode;
677         oidcpy(&o.oid, &stage_data->stages[1].oid);
678
679         a.mode = stage_data->stages[2].mode;
680         oidcpy(&a.oid, &stage_data->stages[2].oid);
681
682         b.mode = stage_data->stages[3].mode;
683         oidcpy(&b.oid, &stage_data->stages[3].oid);
684
685         return update_stages(opt, path,
686                              is_null_oid(&o.oid) ? NULL : &o,
687                              is_null_oid(&a.oid) ? NULL : &a,
688                              is_null_oid(&b.oid) ? NULL : &b);
689 }
690
691 static void update_entry(struct stage_data *entry,
692                          struct diff_filespec *o,
693                          struct diff_filespec *a,
694                          struct diff_filespec *b)
695 {
696         entry->processed = 0;
697         entry->stages[1].mode = o->mode;
698         entry->stages[2].mode = a->mode;
699         entry->stages[3].mode = b->mode;
700         oidcpy(&entry->stages[1].oid, &o->oid);
701         oidcpy(&entry->stages[2].oid, &a->oid);
702         oidcpy(&entry->stages[3].oid, &b->oid);
703 }
704
705 static int remove_file(struct merge_options *o, int clean,
706                        const char *path, int no_wd)
707 {
708         int update_cache = o->call_depth || clean;
709         int update_working_directory = !o->call_depth && !no_wd;
710
711         if (update_cache) {
712                 if (remove_file_from_cache(path))
713                         return -1;
714         }
715         if (update_working_directory) {
716                 if (ignore_case) {
717                         struct cache_entry *ce;
718                         ce = cache_file_exists(path, strlen(path), ignore_case);
719                         if (ce && ce_stage(ce) == 0 && strcmp(path, ce->name))
720                                 return 0;
721                 }
722                 if (remove_path(path))
723                         return -1;
724         }
725         return 0;
726 }
727
728 /* add a string to a strbuf, but converting "/" to "_" */
729 static void add_flattened_path(struct strbuf *out, const char *s)
730 {
731         size_t i = out->len;
732         strbuf_addstr(out, s);
733         for (; i < out->len; i++)
734                 if (out->buf[i] == '/')
735                         out->buf[i] = '_';
736 }
737
738 static char *unique_path(struct merge_options *o, const char *path, const char *branch)
739 {
740         struct path_hashmap_entry *entry;
741         struct strbuf newpath = STRBUF_INIT;
742         int suffix = 0;
743         size_t base_len;
744
745         strbuf_addf(&newpath, "%s~", path);
746         add_flattened_path(&newpath, branch);
747
748         base_len = newpath.len;
749         while (hashmap_get_from_hash(&o->current_file_dir_set,
750                                      path_hash(newpath.buf), newpath.buf) ||
751                (!o->call_depth && file_exists(newpath.buf))) {
752                 strbuf_setlen(&newpath, base_len);
753                 strbuf_addf(&newpath, "_%d", suffix++);
754         }
755
756         FLEX_ALLOC_MEM(entry, path, newpath.buf, newpath.len);
757         hashmap_entry_init(entry, path_hash(entry->path));
758         hashmap_add(&o->current_file_dir_set, entry);
759         return strbuf_detach(&newpath, NULL);
760 }
761
762 /**
763  * Check whether a directory in the index is in the way of an incoming
764  * file.  Return 1 if so.  If check_working_copy is non-zero, also
765  * check the working directory.  If empty_ok is non-zero, also return
766  * 0 in the case where the working-tree dir exists but is empty.
767  */
768 static int dir_in_way(const char *path, int check_working_copy, int empty_ok)
769 {
770         int pos;
771         struct strbuf dirpath = STRBUF_INIT;
772         struct stat st;
773
774         strbuf_addstr(&dirpath, path);
775         strbuf_addch(&dirpath, '/');
776
777         pos = cache_name_pos(dirpath.buf, dirpath.len);
778
779         if (pos < 0)
780                 pos = -1 - pos;
781         if (pos < active_nr &&
782             !strncmp(dirpath.buf, active_cache[pos]->name, dirpath.len)) {
783                 strbuf_release(&dirpath);
784                 return 1;
785         }
786
787         strbuf_release(&dirpath);
788         return check_working_copy && !lstat(path, &st) && S_ISDIR(st.st_mode) &&
789                 !(empty_ok && is_empty_dir(path));
790 }
791
792 /*
793  * Returns whether path was tracked in the index before the merge started,
794  * and its oid and mode match the specified values
795  */
796 static int was_tracked_and_matches(struct merge_options *o, const char *path,
797                                    const struct object_id *oid, unsigned mode)
798 {
799         int pos = index_name_pos(&o->orig_index, path, strlen(path));
800         struct cache_entry *ce;
801
802         if (0 > pos)
803                 /* we were not tracking this path before the merge */
804                 return 0;
805
806         /* See if the file we were tracking before matches */
807         ce = o->orig_index.cache[pos];
808         return (oid_eq(&ce->oid, oid) && ce->ce_mode == mode);
809 }
810
811 /*
812  * Returns whether path was tracked in the index before the merge started
813  */
814 static int was_tracked(struct merge_options *o, const char *path)
815 {
816         int pos = index_name_pos(&o->orig_index, path, strlen(path));
817
818         if (0 <= pos)
819                 /* we were tracking this path before the merge */
820                 return 1;
821
822         return 0;
823 }
824
825 static int would_lose_untracked(const char *path)
826 {
827         /*
828          * This may look like it can be simplified to:
829          *   return !was_tracked(o, path) && file_exists(path)
830          * but it can't.  This function needs to know whether path was in
831          * the working tree due to EITHER having been tracked in the index
832          * before the merge OR having been put into the working copy and
833          * index by unpack_trees().  Due to that either-or requirement, we
834          * check the current index instead of the original one.
835          *
836          * Note that we do not need to worry about merge-recursive itself
837          * updating the index after unpack_trees() and before calling this
838          * function, because we strictly require all code paths in
839          * merge-recursive to update the working tree first and the index
840          * second.  Doing otherwise would break
841          * update_file()/would_lose_untracked(); see every comment in this
842          * file which mentions "update_stages".
843          */
844         int pos = cache_name_pos(path, strlen(path));
845
846         if (pos < 0)
847                 pos = -1 - pos;
848         while (pos < active_nr &&
849                !strcmp(path, active_cache[pos]->name)) {
850                 /*
851                  * If stage #0, it is definitely tracked.
852                  * If it has stage #2 then it was tracked
853                  * before this merge started.  All other
854                  * cases the path was not tracked.
855                  */
856                 switch (ce_stage(active_cache[pos])) {
857                 case 0:
858                 case 2:
859                         return 0;
860                 }
861                 pos++;
862         }
863         return file_exists(path);
864 }
865
866 static int was_dirty(struct merge_options *o, const char *path)
867 {
868         struct cache_entry *ce;
869         int dirty = 1;
870
871         if (o->call_depth || !was_tracked(o, path))
872                 return !dirty;
873
874         ce = index_file_exists(o->unpack_opts.src_index,
875                                path, strlen(path), ignore_case);
876         dirty = verify_uptodate(ce, &o->unpack_opts) != 0;
877         return dirty;
878 }
879
880 static int make_room_for_path(struct merge_options *o, const char *path)
881 {
882         int status, i;
883         const char *msg = _("failed to create path '%s'%s");
884
885         /* Unlink any D/F conflict files that are in the way */
886         for (i = 0; i < o->df_conflict_file_set.nr; i++) {
887                 const char *df_path = o->df_conflict_file_set.items[i].string;
888                 size_t pathlen = strlen(path);
889                 size_t df_pathlen = strlen(df_path);
890                 if (df_pathlen < pathlen &&
891                     path[df_pathlen] == '/' &&
892                     strncmp(path, df_path, df_pathlen) == 0) {
893                         output(o, 3,
894                                _("Removing %s to make room for subdirectory\n"),
895                                df_path);
896                         unlink(df_path);
897                         unsorted_string_list_delete_item(&o->df_conflict_file_set,
898                                                          i, 0);
899                         break;
900                 }
901         }
902
903         /* Make sure leading directories are created */
904         status = safe_create_leading_directories_const(path);
905         if (status) {
906                 if (status == SCLD_EXISTS)
907                         /* something else exists */
908                         return err(o, msg, path, _(": perhaps a D/F conflict?"));
909                 return err(o, msg, path, "");
910         }
911
912         /*
913          * Do not unlink a file in the work tree if we are not
914          * tracking it.
915          */
916         if (would_lose_untracked(path))
917                 return err(o, _("refusing to lose untracked file at '%s'"),
918                              path);
919
920         /* Successful unlink is good.. */
921         if (!unlink(path))
922                 return 0;
923         /* .. and so is no existing file */
924         if (errno == ENOENT)
925                 return 0;
926         /* .. but not some other error (who really cares what?) */
927         return err(o, msg, path, _(": perhaps a D/F conflict?"));
928 }
929
930 static int update_file_flags(struct merge_options *o,
931                              const struct object_id *oid,
932                              unsigned mode,
933                              const char *path,
934                              int update_cache,
935                              int update_wd)
936 {
937         int ret = 0;
938
939         if (o->call_depth)
940                 update_wd = 0;
941
942         if (update_wd) {
943                 enum object_type type;
944                 void *buf;
945                 unsigned long size;
946
947                 if (S_ISGITLINK(mode)) {
948                         /*
949                          * We may later decide to recursively descend into
950                          * the submodule directory and update its index
951                          * and/or work tree, but we do not do that now.
952                          */
953                         update_wd = 0;
954                         goto update_index;
955                 }
956
957                 buf = read_object_file(oid, &type, &size);
958                 if (!buf)
959                         return err(o, _("cannot read object %s '%s'"), oid_to_hex(oid), path);
960                 if (type != OBJ_BLOB) {
961                         ret = err(o, _("blob expected for %s '%s'"), oid_to_hex(oid), path);
962                         goto free_buf;
963                 }
964                 if (S_ISREG(mode)) {
965                         struct strbuf strbuf = STRBUF_INIT;
966                         if (convert_to_working_tree(path, buf, size, &strbuf)) {
967                                 free(buf);
968                                 size = strbuf.len;
969                                 buf = strbuf_detach(&strbuf, NULL);
970                         }
971                 }
972
973                 if (make_room_for_path(o, path) < 0) {
974                         update_wd = 0;
975                         goto free_buf;
976                 }
977                 if (S_ISREG(mode) || (!has_symlinks && S_ISLNK(mode))) {
978                         int fd;
979                         if (mode & 0100)
980                                 mode = 0777;
981                         else
982                                 mode = 0666;
983                         fd = open(path, O_WRONLY | O_TRUNC | O_CREAT, mode);
984                         if (fd < 0) {
985                                 ret = err(o, _("failed to open '%s': %s"),
986                                           path, strerror(errno));
987                                 goto free_buf;
988                         }
989                         write_in_full(fd, buf, size);
990                         close(fd);
991                 } else if (S_ISLNK(mode)) {
992                         char *lnk = xmemdupz(buf, size);
993                         safe_create_leading_directories_const(path);
994                         unlink(path);
995                         if (symlink(lnk, path))
996                                 ret = err(o, _("failed to symlink '%s': %s"),
997                                         path, strerror(errno));
998                         free(lnk);
999                 } else
1000                         ret = err(o,
1001                                   _("do not know what to do with %06o %s '%s'"),
1002                                   mode, oid_to_hex(oid), path);
1003  free_buf:
1004                 free(buf);
1005         }
1006  update_index:
1007         if (!ret && update_cache)
1008                 if (add_cacheinfo(o, mode, oid, path, 0, update_wd,
1009                                   ADD_CACHE_OK_TO_ADD))
1010                         return -1;
1011         return ret;
1012 }
1013
1014 static int update_file(struct merge_options *o,
1015                        int clean,
1016                        const struct object_id *oid,
1017                        unsigned mode,
1018                        const char *path)
1019 {
1020         return update_file_flags(o, oid, mode, path, o->call_depth || clean, !o->call_depth);
1021 }
1022
1023 /* Low level file merging, update and removal */
1024
1025 struct merge_file_info {
1026         struct object_id oid;
1027         unsigned mode;
1028         unsigned clean:1,
1029                  merge:1;
1030 };
1031
1032 static int merge_3way(struct merge_options *o,
1033                       mmbuffer_t *result_buf,
1034                       const struct diff_filespec *one,
1035                       const struct diff_filespec *a,
1036                       const struct diff_filespec *b,
1037                       const char *branch1,
1038                       const char *branch2)
1039 {
1040         mmfile_t orig, src1, src2;
1041         struct ll_merge_options ll_opts = {0};
1042         char *base_name, *name1, *name2;
1043         int merge_status;
1044
1045         ll_opts.renormalize = o->renormalize;
1046         ll_opts.xdl_opts = o->xdl_opts;
1047
1048         if (o->call_depth) {
1049                 ll_opts.virtual_ancestor = 1;
1050                 ll_opts.variant = 0;
1051         } else {
1052                 switch (o->recursive_variant) {
1053                 case MERGE_RECURSIVE_OURS:
1054                         ll_opts.variant = XDL_MERGE_FAVOR_OURS;
1055                         break;
1056                 case MERGE_RECURSIVE_THEIRS:
1057                         ll_opts.variant = XDL_MERGE_FAVOR_THEIRS;
1058                         break;
1059                 default:
1060                         ll_opts.variant = 0;
1061                         break;
1062                 }
1063         }
1064
1065         if (strcmp(a->path, b->path) ||
1066             (o->ancestor != NULL && strcmp(a->path, one->path) != 0)) {
1067                 base_name = o->ancestor == NULL ? NULL :
1068                         mkpathdup("%s:%s", o->ancestor, one->path);
1069                 name1 = mkpathdup("%s:%s", branch1, a->path);
1070                 name2 = mkpathdup("%s:%s", branch2, b->path);
1071         } else {
1072                 base_name = o->ancestor == NULL ? NULL :
1073                         mkpathdup("%s", o->ancestor);
1074                 name1 = mkpathdup("%s", branch1);
1075                 name2 = mkpathdup("%s", branch2);
1076         }
1077
1078         read_mmblob(&orig, &one->oid);
1079         read_mmblob(&src1, &a->oid);
1080         read_mmblob(&src2, &b->oid);
1081
1082         merge_status = ll_merge(result_buf, a->path, &orig, base_name,
1083                                 &src1, name1, &src2, name2, &ll_opts);
1084
1085         free(base_name);
1086         free(name1);
1087         free(name2);
1088         free(orig.ptr);
1089         free(src1.ptr);
1090         free(src2.ptr);
1091         return merge_status;
1092 }
1093
1094 static int find_first_merges(struct object_array *result, const char *path,
1095                 struct commit *a, struct commit *b)
1096 {
1097         int i, j;
1098         struct object_array merges = OBJECT_ARRAY_INIT;
1099         struct commit *commit;
1100         int contains_another;
1101
1102         char merged_revision[42];
1103         const char *rev_args[] = { "rev-list", "--merges", "--ancestry-path",
1104                                    "--all", merged_revision, NULL };
1105         struct rev_info revs;
1106         struct setup_revision_opt rev_opts;
1107
1108         memset(result, 0, sizeof(struct object_array));
1109         memset(&rev_opts, 0, sizeof(rev_opts));
1110
1111         /* get all revisions that merge commit a */
1112         xsnprintf(merged_revision, sizeof(merged_revision), "^%s",
1113                         oid_to_hex(&a->object.oid));
1114         init_revisions(&revs, NULL);
1115         rev_opts.submodule = path;
1116         /* FIXME: can't handle linked worktrees in submodules yet */
1117         revs.single_worktree = path != NULL;
1118         setup_revisions(ARRAY_SIZE(rev_args)-1, rev_args, &revs, &rev_opts);
1119
1120         /* save all revisions from the above list that contain b */
1121         if (prepare_revision_walk(&revs))
1122                 die("revision walk setup failed");
1123         while ((commit = get_revision(&revs)) != NULL) {
1124                 struct object *o = &(commit->object);
1125                 if (in_merge_bases(b, commit))
1126                         add_object_array(o, NULL, &merges);
1127         }
1128         reset_revision_walk();
1129
1130         /* Now we've got all merges that contain a and b. Prune all
1131          * merges that contain another found merge and save them in
1132          * result.
1133          */
1134         for (i = 0; i < merges.nr; i++) {
1135                 struct commit *m1 = (struct commit *) merges.objects[i].item;
1136
1137                 contains_another = 0;
1138                 for (j = 0; j < merges.nr; j++) {
1139                         struct commit *m2 = (struct commit *) merges.objects[j].item;
1140                         if (i != j && in_merge_bases(m2, m1)) {
1141                                 contains_another = 1;
1142                                 break;
1143                         }
1144                 }
1145
1146                 if (!contains_another)
1147                         add_object_array(merges.objects[i].item, NULL, result);
1148         }
1149
1150         object_array_clear(&merges);
1151         return result->nr;
1152 }
1153
1154 static void print_commit(struct commit *commit)
1155 {
1156         struct strbuf sb = STRBUF_INIT;
1157         struct pretty_print_context ctx = {0};
1158         ctx.date_mode.type = DATE_NORMAL;
1159         format_commit_message(commit, " %h: %m %s", &sb, &ctx);
1160         fprintf(stderr, "%s\n", sb.buf);
1161         strbuf_release(&sb);
1162 }
1163
1164 static int merge_submodule(struct merge_options *o,
1165                            struct object_id *result, const char *path,
1166                            const struct object_id *base, const struct object_id *a,
1167                            const struct object_id *b)
1168 {
1169         struct commit *commit_base, *commit_a, *commit_b;
1170         int parent_count;
1171         struct object_array merges;
1172
1173         int i;
1174         int search = !o->call_depth;
1175
1176         /* store a in result in case we fail */
1177         oidcpy(result, a);
1178
1179         /* we can not handle deletion conflicts */
1180         if (is_null_oid(base))
1181                 return 0;
1182         if (is_null_oid(a))
1183                 return 0;
1184         if (is_null_oid(b))
1185                 return 0;
1186
1187         if (add_submodule_odb(path)) {
1188                 output(o, 1, _("Failed to merge submodule %s (not checked out)"), path);
1189                 return 0;
1190         }
1191
1192         if (!(commit_base = lookup_commit_reference(base)) ||
1193             !(commit_a = lookup_commit_reference(a)) ||
1194             !(commit_b = lookup_commit_reference(b))) {
1195                 output(o, 1, _("Failed to merge submodule %s (commits not present)"), path);
1196                 return 0;
1197         }
1198
1199         /* check whether both changes are forward */
1200         if (!in_merge_bases(commit_base, commit_a) ||
1201             !in_merge_bases(commit_base, commit_b)) {
1202                 output(o, 1, _("Failed to merge submodule %s (commits don't follow merge-base)"), path);
1203                 return 0;
1204         }
1205
1206         /* Case #1: a is contained in b or vice versa */
1207         if (in_merge_bases(commit_a, commit_b)) {
1208                 oidcpy(result, b);
1209                 if (show(o, 3)) {
1210                         output(o, 3, _("Fast-forwarding submodule %s to the following commit:"), path);
1211                         output_commit_title(o, commit_b);
1212                 } else if (show(o, 2))
1213                         output(o, 2, _("Fast-forwarding submodule %s"), path);
1214                 else
1215                         ; /* no output */
1216
1217                 return 1;
1218         }
1219         if (in_merge_bases(commit_b, commit_a)) {
1220                 oidcpy(result, a);
1221                 if (show(o, 3)) {
1222                         output(o, 3, _("Fast-forwarding submodule %s to the following commit:"), path);
1223                         output_commit_title(o, commit_a);
1224                 } else if (show(o, 2))
1225                         output(o, 2, _("Fast-forwarding submodule %s"), path);
1226                 else
1227                         ; /* no output */
1228
1229                 return 1;
1230         }
1231
1232         /*
1233          * Case #2: There are one or more merges that contain a and b in
1234          * the submodule. If there is only one, then present it as a
1235          * suggestion to the user, but leave it marked unmerged so the
1236          * user needs to confirm the resolution.
1237          */
1238
1239         /* Skip the search if makes no sense to the calling context.  */
1240         if (!search)
1241                 return 0;
1242
1243         /* find commit which merges them */
1244         parent_count = find_first_merges(&merges, path, commit_a, commit_b);
1245         switch (parent_count) {
1246         case 0:
1247                 output(o, 1, _("Failed to merge submodule %s (merge following commits not found)"), path);
1248                 break;
1249
1250         case 1:
1251                 output(o, 1, _("Failed to merge submodule %s (not fast-forward)"), path);
1252                 output(o, 2, _("Found a possible merge resolution for the submodule:\n"));
1253                 print_commit((struct commit *) merges.objects[0].item);
1254                 output(o, 2, _(
1255                         "If this is correct simply add it to the index "
1256                         "for example\n"
1257                         "by using:\n\n"
1258                         "  git update-index --cacheinfo 160000 %s \"%s\"\n\n"
1259                         "which will accept this suggestion.\n"),
1260                         oid_to_hex(&merges.objects[0].item->oid), path);
1261                 break;
1262
1263         default:
1264                 output(o, 1, _("Failed to merge submodule %s (multiple merges found)"), path);
1265                 for (i = 0; i < merges.nr; i++)
1266                         print_commit((struct commit *) merges.objects[i].item);
1267         }
1268
1269         object_array_clear(&merges);
1270         return 0;
1271 }
1272
1273 static int merge_file_1(struct merge_options *o,
1274                         const struct diff_filespec *one,
1275                         const struct diff_filespec *a,
1276                         const struct diff_filespec *b,
1277                         const char *filename,
1278                         const char *branch1,
1279                         const char *branch2,
1280                         struct merge_file_info *result)
1281 {
1282         result->merge = 0;
1283         result->clean = 1;
1284
1285         if ((S_IFMT & a->mode) != (S_IFMT & b->mode)) {
1286                 result->clean = 0;
1287                 if (S_ISREG(a->mode)) {
1288                         result->mode = a->mode;
1289                         oidcpy(&result->oid, &a->oid);
1290                 } else {
1291                         result->mode = b->mode;
1292                         oidcpy(&result->oid, &b->oid);
1293                 }
1294         } else {
1295                 if (!oid_eq(&a->oid, &one->oid) && !oid_eq(&b->oid, &one->oid))
1296                         result->merge = 1;
1297
1298                 /*
1299                  * Merge modes
1300                  */
1301                 if (a->mode == b->mode || a->mode == one->mode)
1302                         result->mode = b->mode;
1303                 else {
1304                         result->mode = a->mode;
1305                         if (b->mode != one->mode) {
1306                                 result->clean = 0;
1307                                 result->merge = 1;
1308                         }
1309                 }
1310
1311                 if (oid_eq(&a->oid, &b->oid) || oid_eq(&a->oid, &one->oid))
1312                         oidcpy(&result->oid, &b->oid);
1313                 else if (oid_eq(&b->oid, &one->oid))
1314                         oidcpy(&result->oid, &a->oid);
1315                 else if (S_ISREG(a->mode)) {
1316                         mmbuffer_t result_buf;
1317                         int ret = 0, merge_status;
1318
1319                         merge_status = merge_3way(o, &result_buf, one, a, b,
1320                                                   branch1, branch2);
1321
1322                         if ((merge_status < 0) || !result_buf.ptr)
1323                                 ret = err(o, _("Failed to execute internal merge"));
1324
1325                         if (!ret &&
1326                             write_object_file(result_buf.ptr, result_buf.size,
1327                                               blob_type, &result->oid))
1328                                 ret = err(o, _("Unable to add %s to database"),
1329                                           a->path);
1330
1331                         free(result_buf.ptr);
1332                         if (ret)
1333                                 return ret;
1334                         result->clean = (merge_status == 0);
1335                 } else if (S_ISGITLINK(a->mode)) {
1336                         result->clean = merge_submodule(o, &result->oid,
1337                                                        one->path,
1338                                                        &one->oid,
1339                                                        &a->oid,
1340                                                        &b->oid);
1341                 } else if (S_ISLNK(a->mode)) {
1342                         switch (o->recursive_variant) {
1343                         case MERGE_RECURSIVE_NORMAL:
1344                                 oidcpy(&result->oid, &a->oid);
1345                                 if (!oid_eq(&a->oid, &b->oid))
1346                                         result->clean = 0;
1347                                 break;
1348                         case MERGE_RECURSIVE_OURS:
1349                                 oidcpy(&result->oid, &a->oid);
1350                                 break;
1351                         case MERGE_RECURSIVE_THEIRS:
1352                                 oidcpy(&result->oid, &b->oid);
1353                                 break;
1354                         }
1355                 } else
1356                         BUG("unsupported object type in the tree");
1357         }
1358
1359         if (result->merge)
1360                 output(o, 2, _("Auto-merging %s"), filename);
1361
1362         return 0;
1363 }
1364
1365 static int merge_file_special_markers(struct merge_options *o,
1366                                       const struct diff_filespec *one,
1367                                       const struct diff_filespec *a,
1368                                       const struct diff_filespec *b,
1369                                       const char *target_filename,
1370                                       const char *branch1,
1371                                       const char *filename1,
1372                                       const char *branch2,
1373                                       const char *filename2,
1374                                       struct merge_file_info *mfi)
1375 {
1376         char *side1 = NULL;
1377         char *side2 = NULL;
1378         int ret;
1379
1380         if (filename1)
1381                 side1 = xstrfmt("%s:%s", branch1, filename1);
1382         if (filename2)
1383                 side2 = xstrfmt("%s:%s", branch2, filename2);
1384
1385         ret = merge_file_1(o, one, a, b, target_filename,
1386                            side1 ? side1 : branch1,
1387                            side2 ? side2 : branch2, mfi);
1388
1389         free(side1);
1390         free(side2);
1391         return ret;
1392 }
1393
1394 static int merge_file_one(struct merge_options *o,
1395                           const char *path,
1396                           const struct object_id *o_oid, int o_mode,
1397                           const struct object_id *a_oid, int a_mode,
1398                           const struct object_id *b_oid, int b_mode,
1399                           const char *branch1,
1400                           const char *branch2,
1401                           struct merge_file_info *mfi)
1402 {
1403         struct diff_filespec one, a, b;
1404
1405         one.path = a.path = b.path = (char *)path;
1406         oidcpy(&one.oid, o_oid);
1407         one.mode = o_mode;
1408         oidcpy(&a.oid, a_oid);
1409         a.mode = a_mode;
1410         oidcpy(&b.oid, b_oid);
1411         b.mode = b_mode;
1412         return merge_file_1(o, &one, &a, &b, path, branch1, branch2, mfi);
1413 }
1414
1415 static int conflict_rename_dir(struct merge_options *o,
1416                                struct diff_filepair *pair,
1417                                const char *rename_branch,
1418                                const char *other_branch)
1419 {
1420         const struct diff_filespec *dest = pair->two;
1421
1422         if (!o->call_depth && would_lose_untracked(dest->path)) {
1423                 char *alt_path = unique_path(o, dest->path, rename_branch);
1424
1425                 output(o, 1, _("Error: Refusing to lose untracked file at %s; "
1426                                "writing to %s instead."),
1427                        dest->path, alt_path);
1428                 /*
1429                  * Write the file in worktree at alt_path, but not in the
1430                  * index.  Instead, write to dest->path for the index but
1431                  * only at the higher appropriate stage.
1432                  */
1433                 if (update_file(o, 0, &dest->oid, dest->mode, alt_path))
1434                         return -1;
1435                 free(alt_path);
1436                 return update_stages(o, dest->path, NULL,
1437                                      rename_branch == o->branch1 ? dest : NULL,
1438                                      rename_branch == o->branch1 ? NULL : dest);
1439         }
1440
1441         /* Update dest->path both in index and in worktree */
1442         if (update_file(o, 1, &dest->oid, dest->mode, dest->path))
1443                 return -1;
1444         return 0;
1445 }
1446
1447 static int handle_change_delete(struct merge_options *o,
1448                                  const char *path, const char *old_path,
1449                                  const struct object_id *o_oid, int o_mode,
1450                                  const struct object_id *changed_oid,
1451                                  int changed_mode,
1452                                  const char *change_branch,
1453                                  const char *delete_branch,
1454                                  const char *change, const char *change_past)
1455 {
1456         char *alt_path = NULL;
1457         const char *update_path = path;
1458         int ret = 0;
1459
1460         if (dir_in_way(path, !o->call_depth, 0) ||
1461             (!o->call_depth && would_lose_untracked(path))) {
1462                 update_path = alt_path = unique_path(o, path, change_branch);
1463         }
1464
1465         if (o->call_depth) {
1466                 /*
1467                  * We cannot arbitrarily accept either a_sha or b_sha as
1468                  * correct; since there is no true "middle point" between
1469                  * them, simply reuse the base version for virtual merge base.
1470                  */
1471                 ret = remove_file_from_cache(path);
1472                 if (!ret)
1473                         ret = update_file(o, 0, o_oid, o_mode, update_path);
1474         } else {
1475                 if (!alt_path) {
1476                         if (!old_path) {
1477                                 output(o, 1, _("CONFLICT (%s/delete): %s deleted in %s "
1478                                        "and %s in %s. Version %s of %s left in tree."),
1479                                        change, path, delete_branch, change_past,
1480                                        change_branch, change_branch, path);
1481                         } else {
1482                                 output(o, 1, _("CONFLICT (%s/delete): %s deleted in %s "
1483                                        "and %s to %s in %s. Version %s of %s left in tree."),
1484                                        change, old_path, delete_branch, change_past, path,
1485                                        change_branch, change_branch, path);
1486                         }
1487                 } else {
1488                         if (!old_path) {
1489                                 output(o, 1, _("CONFLICT (%s/delete): %s deleted in %s "
1490                                        "and %s in %s. Version %s of %s left in tree at %s."),
1491                                        change, path, delete_branch, change_past,
1492                                        change_branch, change_branch, path, alt_path);
1493                         } else {
1494                                 output(o, 1, _("CONFLICT (%s/delete): %s deleted in %s "
1495                                        "and %s to %s in %s. Version %s of %s left in tree at %s."),
1496                                        change, old_path, delete_branch, change_past, path,
1497                                        change_branch, change_branch, path, alt_path);
1498                         }
1499                 }
1500                 /*
1501                  * No need to call update_file() on path when change_branch ==
1502                  * o->branch1 && !alt_path, since that would needlessly touch
1503                  * path.  We could call update_file_flags() with update_cache=0
1504                  * and update_wd=0, but that's a no-op.
1505                  */
1506                 if (change_branch != o->branch1 || alt_path)
1507                         ret = update_file(o, 0, changed_oid, changed_mode, update_path);
1508         }
1509         free(alt_path);
1510
1511         return ret;
1512 }
1513
1514 static int conflict_rename_delete(struct merge_options *o,
1515                                    struct diff_filepair *pair,
1516                                    const char *rename_branch,
1517                                    const char *delete_branch)
1518 {
1519         const struct diff_filespec *orig = pair->one;
1520         const struct diff_filespec *dest = pair->two;
1521
1522         if (handle_change_delete(o,
1523                                  o->call_depth ? orig->path : dest->path,
1524                                  o->call_depth ? NULL : orig->path,
1525                                  &orig->oid, orig->mode,
1526                                  &dest->oid, dest->mode,
1527                                  rename_branch, delete_branch,
1528                                  _("rename"), _("renamed")))
1529                 return -1;
1530
1531         if (o->call_depth)
1532                 return remove_file_from_cache(dest->path);
1533         else
1534                 return update_stages(o, dest->path, NULL,
1535                                      rename_branch == o->branch1 ? dest : NULL,
1536                                      rename_branch == o->branch1 ? NULL : dest);
1537 }
1538
1539 static struct diff_filespec *filespec_from_entry(struct diff_filespec *target,
1540                                                  struct stage_data *entry,
1541                                                  int stage)
1542 {
1543         struct object_id *oid = &entry->stages[stage].oid;
1544         unsigned mode = entry->stages[stage].mode;
1545         if (mode == 0 || is_null_oid(oid))
1546                 return NULL;
1547         oidcpy(&target->oid, oid);
1548         target->mode = mode;
1549         return target;
1550 }
1551
1552 static int handle_file(struct merge_options *o,
1553                         struct diff_filespec *rename,
1554                         int stage,
1555                         struct rename_conflict_info *ci)
1556 {
1557         char *dst_name = rename->path;
1558         struct stage_data *dst_entry;
1559         const char *cur_branch, *other_branch;
1560         struct diff_filespec other;
1561         struct diff_filespec *add;
1562         int ret;
1563
1564         if (stage == 2) {
1565                 dst_entry = ci->dst_entry1;
1566                 cur_branch = ci->branch1;
1567                 other_branch = ci->branch2;
1568         } else {
1569                 dst_entry = ci->dst_entry2;
1570                 cur_branch = ci->branch2;
1571                 other_branch = ci->branch1;
1572         }
1573
1574         add = filespec_from_entry(&other, dst_entry, stage ^ 1);
1575         if (add) {
1576                 int ren_src_was_dirty = was_dirty(o, rename->path);
1577                 char *add_name = unique_path(o, rename->path, other_branch);
1578                 if (update_file(o, 0, &add->oid, add->mode, add_name))
1579                         return -1;
1580
1581                 if (ren_src_was_dirty) {
1582                         output(o, 1, _("Refusing to lose dirty file at %s"),
1583                                rename->path);
1584                 }
1585                 /*
1586                  * Because the double negatives somehow keep confusing me...
1587                  *    1) update_wd iff !ren_src_was_dirty.
1588                  *    2) no_wd iff !update_wd
1589                  *    3) so, no_wd == !!ren_src_was_dirty == ren_src_was_dirty
1590                  */
1591                 remove_file(o, 0, rename->path, ren_src_was_dirty);
1592                 dst_name = unique_path(o, rename->path, cur_branch);
1593         } else {
1594                 if (dir_in_way(rename->path, !o->call_depth, 0)) {
1595                         dst_name = unique_path(o, rename->path, cur_branch);
1596                         output(o, 1, _("%s is a directory in %s adding as %s instead"),
1597                                rename->path, other_branch, dst_name);
1598                 } else if (!o->call_depth &&
1599                            would_lose_untracked(rename->path)) {
1600                         dst_name = unique_path(o, rename->path, cur_branch);
1601                         output(o, 1, _("Refusing to lose untracked file at %s; "
1602                                        "adding as %s instead"),
1603                                rename->path, dst_name);
1604                 }
1605         }
1606         if ((ret = update_file(o, 0, &rename->oid, rename->mode, dst_name)))
1607                 ; /* fall through, do allow dst_name to be released */
1608         else if (stage == 2)
1609                 ret = update_stages(o, rename->path, NULL, rename, add);
1610         else
1611                 ret = update_stages(o, rename->path, NULL, add, rename);
1612
1613         if (dst_name != rename->path)
1614                 free(dst_name);
1615
1616         return ret;
1617 }
1618
1619 static int conflict_rename_rename_1to2(struct merge_options *o,
1620                                         struct rename_conflict_info *ci)
1621 {
1622         /* One file was renamed in both branches, but to different names. */
1623         struct diff_filespec *one = ci->pair1->one;
1624         struct diff_filespec *a = ci->pair1->two;
1625         struct diff_filespec *b = ci->pair2->two;
1626
1627         output(o, 1, _("CONFLICT (rename/rename): "
1628                "Rename \"%s\"->\"%s\" in branch \"%s\" "
1629                "rename \"%s\"->\"%s\" in \"%s\"%s"),
1630                one->path, a->path, ci->branch1,
1631                one->path, b->path, ci->branch2,
1632                o->call_depth ? _(" (left unresolved)") : "");
1633         if (o->call_depth) {
1634                 struct merge_file_info mfi;
1635                 struct diff_filespec other;
1636                 struct diff_filespec *add;
1637                 if (merge_file_one(o, one->path,
1638                                  &one->oid, one->mode,
1639                                  &a->oid, a->mode,
1640                                  &b->oid, b->mode,
1641                                  ci->branch1, ci->branch2, &mfi))
1642                         return -1;
1643
1644                 /*
1645                  * FIXME: For rename/add-source conflicts (if we could detect
1646                  * such), this is wrong.  We should instead find a unique
1647                  * pathname and then either rename the add-source file to that
1648                  * unique path, or use that unique path instead of src here.
1649                  */
1650                 if (update_file(o, 0, &mfi.oid, mfi.mode, one->path))
1651                         return -1;
1652
1653                 /*
1654                  * Above, we put the merged content at the merge-base's
1655                  * path.  Now we usually need to delete both a->path and
1656                  * b->path.  However, the rename on each side of the merge
1657                  * could also be involved in a rename/add conflict.  In
1658                  * such cases, we should keep the added file around,
1659                  * resolving the conflict at that path in its favor.
1660                  */
1661                 add = filespec_from_entry(&other, ci->dst_entry1, 2 ^ 1);
1662                 if (add) {
1663                         if (update_file(o, 0, &add->oid, add->mode, a->path))
1664                                 return -1;
1665                 }
1666                 else
1667                         remove_file_from_cache(a->path);
1668                 add = filespec_from_entry(&other, ci->dst_entry2, 3 ^ 1);
1669                 if (add) {
1670                         if (update_file(o, 0, &add->oid, add->mode, b->path))
1671                                 return -1;
1672                 }
1673                 else
1674                         remove_file_from_cache(b->path);
1675         } else if (handle_file(o, a, 2, ci) || handle_file(o, b, 3, ci))
1676                 return -1;
1677
1678         return 0;
1679 }
1680
1681 static int conflict_rename_rename_2to1(struct merge_options *o,
1682                                         struct rename_conflict_info *ci)
1683 {
1684         /* Two files, a & b, were renamed to the same thing, c. */
1685         struct diff_filespec *a = ci->pair1->one;
1686         struct diff_filespec *b = ci->pair2->one;
1687         struct diff_filespec *c1 = ci->pair1->two;
1688         struct diff_filespec *c2 = ci->pair2->two;
1689         char *path = c1->path; /* == c2->path */
1690         char *path_side_1_desc;
1691         char *path_side_2_desc;
1692         struct merge_file_info mfi_c1;
1693         struct merge_file_info mfi_c2;
1694         int ret;
1695
1696         output(o, 1, _("CONFLICT (rename/rename): "
1697                "Rename %s->%s in %s. "
1698                "Rename %s->%s in %s"),
1699                a->path, c1->path, ci->branch1,
1700                b->path, c2->path, ci->branch2);
1701
1702         remove_file(o, 1, a->path, o->call_depth || would_lose_untracked(a->path));
1703         remove_file(o, 1, b->path, o->call_depth || would_lose_untracked(b->path));
1704
1705         path_side_1_desc = xstrfmt("%s (was %s)", path, a->path);
1706         path_side_2_desc = xstrfmt("%s (was %s)", path, b->path);
1707         if (merge_file_special_markers(o, a, c1, &ci->ren1_other,
1708                                        path_side_1_desc,
1709                                        o->branch1, c1->path,
1710                                        o->branch2, ci->ren1_other.path, &mfi_c1) ||
1711             merge_file_special_markers(o, b, &ci->ren2_other, c2,
1712                                        path_side_2_desc,
1713                                        o->branch1, ci->ren2_other.path,
1714                                        o->branch2, c2->path, &mfi_c2))
1715                 return -1;
1716         free(path_side_1_desc);
1717         free(path_side_2_desc);
1718
1719         if (o->call_depth) {
1720                 /*
1721                  * If mfi_c1.clean && mfi_c2.clean, then it might make
1722                  * sense to do a two-way merge of those results.  But, I
1723                  * think in all cases, it makes sense to have the virtual
1724                  * merge base just undo the renames; they can be detected
1725                  * again later for the non-recursive merge.
1726                  */
1727                 remove_file(o, 0, path, 0);
1728                 ret = update_file(o, 0, &mfi_c1.oid, mfi_c1.mode, a->path);
1729                 if (!ret)
1730                         ret = update_file(o, 0, &mfi_c2.oid, mfi_c2.mode,
1731                                           b->path);
1732         } else {
1733                 char *new_path1 = unique_path(o, path, ci->branch1);
1734                 char *new_path2 = unique_path(o, path, ci->branch2);
1735                 output(o, 1, _("Renaming %s to %s and %s to %s instead"),
1736                        a->path, new_path1, b->path, new_path2);
1737                 if (was_dirty(o, path))
1738                         output(o, 1, _("Refusing to lose dirty file at %s"),
1739                                path);
1740                 else if (would_lose_untracked(path))
1741                         /*
1742                          * Only way we get here is if both renames were from
1743                          * a directory rename AND user had an untracked file
1744                          * at the location where both files end up after the
1745                          * two directory renames.  See testcase 10d of t6043.
1746                          */
1747                         output(o, 1, _("Refusing to lose untracked file at "
1748                                        "%s, even though it's in the way."),
1749                                path);
1750                 else
1751                         remove_file(o, 0, path, 0);
1752                 ret = update_file(o, 0, &mfi_c1.oid, mfi_c1.mode, new_path1);
1753                 if (!ret)
1754                         ret = update_file(o, 0, &mfi_c2.oid, mfi_c2.mode,
1755                                           new_path2);
1756                 /*
1757                  * unpack_trees() actually populates the index for us for
1758                  * "normal" rename/rename(2to1) situtations so that the
1759                  * correct entries are at the higher stages, which would
1760                  * make the call below to update_stages_for_stage_data
1761                  * unnecessary.  However, if either of the renames came
1762                  * from a directory rename, then unpack_trees() will not
1763                  * have gotten the right data loaded into the index, so we
1764                  * need to do so now.  (While it'd be tempting to move this
1765                  * call to update_stages_for_stage_data() to
1766                  * apply_directory_rename_modifications(), that would break
1767                  * our intermediate calls to would_lose_untracked() since
1768                  * those rely on the current in-memory index.  See also the
1769                  * big "NOTE" in update_stages()).
1770                  */
1771                 if (update_stages_for_stage_data(o, path, ci->dst_entry1))
1772                         ret = -1;
1773
1774                 free(new_path2);
1775                 free(new_path1);
1776         }
1777
1778         return ret;
1779 }
1780
1781 /*
1782  * Get the diff_filepairs changed between o_tree and tree.
1783  */
1784 static struct diff_queue_struct *get_diffpairs(struct merge_options *o,
1785                                                struct tree *o_tree,
1786                                                struct tree *tree)
1787 {
1788         struct diff_queue_struct *ret;
1789         struct diff_options opts;
1790
1791         diff_setup(&opts);
1792         opts.flags.recursive = 1;
1793         opts.flags.rename_empty = 0;
1794         opts.detect_rename = merge_detect_rename(o);
1795         /*
1796          * We do not have logic to handle the detection of copies.  In
1797          * fact, it may not even make sense to add such logic: would we
1798          * really want a change to a base file to be propagated through
1799          * multiple other files by a merge?
1800          */
1801         if (opts.detect_rename > DIFF_DETECT_RENAME)
1802                 opts.detect_rename = DIFF_DETECT_RENAME;
1803         opts.rename_limit = o->merge_rename_limit >= 0 ? o->merge_rename_limit :
1804                             o->diff_rename_limit >= 0 ? o->diff_rename_limit :
1805                             1000;
1806         opts.rename_score = o->rename_score;
1807         opts.show_rename_progress = o->show_rename_progress;
1808         opts.output_format = DIFF_FORMAT_NO_OUTPUT;
1809         diff_setup_done(&opts);
1810         diff_tree_oid(&o_tree->object.oid, &tree->object.oid, "", &opts);
1811         diffcore_std(&opts);
1812         if (opts.needed_rename_limit > o->needed_rename_limit)
1813                 o->needed_rename_limit = opts.needed_rename_limit;
1814
1815         ret = xmalloc(sizeof(*ret));
1816         *ret = diff_queued_diff;
1817
1818         opts.output_format = DIFF_FORMAT_NO_OUTPUT;
1819         diff_queued_diff.nr = 0;
1820         diff_queued_diff.queue = NULL;
1821         diff_flush(&opts);
1822         return ret;
1823 }
1824
1825 static int tree_has_path(struct tree *tree, const char *path)
1826 {
1827         struct object_id hashy;
1828         unsigned int mode_o;
1829
1830         return !get_tree_entry(&tree->object.oid, path,
1831                                &hashy, &mode_o);
1832 }
1833
1834 /*
1835  * Return a new string that replaces the beginning portion (which matches
1836  * entry->dir), with entry->new_dir.  In perl-speak:
1837  *   new_path_name = (old_path =~ s/entry->dir/entry->new_dir/);
1838  * NOTE:
1839  *   Caller must ensure that old_path starts with entry->dir + '/'.
1840  */
1841 static char *apply_dir_rename(struct dir_rename_entry *entry,
1842                               const char *old_path)
1843 {
1844         struct strbuf new_path = STRBUF_INIT;
1845         int oldlen, newlen;
1846
1847         if (entry->non_unique_new_dir)
1848                 return NULL;
1849
1850         oldlen = strlen(entry->dir);
1851         newlen = entry->new_dir.len + (strlen(old_path) - oldlen) + 1;
1852         strbuf_grow(&new_path, newlen);
1853         strbuf_addbuf(&new_path, &entry->new_dir);
1854         strbuf_addstr(&new_path, &old_path[oldlen]);
1855
1856         return strbuf_detach(&new_path, NULL);
1857 }
1858
1859 static void get_renamed_dir_portion(const char *old_path, const char *new_path,
1860                                     char **old_dir, char **new_dir)
1861 {
1862         char *end_of_old, *end_of_new;
1863         int old_len, new_len;
1864
1865         *old_dir = NULL;
1866         *new_dir = NULL;
1867
1868         /*
1869          * For
1870          *    "a/b/c/d/e/foo.c" -> "a/b/some/thing/else/e/foo.c"
1871          * the "e/foo.c" part is the same, we just want to know that
1872          *    "a/b/c/d" was renamed to "a/b/some/thing/else"
1873          * so, for this example, this function returns "a/b/c/d" in
1874          * *old_dir and "a/b/some/thing/else" in *new_dir.
1875          *
1876          * Also, if the basename of the file changed, we don't care.  We
1877          * want to know which portion of the directory, if any, changed.
1878          */
1879         end_of_old = strrchr(old_path, '/');
1880         end_of_new = strrchr(new_path, '/');
1881
1882         if (end_of_old == NULL || end_of_new == NULL)
1883                 return;
1884         while (*--end_of_new == *--end_of_old &&
1885                end_of_old != old_path &&
1886                end_of_new != new_path)
1887                 ; /* Do nothing; all in the while loop */
1888         /*
1889          * We've found the first non-matching character in the directory
1890          * paths.  That means the current directory we were comparing
1891          * represents the rename.  Move end_of_old and end_of_new back
1892          * to the full directory name.
1893          */
1894         if (*end_of_old == '/')
1895                 end_of_old++;
1896         if (*end_of_old != '/')
1897                 end_of_new++;
1898         end_of_old = strchr(end_of_old, '/');
1899         end_of_new = strchr(end_of_new, '/');
1900
1901         /*
1902          * It may have been the case that old_path and new_path were the same
1903          * directory all along.  Don't claim a rename if they're the same.
1904          */
1905         old_len = end_of_old - old_path;
1906         new_len = end_of_new - new_path;
1907
1908         if (old_len != new_len || strncmp(old_path, new_path, old_len)) {
1909                 *old_dir = xstrndup(old_path, old_len);
1910                 *new_dir = xstrndup(new_path, new_len);
1911         }
1912 }
1913
1914 static void remove_hashmap_entries(struct hashmap *dir_renames,
1915                                    struct string_list *items_to_remove)
1916 {
1917         int i;
1918         struct dir_rename_entry *entry;
1919
1920         for (i = 0; i < items_to_remove->nr; i++) {
1921                 entry = items_to_remove->items[i].util;
1922                 hashmap_remove(dir_renames, entry, NULL);
1923         }
1924         string_list_clear(items_to_remove, 0);
1925 }
1926
1927 /*
1928  * See if there is a directory rename for path, and if there are any file
1929  * level conflicts for the renamed location.  If there is a rename and
1930  * there are no conflicts, return the new name.  Otherwise, return NULL.
1931  */
1932 static char *handle_path_level_conflicts(struct merge_options *o,
1933                                          const char *path,
1934                                          struct dir_rename_entry *entry,
1935                                          struct hashmap *collisions,
1936                                          struct tree *tree)
1937 {
1938         char *new_path = NULL;
1939         struct collision_entry *collision_ent;
1940         int clean = 1;
1941         struct strbuf collision_paths = STRBUF_INIT;
1942
1943         /*
1944          * entry has the mapping of old directory name to new directory name
1945          * that we want to apply to path.
1946          */
1947         new_path = apply_dir_rename(entry, path);
1948
1949         if (!new_path) {
1950                 /* This should only happen when entry->non_unique_new_dir set */
1951                 if (!entry->non_unique_new_dir)
1952                         BUG("entry->non_unqiue_dir not set and !new_path");
1953                 output(o, 1, _("CONFLICT (directory rename split): "
1954                                "Unclear where to place %s because directory "
1955                                "%s was renamed to multiple other directories, "
1956                                "with no destination getting a majority of the "
1957                                "files."),
1958                        path, entry->dir);
1959                 clean = 0;
1960                 return NULL;
1961         }
1962
1963         /*
1964          * The caller needs to have ensured that it has pre-populated
1965          * collisions with all paths that map to new_path.  Do a quick check
1966          * to ensure that's the case.
1967          */
1968         collision_ent = collision_find_entry(collisions, new_path);
1969         if (collision_ent == NULL)
1970                 BUG("collision_ent is NULL");
1971
1972         /*
1973          * Check for one-sided add/add/.../add conflicts, i.e.
1974          * where implicit renames from the other side doing
1975          * directory rename(s) can affect this side of history
1976          * to put multiple paths into the same location.  Warn
1977          * and bail on directory renames for such paths.
1978          */
1979         if (collision_ent->reported_already) {
1980                 clean = 0;
1981         } else if (tree_has_path(tree, new_path)) {
1982                 collision_ent->reported_already = 1;
1983                 strbuf_add_separated_string_list(&collision_paths, ", ",
1984                                                  &collision_ent->source_files);
1985                 output(o, 1, _("CONFLICT (implicit dir rename): Existing "
1986                                "file/dir at %s in the way of implicit "
1987                                "directory rename(s) putting the following "
1988                                "path(s) there: %s."),
1989                        new_path, collision_paths.buf);
1990                 clean = 0;
1991         } else if (collision_ent->source_files.nr > 1) {
1992                 collision_ent->reported_already = 1;
1993                 strbuf_add_separated_string_list(&collision_paths, ", ",
1994                                                  &collision_ent->source_files);
1995                 output(o, 1, _("CONFLICT (implicit dir rename): Cannot map "
1996                                "more than one path to %s; implicit directory "
1997                                "renames tried to put these paths there: %s"),
1998                        new_path, collision_paths.buf);
1999                 clean = 0;
2000         }
2001
2002         /* Free memory we no longer need */
2003         strbuf_release(&collision_paths);
2004         if (!clean && new_path) {
2005                 free(new_path);
2006                 return NULL;
2007         }
2008
2009         return new_path;
2010 }
2011
2012 /*
2013  * There are a couple things we want to do at the directory level:
2014  *   1. Check for both sides renaming to the same thing, in order to avoid
2015  *      implicit renaming of files that should be left in place.  (See
2016  *      testcase 6b in t6043 for details.)
2017  *   2. Prune directory renames if there are still files left in the
2018  *      the original directory.  These represent a partial directory rename,
2019  *      i.e. a rename where only some of the files within the directory
2020  *      were renamed elsewhere.  (Technically, this could be done earlier
2021  *      in get_directory_renames(), except that would prevent us from
2022  *      doing the previous check and thus failing testcase 6b.)
2023  *   3. Check for rename/rename(1to2) conflicts (at the directory level).
2024  *      In the future, we could potentially record this info as well and
2025  *      omit reporting rename/rename(1to2) conflicts for each path within
2026  *      the affected directories, thus cleaning up the merge output.
2027  *   NOTE: We do NOT check for rename/rename(2to1) conflicts at the
2028  *         directory level, because merging directories is fine.  If it
2029  *         causes conflicts for files within those merged directories, then
2030  *         that should be detected at the individual path level.
2031  */
2032 static void handle_directory_level_conflicts(struct merge_options *o,
2033                                              struct hashmap *dir_re_head,
2034                                              struct tree *head,
2035                                              struct hashmap *dir_re_merge,
2036                                              struct tree *merge)
2037 {
2038         struct hashmap_iter iter;
2039         struct dir_rename_entry *head_ent;
2040         struct dir_rename_entry *merge_ent;
2041
2042         struct string_list remove_from_head = STRING_LIST_INIT_NODUP;
2043         struct string_list remove_from_merge = STRING_LIST_INIT_NODUP;
2044
2045         hashmap_iter_init(dir_re_head, &iter);
2046         while ((head_ent = hashmap_iter_next(&iter))) {
2047                 merge_ent = dir_rename_find_entry(dir_re_merge, head_ent->dir);
2048                 if (merge_ent &&
2049                     !head_ent->non_unique_new_dir &&
2050                     !merge_ent->non_unique_new_dir &&
2051                     !strbuf_cmp(&head_ent->new_dir, &merge_ent->new_dir)) {
2052                         /* 1. Renamed identically; remove it from both sides */
2053                         string_list_append(&remove_from_head,
2054                                            head_ent->dir)->util = head_ent;
2055                         strbuf_release(&head_ent->new_dir);
2056                         string_list_append(&remove_from_merge,
2057                                            merge_ent->dir)->util = merge_ent;
2058                         strbuf_release(&merge_ent->new_dir);
2059                 } else if (tree_has_path(head, head_ent->dir)) {
2060                         /* 2. This wasn't a directory rename after all */
2061                         string_list_append(&remove_from_head,
2062                                            head_ent->dir)->util = head_ent;
2063                         strbuf_release(&head_ent->new_dir);
2064                 }
2065         }
2066
2067         remove_hashmap_entries(dir_re_head, &remove_from_head);
2068         remove_hashmap_entries(dir_re_merge, &remove_from_merge);
2069
2070         hashmap_iter_init(dir_re_merge, &iter);
2071         while ((merge_ent = hashmap_iter_next(&iter))) {
2072                 head_ent = dir_rename_find_entry(dir_re_head, merge_ent->dir);
2073                 if (tree_has_path(merge, merge_ent->dir)) {
2074                         /* 2. This wasn't a directory rename after all */
2075                         string_list_append(&remove_from_merge,
2076                                            merge_ent->dir)->util = merge_ent;
2077                 } else if (head_ent &&
2078                            !head_ent->non_unique_new_dir &&
2079                            !merge_ent->non_unique_new_dir) {
2080                         /* 3. rename/rename(1to2) */
2081                         /*
2082                          * We can assume it's not rename/rename(1to1) because
2083                          * that was case (1), already checked above.  So we
2084                          * know that head_ent->new_dir and merge_ent->new_dir
2085                          * are different strings.
2086                          */
2087                         output(o, 1, _("CONFLICT (rename/rename): "
2088                                        "Rename directory %s->%s in %s. "
2089                                        "Rename directory %s->%s in %s"),
2090                                head_ent->dir, head_ent->new_dir.buf, o->branch1,
2091                                head_ent->dir, merge_ent->new_dir.buf, o->branch2);
2092                         string_list_append(&remove_from_head,
2093                                            head_ent->dir)->util = head_ent;
2094                         strbuf_release(&head_ent->new_dir);
2095                         string_list_append(&remove_from_merge,
2096                                            merge_ent->dir)->util = merge_ent;
2097                         strbuf_release(&merge_ent->new_dir);
2098                 }
2099         }
2100
2101         remove_hashmap_entries(dir_re_head, &remove_from_head);
2102         remove_hashmap_entries(dir_re_merge, &remove_from_merge);
2103 }
2104
2105 static struct hashmap *get_directory_renames(struct diff_queue_struct *pairs,
2106                                              struct tree *tree)
2107 {
2108         struct hashmap *dir_renames;
2109         struct hashmap_iter iter;
2110         struct dir_rename_entry *entry;
2111         int i;
2112
2113         /*
2114          * Typically, we think of a directory rename as all files from a
2115          * certain directory being moved to a target directory.  However,
2116          * what if someone first moved two files from the original
2117          * directory in one commit, and then renamed the directory
2118          * somewhere else in a later commit?  At merge time, we just know
2119          * that files from the original directory went to two different
2120          * places, and that the bulk of them ended up in the same place.
2121          * We want each directory rename to represent where the bulk of the
2122          * files from that directory end up; this function exists to find
2123          * where the bulk of the files went.
2124          *
2125          * The first loop below simply iterates through the list of file
2126          * renames, finding out how often each directory rename pair
2127          * possibility occurs.
2128          */
2129         dir_renames = xmalloc(sizeof(*dir_renames));
2130         dir_rename_init(dir_renames);
2131         for (i = 0; i < pairs->nr; ++i) {
2132                 struct string_list_item *item;
2133                 int *count;
2134                 struct diff_filepair *pair = pairs->queue[i];
2135                 char *old_dir, *new_dir;
2136
2137                 /* File not part of directory rename if it wasn't renamed */
2138                 if (pair->status != 'R')
2139                         continue;
2140
2141                 get_renamed_dir_portion(pair->one->path, pair->two->path,
2142                                         &old_dir,        &new_dir);
2143                 if (!old_dir)
2144                         /* Directory didn't change at all; ignore this one. */
2145                         continue;
2146
2147                 entry = dir_rename_find_entry(dir_renames, old_dir);
2148                 if (!entry) {
2149                         entry = xmalloc(sizeof(*entry));
2150                         dir_rename_entry_init(entry, old_dir);
2151                         hashmap_put(dir_renames, entry);
2152                 } else {
2153                         free(old_dir);
2154                 }
2155                 item = string_list_lookup(&entry->possible_new_dirs, new_dir);
2156                 if (!item) {
2157                         item = string_list_insert(&entry->possible_new_dirs,
2158                                                   new_dir);
2159                         item->util = xcalloc(1, sizeof(int));
2160                 } else {
2161                         free(new_dir);
2162                 }
2163                 count = item->util;
2164                 *count += 1;
2165         }
2166
2167         /*
2168          * For each directory with files moved out of it, we find out which
2169          * target directory received the most files so we can declare it to
2170          * be the "winning" target location for the directory rename.  This
2171          * winner gets recorded in new_dir.  If there is no winner
2172          * (multiple target directories received the same number of files),
2173          * we set non_unique_new_dir.  Once we've determined the winner (or
2174          * that there is no winner), we no longer need possible_new_dirs.
2175          */
2176         hashmap_iter_init(dir_renames, &iter);
2177         while ((entry = hashmap_iter_next(&iter))) {
2178                 int max = 0;
2179                 int bad_max = 0;
2180                 char *best = NULL;
2181
2182                 for (i = 0; i < entry->possible_new_dirs.nr; i++) {
2183                         int *count = entry->possible_new_dirs.items[i].util;
2184
2185                         if (*count == max)
2186                                 bad_max = max;
2187                         else if (*count > max) {
2188                                 max = *count;
2189                                 best = entry->possible_new_dirs.items[i].string;
2190                         }
2191                 }
2192                 if (bad_max == max)
2193                         entry->non_unique_new_dir = 1;
2194                 else {
2195                         assert(entry->new_dir.len == 0);
2196                         strbuf_addstr(&entry->new_dir, best);
2197                 }
2198                 /*
2199                  * The relevant directory sub-portion of the original full
2200                  * filepaths were xstrndup'ed before inserting into
2201                  * possible_new_dirs, and instead of manually iterating the
2202                  * list and free'ing each, just lie and tell
2203                  * possible_new_dirs that it did the strdup'ing so that it
2204                  * will free them for us.
2205                  */
2206                 entry->possible_new_dirs.strdup_strings = 1;
2207                 string_list_clear(&entry->possible_new_dirs, 1);
2208         }
2209
2210         return dir_renames;
2211 }
2212
2213 static struct dir_rename_entry *check_dir_renamed(const char *path,
2214                                                   struct hashmap *dir_renames)
2215 {
2216         char *temp = xstrdup(path);
2217         char *end;
2218         struct dir_rename_entry *entry = NULL;;
2219
2220         while ((end = strrchr(temp, '/'))) {
2221                 *end = '\0';
2222                 entry = dir_rename_find_entry(dir_renames, temp);
2223                 if (entry)
2224                         break;
2225         }
2226         free(temp);
2227         return entry;
2228 }
2229
2230 static void compute_collisions(struct hashmap *collisions,
2231                                struct hashmap *dir_renames,
2232                                struct diff_queue_struct *pairs)
2233 {
2234         int i;
2235
2236         /*
2237          * Multiple files can be mapped to the same path due to directory
2238          * renames done by the other side of history.  Since that other
2239          * side of history could have merged multiple directories into one,
2240          * if our side of history added the same file basename to each of
2241          * those directories, then all N of them would get implicitly
2242          * renamed by the directory rename detection into the same path,
2243          * and we'd get an add/add/.../add conflict, and all those adds
2244          * from *this* side of history.  This is not representable in the
2245          * index, and users aren't going to easily be able to make sense of
2246          * it.  So we need to provide a good warning about what's
2247          * happening, and fall back to no-directory-rename detection
2248          * behavior for those paths.
2249          *
2250          * See testcases 9e and all of section 5 from t6043 for examples.
2251          */
2252         collision_init(collisions);
2253
2254         for (i = 0; i < pairs->nr; ++i) {
2255                 struct dir_rename_entry *dir_rename_ent;
2256                 struct collision_entry *collision_ent;
2257                 char *new_path;
2258                 struct diff_filepair *pair = pairs->queue[i];
2259
2260                 if (pair->status != 'A' && pair->status != 'R')
2261                         continue;
2262                 dir_rename_ent = check_dir_renamed(pair->two->path,
2263                                                    dir_renames);
2264                 if (!dir_rename_ent)
2265                         continue;
2266
2267                 new_path = apply_dir_rename(dir_rename_ent, pair->two->path);
2268                 if (!new_path)
2269                         /*
2270                          * dir_rename_ent->non_unique_new_path is true, which
2271                          * means there is no directory rename for us to use,
2272                          * which means it won't cause us any additional
2273                          * collisions.
2274                          */
2275                         continue;
2276                 collision_ent = collision_find_entry(collisions, new_path);
2277                 if (!collision_ent) {
2278                         collision_ent = xcalloc(1,
2279                                                 sizeof(struct collision_entry));
2280                         hashmap_entry_init(collision_ent, strhash(new_path));
2281                         hashmap_put(collisions, collision_ent);
2282                         collision_ent->target_file = new_path;
2283                 } else {
2284                         free(new_path);
2285                 }
2286                 string_list_insert(&collision_ent->source_files,
2287                                    pair->two->path);
2288         }
2289 }
2290
2291 static char *check_for_directory_rename(struct merge_options *o,
2292                                         const char *path,
2293                                         struct tree *tree,
2294                                         struct hashmap *dir_renames,
2295                                         struct hashmap *dir_rename_exclusions,
2296                                         struct hashmap *collisions,
2297                                         int *clean_merge)
2298 {
2299         char *new_path = NULL;
2300         struct dir_rename_entry *entry = check_dir_renamed(path, dir_renames);
2301         struct dir_rename_entry *oentry = NULL;
2302
2303         if (!entry)
2304                 return new_path;
2305
2306         /*
2307          * This next part is a little weird.  We do not want to do an
2308          * implicit rename into a directory we renamed on our side, because
2309          * that will result in a spurious rename/rename(1to2) conflict.  An
2310          * example:
2311          *   Base commit: dumbdir/afile, otherdir/bfile
2312          *   Side 1:      smrtdir/afile, otherdir/bfile
2313          *   Side 2:      dumbdir/afile, dumbdir/bfile
2314          * Here, while working on Side 1, we could notice that otherdir was
2315          * renamed/merged to dumbdir, and change the diff_filepair for
2316          * otherdir/bfile into a rename into dumbdir/bfile.  However, Side
2317          * 2 will notice the rename from dumbdir to smrtdir, and do the
2318          * transitive rename to move it from dumbdir/bfile to
2319          * smrtdir/bfile.  That gives us bfile in dumbdir vs being in
2320          * smrtdir, a rename/rename(1to2) conflict.  We really just want
2321          * the file to end up in smrtdir.  And the way to achieve that is
2322          * to not let Side1 do the rename to dumbdir, since we know that is
2323          * the source of one of our directory renames.
2324          *
2325          * That's why oentry and dir_rename_exclusions is here.
2326          *
2327          * As it turns out, this also prevents N-way transient rename
2328          * confusion; See testcases 9c and 9d of t6043.
2329          */
2330         oentry = dir_rename_find_entry(dir_rename_exclusions, entry->new_dir.buf);
2331         if (oentry) {
2332                 output(o, 1, _("WARNING: Avoiding applying %s -> %s rename "
2333                                "to %s, because %s itself was renamed."),
2334                        entry->dir, entry->new_dir.buf, path, entry->new_dir.buf);
2335         } else {
2336                 new_path = handle_path_level_conflicts(o, path, entry,
2337                                                        collisions, tree);
2338                 *clean_merge &= (new_path != NULL);
2339         }
2340
2341         return new_path;
2342 }
2343
2344 static void apply_directory_rename_modifications(struct merge_options *o,
2345                                                  struct diff_filepair *pair,
2346                                                  char *new_path,
2347                                                  struct rename *re,
2348                                                  struct tree *tree,
2349                                                  struct tree *o_tree,
2350                                                  struct tree *a_tree,
2351                                                  struct tree *b_tree,
2352                                                  struct string_list *entries,
2353                                                  int *clean)
2354 {
2355         struct string_list_item *item;
2356         int stage = (tree == a_tree ? 2 : 3);
2357         int update_wd;
2358
2359         /*
2360          * In all cases where we can do directory rename detection,
2361          * unpack_trees() will have read pair->two->path into the
2362          * index and the working copy.  We need to remove it so that
2363          * we can instead place it at new_path.  It is guaranteed to
2364          * not be untracked (unpack_trees() would have errored out
2365          * saying the file would have been overwritten), but it might
2366          * be dirty, though.
2367          */
2368         update_wd = !was_dirty(o, pair->two->path);
2369         if (!update_wd)
2370                 output(o, 1, _("Refusing to lose dirty file at %s"),
2371                        pair->two->path);
2372         remove_file(o, 1, pair->two->path, !update_wd);
2373
2374         /* Find or create a new re->dst_entry */
2375         item = string_list_lookup(entries, new_path);
2376         if (item) {
2377                 /*
2378                  * Since we're renaming on this side of history, and it's
2379                  * due to a directory rename on the other side of history
2380                  * (which we only allow when the directory in question no
2381                  * longer exists on the other side of history), the
2382                  * original entry for re->dst_entry is no longer
2383                  * necessary...
2384                  */
2385                 re->dst_entry->processed = 1;
2386
2387                 /*
2388                  * ...because we'll be using this new one.
2389                  */
2390                 re->dst_entry = item->util;
2391         } else {
2392                 /*
2393                  * re->dst_entry is for the before-dir-rename path, and we
2394                  * need it to hold information for the after-dir-rename
2395                  * path.  Before creating a new entry, we need to mark the
2396                  * old one as unnecessary (...unless it is shared by
2397                  * src_entry, i.e. this didn't use to be a rename, in which
2398                  * case we can just allow the normal processing to happen
2399                  * for it).
2400                  */
2401                 if (pair->status == 'R')
2402                         re->dst_entry->processed = 1;
2403
2404                 re->dst_entry = insert_stage_data(new_path,
2405                                                   o_tree, a_tree, b_tree,
2406                                                   entries);
2407                 item = string_list_insert(entries, new_path);
2408                 item->util = re->dst_entry;
2409         }
2410
2411         /*
2412          * Update the stage_data with the information about the path we are
2413          * moving into place.  That slot will be empty and available for us
2414          * to write to because of the collision checks in
2415          * handle_path_level_conflicts().  In other words,
2416          * re->dst_entry->stages[stage].oid will be the null_oid, so it's
2417          * open for us to write to.
2418          *
2419          * It may be tempting to actually update the index at this point as
2420          * well, using update_stages_for_stage_data(), but as per the big
2421          * "NOTE" in update_stages(), doing so will modify the current
2422          * in-memory index which will break calls to would_lose_untracked()
2423          * that we need to make.  Instead, we need to just make sure that
2424          * the various conflict_rename_*() functions update the index
2425          * explicitly rather than relying on unpack_trees() to have done it.
2426          */
2427         get_tree_entry(&tree->object.oid,
2428                        pair->two->path,
2429                        &re->dst_entry->stages[stage].oid,
2430                        &re->dst_entry->stages[stage].mode);
2431
2432         /* Update pair status */
2433         if (pair->status == 'A') {
2434                 /*
2435                  * Recording rename information for this add makes it look
2436                  * like a rename/delete conflict.  Make sure we can
2437                  * correctly handle this as an add that was moved to a new
2438                  * directory instead of reporting a rename/delete conflict.
2439                  */
2440                 re->add_turned_into_rename = 1;
2441         }
2442         /*
2443          * We don't actually look at pair->status again, but it seems
2444          * pedagogically correct to adjust it.
2445          */
2446         pair->status = 'R';
2447
2448         /*
2449          * Finally, record the new location.
2450          */
2451         pair->two->path = new_path;
2452 }
2453
2454 /*
2455  * Get information of all renames which occurred in 'pairs', making use of
2456  * any implicit directory renames inferred from the other side of history.
2457  * We need the three trees in the merge ('o_tree', 'a_tree' and 'b_tree')
2458  * to be able to associate the correct cache entries with the rename
2459  * information; tree is always equal to either a_tree or b_tree.
2460  */
2461 static struct string_list *get_renames(struct merge_options *o,
2462                                        struct diff_queue_struct *pairs,
2463                                        struct hashmap *dir_renames,
2464                                        struct hashmap *dir_rename_exclusions,
2465                                        struct tree *tree,
2466                                        struct tree *o_tree,
2467                                        struct tree *a_tree,
2468                                        struct tree *b_tree,
2469                                        struct string_list *entries,
2470                                        int *clean_merge)
2471 {
2472         int i;
2473         struct hashmap collisions;
2474         struct hashmap_iter iter;
2475         struct collision_entry *e;
2476         struct string_list *renames;
2477
2478         compute_collisions(&collisions, dir_renames, pairs);
2479         renames = xcalloc(1, sizeof(struct string_list));
2480
2481         for (i = 0; i < pairs->nr; ++i) {
2482                 struct string_list_item *item;
2483                 struct rename *re;
2484                 struct diff_filepair *pair = pairs->queue[i];
2485                 char *new_path; /* non-NULL only with directory renames */
2486
2487                 if (pair->status != 'A' && pair->status != 'R') {
2488                         diff_free_filepair(pair);
2489                         continue;
2490                 }
2491                 new_path = check_for_directory_rename(o, pair->two->path, tree,
2492                                                       dir_renames,
2493                                                       dir_rename_exclusions,
2494                                                       &collisions,
2495                                                       clean_merge);
2496                 if (pair->status != 'R' && !new_path) {
2497                         diff_free_filepair(pair);
2498                         continue;
2499                 }
2500
2501                 re = xmalloc(sizeof(*re));
2502                 re->processed = 0;
2503                 re->add_turned_into_rename = 0;
2504                 re->pair = pair;
2505                 item = string_list_lookup(entries, re->pair->one->path);
2506                 if (!item)
2507                         re->src_entry = insert_stage_data(re->pair->one->path,
2508                                         o_tree, a_tree, b_tree, entries);
2509                 else
2510                         re->src_entry = item->util;
2511
2512                 item = string_list_lookup(entries, re->pair->two->path);
2513                 if (!item)
2514                         re->dst_entry = insert_stage_data(re->pair->two->path,
2515                                         o_tree, a_tree, b_tree, entries);
2516                 else
2517                         re->dst_entry = item->util;
2518                 item = string_list_insert(renames, pair->one->path);
2519                 item->util = re;
2520                 if (new_path)
2521                         apply_directory_rename_modifications(o, pair, new_path,
2522                                                              re, tree, o_tree,
2523                                                              a_tree, b_tree,
2524                                                              entries,
2525                                                              clean_merge);
2526         }
2527
2528         hashmap_iter_init(&collisions, &iter);
2529         while ((e = hashmap_iter_next(&iter))) {
2530                 free(e->target_file);
2531                 string_list_clear(&e->source_files, 0);
2532         }
2533         hashmap_free(&collisions, 1);
2534         return renames;
2535 }
2536
2537 static int process_renames(struct merge_options *o,
2538                            struct string_list *a_renames,
2539                            struct string_list *b_renames)
2540 {
2541         int clean_merge = 1, i, j;
2542         struct string_list a_by_dst = STRING_LIST_INIT_NODUP;
2543         struct string_list b_by_dst = STRING_LIST_INIT_NODUP;
2544         const struct rename *sre;
2545
2546         for (i = 0; i < a_renames->nr; i++) {
2547                 sre = a_renames->items[i].util;
2548                 string_list_insert(&a_by_dst, sre->pair->two->path)->util
2549                         = (void *)sre;
2550         }
2551         for (i = 0; i < b_renames->nr; i++) {
2552                 sre = b_renames->items[i].util;
2553                 string_list_insert(&b_by_dst, sre->pair->two->path)->util
2554                         = (void *)sre;
2555         }
2556
2557         for (i = 0, j = 0; i < a_renames->nr || j < b_renames->nr;) {
2558                 struct string_list *renames1, *renames2Dst;
2559                 struct rename *ren1 = NULL, *ren2 = NULL;
2560                 const char *branch1, *branch2;
2561                 const char *ren1_src, *ren1_dst;
2562                 struct string_list_item *lookup;
2563
2564                 if (i >= a_renames->nr) {
2565                         ren2 = b_renames->items[j++].util;
2566                 } else if (j >= b_renames->nr) {
2567                         ren1 = a_renames->items[i++].util;
2568                 } else {
2569                         int compare = strcmp(a_renames->items[i].string,
2570                                              b_renames->items[j].string);
2571                         if (compare <= 0)
2572                                 ren1 = a_renames->items[i++].util;
2573                         if (compare >= 0)
2574                                 ren2 = b_renames->items[j++].util;
2575                 }
2576
2577                 /* TODO: refactor, so that 1/2 are not needed */
2578                 if (ren1) {
2579                         renames1 = a_renames;
2580                         renames2Dst = &b_by_dst;
2581                         branch1 = o->branch1;
2582                         branch2 = o->branch2;
2583                 } else {
2584                         renames1 = b_renames;
2585                         renames2Dst = &a_by_dst;
2586                         branch1 = o->branch2;
2587                         branch2 = o->branch1;
2588                         SWAP(ren2, ren1);
2589                 }
2590
2591                 if (ren1->processed)
2592                         continue;
2593                 ren1->processed = 1;
2594                 ren1->dst_entry->processed = 1;
2595                 /* BUG: We should only mark src_entry as processed if we
2596                  * are not dealing with a rename + add-source case.
2597                  */
2598                 ren1->src_entry->processed = 1;
2599
2600                 ren1_src = ren1->pair->one->path;
2601                 ren1_dst = ren1->pair->two->path;
2602
2603                 if (ren2) {
2604                         /* One file renamed on both sides */
2605                         const char *ren2_src = ren2->pair->one->path;
2606                         const char *ren2_dst = ren2->pair->two->path;
2607                         enum rename_type rename_type;
2608                         if (strcmp(ren1_src, ren2_src) != 0)
2609                                 BUG("ren1_src != ren2_src");
2610                         ren2->dst_entry->processed = 1;
2611                         ren2->processed = 1;
2612                         if (strcmp(ren1_dst, ren2_dst) != 0) {
2613                                 rename_type = RENAME_ONE_FILE_TO_TWO;
2614                                 clean_merge = 0;
2615                         } else {
2616                                 rename_type = RENAME_ONE_FILE_TO_ONE;
2617                                 /* BUG: We should only remove ren1_src in
2618                                  * the base stage (think of rename +
2619                                  * add-source cases).
2620                                  */
2621                                 remove_file(o, 1, ren1_src, 1);
2622                                 update_entry(ren1->dst_entry,
2623                                              ren1->pair->one,
2624                                              ren1->pair->two,
2625                                              ren2->pair->two);
2626                         }
2627                         setup_rename_conflict_info(rename_type,
2628                                                    ren1->pair,
2629                                                    ren2->pair,
2630                                                    branch1,
2631                                                    branch2,
2632                                                    ren1->dst_entry,
2633                                                    ren2->dst_entry,
2634                                                    o,
2635                                                    NULL,
2636                                                    NULL);
2637                 } else if ((lookup = string_list_lookup(renames2Dst, ren1_dst))) {
2638                         /* Two different files renamed to the same thing */
2639                         char *ren2_dst;
2640                         ren2 = lookup->util;
2641                         ren2_dst = ren2->pair->two->path;
2642                         if (strcmp(ren1_dst, ren2_dst) != 0)
2643                                 BUG("ren1_dst != ren2_dst");
2644
2645                         clean_merge = 0;
2646                         ren2->processed = 1;
2647                         /*
2648                          * BUG: We should only mark src_entry as processed
2649                          * if we are not dealing with a rename + add-source
2650                          * case.
2651                          */
2652                         ren2->src_entry->processed = 1;
2653
2654                         setup_rename_conflict_info(RENAME_TWO_FILES_TO_ONE,
2655                                                    ren1->pair,
2656                                                    ren2->pair,
2657                                                    branch1,
2658                                                    branch2,
2659                                                    ren1->dst_entry,
2660                                                    ren2->dst_entry,
2661                                                    o,
2662                                                    ren1->src_entry,
2663                                                    ren2->src_entry);
2664
2665                 } else {
2666                         /* Renamed in 1, maybe changed in 2 */
2667                         /* we only use sha1 and mode of these */
2668                         struct diff_filespec src_other, dst_other;
2669                         int try_merge;
2670
2671                         /*
2672                          * unpack_trees loads entries from common-commit
2673                          * into stage 1, from head-commit into stage 2, and
2674                          * from merge-commit into stage 3.  We keep track
2675                          * of which side corresponds to the rename.
2676                          */
2677                         int renamed_stage = a_renames == renames1 ? 2 : 3;
2678                         int other_stage =   a_renames == renames1 ? 3 : 2;
2679
2680                         /* BUG: We should only remove ren1_src in the base
2681                          * stage and in other_stage (think of rename +
2682                          * add-source case).
2683                          */
2684                         remove_file(o, 1, ren1_src,
2685                                     renamed_stage == 2 || !was_tracked(o, ren1_src));
2686
2687                         oidcpy(&src_other.oid,
2688                                &ren1->src_entry->stages[other_stage].oid);
2689                         src_other.mode = ren1->src_entry->stages[other_stage].mode;
2690                         oidcpy(&dst_other.oid,
2691                                &ren1->dst_entry->stages[other_stage].oid);
2692                         dst_other.mode = ren1->dst_entry->stages[other_stage].mode;
2693                         try_merge = 0;
2694
2695                         if (oid_eq(&src_other.oid, &null_oid) &&
2696                             ren1->add_turned_into_rename) {
2697                                 setup_rename_conflict_info(RENAME_DIR,
2698                                                            ren1->pair,
2699                                                            NULL,
2700                                                            branch1,
2701                                                            branch2,
2702                                                            ren1->dst_entry,
2703                                                            NULL,
2704                                                            o,
2705                                                            NULL,
2706                                                            NULL);
2707                         } else if (oid_eq(&src_other.oid, &null_oid)) {
2708                                 setup_rename_conflict_info(RENAME_DELETE,
2709                                                            ren1->pair,
2710                                                            NULL,
2711                                                            branch1,
2712                                                            branch2,
2713                                                            ren1->dst_entry,
2714                                                            NULL,
2715                                                            o,
2716                                                            NULL,
2717                                                            NULL);
2718                         } else if ((dst_other.mode == ren1->pair->two->mode) &&
2719                                    oid_eq(&dst_other.oid, &ren1->pair->two->oid)) {
2720                                 /*
2721                                  * Added file on the other side identical to
2722                                  * the file being renamed: clean merge.
2723                                  * Also, there is no need to overwrite the
2724                                  * file already in the working copy, so call
2725                                  * update_file_flags() instead of
2726                                  * update_file().
2727                                  */
2728                                 if (update_file_flags(o,
2729                                                       &ren1->pair->two->oid,
2730                                                       ren1->pair->two->mode,
2731                                                       ren1_dst,
2732                                                       1, /* update_cache */
2733                                                       0  /* update_wd    */))
2734                                         clean_merge = -1;
2735                         } else if (!oid_eq(&dst_other.oid, &null_oid)) {
2736                                 clean_merge = 0;
2737                                 try_merge = 1;
2738                                 output(o, 1, _("CONFLICT (rename/add): Rename %s->%s in %s. "
2739                                        "%s added in %s"),
2740                                        ren1_src, ren1_dst, branch1,
2741                                        ren1_dst, branch2);
2742                                 if (o->call_depth) {
2743                                         struct merge_file_info mfi;
2744                                         if (merge_file_one(o, ren1_dst, &null_oid, 0,
2745                                                            &ren1->pair->two->oid,
2746                                                            ren1->pair->two->mode,
2747                                                            &dst_other.oid,
2748                                                            dst_other.mode,
2749                                                            branch1, branch2, &mfi)) {
2750                                                 clean_merge = -1;
2751                                                 goto cleanup_and_return;
2752                                         }
2753                                         output(o, 1, _("Adding merged %s"), ren1_dst);
2754                                         if (update_file(o, 0, &mfi.oid,
2755                                                         mfi.mode, ren1_dst))
2756                                                 clean_merge = -1;
2757                                         try_merge = 0;
2758                                 } else {
2759                                         char *new_path = unique_path(o, ren1_dst, branch2);
2760                                         output(o, 1, _("Adding as %s instead"), new_path);
2761                                         if (update_file(o, 0, &dst_other.oid,
2762                                                         dst_other.mode, new_path))
2763                                                 clean_merge = -1;
2764                                         free(new_path);
2765                                 }
2766                         } else
2767                                 try_merge = 1;
2768
2769                         if (clean_merge < 0)
2770                                 goto cleanup_and_return;
2771                         if (try_merge) {
2772                                 struct diff_filespec *one, *a, *b;
2773                                 src_other.path = (char *)ren1_src;
2774
2775                                 one = ren1->pair->one;
2776                                 if (a_renames == renames1) {
2777                                         a = ren1->pair->two;
2778                                         b = &src_other;
2779                                 } else {
2780                                         b = ren1->pair->two;
2781                                         a = &src_other;
2782                                 }
2783                                 update_entry(ren1->dst_entry, one, a, b);
2784                                 setup_rename_conflict_info(RENAME_NORMAL,
2785                                                            ren1->pair,
2786                                                            NULL,
2787                                                            branch1,
2788                                                            NULL,
2789                                                            ren1->dst_entry,
2790                                                            NULL,
2791                                                            o,
2792                                                            NULL,
2793                                                            NULL);
2794                         }
2795                 }
2796         }
2797 cleanup_and_return:
2798         string_list_clear(&a_by_dst, 0);
2799         string_list_clear(&b_by_dst, 0);
2800
2801         return clean_merge;
2802 }
2803
2804 struct rename_info {
2805         struct string_list *head_renames;
2806         struct string_list *merge_renames;
2807 };
2808
2809 static void initial_cleanup_rename(struct diff_queue_struct *pairs,
2810                                    struct hashmap *dir_renames)
2811 {
2812         struct hashmap_iter iter;
2813         struct dir_rename_entry *e;
2814
2815         hashmap_iter_init(dir_renames, &iter);
2816         while ((e = hashmap_iter_next(&iter))) {
2817                 free(e->dir);
2818                 strbuf_release(&e->new_dir);
2819                 /* possible_new_dirs already cleared in get_directory_renames */
2820         }
2821         hashmap_free(dir_renames, 1);
2822         free(dir_renames);
2823
2824         free(pairs->queue);
2825         free(pairs);
2826 }
2827
2828 static int handle_renames(struct merge_options *o,
2829                           struct tree *common,
2830                           struct tree *head,
2831                           struct tree *merge,
2832                           struct string_list *entries,
2833                           struct rename_info *ri)
2834 {
2835         struct diff_queue_struct *head_pairs, *merge_pairs;
2836         struct hashmap *dir_re_head, *dir_re_merge;
2837         int clean = 1;
2838
2839         ri->head_renames = NULL;
2840         ri->merge_renames = NULL;
2841
2842         if (!merge_detect_rename(o))
2843                 return 1;
2844
2845         head_pairs = get_diffpairs(o, common, head);
2846         merge_pairs = get_diffpairs(o, common, merge);
2847
2848         dir_re_head = get_directory_renames(head_pairs, head);
2849         dir_re_merge = get_directory_renames(merge_pairs, merge);
2850
2851         handle_directory_level_conflicts(o,
2852                                          dir_re_head, head,
2853                                          dir_re_merge, merge);
2854
2855         ri->head_renames  = get_renames(o, head_pairs,
2856                                         dir_re_merge, dir_re_head, head,
2857                                         common, head, merge, entries,
2858                                         &clean);
2859         if (clean < 0)
2860                 goto cleanup;
2861         ri->merge_renames = get_renames(o, merge_pairs,
2862                                         dir_re_head, dir_re_merge, merge,
2863                                         common, head, merge, entries,
2864                                         &clean);
2865         if (clean < 0)
2866                 goto cleanup;
2867         clean &= process_renames(o, ri->head_renames, ri->merge_renames);
2868
2869 cleanup:
2870         /*
2871          * Some cleanup is deferred until cleanup_renames() because the
2872          * data structures are still needed and referenced in
2873          * process_entry().  But there are a few things we can free now.
2874          */
2875         initial_cleanup_rename(head_pairs, dir_re_head);
2876         initial_cleanup_rename(merge_pairs, dir_re_merge);
2877
2878         return clean;
2879 }
2880
2881 static void final_cleanup_rename(struct string_list *rename)
2882 {
2883         const struct rename *re;
2884         int i;
2885
2886         if (rename == NULL)
2887                 return;
2888
2889         for (i = 0; i < rename->nr; i++) {
2890                 re = rename->items[i].util;
2891                 diff_free_filepair(re->pair);
2892         }
2893         string_list_clear(rename, 1);
2894         free(rename);
2895 }
2896
2897 static void final_cleanup_renames(struct rename_info *re_info)
2898 {
2899         final_cleanup_rename(re_info->head_renames);
2900         final_cleanup_rename(re_info->merge_renames);
2901 }
2902
2903 static struct object_id *stage_oid(const struct object_id *oid, unsigned mode)
2904 {
2905         return (is_null_oid(oid) || mode == 0) ? NULL: (struct object_id *)oid;
2906 }
2907
2908 static int read_oid_strbuf(struct merge_options *o,
2909         const struct object_id *oid, struct strbuf *dst)
2910 {
2911         void *buf;
2912         enum object_type type;
2913         unsigned long size;
2914         buf = read_object_file(oid, &type, &size);
2915         if (!buf)
2916                 return err(o, _("cannot read object %s"), oid_to_hex(oid));
2917         if (type != OBJ_BLOB) {
2918                 free(buf);
2919                 return err(o, _("object %s is not a blob"), oid_to_hex(oid));
2920         }
2921         strbuf_attach(dst, buf, size, size + 1);
2922         return 0;
2923 }
2924
2925 static int blob_unchanged(struct merge_options *opt,
2926                           const struct object_id *o_oid,
2927                           unsigned o_mode,
2928                           const struct object_id *a_oid,
2929                           unsigned a_mode,
2930                           int renormalize, const char *path)
2931 {
2932         struct strbuf o = STRBUF_INIT;
2933         struct strbuf a = STRBUF_INIT;
2934         int ret = 0; /* assume changed for safety */
2935
2936         if (a_mode != o_mode)
2937                 return 0;
2938         if (oid_eq(o_oid, a_oid))
2939                 return 1;
2940         if (!renormalize)
2941                 return 0;
2942
2943         assert(o_oid && a_oid);
2944         if (read_oid_strbuf(opt, o_oid, &o) || read_oid_strbuf(opt, a_oid, &a))
2945                 goto error_return;
2946         /*
2947          * Note: binary | is used so that both renormalizations are
2948          * performed.  Comparison can be skipped if both files are
2949          * unchanged since their sha1s have already been compared.
2950          */
2951         if (renormalize_buffer(&the_index, path, o.buf, o.len, &o) |
2952             renormalize_buffer(&the_index, path, a.buf, a.len, &a))
2953                 ret = (o.len == a.len && !memcmp(o.buf, a.buf, o.len));
2954
2955 error_return:
2956         strbuf_release(&o);
2957         strbuf_release(&a);
2958         return ret;
2959 }
2960
2961 static int handle_modify_delete(struct merge_options *o,
2962                                  const char *path,
2963                                  struct object_id *o_oid, int o_mode,
2964                                  struct object_id *a_oid, int a_mode,
2965                                  struct object_id *b_oid, int b_mode)
2966 {
2967         const char *modify_branch, *delete_branch;
2968         struct object_id *changed_oid;
2969         int changed_mode;
2970
2971         if (a_oid) {
2972                 modify_branch = o->branch1;
2973                 delete_branch = o->branch2;
2974                 changed_oid = a_oid;
2975                 changed_mode = a_mode;
2976         } else {
2977                 modify_branch = o->branch2;
2978                 delete_branch = o->branch1;
2979                 changed_oid = b_oid;
2980                 changed_mode = b_mode;
2981         }
2982
2983         return handle_change_delete(o,
2984                                     path, NULL,
2985                                     o_oid, o_mode,
2986                                     changed_oid, changed_mode,
2987                                     modify_branch, delete_branch,
2988                                     _("modify"), _("modified"));
2989 }
2990
2991 static int merge_content(struct merge_options *o,
2992                          const char *path,
2993                          int is_dirty,
2994                          struct object_id *o_oid, int o_mode,
2995                          struct object_id *a_oid, int a_mode,
2996                          struct object_id *b_oid, int b_mode,
2997                          struct rename_conflict_info *rename_conflict_info)
2998 {
2999         const char *reason = _("content");
3000         const char *path1 = NULL, *path2 = NULL;
3001         struct merge_file_info mfi;
3002         struct diff_filespec one, a, b;
3003         unsigned df_conflict_remains = 0;
3004
3005         if (!o_oid) {
3006                 reason = _("add/add");
3007                 o_oid = (struct object_id *)&null_oid;
3008         }
3009         one.path = a.path = b.path = (char *)path;
3010         oidcpy(&one.oid, o_oid);
3011         one.mode = o_mode;
3012         oidcpy(&a.oid, a_oid);
3013         a.mode = a_mode;
3014         oidcpy(&b.oid, b_oid);
3015         b.mode = b_mode;
3016
3017         if (rename_conflict_info) {
3018                 struct diff_filepair *pair1 = rename_conflict_info->pair1;
3019
3020                 path1 = (o->branch1 == rename_conflict_info->branch1) ?
3021                         pair1->two->path : pair1->one->path;
3022                 /* If rename_conflict_info->pair2 != NULL, we are in
3023                  * RENAME_ONE_FILE_TO_ONE case.  Otherwise, we have a
3024                  * normal rename.
3025                  */
3026                 path2 = (rename_conflict_info->pair2 ||
3027                          o->branch2 == rename_conflict_info->branch1) ?
3028                         pair1->two->path : pair1->one->path;
3029
3030                 if (dir_in_way(path, !o->call_depth,
3031                                S_ISGITLINK(pair1->two->mode)))
3032                         df_conflict_remains = 1;
3033         }
3034         if (merge_file_special_markers(o, &one, &a, &b, path,
3035                                        o->branch1, path1,
3036                                        o->branch2, path2, &mfi))
3037                 return -1;
3038
3039         /*
3040          * We can skip updating the working tree file iff:
3041          *   a) The merge is clean
3042          *   b) The merge matches what was in HEAD (content, mode, pathname)
3043          *   c) The target path is usable (i.e. not involved in D/F conflict)
3044          */
3045         if (mfi.clean &&
3046             was_tracked_and_matches(o, path, &mfi.oid, mfi.mode) &&
3047             !df_conflict_remains) {
3048                 output(o, 3, _("Skipped %s (merged same as existing)"), path);
3049                 if (add_cacheinfo(o, mfi.mode, &mfi.oid, path,
3050                                   0, (!o->call_depth && !is_dirty), 0))
3051                         return -1;
3052                 return mfi.clean;
3053         }
3054
3055         if (!mfi.clean) {
3056                 if (S_ISGITLINK(mfi.mode))
3057                         reason = _("submodule");
3058                 output(o, 1, _("CONFLICT (%s): Merge conflict in %s"),
3059                                 reason, path);
3060                 if (rename_conflict_info && !df_conflict_remains)
3061                         if (update_stages(o, path, &one, &a, &b))
3062                                 return -1;
3063         }
3064
3065         if (df_conflict_remains || is_dirty) {
3066                 char *new_path;
3067                 if (o->call_depth) {
3068                         remove_file_from_cache(path);
3069                 } else {
3070                         if (!mfi.clean) {
3071                                 if (update_stages(o, path, &one, &a, &b))
3072                                         return -1;
3073                         } else {
3074                                 int file_from_stage2 = was_tracked(o, path);
3075                                 struct diff_filespec merged;
3076                                 oidcpy(&merged.oid, &mfi.oid);
3077                                 merged.mode = mfi.mode;
3078
3079                                 if (update_stages(o, path, NULL,
3080                                                   file_from_stage2 ? &merged : NULL,
3081                                                   file_from_stage2 ? NULL : &merged))
3082                                         return -1;
3083                         }
3084
3085                 }
3086                 new_path = unique_path(o, path, rename_conflict_info->branch1);
3087                 if (is_dirty) {
3088                         output(o, 1, _("Refusing to lose dirty file at %s"),
3089                                path);
3090                 }
3091                 output(o, 1, _("Adding as %s instead"), new_path);
3092                 if (update_file(o, 0, &mfi.oid, mfi.mode, new_path)) {
3093                         free(new_path);
3094                         return -1;
3095                 }
3096                 free(new_path);
3097                 mfi.clean = 0;
3098         } else if (update_file(o, mfi.clean, &mfi.oid, mfi.mode, path))
3099                 return -1;
3100         return !is_dirty && mfi.clean;
3101 }
3102
3103 static int conflict_rename_normal(struct merge_options *o,
3104                                   const char *path,
3105                                   struct object_id *o_oid, unsigned int o_mode,
3106                                   struct object_id *a_oid, unsigned int a_mode,
3107                                   struct object_id *b_oid, unsigned int b_mode,
3108                                   struct rename_conflict_info *ci)
3109 {
3110         /* Merge the content and write it out */
3111         return merge_content(o, path, was_dirty(o, path),
3112                              o_oid, o_mode, a_oid, a_mode, b_oid, b_mode,
3113                              ci);
3114 }
3115
3116 /* Per entry merge function */
3117 static int process_entry(struct merge_options *o,
3118                          const char *path, struct stage_data *entry)
3119 {
3120         int clean_merge = 1;
3121         int normalize = o->renormalize;
3122         unsigned o_mode = entry->stages[1].mode;
3123         unsigned a_mode = entry->stages[2].mode;
3124         unsigned b_mode = entry->stages[3].mode;
3125         struct object_id *o_oid = stage_oid(&entry->stages[1].oid, o_mode);
3126         struct object_id *a_oid = stage_oid(&entry->stages[2].oid, a_mode);
3127         struct object_id *b_oid = stage_oid(&entry->stages[3].oid, b_mode);
3128
3129         entry->processed = 1;
3130         if (entry->rename_conflict_info) {
3131                 struct rename_conflict_info *conflict_info = entry->rename_conflict_info;
3132                 switch (conflict_info->rename_type) {
3133                 case RENAME_NORMAL:
3134                 case RENAME_ONE_FILE_TO_ONE:
3135                         clean_merge = conflict_rename_normal(o,
3136                                                              path,
3137                                                              o_oid, o_mode,
3138                                                              a_oid, a_mode,
3139                                                              b_oid, b_mode,
3140                                                              conflict_info);
3141                         break;
3142                 case RENAME_DIR:
3143                         clean_merge = 1;
3144                         if (conflict_rename_dir(o,
3145                                                 conflict_info->pair1,
3146                                                 conflict_info->branch1,
3147                                                 conflict_info->branch2))
3148                                 clean_merge = -1;
3149                         break;
3150                 case RENAME_DELETE:
3151                         clean_merge = 0;
3152                         if (conflict_rename_delete(o,
3153                                                    conflict_info->pair1,
3154                                                    conflict_info->branch1,
3155                                                    conflict_info->branch2))
3156                                 clean_merge = -1;
3157                         break;
3158                 case RENAME_ONE_FILE_TO_TWO:
3159                         clean_merge = 0;
3160                         if (conflict_rename_rename_1to2(o, conflict_info))
3161                                 clean_merge = -1;
3162                         break;
3163                 case RENAME_TWO_FILES_TO_ONE:
3164                         clean_merge = 0;
3165                         if (conflict_rename_rename_2to1(o, conflict_info))
3166                                 clean_merge = -1;
3167                         break;
3168                 default:
3169                         entry->processed = 0;
3170                         break;
3171                 }
3172         } else if (o_oid && (!a_oid || !b_oid)) {
3173                 /* Case A: Deleted in one */
3174                 if ((!a_oid && !b_oid) ||
3175                     (!b_oid && blob_unchanged(o, o_oid, o_mode, a_oid, a_mode, normalize, path)) ||
3176                     (!a_oid && blob_unchanged(o, o_oid, o_mode, b_oid, b_mode, normalize, path))) {
3177                         /* Deleted in both or deleted in one and
3178                          * unchanged in the other */
3179                         if (a_oid)
3180                                 output(o, 2, _("Removing %s"), path);
3181                         /* do not touch working file if it did not exist */
3182                         remove_file(o, 1, path, !a_oid);
3183                 } else {
3184                         /* Modify/delete; deleted side may have put a directory in the way */
3185                         clean_merge = 0;
3186                         if (handle_modify_delete(o, path, o_oid, o_mode,
3187                                                  a_oid, a_mode, b_oid, b_mode))
3188                                 clean_merge = -1;
3189                 }
3190         } else if ((!o_oid && a_oid && !b_oid) ||
3191                    (!o_oid && !a_oid && b_oid)) {
3192                 /* Case B: Added in one. */
3193                 /* [nothing|directory] -> ([nothing|directory], file) */
3194
3195                 const char *add_branch;
3196                 const char *other_branch;
3197                 unsigned mode;
3198                 const struct object_id *oid;
3199                 const char *conf;
3200
3201                 if (a_oid) {
3202                         add_branch = o->branch1;
3203                         other_branch = o->branch2;
3204                         mode = a_mode;
3205                         oid = a_oid;
3206                         conf = _("file/directory");
3207                 } else {
3208                         add_branch = o->branch2;
3209                         other_branch = o->branch1;
3210                         mode = b_mode;
3211                         oid = b_oid;
3212                         conf = _("directory/file");
3213                 }
3214                 if (dir_in_way(path,
3215                                !o->call_depth && !S_ISGITLINK(a_mode),
3216                                0)) {
3217                         char *new_path = unique_path(o, path, add_branch);
3218                         clean_merge = 0;
3219                         output(o, 1, _("CONFLICT (%s): There is a directory with name %s in %s. "
3220                                "Adding %s as %s"),
3221                                conf, path, other_branch, path, new_path);
3222                         if (update_file(o, 0, oid, mode, new_path))
3223                                 clean_merge = -1;
3224                         else if (o->call_depth)
3225                                 remove_file_from_cache(path);
3226                         free(new_path);
3227                 } else {
3228                         output(o, 2, _("Adding %s"), path);
3229                         /* do not overwrite file if already present */
3230                         if (update_file_flags(o, oid, mode, path, 1, !a_oid))
3231                                 clean_merge = -1;
3232                 }
3233         } else if (a_oid && b_oid) {
3234                 /* Case C: Added in both (check for same permissions) and */
3235                 /* case D: Modified in both, but differently. */
3236                 int is_dirty = 0; /* unpack_trees would have bailed if dirty */
3237                 clean_merge = merge_content(o, path, is_dirty,
3238                                             o_oid, o_mode, a_oid, a_mode, b_oid, b_mode,
3239                                             NULL);
3240         } else if (!o_oid && !a_oid && !b_oid) {
3241                 /*
3242                  * this entry was deleted altogether. a_mode == 0 means
3243                  * we had that path and want to actively remove it.
3244                  */
3245                 remove_file(o, 1, path, !a_mode);
3246         } else
3247                 BUG("fatal merge failure, shouldn't happen.");
3248
3249         return clean_merge;
3250 }
3251
3252 int merge_trees(struct merge_options *o,
3253                 struct tree *head,
3254                 struct tree *merge,
3255                 struct tree *common,
3256                 struct tree **result)
3257 {
3258         int code, clean;
3259
3260         if (o->subtree_shift) {
3261                 merge = shift_tree_object(head, merge, o->subtree_shift);
3262                 common = shift_tree_object(head, common, o->subtree_shift);
3263         }
3264
3265         if (oid_eq(&common->object.oid, &merge->object.oid)) {
3266                 struct strbuf sb = STRBUF_INIT;
3267
3268                 if (!o->call_depth && index_has_changes(&sb)) {
3269                         err(o, _("Dirty index: cannot merge (dirty: %s)"),
3270                             sb.buf);
3271                         return 0;
3272                 }
3273                 output(o, 0, _("Already up to date!"));
3274                 *result = head;
3275                 return 1;
3276         }
3277
3278         code = unpack_trees_start(o, common, head, merge);
3279
3280         if (code != 0) {
3281                 if (show(o, 4) || o->call_depth)
3282                         err(o, _("merging of trees %s and %s failed"),
3283                             oid_to_hex(&head->object.oid),
3284                             oid_to_hex(&merge->object.oid));
3285                 unpack_trees_finish(o);
3286                 return -1;
3287         }
3288
3289         if (unmerged_cache()) {
3290                 struct string_list *entries;
3291                 struct rename_info re_info;
3292                 int i;
3293                 /*
3294                  * Only need the hashmap while processing entries, so
3295                  * initialize it here and free it when we are done running
3296                  * through the entries. Keeping it in the merge_options as
3297                  * opposed to decaring a local hashmap is for convenience
3298                  * so that we don't have to pass it to around.
3299                  */
3300                 hashmap_init(&o->current_file_dir_set, path_hashmap_cmp, NULL, 512);
3301                 get_files_dirs(o, head);
3302                 get_files_dirs(o, merge);
3303
3304                 entries = get_unmerged();
3305                 clean = handle_renames(o, common, head, merge, entries,
3306                                        &re_info);
3307                 record_df_conflict_files(o, entries);
3308                 if (clean < 0)
3309                         goto cleanup;
3310                 for (i = entries->nr-1; 0 <= i; i--) {
3311                         const char *path = entries->items[i].string;
3312                         struct stage_data *e = entries->items[i].util;
3313                         if (!e->processed) {
3314                                 int ret = process_entry(o, path, e);
3315                                 if (!ret)
3316                                         clean = 0;
3317                                 else if (ret < 0) {
3318                                         clean = ret;
3319                                         goto cleanup;
3320                                 }
3321                         }
3322                 }
3323                 for (i = 0; i < entries->nr; i++) {
3324                         struct stage_data *e = entries->items[i].util;
3325                         if (!e->processed)
3326                                 BUG("unprocessed path??? %s",
3327                                     entries->items[i].string);
3328                 }
3329
3330 cleanup:
3331                 final_cleanup_renames(&re_info);
3332
3333                 string_list_clear(entries, 1);
3334                 free(entries);
3335
3336                 hashmap_free(&o->current_file_dir_set, 1);
3337
3338                 if (clean < 0) {
3339                         unpack_trees_finish(o);
3340                         return clean;
3341                 }
3342         }
3343         else
3344                 clean = 1;
3345
3346         unpack_trees_finish(o);
3347
3348         if (o->call_depth && !(*result = write_tree_from_memory(o)))
3349                 return -1;
3350
3351         return clean;
3352 }
3353
3354 static struct commit_list *reverse_commit_list(struct commit_list *list)
3355 {
3356         struct commit_list *next = NULL, *current, *backup;
3357         for (current = list; current; current = backup) {
3358                 backup = current->next;
3359                 current->next = next;
3360                 next = current;
3361         }
3362         return next;
3363 }
3364
3365 /*
3366  * Merge the commits h1 and h2, return the resulting virtual
3367  * commit object and a flag indicating the cleanness of the merge.
3368  */
3369 int merge_recursive(struct merge_options *o,
3370                     struct commit *h1,
3371                     struct commit *h2,
3372                     struct commit_list *ca,
3373                     struct commit **result)
3374 {
3375         struct commit_list *iter;
3376         struct commit *merged_common_ancestors;
3377         struct tree *mrtree;
3378         int clean;
3379
3380         if (show(o, 4)) {
3381                 output(o, 4, _("Merging:"));
3382                 output_commit_title(o, h1);
3383                 output_commit_title(o, h2);
3384         }
3385
3386         if (!ca) {
3387                 ca = get_merge_bases(h1, h2);
3388                 ca = reverse_commit_list(ca);
3389         }
3390
3391         if (show(o, 5)) {
3392                 unsigned cnt = commit_list_count(ca);
3393
3394                 output(o, 5, Q_("found %u common ancestor:",
3395                                 "found %u common ancestors:", cnt), cnt);
3396                 for (iter = ca; iter; iter = iter->next)
3397                         output_commit_title(o, iter->item);
3398         }
3399
3400         merged_common_ancestors = pop_commit(&ca);
3401         if (merged_common_ancestors == NULL) {
3402                 /* if there is no common ancestor, use an empty tree */
3403                 struct tree *tree;
3404
3405                 tree = lookup_tree(the_hash_algo->empty_tree);
3406                 merged_common_ancestors = make_virtual_commit(tree, "ancestor");
3407         }
3408
3409         for (iter = ca; iter; iter = iter->next) {
3410                 const char *saved_b1, *saved_b2;
3411                 o->call_depth++;
3412                 /*
3413                  * When the merge fails, the result contains files
3414                  * with conflict markers. The cleanness flag is
3415                  * ignored (unless indicating an error), it was never
3416                  * actually used, as result of merge_trees has always
3417                  * overwritten it: the committed "conflicts" were
3418                  * already resolved.
3419                  */
3420                 discard_cache();
3421                 saved_b1 = o->branch1;
3422                 saved_b2 = o->branch2;
3423                 o->branch1 = "Temporary merge branch 1";
3424                 o->branch2 = "Temporary merge branch 2";
3425                 if (merge_recursive(o, merged_common_ancestors, iter->item,
3426                                     NULL, &merged_common_ancestors) < 0)
3427                         return -1;
3428                 o->branch1 = saved_b1;
3429                 o->branch2 = saved_b2;
3430                 o->call_depth--;
3431
3432                 if (!merged_common_ancestors)
3433                         return err(o, _("merge returned no commit"));
3434         }
3435
3436         discard_cache();
3437         if (!o->call_depth)
3438                 read_cache();
3439
3440         o->ancestor = "merged common ancestors";
3441         clean = merge_trees(o, get_commit_tree(h1), get_commit_tree(h2),
3442                             get_commit_tree(merged_common_ancestors),
3443                             &mrtree);
3444         if (clean < 0) {
3445                 flush_output(o);
3446                 return clean;
3447         }
3448
3449         if (o->call_depth) {
3450                 *result = make_virtual_commit(mrtree, "merged tree");
3451                 commit_list_insert(h1, &(*result)->parents);
3452                 commit_list_insert(h2, &(*result)->parents->next);
3453         }
3454         flush_output(o);
3455         if (!o->call_depth && o->buffer_output < 2)
3456                 strbuf_release(&o->obuf);
3457         if (show(o, 2))
3458                 diff_warn_rename_limit("merge.renamelimit",
3459                                        o->needed_rename_limit, 0);
3460         return clean;
3461 }
3462
3463 static struct commit *get_ref(const struct object_id *oid, const char *name)
3464 {
3465         struct object *object;
3466
3467         object = deref_tag(parse_object(oid), name, strlen(name));
3468         if (!object)
3469                 return NULL;
3470         if (object->type == OBJ_TREE)
3471                 return make_virtual_commit((struct tree*)object, name);
3472         if (object->type != OBJ_COMMIT)
3473                 return NULL;
3474         if (parse_commit((struct commit *)object))
3475                 return NULL;
3476         return (struct commit *)object;
3477 }
3478
3479 int merge_recursive_generic(struct merge_options *o,
3480                             const struct object_id *head,
3481                             const struct object_id *merge,
3482                             int num_base_list,
3483                             const struct object_id **base_list,
3484                             struct commit **result)
3485 {
3486         int clean;
3487         struct lock_file lock = LOCK_INIT;
3488         struct commit *head_commit = get_ref(head, o->branch1);
3489         struct commit *next_commit = get_ref(merge, o->branch2);
3490         struct commit_list *ca = NULL;
3491
3492         if (base_list) {
3493                 int i;
3494                 for (i = 0; i < num_base_list; ++i) {
3495                         struct commit *base;
3496                         if (!(base = get_ref(base_list[i], oid_to_hex(base_list[i]))))
3497                                 return err(o, _("Could not parse object '%s'"),
3498                                         oid_to_hex(base_list[i]));
3499                         commit_list_insert(base, &ca);
3500                 }
3501         }
3502
3503         hold_locked_index(&lock, LOCK_DIE_ON_ERROR);
3504         clean = merge_recursive(o, head_commit, next_commit, ca,
3505                         result);
3506         if (clean < 0) {
3507                 rollback_lock_file(&lock);
3508                 return clean;
3509         }
3510
3511         if (write_locked_index(&the_index, &lock,
3512                                COMMIT_LOCK | SKIP_IF_UNCHANGED))
3513                 return err(o, _("Unable to write index."));
3514
3515         return clean ? 0 : 1;
3516 }
3517
3518 static void merge_recursive_config(struct merge_options *o)
3519 {
3520         char *value = NULL;
3521         git_config_get_int("merge.verbosity", &o->verbosity);
3522         git_config_get_int("diff.renamelimit", &o->diff_rename_limit);
3523         git_config_get_int("merge.renamelimit", &o->merge_rename_limit);
3524         if (!git_config_get_string("diff.renames", &value)) {
3525                 o->diff_detect_rename = git_config_rename("diff.renames", value);
3526                 free(value);
3527         }
3528         if (!git_config_get_string("merge.renames", &value)) {
3529                 o->merge_detect_rename = git_config_rename("merge.renames", value);
3530                 free(value);
3531         }
3532         git_config(git_xmerge_config, NULL);
3533 }
3534
3535 void init_merge_options(struct merge_options *o)
3536 {
3537         const char *merge_verbosity;
3538         memset(o, 0, sizeof(struct merge_options));
3539         o->verbosity = 2;
3540         o->buffer_output = 1;
3541         o->diff_rename_limit = -1;
3542         o->merge_rename_limit = -1;
3543         o->renormalize = 0;
3544         o->diff_detect_rename = -1;
3545         o->merge_detect_rename = -1;
3546         merge_recursive_config(o);
3547         merge_verbosity = getenv("GIT_MERGE_VERBOSITY");
3548         if (merge_verbosity)
3549                 o->verbosity = strtol(merge_verbosity, NULL, 10);
3550         if (o->verbosity >= 5)
3551                 o->buffer_output = 0;
3552         strbuf_init(&o->obuf, 0);
3553         string_list_init(&o->df_conflict_file_set, 1);
3554 }
3555
3556 int parse_merge_opt(struct merge_options *o, const char *s)
3557 {
3558         const char *arg;
3559
3560         if (!s || !*s)
3561                 return -1;
3562         if (!strcmp(s, "ours"))
3563                 o->recursive_variant = MERGE_RECURSIVE_OURS;
3564         else if (!strcmp(s, "theirs"))
3565                 o->recursive_variant = MERGE_RECURSIVE_THEIRS;
3566         else if (!strcmp(s, "subtree"))
3567                 o->subtree_shift = "";
3568         else if (skip_prefix(s, "subtree=", &arg))
3569                 o->subtree_shift = arg;
3570         else if (!strcmp(s, "patience"))
3571                 o->xdl_opts = DIFF_WITH_ALG(o, PATIENCE_DIFF);
3572         else if (!strcmp(s, "histogram"))
3573                 o->xdl_opts = DIFF_WITH_ALG(o, HISTOGRAM_DIFF);
3574         else if (skip_prefix(s, "diff-algorithm=", &arg)) {
3575                 long value = parse_algorithm_value(arg);
3576                 if (value < 0)
3577                         return -1;
3578                 /* clear out previous settings */
3579                 DIFF_XDL_CLR(o, NEED_MINIMAL);
3580                 o->xdl_opts &= ~XDF_DIFF_ALGORITHM_MASK;
3581                 o->xdl_opts |= value;
3582         }
3583         else if (!strcmp(s, "ignore-space-change"))
3584                 DIFF_XDL_SET(o, IGNORE_WHITESPACE_CHANGE);
3585         else if (!strcmp(s, "ignore-all-space"))
3586                 DIFF_XDL_SET(o, IGNORE_WHITESPACE);
3587         else if (!strcmp(s, "ignore-space-at-eol"))
3588                 DIFF_XDL_SET(o, IGNORE_WHITESPACE_AT_EOL);
3589         else if (!strcmp(s, "ignore-cr-at-eol"))
3590                 DIFF_XDL_SET(o, IGNORE_CR_AT_EOL);
3591         else if (!strcmp(s, "renormalize"))
3592                 o->renormalize = 1;
3593         else if (!strcmp(s, "no-renormalize"))
3594                 o->renormalize = 0;
3595         else if (!strcmp(s, "no-renames"))
3596                 o->merge_detect_rename = 0;
3597         else if (!strcmp(s, "find-renames")) {
3598                 o->merge_detect_rename = 1;
3599                 o->rename_score = 0;
3600         }
3601         else if (skip_prefix(s, "find-renames=", &arg) ||
3602                  skip_prefix(s, "rename-threshold=", &arg)) {
3603                 if ((o->rename_score = parse_rename_score(&arg)) == -1 || *arg != 0)
3604                         return -1;
3605                 o->merge_detect_rename = 1;
3606         }
3607         else
3608                 return -1;
3609         return 0;
3610 }