merge-recursive: remove ren[12]_other fields from rename_conflict_info
[git] / merge-recursive.c
1 /*
2  * Recursive Merge algorithm stolen from git-merge-recursive.py by
3  * Fredrik Kuivinen.
4  * The thieves were Alex Riesen and Johannes Schindelin, in June/July 2006
5  */
6 #include "cache.h"
7 #include "config.h"
8 #include "advice.h"
9 #include "lockfile.h"
10 #include "cache-tree.h"
11 #include "object-store.h"
12 #include "repository.h"
13 #include "commit.h"
14 #include "blob.h"
15 #include "builtin.h"
16 #include "tree-walk.h"
17 #include "diff.h"
18 #include "diffcore.h"
19 #include "tag.h"
20 #include "alloc.h"
21 #include "unpack-trees.h"
22 #include "string-list.h"
23 #include "xdiff-interface.h"
24 #include "ll-merge.h"
25 #include "attr.h"
26 #include "merge-recursive.h"
27 #include "dir.h"
28 #include "submodule.h"
29 #include "revision.h"
30 #include "commit-reach.h"
31
32 struct path_hashmap_entry {
33         struct hashmap_entry e;
34         char path[FLEX_ARRAY];
35 };
36
37 static int path_hashmap_cmp(const void *cmp_data,
38                             const void *entry,
39                             const void *entry_or_key,
40                             const void *keydata)
41 {
42         const struct path_hashmap_entry *a = entry;
43         const struct path_hashmap_entry *b = entry_or_key;
44         const char *key = keydata;
45
46         if (ignore_case)
47                 return strcasecmp(a->path, key ? key : b->path);
48         else
49                 return strcmp(a->path, key ? key : b->path);
50 }
51
52 static unsigned int path_hash(const char *path)
53 {
54         return ignore_case ? strihash(path) : strhash(path);
55 }
56
57 static struct dir_rename_entry *dir_rename_find_entry(struct hashmap *hashmap,
58                                                       char *dir)
59 {
60         struct dir_rename_entry key;
61
62         if (dir == NULL)
63                 return NULL;
64         hashmap_entry_init(&key, strhash(dir));
65         key.dir = dir;
66         return hashmap_get(hashmap, &key, NULL);
67 }
68
69 static int dir_rename_cmp(const void *unused_cmp_data,
70                           const void *entry,
71                           const void *entry_or_key,
72                           const void *unused_keydata)
73 {
74         const struct dir_rename_entry *e1 = entry;
75         const struct dir_rename_entry *e2 = entry_or_key;
76
77         return strcmp(e1->dir, e2->dir);
78 }
79
80 static void dir_rename_init(struct hashmap *map)
81 {
82         hashmap_init(map, dir_rename_cmp, NULL, 0);
83 }
84
85 static void dir_rename_entry_init(struct dir_rename_entry *entry,
86                                   char *directory)
87 {
88         hashmap_entry_init(entry, strhash(directory));
89         entry->dir = directory;
90         entry->non_unique_new_dir = 0;
91         strbuf_init(&entry->new_dir, 0);
92         string_list_init(&entry->possible_new_dirs, 0);
93 }
94
95 static struct collision_entry *collision_find_entry(struct hashmap *hashmap,
96                                                     char *target_file)
97 {
98         struct collision_entry key;
99
100         hashmap_entry_init(&key, strhash(target_file));
101         key.target_file = target_file;
102         return hashmap_get(hashmap, &key, NULL);
103 }
104
105 static int collision_cmp(void *unused_cmp_data,
106                          const struct collision_entry *e1,
107                          const struct collision_entry *e2,
108                          const void *unused_keydata)
109 {
110         return strcmp(e1->target_file, e2->target_file);
111 }
112
113 static void collision_init(struct hashmap *map)
114 {
115         hashmap_init(map, (hashmap_cmp_fn) collision_cmp, NULL, 0);
116 }
117
118 static void flush_output(struct merge_options *opt)
119 {
120         if (opt->buffer_output < 2 && opt->obuf.len) {
121                 fputs(opt->obuf.buf, stdout);
122                 strbuf_reset(&opt->obuf);
123         }
124 }
125
126 static int err(struct merge_options *opt, const char *err, ...)
127 {
128         va_list params;
129
130         if (opt->buffer_output < 2)
131                 flush_output(opt);
132         else {
133                 strbuf_complete(&opt->obuf, '\n');
134                 strbuf_addstr(&opt->obuf, "error: ");
135         }
136         va_start(params, err);
137         strbuf_vaddf(&opt->obuf, err, params);
138         va_end(params);
139         if (opt->buffer_output > 1)
140                 strbuf_addch(&opt->obuf, '\n');
141         else {
142                 error("%s", opt->obuf.buf);
143                 strbuf_reset(&opt->obuf);
144         }
145
146         return -1;
147 }
148
149 static struct tree *shift_tree_object(struct repository *repo,
150                                       struct tree *one, struct tree *two,
151                                       const char *subtree_shift)
152 {
153         struct object_id shifted;
154
155         if (!*subtree_shift) {
156                 shift_tree(&one->object.oid, &two->object.oid, &shifted, 0);
157         } else {
158                 shift_tree_by(&one->object.oid, &two->object.oid, &shifted,
159                               subtree_shift);
160         }
161         if (oideq(&two->object.oid, &shifted))
162                 return two;
163         return lookup_tree(repo, &shifted);
164 }
165
166 static struct commit *make_virtual_commit(struct repository *repo,
167                                           struct tree *tree,
168                                           const char *comment)
169 {
170         struct commit *commit = alloc_commit_node(repo);
171
172         set_merge_remote_desc(commit, comment, (struct object *)commit);
173         commit->maybe_tree = tree;
174         commit->object.parsed = 1;
175         return commit;
176 }
177
178 /*
179  * Since we use get_tree_entry(), which does not put the read object into
180  * the object pool, we cannot rely on a == b.
181  */
182 static int oid_eq(const struct object_id *a, const struct object_id *b)
183 {
184         if (!a && !b)
185                 return 2;
186         return a && b && oideq(a, b);
187 }
188
189 enum rename_type {
190         RENAME_NORMAL = 0,
191         RENAME_VIA_DIR,
192         RENAME_ADD,
193         RENAME_DELETE,
194         RENAME_ONE_FILE_TO_ONE,
195         RENAME_ONE_FILE_TO_TWO,
196         RENAME_TWO_FILES_TO_ONE
197 };
198
199 /*
200  * Since we want to write the index eventually, we cannot reuse the index
201  * for these (temporary) data.
202  */
203 struct stage_data {
204         struct {
205                 unsigned short mode;
206                 struct object_id oid;
207         } stages[4];
208         struct rename_conflict_info *rename_conflict_info;
209         unsigned processed:1;
210 };
211
212 struct rename {
213         struct diff_filepair *pair;
214         /*
215          * Purpose of src_entry and dst_entry:
216          *
217          * If 'before' is renamed to 'after' then src_entry will contain
218          * the versions of 'before' from the merge_base, HEAD, and MERGE in
219          * stages 1, 2, and 3; dst_entry will contain the respective
220          * versions of 'after' in corresponding locations.  Thus, we have a
221          * total of six modes and oids, though some will be null.  (Stage 0
222          * is ignored; we're interested in handling conflicts.)
223          *
224          * Since we don't turn on break-rewrites by default, neither
225          * src_entry nor dst_entry can have all three of their stages have
226          * non-null oids, meaning at most four of the six will be non-null.
227          * Also, since this is a rename, both src_entry and dst_entry will
228          * have at least one non-null oid, meaning at least two will be
229          * non-null.  Of the six oids, a typical rename will have three be
230          * non-null.  Only two implies a rename/delete, and four implies a
231          * rename/add.
232          */
233         struct stage_data *src_entry;
234         struct stage_data *dst_entry;
235         unsigned add_turned_into_rename:1;
236         unsigned processed:1;
237 };
238
239 struct rename_conflict_info {
240         enum rename_type rename_type;
241         struct rename *ren1;
242         struct rename *ren2;
243         const char *branch1;
244         const char *branch2;
245 };
246
247 static inline void setup_rename_conflict_info(enum rename_type rename_type,
248                                               struct merge_options *opt,
249                                               struct rename *ren1,
250                                               struct rename *ren2,
251                                               const char *branch1,
252                                               const char *branch2)
253 {
254         struct rename_conflict_info *ci;
255
256         /*
257          * When we have two renames involved, it's easiest to get the
258          * correct things into stage 2 and 3, and to make sure that the
259          * content merge puts HEAD before the other branch if we just
260          * ensure that branch1 == opt->branch1.  So, simply flip arguments
261          * around if we don't have that.
262          */
263         if (ren2 && branch1 != opt->branch1) {
264                 setup_rename_conflict_info(rename_type,
265                                            opt,
266                                            ren2,       ren1,
267                                            branch2,    branch1);
268                 return;
269         }
270
271         ci = xcalloc(1, sizeof(struct rename_conflict_info));
272         ci->rename_type = rename_type;
273         ci->ren1 = ren1;
274         ci->ren2 = ren2;
275         ci->branch1 = branch1;
276         ci->branch2 = branch2;
277
278         ci->ren1->dst_entry->processed = 0;
279         ci->ren1->dst_entry->rename_conflict_info = ci;
280
281         if (ren2) {
282                 ci->ren2->dst_entry->rename_conflict_info = ci;
283         }
284 }
285
286 static int show(struct merge_options *opt, int v)
287 {
288         return (!opt->call_depth && opt->verbosity >= v) || opt->verbosity >= 5;
289 }
290
291 __attribute__((format (printf, 3, 4)))
292 static void output(struct merge_options *opt, int v, const char *fmt, ...)
293 {
294         va_list ap;
295
296         if (!show(opt, v))
297                 return;
298
299         strbuf_addchars(&opt->obuf, ' ', opt->call_depth * 2);
300
301         va_start(ap, fmt);
302         strbuf_vaddf(&opt->obuf, fmt, ap);
303         va_end(ap);
304
305         strbuf_addch(&opt->obuf, '\n');
306         if (!opt->buffer_output)
307                 flush_output(opt);
308 }
309
310 static void output_commit_title(struct merge_options *opt, struct commit *commit)
311 {
312         struct merge_remote_desc *desc;
313
314         strbuf_addchars(&opt->obuf, ' ', opt->call_depth * 2);
315         desc = merge_remote_util(commit);
316         if (desc)
317                 strbuf_addf(&opt->obuf, "virtual %s\n", desc->name);
318         else {
319                 strbuf_add_unique_abbrev(&opt->obuf, &commit->object.oid,
320                                          DEFAULT_ABBREV);
321                 strbuf_addch(&opt->obuf, ' ');
322                 if (parse_commit(commit) != 0)
323                         strbuf_addstr(&opt->obuf, _("(bad commit)\n"));
324                 else {
325                         const char *title;
326                         const char *msg = get_commit_buffer(commit, NULL);
327                         int len = find_commit_subject(msg, &title);
328                         if (len)
329                                 strbuf_addf(&opt->obuf, "%.*s\n", len, title);
330                         unuse_commit_buffer(commit, msg);
331                 }
332         }
333         flush_output(opt);
334 }
335
336 static int add_cacheinfo(struct merge_options *opt,
337                          unsigned int mode, const struct object_id *oid,
338                          const char *path, int stage, int refresh, int options)
339 {
340         struct index_state *istate = opt->repo->index;
341         struct cache_entry *ce;
342         int ret;
343
344         ce = make_cache_entry(istate, mode, oid ? oid : &null_oid, path, stage, 0);
345         if (!ce)
346                 return err(opt, _("add_cacheinfo failed for path '%s'; merge aborting."), path);
347
348         ret = add_index_entry(istate, ce, options);
349         if (refresh) {
350                 struct cache_entry *nce;
351
352                 nce = refresh_cache_entry(istate, ce,
353                                           CE_MATCH_REFRESH | CE_MATCH_IGNORE_MISSING);
354                 if (!nce)
355                         return err(opt, _("add_cacheinfo failed to refresh for path '%s'; merge aborting."), path);
356                 if (nce != ce)
357                         ret = add_index_entry(istate, nce, options);
358         }
359         return ret;
360 }
361
362 static void init_tree_desc_from_tree(struct tree_desc *desc, struct tree *tree)
363 {
364         parse_tree(tree);
365         init_tree_desc(desc, tree->buffer, tree->size);
366 }
367
368 static int unpack_trees_start(struct merge_options *opt,
369                               struct tree *common,
370                               struct tree *head,
371                               struct tree *merge)
372 {
373         int rc;
374         struct tree_desc t[3];
375         struct index_state tmp_index = { NULL };
376
377         memset(&opt->unpack_opts, 0, sizeof(opt->unpack_opts));
378         if (opt->call_depth)
379                 opt->unpack_opts.index_only = 1;
380         else
381                 opt->unpack_opts.update = 1;
382         opt->unpack_opts.merge = 1;
383         opt->unpack_opts.head_idx = 2;
384         opt->unpack_opts.fn = threeway_merge;
385         opt->unpack_opts.src_index = opt->repo->index;
386         opt->unpack_opts.dst_index = &tmp_index;
387         opt->unpack_opts.aggressive = !merge_detect_rename(opt);
388         setup_unpack_trees_porcelain(&opt->unpack_opts, "merge");
389
390         init_tree_desc_from_tree(t+0, common);
391         init_tree_desc_from_tree(t+1, head);
392         init_tree_desc_from_tree(t+2, merge);
393
394         rc = unpack_trees(3, t, &opt->unpack_opts);
395         cache_tree_free(&opt->repo->index->cache_tree);
396
397         /*
398          * Update opt->repo->index to match the new results, AFTER saving a copy
399          * in opt->orig_index.  Update src_index to point to the saved copy.
400          * (verify_uptodate() checks src_index, and the original index is
401          * the one that had the necessary modification timestamps.)
402          */
403         opt->orig_index = *opt->repo->index;
404         *opt->repo->index = tmp_index;
405         opt->unpack_opts.src_index = &opt->orig_index;
406
407         return rc;
408 }
409
410 static void unpack_trees_finish(struct merge_options *opt)
411 {
412         discard_index(&opt->orig_index);
413         clear_unpack_trees_porcelain(&opt->unpack_opts);
414 }
415
416 struct tree *write_tree_from_memory(struct merge_options *opt)
417 {
418         struct tree *result = NULL;
419         struct index_state *istate = opt->repo->index;
420
421         if (unmerged_index(istate)) {
422                 int i;
423                 fprintf(stderr, "BUG: There are unmerged index entries:\n");
424                 for (i = 0; i < istate->cache_nr; i++) {
425                         const struct cache_entry *ce = istate->cache[i];
426                         if (ce_stage(ce))
427                                 fprintf(stderr, "BUG: %d %.*s\n", ce_stage(ce),
428                                         (int)ce_namelen(ce), ce->name);
429                 }
430                 BUG("unmerged index entries in merge-recursive.c");
431         }
432
433         if (!istate->cache_tree)
434                 istate->cache_tree = cache_tree();
435
436         if (!cache_tree_fully_valid(istate->cache_tree) &&
437             cache_tree_update(istate, 0) < 0) {
438                 err(opt, _("error building trees"));
439                 return NULL;
440         }
441
442         result = lookup_tree(opt->repo, &istate->cache_tree->oid);
443
444         return result;
445 }
446
447 static int save_files_dirs(const struct object_id *oid,
448                            struct strbuf *base, const char *path,
449                            unsigned int mode, int stage, void *context)
450 {
451         struct path_hashmap_entry *entry;
452         int baselen = base->len;
453         struct merge_options *opt = context;
454
455         strbuf_addstr(base, path);
456
457         FLEX_ALLOC_MEM(entry, path, base->buf, base->len);
458         hashmap_entry_init(entry, path_hash(entry->path));
459         hashmap_add(&opt->current_file_dir_set, entry);
460
461         strbuf_setlen(base, baselen);
462         return (S_ISDIR(mode) ? READ_TREE_RECURSIVE : 0);
463 }
464
465 static void get_files_dirs(struct merge_options *opt, struct tree *tree)
466 {
467         struct pathspec match_all;
468         memset(&match_all, 0, sizeof(match_all));
469         read_tree_recursive(the_repository, tree, "", 0, 0,
470                             &match_all, save_files_dirs, opt);
471 }
472
473 static int get_tree_entry_if_blob(const struct object_id *tree,
474                                   const char *path,
475                                   struct object_id *hashy,
476                                   unsigned short *mode_o)
477 {
478         int ret;
479
480         ret = get_tree_entry(tree, path, hashy, mode_o);
481         if (S_ISDIR(*mode_o)) {
482                 oidcpy(hashy, &null_oid);
483                 *mode_o = 0;
484         }
485         return ret;
486 }
487
488 /*
489  * Returns an index_entry instance which doesn't have to correspond to
490  * a real cache entry in Git's index.
491  */
492 static struct stage_data *insert_stage_data(const char *path,
493                 struct tree *o, struct tree *a, struct tree *b,
494                 struct string_list *entries)
495 {
496         struct string_list_item *item;
497         struct stage_data *e = xcalloc(1, sizeof(struct stage_data));
498         get_tree_entry_if_blob(&o->object.oid, path,
499                                &e->stages[1].oid, &e->stages[1].mode);
500         get_tree_entry_if_blob(&a->object.oid, path,
501                                &e->stages[2].oid, &e->stages[2].mode);
502         get_tree_entry_if_blob(&b->object.oid, path,
503                                &e->stages[3].oid, &e->stages[3].mode);
504         item = string_list_insert(entries, path);
505         item->util = e;
506         return e;
507 }
508
509 /*
510  * Create a dictionary mapping file names to stage_data objects. The
511  * dictionary contains one entry for every path with a non-zero stage entry.
512  */
513 static struct string_list *get_unmerged(struct index_state *istate)
514 {
515         struct string_list *unmerged = xcalloc(1, sizeof(struct string_list));
516         int i;
517
518         unmerged->strdup_strings = 1;
519
520         for (i = 0; i < istate->cache_nr; i++) {
521                 struct string_list_item *item;
522                 struct stage_data *e;
523                 const struct cache_entry *ce = istate->cache[i];
524                 if (!ce_stage(ce))
525                         continue;
526
527                 item = string_list_lookup(unmerged, ce->name);
528                 if (!item) {
529                         item = string_list_insert(unmerged, ce->name);
530                         item->util = xcalloc(1, sizeof(struct stage_data));
531                 }
532                 e = item->util;
533                 e->stages[ce_stage(ce)].mode = ce->ce_mode;
534                 oidcpy(&e->stages[ce_stage(ce)].oid, &ce->oid);
535         }
536
537         return unmerged;
538 }
539
540 static int string_list_df_name_compare(const char *one, const char *two)
541 {
542         int onelen = strlen(one);
543         int twolen = strlen(two);
544         /*
545          * Here we only care that entries for D/F conflicts are
546          * adjacent, in particular with the file of the D/F conflict
547          * appearing before files below the corresponding directory.
548          * The order of the rest of the list is irrelevant for us.
549          *
550          * To achieve this, we sort with df_name_compare and provide
551          * the mode S_IFDIR so that D/F conflicts will sort correctly.
552          * We use the mode S_IFDIR for everything else for simplicity,
553          * since in other cases any changes in their order due to
554          * sorting cause no problems for us.
555          */
556         int cmp = df_name_compare(one, onelen, S_IFDIR,
557                                   two, twolen, S_IFDIR);
558         /*
559          * Now that 'foo' and 'foo/bar' compare equal, we have to make sure
560          * that 'foo' comes before 'foo/bar'.
561          */
562         if (cmp)
563                 return cmp;
564         return onelen - twolen;
565 }
566
567 static void record_df_conflict_files(struct merge_options *opt,
568                                      struct string_list *entries)
569 {
570         /* If there is a D/F conflict and the file for such a conflict
571          * currently exists in the working tree, we want to allow it to be
572          * removed to make room for the corresponding directory if needed.
573          * The files underneath the directories of such D/F conflicts will
574          * be processed before the corresponding file involved in the D/F
575          * conflict.  If the D/F directory ends up being removed by the
576          * merge, then we won't have to touch the D/F file.  If the D/F
577          * directory needs to be written to the working copy, then the D/F
578          * file will simply be removed (in make_room_for_path()) to make
579          * room for the necessary paths.  Note that if both the directory
580          * and the file need to be present, then the D/F file will be
581          * reinstated with a new unique name at the time it is processed.
582          */
583         struct string_list df_sorted_entries = STRING_LIST_INIT_NODUP;
584         const char *last_file = NULL;
585         int last_len = 0;
586         int i;
587
588         /*
589          * If we're merging merge-bases, we don't want to bother with
590          * any working directory changes.
591          */
592         if (opt->call_depth)
593                 return;
594
595         /* Ensure D/F conflicts are adjacent in the entries list. */
596         for (i = 0; i < entries->nr; i++) {
597                 struct string_list_item *next = &entries->items[i];
598                 string_list_append(&df_sorted_entries, next->string)->util =
599                                    next->util;
600         }
601         df_sorted_entries.cmp = string_list_df_name_compare;
602         string_list_sort(&df_sorted_entries);
603
604         string_list_clear(&opt->df_conflict_file_set, 1);
605         for (i = 0; i < df_sorted_entries.nr; i++) {
606                 const char *path = df_sorted_entries.items[i].string;
607                 int len = strlen(path);
608                 struct stage_data *e = df_sorted_entries.items[i].util;
609
610                 /*
611                  * Check if last_file & path correspond to a D/F conflict;
612                  * i.e. whether path is last_file+'/'+<something>.
613                  * If so, record that it's okay to remove last_file to make
614                  * room for path and friends if needed.
615                  */
616                 if (last_file &&
617                     len > last_len &&
618                     memcmp(path, last_file, last_len) == 0 &&
619                     path[last_len] == '/') {
620                         string_list_insert(&opt->df_conflict_file_set, last_file);
621                 }
622
623                 /*
624                  * Determine whether path could exist as a file in the
625                  * working directory as a possible D/F conflict.  This
626                  * will only occur when it exists in stage 2 as a
627                  * file.
628                  */
629                 if (S_ISREG(e->stages[2].mode) || S_ISLNK(e->stages[2].mode)) {
630                         last_file = path;
631                         last_len = len;
632                 } else {
633                         last_file = NULL;
634                 }
635         }
636         string_list_clear(&df_sorted_entries, 0);
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_index(opt->repo->index, 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 void update_entry(struct stage_data *entry,
671                          struct diff_filespec *o,
672                          struct diff_filespec *a,
673                          struct diff_filespec *b)
674 {
675         entry->processed = 0;
676         entry->stages[1].mode = o->mode;
677         entry->stages[2].mode = a->mode;
678         entry->stages[3].mode = b->mode;
679         oidcpy(&entry->stages[1].oid, &o->oid);
680         oidcpy(&entry->stages[2].oid, &a->oid);
681         oidcpy(&entry->stages[3].oid, &b->oid);
682 }
683
684 static int remove_file(struct merge_options *opt, int clean,
685                        const char *path, int no_wd)
686 {
687         int update_cache = opt->call_depth || clean;
688         int update_working_directory = !opt->call_depth && !no_wd;
689
690         if (update_cache) {
691                 if (remove_file_from_index(opt->repo->index, path))
692                         return -1;
693         }
694         if (update_working_directory) {
695                 if (ignore_case) {
696                         struct cache_entry *ce;
697                         ce = index_file_exists(opt->repo->index, path, strlen(path),
698                                                ignore_case);
699                         if (ce && ce_stage(ce) == 0 && strcmp(path, ce->name))
700                                 return 0;
701                 }
702                 if (remove_path(path))
703                         return -1;
704         }
705         return 0;
706 }
707
708 /* add a string to a strbuf, but converting "/" to "_" */
709 static void add_flattened_path(struct strbuf *out, const char *s)
710 {
711         size_t i = out->len;
712         strbuf_addstr(out, s);
713         for (; i < out->len; i++)
714                 if (out->buf[i] == '/')
715                         out->buf[i] = '_';
716 }
717
718 static char *unique_path(struct merge_options *opt, const char *path, const char *branch)
719 {
720         struct path_hashmap_entry *entry;
721         struct strbuf newpath = STRBUF_INIT;
722         int suffix = 0;
723         size_t base_len;
724
725         strbuf_addf(&newpath, "%s~", path);
726         add_flattened_path(&newpath, branch);
727
728         base_len = newpath.len;
729         while (hashmap_get_from_hash(&opt->current_file_dir_set,
730                                      path_hash(newpath.buf), newpath.buf) ||
731                (!opt->call_depth && file_exists(newpath.buf))) {
732                 strbuf_setlen(&newpath, base_len);
733                 strbuf_addf(&newpath, "_%d", suffix++);
734         }
735
736         FLEX_ALLOC_MEM(entry, path, newpath.buf, newpath.len);
737         hashmap_entry_init(entry, path_hash(entry->path));
738         hashmap_add(&opt->current_file_dir_set, entry);
739         return strbuf_detach(&newpath, NULL);
740 }
741
742 /**
743  * Check whether a directory in the index is in the way of an incoming
744  * file.  Return 1 if so.  If check_working_copy is non-zero, also
745  * check the working directory.  If empty_ok is non-zero, also return
746  * 0 in the case where the working-tree dir exists but is empty.
747  */
748 static int dir_in_way(struct index_state *istate, const char *path,
749                       int check_working_copy, int empty_ok)
750 {
751         int pos;
752         struct strbuf dirpath = STRBUF_INIT;
753         struct stat st;
754
755         strbuf_addstr(&dirpath, path);
756         strbuf_addch(&dirpath, '/');
757
758         pos = index_name_pos(istate, dirpath.buf, dirpath.len);
759
760         if (pos < 0)
761                 pos = -1 - pos;
762         if (pos < istate->cache_nr &&
763             !strncmp(dirpath.buf, istate->cache[pos]->name, dirpath.len)) {
764                 strbuf_release(&dirpath);
765                 return 1;
766         }
767
768         strbuf_release(&dirpath);
769         return check_working_copy && !lstat(path, &st) && S_ISDIR(st.st_mode) &&
770                 !(empty_ok && is_empty_dir(path));
771 }
772
773 /*
774  * Returns whether path was tracked in the index before the merge started,
775  * and its oid and mode match the specified values
776  */
777 static int was_tracked_and_matches(struct merge_options *opt, const char *path,
778                                    const struct object_id *oid, unsigned mode)
779 {
780         int pos = index_name_pos(&opt->orig_index, path, strlen(path));
781         struct cache_entry *ce;
782
783         if (0 > pos)
784                 /* we were not tracking this path before the merge */
785                 return 0;
786
787         /* See if the file we were tracking before matches */
788         ce = opt->orig_index.cache[pos];
789         return (oid_eq(&ce->oid, oid) && ce->ce_mode == mode);
790 }
791
792 /*
793  * Returns whether path was tracked in the index before the merge started
794  */
795 static int was_tracked(struct merge_options *opt, const char *path)
796 {
797         int pos = index_name_pos(&opt->orig_index, path, strlen(path));
798
799         if (0 <= pos)
800                 /* we were tracking this path before the merge */
801                 return 1;
802
803         return 0;
804 }
805
806 static int would_lose_untracked(struct merge_options *opt, const char *path)
807 {
808         struct index_state *istate = opt->repo->index;
809
810         /*
811          * This may look like it can be simplified to:
812          *   return !was_tracked(opt, path) && file_exists(path)
813          * but it can't.  This function needs to know whether path was in
814          * the working tree due to EITHER having been tracked in the index
815          * before the merge OR having been put into the working copy and
816          * index by unpack_trees().  Due to that either-or requirement, we
817          * check the current index instead of the original one.
818          *
819          * Note that we do not need to worry about merge-recursive itself
820          * updating the index after unpack_trees() and before calling this
821          * function, because we strictly require all code paths in
822          * merge-recursive to update the working tree first and the index
823          * second.  Doing otherwise would break
824          * update_file()/would_lose_untracked(); see every comment in this
825          * file which mentions "update_stages".
826          */
827         int pos = index_name_pos(istate, path, strlen(path));
828
829         if (pos < 0)
830                 pos = -1 - pos;
831         while (pos < istate->cache_nr &&
832                !strcmp(path, istate->cache[pos]->name)) {
833                 /*
834                  * If stage #0, it is definitely tracked.
835                  * If it has stage #2 then it was tracked
836                  * before this merge started.  All other
837                  * cases the path was not tracked.
838                  */
839                 switch (ce_stage(istate->cache[pos])) {
840                 case 0:
841                 case 2:
842                         return 0;
843                 }
844                 pos++;
845         }
846         return file_exists(path);
847 }
848
849 static int was_dirty(struct merge_options *opt, const char *path)
850 {
851         struct cache_entry *ce;
852         int dirty = 1;
853
854         if (opt->call_depth || !was_tracked(opt, path))
855                 return !dirty;
856
857         ce = index_file_exists(opt->unpack_opts.src_index,
858                                path, strlen(path), ignore_case);
859         dirty = verify_uptodate(ce, &opt->unpack_opts) != 0;
860         return dirty;
861 }
862
863 static int make_room_for_path(struct merge_options *opt, const char *path)
864 {
865         int status, i;
866         const char *msg = _("failed to create path '%s'%s");
867
868         /* Unlink any D/F conflict files that are in the way */
869         for (i = 0; i < opt->df_conflict_file_set.nr; i++) {
870                 const char *df_path = opt->df_conflict_file_set.items[i].string;
871                 size_t pathlen = strlen(path);
872                 size_t df_pathlen = strlen(df_path);
873                 if (df_pathlen < pathlen &&
874                     path[df_pathlen] == '/' &&
875                     strncmp(path, df_path, df_pathlen) == 0) {
876                         output(opt, 3,
877                                _("Removing %s to make room for subdirectory\n"),
878                                df_path);
879                         unlink(df_path);
880                         unsorted_string_list_delete_item(&opt->df_conflict_file_set,
881                                                          i, 0);
882                         break;
883                 }
884         }
885
886         /* Make sure leading directories are created */
887         status = safe_create_leading_directories_const(path);
888         if (status) {
889                 if (status == SCLD_EXISTS)
890                         /* something else exists */
891                         return err(opt, msg, path, _(": perhaps a D/F conflict?"));
892                 return err(opt, msg, path, "");
893         }
894
895         /*
896          * Do not unlink a file in the work tree if we are not
897          * tracking it.
898          */
899         if (would_lose_untracked(opt, path))
900                 return err(opt, _("refusing to lose untracked file at '%s'"),
901                            path);
902
903         /* Successful unlink is good.. */
904         if (!unlink(path))
905                 return 0;
906         /* .. and so is no existing file */
907         if (errno == ENOENT)
908                 return 0;
909         /* .. but not some other error (who really cares what?) */
910         return err(opt, msg, path, _(": perhaps a D/F conflict?"));
911 }
912
913 static int update_file_flags(struct merge_options *opt,
914                              const struct object_id *oid,
915                              unsigned mode,
916                              const char *path,
917                              int update_cache,
918                              int update_wd)
919 {
920         int ret = 0;
921
922         if (opt->call_depth)
923                 update_wd = 0;
924
925         if (update_wd) {
926                 enum object_type type;
927                 void *buf;
928                 unsigned long size;
929
930                 if (S_ISGITLINK(mode)) {
931                         /*
932                          * We may later decide to recursively descend into
933                          * the submodule directory and update its index
934                          * and/or work tree, but we do not do that now.
935                          */
936                         update_wd = 0;
937                         goto update_index;
938                 }
939
940                 buf = read_object_file(oid, &type, &size);
941                 if (!buf)
942                         return err(opt, _("cannot read object %s '%s'"), oid_to_hex(oid), path);
943                 if (type != OBJ_BLOB) {
944                         ret = err(opt, _("blob expected for %s '%s'"), oid_to_hex(oid), path);
945                         goto free_buf;
946                 }
947                 if (S_ISREG(mode)) {
948                         struct strbuf strbuf = STRBUF_INIT;
949                         if (convert_to_working_tree(opt->repo->index, path, buf, size, &strbuf)) {
950                                 free(buf);
951                                 size = strbuf.len;
952                                 buf = strbuf_detach(&strbuf, NULL);
953                         }
954                 }
955
956                 if (make_room_for_path(opt, path) < 0) {
957                         update_wd = 0;
958                         goto free_buf;
959                 }
960                 if (S_ISREG(mode) || (!has_symlinks && S_ISLNK(mode))) {
961                         int fd;
962                         if (mode & 0100)
963                                 mode = 0777;
964                         else
965                                 mode = 0666;
966                         fd = open(path, O_WRONLY | O_TRUNC | O_CREAT, mode);
967                         if (fd < 0) {
968                                 ret = err(opt, _("failed to open '%s': %s"),
969                                           path, strerror(errno));
970                                 goto free_buf;
971                         }
972                         write_in_full(fd, buf, size);
973                         close(fd);
974                 } else if (S_ISLNK(mode)) {
975                         char *lnk = xmemdupz(buf, size);
976                         safe_create_leading_directories_const(path);
977                         unlink(path);
978                         if (symlink(lnk, path))
979                                 ret = err(opt, _("failed to symlink '%s': %s"),
980                                           path, strerror(errno));
981                         free(lnk);
982                 } else
983                         ret = err(opt,
984                                   _("do not know what to do with %06o %s '%s'"),
985                                   mode, oid_to_hex(oid), path);
986         free_buf:
987                 free(buf);
988         }
989 update_index:
990         if (!ret && update_cache)
991                 if (add_cacheinfo(opt, mode, oid, path, 0, update_wd,
992                                   ADD_CACHE_OK_TO_ADD))
993                         return -1;
994         return ret;
995 }
996
997 static int update_file(struct merge_options *opt,
998                        int clean,
999                        const struct object_id *oid,
1000                        unsigned mode,
1001                        const char *path)
1002 {
1003         return update_file_flags(opt, oid, mode, path, opt->call_depth || clean, !opt->call_depth);
1004 }
1005
1006 /* Low level file merging, update and removal */
1007
1008 struct merge_file_info {
1009         struct object_id oid;
1010         unsigned mode;
1011         unsigned clean:1,
1012                  merge:1;
1013 };
1014
1015 static int merge_3way(struct merge_options *opt,
1016                       mmbuffer_t *result_buf,
1017                       const struct diff_filespec *o,
1018                       const struct diff_filespec *a,
1019                       const struct diff_filespec *b,
1020                       const char *branch1,
1021                       const char *branch2,
1022                       const int extra_marker_size)
1023 {
1024         mmfile_t orig, src1, src2;
1025         struct ll_merge_options ll_opts = {0};
1026         char *base_name, *name1, *name2;
1027         int merge_status;
1028
1029         ll_opts.renormalize = opt->renormalize;
1030         ll_opts.extra_marker_size = extra_marker_size;
1031         ll_opts.xdl_opts = opt->xdl_opts;
1032
1033         if (opt->call_depth) {
1034                 ll_opts.virtual_ancestor = 1;
1035                 ll_opts.variant = 0;
1036         } else {
1037                 switch (opt->recursive_variant) {
1038                 case MERGE_RECURSIVE_OURS:
1039                         ll_opts.variant = XDL_MERGE_FAVOR_OURS;
1040                         break;
1041                 case MERGE_RECURSIVE_THEIRS:
1042                         ll_opts.variant = XDL_MERGE_FAVOR_THEIRS;
1043                         break;
1044                 default:
1045                         ll_opts.variant = 0;
1046                         break;
1047                 }
1048         }
1049
1050         if (strcmp(a->path, b->path) ||
1051             (opt->ancestor != NULL && strcmp(a->path, o->path) != 0)) {
1052                 base_name = opt->ancestor == NULL ? NULL :
1053                         mkpathdup("%s:%s", opt->ancestor, o->path);
1054                 name1 = mkpathdup("%s:%s", branch1, a->path);
1055                 name2 = mkpathdup("%s:%s", branch2, b->path);
1056         } else {
1057                 base_name = opt->ancestor == NULL ? NULL :
1058                         mkpathdup("%s", opt->ancestor);
1059                 name1 = mkpathdup("%s", branch1);
1060                 name2 = mkpathdup("%s", branch2);
1061         }
1062
1063         read_mmblob(&orig, &o->oid);
1064         read_mmblob(&src1, &a->oid);
1065         read_mmblob(&src2, &b->oid);
1066
1067         merge_status = ll_merge(result_buf, a->path, &orig, base_name,
1068                                 &src1, name1, &src2, name2,
1069                                 opt->repo->index, &ll_opts);
1070
1071         free(base_name);
1072         free(name1);
1073         free(name2);
1074         free(orig.ptr);
1075         free(src1.ptr);
1076         free(src2.ptr);
1077         return merge_status;
1078 }
1079
1080 static int find_first_merges(struct repository *repo,
1081                              struct object_array *result, const char *path,
1082                              struct commit *a, struct commit *b)
1083 {
1084         int i, j;
1085         struct object_array merges = OBJECT_ARRAY_INIT;
1086         struct commit *commit;
1087         int contains_another;
1088
1089         char merged_revision[42];
1090         const char *rev_args[] = { "rev-list", "--merges", "--ancestry-path",
1091                                    "--all", merged_revision, NULL };
1092         struct rev_info revs;
1093         struct setup_revision_opt rev_opts;
1094
1095         memset(result, 0, sizeof(struct object_array));
1096         memset(&rev_opts, 0, sizeof(rev_opts));
1097
1098         /* get all revisions that merge commit a */
1099         xsnprintf(merged_revision, sizeof(merged_revision), "^%s",
1100                   oid_to_hex(&a->object.oid));
1101         repo_init_revisions(repo, &revs, NULL);
1102         rev_opts.submodule = path;
1103         /* FIXME: can't handle linked worktrees in submodules yet */
1104         revs.single_worktree = path != NULL;
1105         setup_revisions(ARRAY_SIZE(rev_args)-1, rev_args, &revs, &rev_opts);
1106
1107         /* save all revisions from the above list that contain b */
1108         if (prepare_revision_walk(&revs))
1109                 die("revision walk setup failed");
1110         while ((commit = get_revision(&revs)) != NULL) {
1111                 struct object *o = &(commit->object);
1112                 if (in_merge_bases(b, commit))
1113                         add_object_array(o, NULL, &merges);
1114         }
1115         reset_revision_walk();
1116
1117         /* Now we've got all merges that contain a and b. Prune all
1118          * merges that contain another found merge and save them in
1119          * result.
1120          */
1121         for (i = 0; i < merges.nr; i++) {
1122                 struct commit *m1 = (struct commit *) merges.objects[i].item;
1123
1124                 contains_another = 0;
1125                 for (j = 0; j < merges.nr; j++) {
1126                         struct commit *m2 = (struct commit *) merges.objects[j].item;
1127                         if (i != j && in_merge_bases(m2, m1)) {
1128                                 contains_another = 1;
1129                                 break;
1130                         }
1131                 }
1132
1133                 if (!contains_another)
1134                         add_object_array(merges.objects[i].item, NULL, result);
1135         }
1136
1137         object_array_clear(&merges);
1138         return result->nr;
1139 }
1140
1141 static void print_commit(struct commit *commit)
1142 {
1143         struct strbuf sb = STRBUF_INIT;
1144         struct pretty_print_context ctx = {0};
1145         ctx.date_mode.type = DATE_NORMAL;
1146         format_commit_message(commit, " %h: %m %s", &sb, &ctx);
1147         fprintf(stderr, "%s\n", sb.buf);
1148         strbuf_release(&sb);
1149 }
1150
1151 static int merge_submodule(struct merge_options *opt,
1152                            struct object_id *result, const char *path,
1153                            const struct object_id *base, const struct object_id *a,
1154                            const struct object_id *b)
1155 {
1156         struct commit *commit_base, *commit_a, *commit_b;
1157         int parent_count;
1158         struct object_array merges;
1159
1160         int i;
1161         int search = !opt->call_depth;
1162
1163         /* store a in result in case we fail */
1164         oidcpy(result, a);
1165
1166         /* we can not handle deletion conflicts */
1167         if (is_null_oid(base))
1168                 return 0;
1169         if (is_null_oid(a))
1170                 return 0;
1171         if (is_null_oid(b))
1172                 return 0;
1173
1174         if (add_submodule_odb(path)) {
1175                 output(opt, 1, _("Failed to merge submodule %s (not checked out)"), path);
1176                 return 0;
1177         }
1178
1179         if (!(commit_base = lookup_commit_reference(opt->repo, base)) ||
1180             !(commit_a = lookup_commit_reference(opt->repo, a)) ||
1181             !(commit_b = lookup_commit_reference(opt->repo, b))) {
1182                 output(opt, 1, _("Failed to merge submodule %s (commits not present)"), path);
1183                 return 0;
1184         }
1185
1186         /* check whether both changes are forward */
1187         if (!in_merge_bases(commit_base, commit_a) ||
1188             !in_merge_bases(commit_base, commit_b)) {
1189                 output(opt, 1, _("Failed to merge submodule %s (commits don't follow merge-base)"), path);
1190                 return 0;
1191         }
1192
1193         /* Case #1: a is contained in b or vice versa */
1194         if (in_merge_bases(commit_a, commit_b)) {
1195                 oidcpy(result, b);
1196                 if (show(opt, 3)) {
1197                         output(opt, 3, _("Fast-forwarding submodule %s to the following commit:"), path);
1198                         output_commit_title(opt, commit_b);
1199                 } else if (show(opt, 2))
1200                         output(opt, 2, _("Fast-forwarding submodule %s"), path);
1201                 else
1202                         ; /* no output */
1203
1204                 return 1;
1205         }
1206         if (in_merge_bases(commit_b, commit_a)) {
1207                 oidcpy(result, a);
1208                 if (show(opt, 3)) {
1209                         output(opt, 3, _("Fast-forwarding submodule %s to the following commit:"), path);
1210                         output_commit_title(opt, commit_a);
1211                 } else if (show(opt, 2))
1212                         output(opt, 2, _("Fast-forwarding submodule %s"), path);
1213                 else
1214                         ; /* no output */
1215
1216                 return 1;
1217         }
1218
1219         /*
1220          * Case #2: There are one or more merges that contain a and b in
1221          * the submodule. If there is only one, then present it as a
1222          * suggestion to the user, but leave it marked unmerged so the
1223          * user needs to confirm the resolution.
1224          */
1225
1226         /* Skip the search if makes no sense to the calling context.  */
1227         if (!search)
1228                 return 0;
1229
1230         /* find commit which merges them */
1231         parent_count = find_first_merges(opt->repo, &merges, path,
1232                                          commit_a, commit_b);
1233         switch (parent_count) {
1234         case 0:
1235                 output(opt, 1, _("Failed to merge submodule %s (merge following commits not found)"), path);
1236                 break;
1237
1238         case 1:
1239                 output(opt, 1, _("Failed to merge submodule %s (not fast-forward)"), path);
1240                 output(opt, 2, _("Found a possible merge resolution for the submodule:\n"));
1241                 print_commit((struct commit *) merges.objects[0].item);
1242                 output(opt, 2, _(
1243                        "If this is correct simply add it to the index "
1244                        "for example\n"
1245                        "by using:\n\n"
1246                        "  git update-index --cacheinfo 160000 %s \"%s\"\n\n"
1247                        "which will accept this suggestion.\n"),
1248                        oid_to_hex(&merges.objects[0].item->oid), path);
1249                 break;
1250
1251         default:
1252                 output(opt, 1, _("Failed to merge submodule %s (multiple merges found)"), path);
1253                 for (i = 0; i < merges.nr; i++)
1254                         print_commit((struct commit *) merges.objects[i].item);
1255         }
1256
1257         object_array_clear(&merges);
1258         return 0;
1259 }
1260
1261 static int merge_mode_and_contents(struct merge_options *opt,
1262                                    const struct diff_filespec *o,
1263                                    const struct diff_filespec *a,
1264                                    const struct diff_filespec *b,
1265                                    const char *filename,
1266                                    const char *branch1,
1267                                    const char *branch2,
1268                                    const int extra_marker_size,
1269                                    struct merge_file_info *result)
1270 {
1271         if (opt->branch1 != branch1) {
1272                 /*
1273                  * It's weird getting a reverse merge with HEAD on the bottom
1274                  * side of the conflict markers and the other branch on the
1275                  * top.  Fix that.
1276                  */
1277                 return merge_mode_and_contents(opt, o, b, a,
1278                                                filename,
1279                                                branch2, branch1,
1280                                                extra_marker_size, result);
1281         }
1282
1283         result->merge = 0;
1284         result->clean = 1;
1285
1286         if ((S_IFMT & a->mode) != (S_IFMT & b->mode)) {
1287                 result->clean = 0;
1288                 if (S_ISREG(a->mode)) {
1289                         result->mode = a->mode;
1290                         oidcpy(&result->oid, &a->oid);
1291                 } else {
1292                         result->mode = b->mode;
1293                         oidcpy(&result->oid, &b->oid);
1294                 }
1295         } else {
1296                 if (!oid_eq(&a->oid, &o->oid) && !oid_eq(&b->oid, &o->oid))
1297                         result->merge = 1;
1298
1299                 /*
1300                  * Merge modes
1301                  */
1302                 if (a->mode == b->mode || a->mode == o->mode)
1303                         result->mode = b->mode;
1304                 else {
1305                         result->mode = a->mode;
1306                         if (b->mode != o->mode) {
1307                                 result->clean = 0;
1308                                 result->merge = 1;
1309                         }
1310                 }
1311
1312                 if (oid_eq(&a->oid, &b->oid) || oid_eq(&a->oid, &o->oid))
1313                         oidcpy(&result->oid, &b->oid);
1314                 else if (oid_eq(&b->oid, &o->oid))
1315                         oidcpy(&result->oid, &a->oid);
1316                 else if (S_ISREG(a->mode)) {
1317                         mmbuffer_t result_buf;
1318                         int ret = 0, merge_status;
1319
1320                         merge_status = merge_3way(opt, &result_buf, o, a, b,
1321                                                   branch1, branch2,
1322                                                   extra_marker_size);
1323
1324                         if ((merge_status < 0) || !result_buf.ptr)
1325                                 ret = err(opt, _("Failed to execute internal merge"));
1326
1327                         if (!ret &&
1328                             write_object_file(result_buf.ptr, result_buf.size,
1329                                               blob_type, &result->oid))
1330                                 ret = err(opt, _("Unable to add %s to database"),
1331                                           a->path);
1332
1333                         free(result_buf.ptr);
1334                         if (ret)
1335                                 return ret;
1336                         result->clean = (merge_status == 0);
1337                 } else if (S_ISGITLINK(a->mode)) {
1338                         result->clean = merge_submodule(opt, &result->oid,
1339                                                         o->path,
1340                                                         &o->oid,
1341                                                         &a->oid,
1342                                                         &b->oid);
1343                 } else if (S_ISLNK(a->mode)) {
1344                         switch (opt->recursive_variant) {
1345                         case MERGE_RECURSIVE_NORMAL:
1346                                 oidcpy(&result->oid, &a->oid);
1347                                 if (!oid_eq(&a->oid, &b->oid))
1348                                         result->clean = 0;
1349                                 break;
1350                         case MERGE_RECURSIVE_OURS:
1351                                 oidcpy(&result->oid, &a->oid);
1352                                 break;
1353                         case MERGE_RECURSIVE_THEIRS:
1354                                 oidcpy(&result->oid, &b->oid);
1355                                 break;
1356                         }
1357                 } else
1358                         BUG("unsupported object type in the tree");
1359         }
1360
1361         if (result->merge)
1362                 output(opt, 2, _("Auto-merging %s"), filename);
1363
1364         return 0;
1365 }
1366
1367 static int handle_rename_via_dir(struct merge_options *opt,
1368                                  struct diff_filepair *pair,
1369                                  const char *rename_branch)
1370 {
1371         /*
1372          * Handle file adds that need to be renamed due to directory rename
1373          * detection.  This differs from handle_rename_normal, because
1374          * there is no content merge to do; just move the file into the
1375          * desired final location.
1376          */
1377         const struct diff_filespec *dest = pair->two;
1378
1379         if (!opt->call_depth && would_lose_untracked(opt, dest->path)) {
1380                 char *alt_path = unique_path(opt, dest->path, rename_branch);
1381
1382                 output(opt, 1, _("Error: Refusing to lose untracked file at %s; "
1383                                "writing to %s instead."),
1384                        dest->path, alt_path);
1385                 /*
1386                  * Write the file in worktree at alt_path, but not in the
1387                  * index.  Instead, write to dest->path for the index but
1388                  * only at the higher appropriate stage.
1389                  */
1390                 if (update_file(opt, 0, &dest->oid, dest->mode, alt_path))
1391                         return -1;
1392                 free(alt_path);
1393                 return update_stages(opt, dest->path, NULL,
1394                                      rename_branch == opt->branch1 ? dest : NULL,
1395                                      rename_branch == opt->branch1 ? NULL : dest);
1396         }
1397
1398         /* Update dest->path both in index and in worktree */
1399         if (update_file(opt, 1, &dest->oid, dest->mode, dest->path))
1400                 return -1;
1401         return 0;
1402 }
1403
1404 static int handle_change_delete(struct merge_options *opt,
1405                                 const char *path, const char *old_path,
1406                                 const struct object_id *o_oid, int o_mode,
1407                                 const struct object_id *changed_oid,
1408                                 int changed_mode,
1409                                 const char *change_branch,
1410                                 const char *delete_branch,
1411                                 const char *change, const char *change_past)
1412 {
1413         char *alt_path = NULL;
1414         const char *update_path = path;
1415         int ret = 0;
1416
1417         if (dir_in_way(opt->repo->index, path, !opt->call_depth, 0) ||
1418             (!opt->call_depth && would_lose_untracked(opt, path))) {
1419                 update_path = alt_path = unique_path(opt, path, change_branch);
1420         }
1421
1422         if (opt->call_depth) {
1423                 /*
1424                  * We cannot arbitrarily accept either a_sha or b_sha as
1425                  * correct; since there is no true "middle point" between
1426                  * them, simply reuse the base version for virtual merge base.
1427                  */
1428                 ret = remove_file_from_index(opt->repo->index, path);
1429                 if (!ret)
1430                         ret = update_file(opt, 0, o_oid, o_mode, update_path);
1431         } else {
1432                 /*
1433                  * Despite the four nearly duplicate messages and argument
1434                  * lists below and the ugliness of the nested if-statements,
1435                  * having complete messages makes the job easier for
1436                  * translators.
1437                  *
1438                  * The slight variance among the cases is due to the fact
1439                  * that:
1440                  *   1) directory/file conflicts (in effect if
1441                  *      !alt_path) could cause us to need to write the
1442                  *      file to a different path.
1443                  *   2) renames (in effect if !old_path) could mean that
1444                  *      there are two names for the path that the user
1445                  *      may know the file by.
1446                  */
1447                 if (!alt_path) {
1448                         if (!old_path) {
1449                                 output(opt, 1, _("CONFLICT (%s/delete): %s deleted in %s "
1450                                        "and %s in %s. Version %s of %s left in tree."),
1451                                        change, path, delete_branch, change_past,
1452                                        change_branch, change_branch, path);
1453                         } else {
1454                                 output(opt, 1, _("CONFLICT (%s/delete): %s deleted in %s "
1455                                        "and %s to %s in %s. Version %s of %s left in tree."),
1456                                        change, old_path, delete_branch, change_past, path,
1457                                        change_branch, change_branch, path);
1458                         }
1459                 } else {
1460                         if (!old_path) {
1461                                 output(opt, 1, _("CONFLICT (%s/delete): %s deleted in %s "
1462                                        "and %s in %s. Version %s of %s left in tree at %s."),
1463                                        change, path, delete_branch, change_past,
1464                                        change_branch, change_branch, path, alt_path);
1465                         } else {
1466                                 output(opt, 1, _("CONFLICT (%s/delete): %s deleted in %s "
1467                                        "and %s to %s in %s. Version %s of %s left in tree at %s."),
1468                                        change, old_path, delete_branch, change_past, path,
1469                                        change_branch, change_branch, path, alt_path);
1470                         }
1471                 }
1472                 /*
1473                  * No need to call update_file() on path when change_branch ==
1474                  * opt->branch1 && !alt_path, since that would needlessly touch
1475                  * path.  We could call update_file_flags() with update_cache=0
1476                  * and update_wd=0, but that's a no-op.
1477                  */
1478                 if (change_branch != opt->branch1 || alt_path)
1479                         ret = update_file(opt, 0, changed_oid, changed_mode, update_path);
1480         }
1481         free(alt_path);
1482
1483         return ret;
1484 }
1485
1486 static int handle_rename_delete(struct merge_options *opt,
1487                                 struct diff_filepair *pair,
1488                                 const char *rename_branch,
1489                                 const char *delete_branch)
1490 {
1491         const struct diff_filespec *orig = pair->one;
1492         const struct diff_filespec *dest = pair->two;
1493
1494         if (handle_change_delete(opt,
1495                                  opt->call_depth ? orig->path : dest->path,
1496                                  opt->call_depth ? NULL : orig->path,
1497                                  &orig->oid, orig->mode,
1498                                  &dest->oid, dest->mode,
1499                                  rename_branch, delete_branch,
1500                                  _("rename"), _("renamed")))
1501                 return -1;
1502
1503         if (opt->call_depth)
1504                 return remove_file_from_index(opt->repo->index, dest->path);
1505         else
1506                 return update_stages(opt, dest->path, NULL,
1507                                      rename_branch == opt->branch1 ? dest : NULL,
1508                                      rename_branch == opt->branch1 ? NULL : dest);
1509 }
1510
1511 static struct diff_filespec *filespec_from_entry(struct diff_filespec *target,
1512                                                  struct stage_data *entry,
1513                                                  int stage)
1514 {
1515         struct object_id *oid = &entry->stages[stage].oid;
1516         unsigned mode = entry->stages[stage].mode;
1517         if (mode == 0 || is_null_oid(oid))
1518                 return NULL;
1519         oidcpy(&target->oid, oid);
1520         target->mode = mode;
1521         return target;
1522 }
1523
1524 static int handle_file_collision(struct merge_options *opt,
1525                                  const char *collide_path,
1526                                  const char *prev_path1,
1527                                  const char *prev_path2,
1528                                  const char *branch1, const char *branch2,
1529                                  const struct object_id *a_oid,
1530                                  unsigned int a_mode,
1531                                  const struct object_id *b_oid,
1532                                  unsigned int b_mode)
1533 {
1534         struct merge_file_info mfi;
1535         struct diff_filespec null, a, b;
1536         char *alt_path = NULL;
1537         const char *update_path = collide_path;
1538
1539         /*
1540          * It's easiest to get the correct things into stage 2 and 3, and
1541          * to make sure that the content merge puts HEAD before the other
1542          * branch if we just ensure that branch1 == opt->branch1.  So, simply
1543          * flip arguments around if we don't have that.
1544          */
1545         if (branch1 != opt->branch1) {
1546                 return handle_file_collision(opt, collide_path,
1547                                              prev_path2, prev_path1,
1548                                              branch2, branch1,
1549                                              b_oid, b_mode,
1550                                              a_oid, a_mode);
1551         }
1552
1553         /*
1554          * In the recursive case, we just opt to undo renames
1555          */
1556         if (opt->call_depth && (prev_path1 || prev_path2)) {
1557                 /* Put first file (a_oid, a_mode) in its original spot */
1558                 if (prev_path1) {
1559                         if (update_file(opt, 1, a_oid, a_mode, prev_path1))
1560                                 return -1;
1561                 } else {
1562                         if (update_file(opt, 1, a_oid, a_mode, collide_path))
1563                                 return -1;
1564                 }
1565
1566                 /* Put second file (b_oid, b_mode) in its original spot */
1567                 if (prev_path2) {
1568                         if (update_file(opt, 1, b_oid, b_mode, prev_path2))
1569                                 return -1;
1570                 } else {
1571                         if (update_file(opt, 1, b_oid, b_mode, collide_path))
1572                                 return -1;
1573                 }
1574
1575                 /* Don't leave something at collision path if unrenaming both */
1576                 if (prev_path1 && prev_path2)
1577                         remove_file(opt, 1, collide_path, 0);
1578
1579                 return 0;
1580         }
1581
1582         /* Remove rename sources if rename/add or rename/rename(2to1) */
1583         if (prev_path1)
1584                 remove_file(opt, 1, prev_path1,
1585                             opt->call_depth || would_lose_untracked(opt, prev_path1));
1586         if (prev_path2)
1587                 remove_file(opt, 1, prev_path2,
1588                             opt->call_depth || would_lose_untracked(opt, prev_path2));
1589
1590         /*
1591          * Remove the collision path, if it wouldn't cause dirty contents
1592          * or an untracked file to get lost.  We'll either overwrite with
1593          * merged contents, or just write out to differently named files.
1594          */
1595         if (was_dirty(opt, collide_path)) {
1596                 output(opt, 1, _("Refusing to lose dirty file at %s"),
1597                        collide_path);
1598                 update_path = alt_path = unique_path(opt, collide_path, "merged");
1599         } else if (would_lose_untracked(opt, collide_path)) {
1600                 /*
1601                  * Only way we get here is if both renames were from
1602                  * a directory rename AND user had an untracked file
1603                  * at the location where both files end up after the
1604                  * two directory renames.  See testcase 10d of t6043.
1605                  */
1606                 output(opt, 1, _("Refusing to lose untracked file at "
1607                                "%s, even though it's in the way."),
1608                        collide_path);
1609                 update_path = alt_path = unique_path(opt, collide_path, "merged");
1610         } else {
1611                 /*
1612                  * FIXME: It's possible that the two files are identical
1613                  * and that the current working copy happens to match, in
1614                  * which case we are unnecessarily touching the working
1615                  * tree file.  It's not a likely enough scenario that I
1616                  * want to code up the checks for it and a better fix is
1617                  * available if we restructure how unpack_trees() and
1618                  * merge-recursive interoperate anyway, so punting for
1619                  * now...
1620                  */
1621                 remove_file(opt, 0, collide_path, 0);
1622         }
1623
1624         /* Store things in diff_filespecs for functions that need it */
1625         memset(&a, 0, sizeof(struct diff_filespec));
1626         memset(&b, 0, sizeof(struct diff_filespec));
1627         null.path = a.path = b.path = (char *)collide_path;
1628         oidcpy(&null.oid, &null_oid);
1629         null.mode = 0;
1630         oidcpy(&a.oid, a_oid);
1631         a.mode = a_mode;
1632         a.oid_valid = 1;
1633         oidcpy(&b.oid, b_oid);
1634         b.mode = b_mode;
1635         b.oid_valid = 1;
1636
1637         if (merge_mode_and_contents(opt, &null, &a, &b, collide_path,
1638                                     branch1, branch2, opt->call_depth * 2, &mfi))
1639                 return -1;
1640         mfi.clean &= !alt_path;
1641         if (update_file(opt, mfi.clean, &mfi.oid, mfi.mode, update_path))
1642                 return -1;
1643         if (!mfi.clean && !opt->call_depth &&
1644             update_stages(opt, collide_path, NULL, &a, &b))
1645                 return -1;
1646         free(alt_path);
1647         /*
1648          * FIXME: If both a & b both started with conflicts (only possible
1649          * if they came from a rename/rename(2to1)), but had IDENTICAL
1650          * contents including those conflicts, then in the next line we claim
1651          * it was clean.  If someone cares about this case, we should have the
1652          * caller notify us if we started with conflicts.
1653          */
1654         return mfi.clean;
1655 }
1656
1657 static int handle_rename_add(struct merge_options *opt,
1658                              struct rename_conflict_info *ci)
1659 {
1660         /* a was renamed to c, and a separate c was added. */
1661         struct diff_filespec *a = ci->ren1->pair->one;
1662         struct diff_filespec *c = ci->ren1->pair->two;
1663         struct diff_filespec tmp;
1664         char *path = c->path;
1665         char *prev_path_desc;
1666         struct merge_file_info mfi;
1667
1668         int other_stage = (ci->branch1 == opt->branch1 ? 3 : 2);
1669
1670         output(opt, 1, _("CONFLICT (rename/add): "
1671                "Rename %s->%s in %s.  Added %s in %s"),
1672                a->path, c->path, ci->branch1,
1673                c->path, ci->branch2);
1674
1675         filespec_from_entry(&tmp, ci->ren1->src_entry, other_stage);
1676         tmp.path = a->path;
1677
1678         prev_path_desc = xstrfmt("version of %s from %s", path, a->path);
1679         if (merge_mode_and_contents(opt, a, c, &tmp,
1680                                     prev_path_desc,
1681                                     opt->branch1, opt->branch2,
1682                                     1 + opt->call_depth * 2, &mfi))
1683                 return -1;
1684         free(prev_path_desc);
1685
1686         return handle_file_collision(opt,
1687                                      c->path, a->path, NULL,
1688                                      ci->branch1, ci->branch2,
1689                                      &mfi.oid, mfi.mode,
1690                                      &ci->ren1->dst_entry->stages[other_stage].oid,
1691                                      ci->ren1->dst_entry->stages[other_stage].mode);
1692 }
1693
1694 static char *find_path_for_conflict(struct merge_options *opt,
1695                                     const char *path,
1696                                     const char *branch1,
1697                                     const char *branch2)
1698 {
1699         char *new_path = NULL;
1700         if (dir_in_way(opt->repo->index, path, !opt->call_depth, 0)) {
1701                 new_path = unique_path(opt, path, branch1);
1702                 output(opt, 1, _("%s is a directory in %s adding "
1703                                "as %s instead"),
1704                        path, branch2, new_path);
1705         } else if (would_lose_untracked(opt, path)) {
1706                 new_path = unique_path(opt, path, branch1);
1707                 output(opt, 1, _("Refusing to lose untracked file"
1708                                " at %s; adding as %s instead"),
1709                        path, new_path);
1710         }
1711
1712         return new_path;
1713 }
1714
1715 static int handle_rename_rename_1to2(struct merge_options *opt,
1716                                      struct rename_conflict_info *ci)
1717 {
1718         /* One file was renamed in both branches, but to different names. */
1719         struct merge_file_info mfi;
1720         struct diff_filespec other;
1721         struct diff_filespec *add;
1722         struct diff_filespec *o = ci->ren1->pair->one;
1723         struct diff_filespec *a = ci->ren1->pair->two;
1724         struct diff_filespec *b = ci->ren2->pair->two;
1725         char *path_desc;
1726
1727         output(opt, 1, _("CONFLICT (rename/rename): "
1728                "Rename \"%s\"->\"%s\" in branch \"%s\" "
1729                "rename \"%s\"->\"%s\" in \"%s\"%s"),
1730                o->path, a->path, ci->branch1,
1731                o->path, b->path, ci->branch2,
1732                opt->call_depth ? _(" (left unresolved)") : "");
1733
1734         path_desc = xstrfmt("%s and %s, both renamed from %s",
1735                             a->path, b->path, o->path);
1736         if (merge_mode_and_contents(opt, o, a, b, path_desc,
1737                                     ci->branch1, ci->branch2,
1738                                     opt->call_depth * 2, &mfi))
1739                 return -1;
1740         free(path_desc);
1741
1742         if (opt->call_depth) {
1743                 /*
1744                  * FIXME: For rename/add-source conflicts (if we could detect
1745                  * such), this is wrong.  We should instead find a unique
1746                  * pathname and then either rename the add-source file to that
1747                  * unique path, or use that unique path instead of src here.
1748                  */
1749                 if (update_file(opt, 0, &mfi.oid, mfi.mode, o->path))
1750                         return -1;
1751
1752                 /*
1753                  * Above, we put the merged content at the merge-base's
1754                  * path.  Now we usually need to delete both a->path and
1755                  * b->path.  However, the rename on each side of the merge
1756                  * could also be involved in a rename/add conflict.  In
1757                  * such cases, we should keep the added file around,
1758                  * resolving the conflict at that path in its favor.
1759                  */
1760                 add = filespec_from_entry(&other, ci->ren1->dst_entry, 2 ^ 1);
1761                 if (add) {
1762                         if (update_file(opt, 0, &add->oid, add->mode, a->path))
1763                                 return -1;
1764                 }
1765                 else
1766                         remove_file_from_index(opt->repo->index, a->path);
1767                 add = filespec_from_entry(&other, ci->ren2->dst_entry, 3 ^ 1);
1768                 if (add) {
1769                         if (update_file(opt, 0, &add->oid, add->mode, b->path))
1770                                 return -1;
1771                 }
1772                 else
1773                         remove_file_from_index(opt->repo->index, b->path);
1774         } else {
1775                 /*
1776                  * For each destination path, we need to see if there is a
1777                  * rename/add collision.  If not, we can write the file out
1778                  * to the specified location.
1779                  */
1780                 add = filespec_from_entry(&other, ci->ren1->dst_entry, 2 ^ 1);
1781                 if (add) {
1782                         if (handle_file_collision(opt, a->path,
1783                                                   NULL, NULL,
1784                                                   ci->branch1, ci->branch2,
1785                                                   &mfi.oid, mfi.mode,
1786                                                   &add->oid, add->mode) < 0)
1787                                 return -1;
1788                 } else {
1789                         char *new_path = find_path_for_conflict(opt, a->path,
1790                                                                 ci->branch1,
1791                                                                 ci->branch2);
1792                         if (update_file(opt, 0, &mfi.oid, mfi.mode, new_path ? new_path : a->path))
1793                                 return -1;
1794                         free(new_path);
1795                         if (update_stages(opt, a->path, NULL, a, NULL))
1796                                 return -1;
1797                 }
1798
1799                 add = filespec_from_entry(&other, ci->ren2->dst_entry, 3 ^ 1);
1800                 if (add) {
1801                         if (handle_file_collision(opt, b->path,
1802                                                   NULL, NULL,
1803                                                   ci->branch1, ci->branch2,
1804                                                   &add->oid, add->mode,
1805                                                   &mfi.oid, mfi.mode) < 0)
1806                                 return -1;
1807                 } else {
1808                         char *new_path = find_path_for_conflict(opt, b->path,
1809                                                                 ci->branch2,
1810                                                                 ci->branch1);
1811                         if (update_file(opt, 0, &mfi.oid, mfi.mode, new_path ? new_path : b->path))
1812                                 return -1;
1813                         free(new_path);
1814                         if (update_stages(opt, b->path, NULL, NULL, b))
1815                                 return -1;
1816                 }
1817         }
1818
1819         return 0;
1820 }
1821
1822 static int handle_rename_rename_2to1(struct merge_options *opt,
1823                                      struct rename_conflict_info *ci)
1824 {
1825         /* Two files, a & b, were renamed to the same thing, c. */
1826         struct diff_filespec *a = ci->ren1->pair->one;
1827         struct diff_filespec *b = ci->ren2->pair->one;
1828         struct diff_filespec *c1 = ci->ren1->pair->two;
1829         struct diff_filespec *c2 = ci->ren2->pair->two;
1830         struct diff_filespec tmp1, tmp2;
1831         char *path = c1->path; /* == c2->path */
1832         char *path_side_1_desc;
1833         char *path_side_2_desc;
1834         struct merge_file_info mfi_c1;
1835         struct merge_file_info mfi_c2;
1836
1837         output(opt, 1, _("CONFLICT (rename/rename): "
1838                "Rename %s->%s in %s. "
1839                "Rename %s->%s in %s"),
1840                a->path, c1->path, ci->branch1,
1841                b->path, c2->path, ci->branch2);
1842
1843         filespec_from_entry(&tmp1, ci->ren1->src_entry, 3);
1844         tmp1.path = a->path;
1845         filespec_from_entry(&tmp2, ci->ren2->src_entry, 2);
1846         tmp2.path = b->path;
1847
1848         path_side_1_desc = xstrfmt("version of %s from %s", path, a->path);
1849         path_side_2_desc = xstrfmt("version of %s from %s", path, b->path);
1850         if (merge_mode_and_contents(opt, a, c1, &tmp1, path_side_1_desc,
1851                                     opt->branch1, opt->branch2,
1852                                     1 + opt->call_depth * 2, &mfi_c1) ||
1853             merge_mode_and_contents(opt, b, &tmp2, c2, path_side_2_desc,
1854                                     opt->branch1, opt->branch2,
1855                                     1 + opt->call_depth * 2, &mfi_c2))
1856                 return -1;
1857         free(path_side_1_desc);
1858         free(path_side_2_desc);
1859
1860         return handle_file_collision(opt, path, a->path, b->path,
1861                                      ci->branch1, ci->branch2,
1862                                      &mfi_c1.oid, mfi_c1.mode,
1863                                      &mfi_c2.oid, mfi_c2.mode);
1864 }
1865
1866 /*
1867  * Get the diff_filepairs changed between o_tree and tree.
1868  */
1869 static struct diff_queue_struct *get_diffpairs(struct merge_options *opt,
1870                                                struct tree *o_tree,
1871                                                struct tree *tree)
1872 {
1873         struct diff_queue_struct *ret;
1874         struct diff_options opts;
1875
1876         repo_diff_setup(opt->repo, &opts);
1877         opts.flags.recursive = 1;
1878         opts.flags.rename_empty = 0;
1879         opts.detect_rename = merge_detect_rename(opt);
1880         /*
1881          * We do not have logic to handle the detection of copies.  In
1882          * fact, it may not even make sense to add such logic: would we
1883          * really want a change to a base file to be propagated through
1884          * multiple other files by a merge?
1885          */
1886         if (opts.detect_rename > DIFF_DETECT_RENAME)
1887                 opts.detect_rename = DIFF_DETECT_RENAME;
1888         opts.rename_limit = opt->merge_rename_limit >= 0 ? opt->merge_rename_limit :
1889                             opt->diff_rename_limit >= 0 ? opt->diff_rename_limit :
1890                             1000;
1891         opts.rename_score = opt->rename_score;
1892         opts.show_rename_progress = opt->show_rename_progress;
1893         opts.output_format = DIFF_FORMAT_NO_OUTPUT;
1894         diff_setup_done(&opts);
1895         diff_tree_oid(&o_tree->object.oid, &tree->object.oid, "", &opts);
1896         diffcore_std(&opts);
1897         if (opts.needed_rename_limit > opt->needed_rename_limit)
1898                 opt->needed_rename_limit = opts.needed_rename_limit;
1899
1900         ret = xmalloc(sizeof(*ret));
1901         *ret = diff_queued_diff;
1902
1903         opts.output_format = DIFF_FORMAT_NO_OUTPUT;
1904         diff_queued_diff.nr = 0;
1905         diff_queued_diff.queue = NULL;
1906         diff_flush(&opts);
1907         return ret;
1908 }
1909
1910 static int tree_has_path(struct tree *tree, const char *path)
1911 {
1912         struct object_id hashy;
1913         unsigned short mode_o;
1914
1915         return !get_tree_entry(&tree->object.oid, path,
1916                                &hashy, &mode_o);
1917 }
1918
1919 /*
1920  * Return a new string that replaces the beginning portion (which matches
1921  * entry->dir), with entry->new_dir.  In perl-speak:
1922  *   new_path_name = (old_path =~ s/entry->dir/entry->new_dir/);
1923  * NOTE:
1924  *   Caller must ensure that old_path starts with entry->dir + '/'.
1925  */
1926 static char *apply_dir_rename(struct dir_rename_entry *entry,
1927                               const char *old_path)
1928 {
1929         struct strbuf new_path = STRBUF_INIT;
1930         int oldlen, newlen;
1931
1932         if (entry->non_unique_new_dir)
1933                 return NULL;
1934
1935         oldlen = strlen(entry->dir);
1936         newlen = entry->new_dir.len + (strlen(old_path) - oldlen) + 1;
1937         strbuf_grow(&new_path, newlen);
1938         strbuf_addbuf(&new_path, &entry->new_dir);
1939         strbuf_addstr(&new_path, &old_path[oldlen]);
1940
1941         return strbuf_detach(&new_path, NULL);
1942 }
1943
1944 static void get_renamed_dir_portion(const char *old_path, const char *new_path,
1945                                     char **old_dir, char **new_dir)
1946 {
1947         char *end_of_old, *end_of_new;
1948         int old_len, new_len;
1949
1950         *old_dir = NULL;
1951         *new_dir = NULL;
1952
1953         /*
1954          * For
1955          *    "a/b/c/d/e/foo.c" -> "a/b/some/thing/else/e/foo.c"
1956          * the "e/foo.c" part is the same, we just want to know that
1957          *    "a/b/c/d" was renamed to "a/b/some/thing/else"
1958          * so, for this example, this function returns "a/b/c/d" in
1959          * *old_dir and "a/b/some/thing/else" in *new_dir.
1960          *
1961          * Also, if the basename of the file changed, we don't care.  We
1962          * want to know which portion of the directory, if any, changed.
1963          */
1964         end_of_old = strrchr(old_path, '/');
1965         end_of_new = strrchr(new_path, '/');
1966
1967         if (end_of_old == NULL || end_of_new == NULL)
1968                 return;
1969         while (*--end_of_new == *--end_of_old &&
1970                end_of_old != old_path &&
1971                end_of_new != new_path)
1972                 ; /* Do nothing; all in the while loop */
1973         /*
1974          * We've found the first non-matching character in the directory
1975          * paths.  That means the current directory we were comparing
1976          * represents the rename.  Move end_of_old and end_of_new back
1977          * to the full directory name.
1978          */
1979         if (*end_of_old == '/')
1980                 end_of_old++;
1981         if (*end_of_old != '/')
1982                 end_of_new++;
1983         end_of_old = strchr(end_of_old, '/');
1984         end_of_new = strchr(end_of_new, '/');
1985
1986         /*
1987          * It may have been the case that old_path and new_path were the same
1988          * directory all along.  Don't claim a rename if they're the same.
1989          */
1990         old_len = end_of_old - old_path;
1991         new_len = end_of_new - new_path;
1992
1993         if (old_len != new_len || strncmp(old_path, new_path, old_len)) {
1994                 *old_dir = xstrndup(old_path, old_len);
1995                 *new_dir = xstrndup(new_path, new_len);
1996         }
1997 }
1998
1999 static void remove_hashmap_entries(struct hashmap *dir_renames,
2000                                    struct string_list *items_to_remove)
2001 {
2002         int i;
2003         struct dir_rename_entry *entry;
2004
2005         for (i = 0; i < items_to_remove->nr; i++) {
2006                 entry = items_to_remove->items[i].util;
2007                 hashmap_remove(dir_renames, entry, NULL);
2008         }
2009         string_list_clear(items_to_remove, 0);
2010 }
2011
2012 /*
2013  * See if there is a directory rename for path, and if there are any file
2014  * level conflicts for the renamed location.  If there is a rename and
2015  * there are no conflicts, return the new name.  Otherwise, return NULL.
2016  */
2017 static char *handle_path_level_conflicts(struct merge_options *opt,
2018                                          const char *path,
2019                                          struct dir_rename_entry *entry,
2020                                          struct hashmap *collisions,
2021                                          struct tree *tree)
2022 {
2023         char *new_path = NULL;
2024         struct collision_entry *collision_ent;
2025         int clean = 1;
2026         struct strbuf collision_paths = STRBUF_INIT;
2027
2028         /*
2029          * entry has the mapping of old directory name to new directory name
2030          * that we want to apply to path.
2031          */
2032         new_path = apply_dir_rename(entry, path);
2033
2034         if (!new_path) {
2035                 /* This should only happen when entry->non_unique_new_dir set */
2036                 if (!entry->non_unique_new_dir)
2037                         BUG("entry->non_unqiue_dir not set and !new_path");
2038                 output(opt, 1, _("CONFLICT (directory rename split): "
2039                                "Unclear where to place %s because directory "
2040                                "%s was renamed to multiple other directories, "
2041                                "with no destination getting a majority of the "
2042                                "files."),
2043                        path, entry->dir);
2044                 clean = 0;
2045                 return NULL;
2046         }
2047
2048         /*
2049          * The caller needs to have ensured that it has pre-populated
2050          * collisions with all paths that map to new_path.  Do a quick check
2051          * to ensure that's the case.
2052          */
2053         collision_ent = collision_find_entry(collisions, new_path);
2054         if (collision_ent == NULL)
2055                 BUG("collision_ent is NULL");
2056
2057         /*
2058          * Check for one-sided add/add/.../add conflicts, i.e.
2059          * where implicit renames from the other side doing
2060          * directory rename(s) can affect this side of history
2061          * to put multiple paths into the same location.  Warn
2062          * and bail on directory renames for such paths.
2063          */
2064         if (collision_ent->reported_already) {
2065                 clean = 0;
2066         } else if (tree_has_path(tree, new_path)) {
2067                 collision_ent->reported_already = 1;
2068                 strbuf_add_separated_string_list(&collision_paths, ", ",
2069                                                  &collision_ent->source_files);
2070                 output(opt, 1, _("CONFLICT (implicit dir rename): Existing "
2071                                "file/dir at %s in the way of implicit "
2072                                "directory rename(s) putting the following "
2073                                "path(s) there: %s."),
2074                        new_path, collision_paths.buf);
2075                 clean = 0;
2076         } else if (collision_ent->source_files.nr > 1) {
2077                 collision_ent->reported_already = 1;
2078                 strbuf_add_separated_string_list(&collision_paths, ", ",
2079                                                  &collision_ent->source_files);
2080                 output(opt, 1, _("CONFLICT (implicit dir rename): Cannot map "
2081                                "more than one path to %s; implicit directory "
2082                                "renames tried to put these paths there: %s"),
2083                        new_path, collision_paths.buf);
2084                 clean = 0;
2085         }
2086
2087         /* Free memory we no longer need */
2088         strbuf_release(&collision_paths);
2089         if (!clean && new_path) {
2090                 free(new_path);
2091                 return NULL;
2092         }
2093
2094         return new_path;
2095 }
2096
2097 /*
2098  * There are a couple things we want to do at the directory level:
2099  *   1. Check for both sides renaming to the same thing, in order to avoid
2100  *      implicit renaming of files that should be left in place.  (See
2101  *      testcase 6b in t6043 for details.)
2102  *   2. Prune directory renames if there are still files left in the
2103  *      the original directory.  These represent a partial directory rename,
2104  *      i.e. a rename where only some of the files within the directory
2105  *      were renamed elsewhere.  (Technically, this could be done earlier
2106  *      in get_directory_renames(), except that would prevent us from
2107  *      doing the previous check and thus failing testcase 6b.)
2108  *   3. Check for rename/rename(1to2) conflicts (at the directory level).
2109  *      In the future, we could potentially record this info as well and
2110  *      omit reporting rename/rename(1to2) conflicts for each path within
2111  *      the affected directories, thus cleaning up the merge output.
2112  *   NOTE: We do NOT check for rename/rename(2to1) conflicts at the
2113  *         directory level, because merging directories is fine.  If it
2114  *         causes conflicts for files within those merged directories, then
2115  *         that should be detected at the individual path level.
2116  */
2117 static void handle_directory_level_conflicts(struct merge_options *opt,
2118                                              struct hashmap *dir_re_head,
2119                                              struct tree *head,
2120                                              struct hashmap *dir_re_merge,
2121                                              struct tree *merge)
2122 {
2123         struct hashmap_iter iter;
2124         struct dir_rename_entry *head_ent;
2125         struct dir_rename_entry *merge_ent;
2126
2127         struct string_list remove_from_head = STRING_LIST_INIT_NODUP;
2128         struct string_list remove_from_merge = STRING_LIST_INIT_NODUP;
2129
2130         hashmap_iter_init(dir_re_head, &iter);
2131         while ((head_ent = hashmap_iter_next(&iter))) {
2132                 merge_ent = dir_rename_find_entry(dir_re_merge, head_ent->dir);
2133                 if (merge_ent &&
2134                     !head_ent->non_unique_new_dir &&
2135                     !merge_ent->non_unique_new_dir &&
2136                     !strbuf_cmp(&head_ent->new_dir, &merge_ent->new_dir)) {
2137                         /* 1. Renamed identically; remove it from both sides */
2138                         string_list_append(&remove_from_head,
2139                                            head_ent->dir)->util = head_ent;
2140                         strbuf_release(&head_ent->new_dir);
2141                         string_list_append(&remove_from_merge,
2142                                            merge_ent->dir)->util = merge_ent;
2143                         strbuf_release(&merge_ent->new_dir);
2144                 } else if (tree_has_path(head, head_ent->dir)) {
2145                         /* 2. This wasn't a directory rename after all */
2146                         string_list_append(&remove_from_head,
2147                                            head_ent->dir)->util = head_ent;
2148                         strbuf_release(&head_ent->new_dir);
2149                 }
2150         }
2151
2152         remove_hashmap_entries(dir_re_head, &remove_from_head);
2153         remove_hashmap_entries(dir_re_merge, &remove_from_merge);
2154
2155         hashmap_iter_init(dir_re_merge, &iter);
2156         while ((merge_ent = hashmap_iter_next(&iter))) {
2157                 head_ent = dir_rename_find_entry(dir_re_head, merge_ent->dir);
2158                 if (tree_has_path(merge, merge_ent->dir)) {
2159                         /* 2. This wasn't a directory rename after all */
2160                         string_list_append(&remove_from_merge,
2161                                            merge_ent->dir)->util = merge_ent;
2162                 } else if (head_ent &&
2163                            !head_ent->non_unique_new_dir &&
2164                            !merge_ent->non_unique_new_dir) {
2165                         /* 3. rename/rename(1to2) */
2166                         /*
2167                          * We can assume it's not rename/rename(1to1) because
2168                          * that was case (1), already checked above.  So we
2169                          * know that head_ent->new_dir and merge_ent->new_dir
2170                          * are different strings.
2171                          */
2172                         output(opt, 1, _("CONFLICT (rename/rename): "
2173                                        "Rename directory %s->%s in %s. "
2174                                        "Rename directory %s->%s in %s"),
2175                                head_ent->dir, head_ent->new_dir.buf, opt->branch1,
2176                                head_ent->dir, merge_ent->new_dir.buf, opt->branch2);
2177                         string_list_append(&remove_from_head,
2178                                            head_ent->dir)->util = head_ent;
2179                         strbuf_release(&head_ent->new_dir);
2180                         string_list_append(&remove_from_merge,
2181                                            merge_ent->dir)->util = merge_ent;
2182                         strbuf_release(&merge_ent->new_dir);
2183                 }
2184         }
2185
2186         remove_hashmap_entries(dir_re_head, &remove_from_head);
2187         remove_hashmap_entries(dir_re_merge, &remove_from_merge);
2188 }
2189
2190 static struct hashmap *get_directory_renames(struct diff_queue_struct *pairs)
2191 {
2192         struct hashmap *dir_renames;
2193         struct hashmap_iter iter;
2194         struct dir_rename_entry *entry;
2195         int i;
2196
2197         /*
2198          * Typically, we think of a directory rename as all files from a
2199          * certain directory being moved to a target directory.  However,
2200          * what if someone first moved two files from the original
2201          * directory in one commit, and then renamed the directory
2202          * somewhere else in a later commit?  At merge time, we just know
2203          * that files from the original directory went to two different
2204          * places, and that the bulk of them ended up in the same place.
2205          * We want each directory rename to represent where the bulk of the
2206          * files from that directory end up; this function exists to find
2207          * where the bulk of the files went.
2208          *
2209          * The first loop below simply iterates through the list of file
2210          * renames, finding out how often each directory rename pair
2211          * possibility occurs.
2212          */
2213         dir_renames = xmalloc(sizeof(*dir_renames));
2214         dir_rename_init(dir_renames);
2215         for (i = 0; i < pairs->nr; ++i) {
2216                 struct string_list_item *item;
2217                 int *count;
2218                 struct diff_filepair *pair = pairs->queue[i];
2219                 char *old_dir, *new_dir;
2220
2221                 /* File not part of directory rename if it wasn't renamed */
2222                 if (pair->status != 'R')
2223                         continue;
2224
2225                 get_renamed_dir_portion(pair->one->path, pair->two->path,
2226                                         &old_dir,        &new_dir);
2227                 if (!old_dir)
2228                         /* Directory didn't change at all; ignore this one. */
2229                         continue;
2230
2231                 entry = dir_rename_find_entry(dir_renames, old_dir);
2232                 if (!entry) {
2233                         entry = xmalloc(sizeof(*entry));
2234                         dir_rename_entry_init(entry, old_dir);
2235                         hashmap_put(dir_renames, entry);
2236                 } else {
2237                         free(old_dir);
2238                 }
2239                 item = string_list_lookup(&entry->possible_new_dirs, new_dir);
2240                 if (!item) {
2241                         item = string_list_insert(&entry->possible_new_dirs,
2242                                                   new_dir);
2243                         item->util = xcalloc(1, sizeof(int));
2244                 } else {
2245                         free(new_dir);
2246                 }
2247                 count = item->util;
2248                 *count += 1;
2249         }
2250
2251         /*
2252          * For each directory with files moved out of it, we find out which
2253          * target directory received the most files so we can declare it to
2254          * be the "winning" target location for the directory rename.  This
2255          * winner gets recorded in new_dir.  If there is no winner
2256          * (multiple target directories received the same number of files),
2257          * we set non_unique_new_dir.  Once we've determined the winner (or
2258          * that there is no winner), we no longer need possible_new_dirs.
2259          */
2260         hashmap_iter_init(dir_renames, &iter);
2261         while ((entry = hashmap_iter_next(&iter))) {
2262                 int max = 0;
2263                 int bad_max = 0;
2264                 char *best = NULL;
2265
2266                 for (i = 0; i < entry->possible_new_dirs.nr; i++) {
2267                         int *count = entry->possible_new_dirs.items[i].util;
2268
2269                         if (*count == max)
2270                                 bad_max = max;
2271                         else if (*count > max) {
2272                                 max = *count;
2273                                 best = entry->possible_new_dirs.items[i].string;
2274                         }
2275                 }
2276                 if (bad_max == max)
2277                         entry->non_unique_new_dir = 1;
2278                 else {
2279                         assert(entry->new_dir.len == 0);
2280                         strbuf_addstr(&entry->new_dir, best);
2281                 }
2282                 /*
2283                  * The relevant directory sub-portion of the original full
2284                  * filepaths were xstrndup'ed before inserting into
2285                  * possible_new_dirs, and instead of manually iterating the
2286                  * list and free'ing each, just lie and tell
2287                  * possible_new_dirs that it did the strdup'ing so that it
2288                  * will free them for us.
2289                  */
2290                 entry->possible_new_dirs.strdup_strings = 1;
2291                 string_list_clear(&entry->possible_new_dirs, 1);
2292         }
2293
2294         return dir_renames;
2295 }
2296
2297 static struct dir_rename_entry *check_dir_renamed(const char *path,
2298                                                   struct hashmap *dir_renames)
2299 {
2300         char *temp = xstrdup(path);
2301         char *end;
2302         struct dir_rename_entry *entry = NULL;
2303
2304         while ((end = strrchr(temp, '/'))) {
2305                 *end = '\0';
2306                 entry = dir_rename_find_entry(dir_renames, temp);
2307                 if (entry)
2308                         break;
2309         }
2310         free(temp);
2311         return entry;
2312 }
2313
2314 static void compute_collisions(struct hashmap *collisions,
2315                                struct hashmap *dir_renames,
2316                                struct diff_queue_struct *pairs)
2317 {
2318         int i;
2319
2320         /*
2321          * Multiple files can be mapped to the same path due to directory
2322          * renames done by the other side of history.  Since that other
2323          * side of history could have merged multiple directories into one,
2324          * if our side of history added the same file basename to each of
2325          * those directories, then all N of them would get implicitly
2326          * renamed by the directory rename detection into the same path,
2327          * and we'd get an add/add/.../add conflict, and all those adds
2328          * from *this* side of history.  This is not representable in the
2329          * index, and users aren't going to easily be able to make sense of
2330          * it.  So we need to provide a good warning about what's
2331          * happening, and fall back to no-directory-rename detection
2332          * behavior for those paths.
2333          *
2334          * See testcases 9e and all of section 5 from t6043 for examples.
2335          */
2336         collision_init(collisions);
2337
2338         for (i = 0; i < pairs->nr; ++i) {
2339                 struct dir_rename_entry *dir_rename_ent;
2340                 struct collision_entry *collision_ent;
2341                 char *new_path;
2342                 struct diff_filepair *pair = pairs->queue[i];
2343
2344                 if (pair->status != 'A' && pair->status != 'R')
2345                         continue;
2346                 dir_rename_ent = check_dir_renamed(pair->two->path,
2347                                                    dir_renames);
2348                 if (!dir_rename_ent)
2349                         continue;
2350
2351                 new_path = apply_dir_rename(dir_rename_ent, pair->two->path);
2352                 if (!new_path)
2353                         /*
2354                          * dir_rename_ent->non_unique_new_path is true, which
2355                          * means there is no directory rename for us to use,
2356                          * which means it won't cause us any additional
2357                          * collisions.
2358                          */
2359                         continue;
2360                 collision_ent = collision_find_entry(collisions, new_path);
2361                 if (!collision_ent) {
2362                         collision_ent = xcalloc(1,
2363                                                 sizeof(struct collision_entry));
2364                         hashmap_entry_init(collision_ent, strhash(new_path));
2365                         hashmap_put(collisions, collision_ent);
2366                         collision_ent->target_file = new_path;
2367                 } else {
2368                         free(new_path);
2369                 }
2370                 string_list_insert(&collision_ent->source_files,
2371                                    pair->two->path);
2372         }
2373 }
2374
2375 static char *check_for_directory_rename(struct merge_options *opt,
2376                                         const char *path,
2377                                         struct tree *tree,
2378                                         struct hashmap *dir_renames,
2379                                         struct hashmap *dir_rename_exclusions,
2380                                         struct hashmap *collisions,
2381                                         int *clean_merge)
2382 {
2383         char *new_path = NULL;
2384         struct dir_rename_entry *entry = check_dir_renamed(path, dir_renames);
2385         struct dir_rename_entry *oentry = NULL;
2386
2387         if (!entry)
2388                 return new_path;
2389
2390         /*
2391          * This next part is a little weird.  We do not want to do an
2392          * implicit rename into a directory we renamed on our side, because
2393          * that will result in a spurious rename/rename(1to2) conflict.  An
2394          * example:
2395          *   Base commit: dumbdir/afile, otherdir/bfile
2396          *   Side 1:      smrtdir/afile, otherdir/bfile
2397          *   Side 2:      dumbdir/afile, dumbdir/bfile
2398          * Here, while working on Side 1, we could notice that otherdir was
2399          * renamed/merged to dumbdir, and change the diff_filepair for
2400          * otherdir/bfile into a rename into dumbdir/bfile.  However, Side
2401          * 2 will notice the rename from dumbdir to smrtdir, and do the
2402          * transitive rename to move it from dumbdir/bfile to
2403          * smrtdir/bfile.  That gives us bfile in dumbdir vs being in
2404          * smrtdir, a rename/rename(1to2) conflict.  We really just want
2405          * the file to end up in smrtdir.  And the way to achieve that is
2406          * to not let Side1 do the rename to dumbdir, since we know that is
2407          * the source of one of our directory renames.
2408          *
2409          * That's why oentry and dir_rename_exclusions is here.
2410          *
2411          * As it turns out, this also prevents N-way transient rename
2412          * confusion; See testcases 9c and 9d of t6043.
2413          */
2414         oentry = dir_rename_find_entry(dir_rename_exclusions, entry->new_dir.buf);
2415         if (oentry) {
2416                 output(opt, 1, _("WARNING: Avoiding applying %s -> %s rename "
2417                                "to %s, because %s itself was renamed."),
2418                        entry->dir, entry->new_dir.buf, path, entry->new_dir.buf);
2419         } else {
2420                 new_path = handle_path_level_conflicts(opt, path, entry,
2421                                                        collisions, tree);
2422                 *clean_merge &= (new_path != NULL);
2423         }
2424
2425         return new_path;
2426 }
2427
2428 static void apply_directory_rename_modifications(struct merge_options *opt,
2429                                                  struct diff_filepair *pair,
2430                                                  char *new_path,
2431                                                  struct rename *re,
2432                                                  struct tree *tree,
2433                                                  struct tree *o_tree,
2434                                                  struct tree *a_tree,
2435                                                  struct tree *b_tree,
2436                                                  struct string_list *entries)
2437 {
2438         struct string_list_item *item;
2439         int stage = (tree == a_tree ? 2 : 3);
2440         int update_wd;
2441
2442         /*
2443          * In all cases where we can do directory rename detection,
2444          * unpack_trees() will have read pair->two->path into the
2445          * index and the working copy.  We need to remove it so that
2446          * we can instead place it at new_path.  It is guaranteed to
2447          * not be untracked (unpack_trees() would have errored out
2448          * saying the file would have been overwritten), but it might
2449          * be dirty, though.
2450          */
2451         update_wd = !was_dirty(opt, pair->two->path);
2452         if (!update_wd)
2453                 output(opt, 1, _("Refusing to lose dirty file at %s"),
2454                        pair->two->path);
2455         remove_file(opt, 1, pair->two->path, !update_wd);
2456
2457         /* Find or create a new re->dst_entry */
2458         item = string_list_lookup(entries, new_path);
2459         if (item) {
2460                 /*
2461                  * Since we're renaming on this side of history, and it's
2462                  * due to a directory rename on the other side of history
2463                  * (which we only allow when the directory in question no
2464                  * longer exists on the other side of history), the
2465                  * original entry for re->dst_entry is no longer
2466                  * necessary...
2467                  */
2468                 re->dst_entry->processed = 1;
2469
2470                 /*
2471                  * ...because we'll be using this new one.
2472                  */
2473                 re->dst_entry = item->util;
2474         } else {
2475                 /*
2476                  * re->dst_entry is for the before-dir-rename path, and we
2477                  * need it to hold information for the after-dir-rename
2478                  * path.  Before creating a new entry, we need to mark the
2479                  * old one as unnecessary (...unless it is shared by
2480                  * src_entry, i.e. this didn't use to be a rename, in which
2481                  * case we can just allow the normal processing to happen
2482                  * for it).
2483                  */
2484                 if (pair->status == 'R')
2485                         re->dst_entry->processed = 1;
2486
2487                 re->dst_entry = insert_stage_data(new_path,
2488                                                   o_tree, a_tree, b_tree,
2489                                                   entries);
2490                 item = string_list_insert(entries, new_path);
2491                 item->util = re->dst_entry;
2492         }
2493
2494         /*
2495          * Update the stage_data with the information about the path we are
2496          * moving into place.  That slot will be empty and available for us
2497          * to write to because of the collision checks in
2498          * handle_path_level_conflicts().  In other words,
2499          * re->dst_entry->stages[stage].oid will be the null_oid, so it's
2500          * open for us to write to.
2501          *
2502          * It may be tempting to actually update the index at this point as
2503          * well, using update_stages_for_stage_data(), but as per the big
2504          * "NOTE" in update_stages(), doing so will modify the current
2505          * in-memory index which will break calls to would_lose_untracked()
2506          * that we need to make.  Instead, we need to just make sure that
2507          * the various handle_rename_*() functions update the index
2508          * explicitly rather than relying on unpack_trees() to have done it.
2509          */
2510         get_tree_entry(&tree->object.oid,
2511                        pair->two->path,
2512                        &re->dst_entry->stages[stage].oid,
2513                        &re->dst_entry->stages[stage].mode);
2514
2515         /* Update pair status */
2516         if (pair->status == 'A') {
2517                 /*
2518                  * Recording rename information for this add makes it look
2519                  * like a rename/delete conflict.  Make sure we can
2520                  * correctly handle this as an add that was moved to a new
2521                  * directory instead of reporting a rename/delete conflict.
2522                  */
2523                 re->add_turned_into_rename = 1;
2524         }
2525         /*
2526          * We don't actually look at pair->status again, but it seems
2527          * pedagogically correct to adjust it.
2528          */
2529         pair->status = 'R';
2530
2531         /*
2532          * Finally, record the new location.
2533          */
2534         pair->two->path = new_path;
2535 }
2536
2537 /*
2538  * Get information of all renames which occurred in 'pairs', making use of
2539  * any implicit directory renames inferred from the other side of history.
2540  * We need the three trees in the merge ('o_tree', 'a_tree' and 'b_tree')
2541  * to be able to associate the correct cache entries with the rename
2542  * information; tree is always equal to either a_tree or b_tree.
2543  */
2544 static struct string_list *get_renames(struct merge_options *opt,
2545                                        struct diff_queue_struct *pairs,
2546                                        struct hashmap *dir_renames,
2547                                        struct hashmap *dir_rename_exclusions,
2548                                        struct tree *tree,
2549                                        struct tree *o_tree,
2550                                        struct tree *a_tree,
2551                                        struct tree *b_tree,
2552                                        struct string_list *entries,
2553                                        int *clean_merge)
2554 {
2555         int i;
2556         struct hashmap collisions;
2557         struct hashmap_iter iter;
2558         struct collision_entry *e;
2559         struct string_list *renames;
2560
2561         compute_collisions(&collisions, dir_renames, pairs);
2562         renames = xcalloc(1, sizeof(struct string_list));
2563
2564         for (i = 0; i < pairs->nr; ++i) {
2565                 struct string_list_item *item;
2566                 struct rename *re;
2567                 struct diff_filepair *pair = pairs->queue[i];
2568                 char *new_path; /* non-NULL only with directory renames */
2569
2570                 if (pair->status != 'A' && pair->status != 'R') {
2571                         diff_free_filepair(pair);
2572                         continue;
2573                 }
2574                 new_path = check_for_directory_rename(opt, pair->two->path, tree,
2575                                                       dir_renames,
2576                                                       dir_rename_exclusions,
2577                                                       &collisions,
2578                                                       clean_merge);
2579                 if (pair->status != 'R' && !new_path) {
2580                         diff_free_filepair(pair);
2581                         continue;
2582                 }
2583
2584                 re = xmalloc(sizeof(*re));
2585                 re->processed = 0;
2586                 re->add_turned_into_rename = 0;
2587                 re->pair = pair;
2588                 item = string_list_lookup(entries, re->pair->one->path);
2589                 if (!item)
2590                         re->src_entry = insert_stage_data(re->pair->one->path,
2591                                         o_tree, a_tree, b_tree, entries);
2592                 else
2593                         re->src_entry = item->util;
2594
2595                 item = string_list_lookup(entries, re->pair->two->path);
2596                 if (!item)
2597                         re->dst_entry = insert_stage_data(re->pair->two->path,
2598                                         o_tree, a_tree, b_tree, entries);
2599                 else
2600                         re->dst_entry = item->util;
2601                 item = string_list_insert(renames, pair->one->path);
2602                 item->util = re;
2603                 if (new_path)
2604                         apply_directory_rename_modifications(opt, pair, new_path,
2605                                                              re, tree, o_tree,
2606                                                              a_tree, b_tree,
2607                                                              entries);
2608         }
2609
2610         hashmap_iter_init(&collisions, &iter);
2611         while ((e = hashmap_iter_next(&iter))) {
2612                 free(e->target_file);
2613                 string_list_clear(&e->source_files, 0);
2614         }
2615         hashmap_free(&collisions, 1);
2616         return renames;
2617 }
2618
2619 static int process_renames(struct merge_options *opt,
2620                            struct string_list *a_renames,
2621                            struct string_list *b_renames)
2622 {
2623         int clean_merge = 1, i, j;
2624         struct string_list a_by_dst = STRING_LIST_INIT_NODUP;
2625         struct string_list b_by_dst = STRING_LIST_INIT_NODUP;
2626         const struct rename *sre;
2627
2628         for (i = 0; i < a_renames->nr; i++) {
2629                 sre = a_renames->items[i].util;
2630                 string_list_insert(&a_by_dst, sre->pair->two->path)->util
2631                         = (void *)sre;
2632         }
2633         for (i = 0; i < b_renames->nr; i++) {
2634                 sre = b_renames->items[i].util;
2635                 string_list_insert(&b_by_dst, sre->pair->two->path)->util
2636                         = (void *)sre;
2637         }
2638
2639         for (i = 0, j = 0; i < a_renames->nr || j < b_renames->nr;) {
2640                 struct string_list *renames1, *renames2Dst;
2641                 struct rename *ren1 = NULL, *ren2 = NULL;
2642                 const char *branch1, *branch2;
2643                 const char *ren1_src, *ren1_dst;
2644                 struct string_list_item *lookup;
2645
2646                 if (i >= a_renames->nr) {
2647                         ren2 = b_renames->items[j++].util;
2648                 } else if (j >= b_renames->nr) {
2649                         ren1 = a_renames->items[i++].util;
2650                 } else {
2651                         int compare = strcmp(a_renames->items[i].string,
2652                                              b_renames->items[j].string);
2653                         if (compare <= 0)
2654                                 ren1 = a_renames->items[i++].util;
2655                         if (compare >= 0)
2656                                 ren2 = b_renames->items[j++].util;
2657                 }
2658
2659                 /* TODO: refactor, so that 1/2 are not needed */
2660                 if (ren1) {
2661                         renames1 = a_renames;
2662                         renames2Dst = &b_by_dst;
2663                         branch1 = opt->branch1;
2664                         branch2 = opt->branch2;
2665                 } else {
2666                         renames1 = b_renames;
2667                         renames2Dst = &a_by_dst;
2668                         branch1 = opt->branch2;
2669                         branch2 = opt->branch1;
2670                         SWAP(ren2, ren1);
2671                 }
2672
2673                 if (ren1->processed)
2674                         continue;
2675                 ren1->processed = 1;
2676                 ren1->dst_entry->processed = 1;
2677                 /* BUG: We should only mark src_entry as processed if we
2678                  * are not dealing with a rename + add-source case.
2679                  */
2680                 ren1->src_entry->processed = 1;
2681
2682                 ren1_src = ren1->pair->one->path;
2683                 ren1_dst = ren1->pair->two->path;
2684
2685                 if (ren2) {
2686                         /* One file renamed on both sides */
2687                         const char *ren2_src = ren2->pair->one->path;
2688                         const char *ren2_dst = ren2->pair->two->path;
2689                         enum rename_type rename_type;
2690                         if (strcmp(ren1_src, ren2_src) != 0)
2691                                 BUG("ren1_src != ren2_src");
2692                         ren2->dst_entry->processed = 1;
2693                         ren2->processed = 1;
2694                         if (strcmp(ren1_dst, ren2_dst) != 0) {
2695                                 rename_type = RENAME_ONE_FILE_TO_TWO;
2696                                 clean_merge = 0;
2697                         } else {
2698                                 rename_type = RENAME_ONE_FILE_TO_ONE;
2699                                 /* BUG: We should only remove ren1_src in
2700                                  * the base stage (think of rename +
2701                                  * add-source cases).
2702                                  */
2703                                 remove_file(opt, 1, ren1_src, 1);
2704                                 update_entry(ren1->dst_entry,
2705                                              ren1->pair->one,
2706                                              ren1->pair->two,
2707                                              ren2->pair->two);
2708                         }
2709                         setup_rename_conflict_info(rename_type,
2710                                                    opt,
2711                                                    ren1,
2712                                                    ren2,
2713                                                    branch1,
2714                                                    branch2);
2715                 } else if ((lookup = string_list_lookup(renames2Dst, ren1_dst))) {
2716                         /* Two different files renamed to the same thing */
2717                         char *ren2_dst;
2718                         ren2 = lookup->util;
2719                         ren2_dst = ren2->pair->two->path;
2720                         if (strcmp(ren1_dst, ren2_dst) != 0)
2721                                 BUG("ren1_dst != ren2_dst");
2722
2723                         clean_merge = 0;
2724                         ren2->processed = 1;
2725                         /*
2726                          * BUG: We should only mark src_entry as processed
2727                          * if we are not dealing with a rename + add-source
2728                          * case.
2729                          */
2730                         ren2->src_entry->processed = 1;
2731
2732                         setup_rename_conflict_info(RENAME_TWO_FILES_TO_ONE,
2733                                                    opt,
2734                                                    ren1,
2735                                                    ren2,
2736                                                    branch1,
2737                                                    branch2);
2738
2739                 } else {
2740                         /* Renamed in 1, maybe changed in 2 */
2741                         /* we only use sha1 and mode of these */
2742                         struct diff_filespec src_other, dst_other;
2743                         int try_merge;
2744
2745                         /*
2746                          * unpack_trees loads entries from common-commit
2747                          * into stage 1, from head-commit into stage 2, and
2748                          * from merge-commit into stage 3.  We keep track
2749                          * of which side corresponds to the rename.
2750                          */
2751                         int renamed_stage = a_renames == renames1 ? 2 : 3;
2752                         int other_stage =   a_renames == renames1 ? 3 : 2;
2753
2754                         /* BUG: We should only remove ren1_src in the base
2755                          * stage and in other_stage (think of rename +
2756                          * add-source case).
2757                          */
2758                         remove_file(opt, 1, ren1_src,
2759                                     renamed_stage == 2 || !was_tracked(opt, ren1_src));
2760
2761                         oidcpy(&src_other.oid,
2762                                &ren1->src_entry->stages[other_stage].oid);
2763                         src_other.mode = ren1->src_entry->stages[other_stage].mode;
2764                         oidcpy(&dst_other.oid,
2765                                &ren1->dst_entry->stages[other_stage].oid);
2766                         dst_other.mode = ren1->dst_entry->stages[other_stage].mode;
2767                         try_merge = 0;
2768
2769                         if (oid_eq(&src_other.oid, &null_oid) &&
2770                             ren1->add_turned_into_rename) {
2771                                 setup_rename_conflict_info(RENAME_VIA_DIR,
2772                                                            opt,
2773                                                            ren1,
2774                                                            NULL,
2775                                                            branch1,
2776                                                            branch2);
2777                         } else if (oid_eq(&src_other.oid, &null_oid)) {
2778                                 setup_rename_conflict_info(RENAME_DELETE,
2779                                                            opt,
2780                                                            ren1,
2781                                                            NULL,
2782                                                            branch1,
2783                                                            branch2);
2784                         } else if ((dst_other.mode == ren1->pair->two->mode) &&
2785                                    oid_eq(&dst_other.oid, &ren1->pair->two->oid)) {
2786                                 /*
2787                                  * Added file on the other side identical to
2788                                  * the file being renamed: clean merge.
2789                                  * Also, there is no need to overwrite the
2790                                  * file already in the working copy, so call
2791                                  * update_file_flags() instead of
2792                                  * update_file().
2793                                  */
2794                                 if (update_file_flags(opt,
2795                                                       &ren1->pair->two->oid,
2796                                                       ren1->pair->two->mode,
2797                                                       ren1_dst,
2798                                                       1, /* update_cache */
2799                                                       0  /* update_wd    */))
2800                                         clean_merge = -1;
2801                         } else if (!oid_eq(&dst_other.oid, &null_oid)) {
2802                                 /*
2803                                  * Probably not a clean merge, but it's
2804                                  * premature to set clean_merge to 0 here,
2805                                  * because if the rename merges cleanly and
2806                                  * the merge exactly matches the newly added
2807                                  * file, then the merge will be clean.
2808                                  */
2809                                 setup_rename_conflict_info(RENAME_ADD,
2810                                                            opt,
2811                                                            ren1,
2812                                                            NULL,
2813                                                            branch1,
2814                                                            branch2);
2815                         } else
2816                                 try_merge = 1;
2817
2818                         if (clean_merge < 0)
2819                                 goto cleanup_and_return;
2820                         if (try_merge) {
2821                                 struct diff_filespec *o, *a, *b;
2822                                 src_other.path = (char *)ren1_src;
2823
2824                                 o = ren1->pair->one;
2825                                 if (a_renames == renames1) {
2826                                         a = ren1->pair->two;
2827                                         b = &src_other;
2828                                 } else {
2829                                         b = ren1->pair->two;
2830                                         a = &src_other;
2831                                 }
2832                                 update_entry(ren1->dst_entry, o, a, b);
2833                                 setup_rename_conflict_info(RENAME_NORMAL,
2834                                                            opt,
2835                                                            ren1,
2836                                                            NULL,
2837                                                            branch1,
2838                                                            NULL);
2839                         }
2840                 }
2841         }
2842 cleanup_and_return:
2843         string_list_clear(&a_by_dst, 0);
2844         string_list_clear(&b_by_dst, 0);
2845
2846         return clean_merge;
2847 }
2848
2849 struct rename_info {
2850         struct string_list *head_renames;
2851         struct string_list *merge_renames;
2852 };
2853
2854 static void initial_cleanup_rename(struct diff_queue_struct *pairs,
2855                                    struct hashmap *dir_renames)
2856 {
2857         struct hashmap_iter iter;
2858         struct dir_rename_entry *e;
2859
2860         hashmap_iter_init(dir_renames, &iter);
2861         while ((e = hashmap_iter_next(&iter))) {
2862                 free(e->dir);
2863                 strbuf_release(&e->new_dir);
2864                 /* possible_new_dirs already cleared in get_directory_renames */
2865         }
2866         hashmap_free(dir_renames, 1);
2867         free(dir_renames);
2868
2869         free(pairs->queue);
2870         free(pairs);
2871 }
2872
2873 static int detect_and_process_renames(struct merge_options *opt,
2874                                       struct tree *common,
2875                                       struct tree *head,
2876                                       struct tree *merge,
2877                                       struct string_list *entries,
2878                                       struct rename_info *ri)
2879 {
2880         struct diff_queue_struct *head_pairs, *merge_pairs;
2881         struct hashmap *dir_re_head, *dir_re_merge;
2882         int clean = 1;
2883
2884         ri->head_renames = NULL;
2885         ri->merge_renames = NULL;
2886
2887         if (!merge_detect_rename(opt))
2888                 return 1;
2889
2890         head_pairs = get_diffpairs(opt, common, head);
2891         merge_pairs = get_diffpairs(opt, common, merge);
2892
2893         if (opt->detect_directory_renames) {
2894                 dir_re_head = get_directory_renames(head_pairs);
2895                 dir_re_merge = get_directory_renames(merge_pairs);
2896
2897                 handle_directory_level_conflicts(opt,
2898                                                  dir_re_head, head,
2899                                                  dir_re_merge, merge);
2900         } else {
2901                 dir_re_head  = xmalloc(sizeof(*dir_re_head));
2902                 dir_re_merge = xmalloc(sizeof(*dir_re_merge));
2903                 dir_rename_init(dir_re_head);
2904                 dir_rename_init(dir_re_merge);
2905         }
2906
2907         ri->head_renames  = get_renames(opt, head_pairs,
2908                                         dir_re_merge, dir_re_head, head,
2909                                         common, head, merge, entries,
2910                                         &clean);
2911         if (clean < 0)
2912                 goto cleanup;
2913         ri->merge_renames = get_renames(opt, merge_pairs,
2914                                         dir_re_head, dir_re_merge, merge,
2915                                         common, head, merge, entries,
2916                                         &clean);
2917         if (clean < 0)
2918                 goto cleanup;
2919         clean &= process_renames(opt, ri->head_renames, ri->merge_renames);
2920
2921 cleanup:
2922         /*
2923          * Some cleanup is deferred until cleanup_renames() because the
2924          * data structures are still needed and referenced in
2925          * process_entry().  But there are a few things we can free now.
2926          */
2927         initial_cleanup_rename(head_pairs, dir_re_head);
2928         initial_cleanup_rename(merge_pairs, dir_re_merge);
2929
2930         return clean;
2931 }
2932
2933 static void final_cleanup_rename(struct string_list *rename)
2934 {
2935         const struct rename *re;
2936         int i;
2937
2938         if (rename == NULL)
2939                 return;
2940
2941         for (i = 0; i < rename->nr; i++) {
2942                 re = rename->items[i].util;
2943                 diff_free_filepair(re->pair);
2944         }
2945         string_list_clear(rename, 1);
2946         free(rename);
2947 }
2948
2949 static void final_cleanup_renames(struct rename_info *re_info)
2950 {
2951         final_cleanup_rename(re_info->head_renames);
2952         final_cleanup_rename(re_info->merge_renames);
2953 }
2954
2955 static struct object_id *stage_oid(const struct object_id *oid, unsigned mode)
2956 {
2957         return (is_null_oid(oid) || mode == 0) ? NULL: (struct object_id *)oid;
2958 }
2959
2960 static int read_oid_strbuf(struct merge_options *opt,
2961                            const struct object_id *oid,
2962                            struct strbuf *dst)
2963 {
2964         void *buf;
2965         enum object_type type;
2966         unsigned long size;
2967         buf = read_object_file(oid, &type, &size);
2968         if (!buf)
2969                 return err(opt, _("cannot read object %s"), oid_to_hex(oid));
2970         if (type != OBJ_BLOB) {
2971                 free(buf);
2972                 return err(opt, _("object %s is not a blob"), oid_to_hex(oid));
2973         }
2974         strbuf_attach(dst, buf, size, size + 1);
2975         return 0;
2976 }
2977
2978 static int blob_unchanged(struct merge_options *opt,
2979                           const struct object_id *o_oid,
2980                           unsigned o_mode,
2981                           const struct object_id *a_oid,
2982                           unsigned a_mode,
2983                           int renormalize, const char *path)
2984 {
2985         struct strbuf obuf = STRBUF_INIT;
2986         struct strbuf abuf = STRBUF_INIT;
2987         int ret = 0; /* assume changed for safety */
2988         const struct index_state *idx = opt->repo->index;
2989
2990         if (a_mode != o_mode)
2991                 return 0;
2992         if (oid_eq(o_oid, a_oid))
2993                 return 1;
2994         if (!renormalize)
2995                 return 0;
2996
2997         assert(o_oid && a_oid);
2998         if (read_oid_strbuf(opt, o_oid, &obuf) ||
2999             read_oid_strbuf(opt, a_oid, &abuf))
3000                 goto error_return;
3001         /*
3002          * Note: binary | is used so that both renormalizations are
3003          * performed.  Comparison can be skipped if both files are
3004          * unchanged since their sha1s have already been compared.
3005          */
3006         if (renormalize_buffer(idx, path, obuf.buf, obuf.len, &obuf) |
3007             renormalize_buffer(idx, path, abuf.buf, abuf.len, &abuf))
3008                 ret = (obuf.len == abuf.len && !memcmp(obuf.buf, abuf.buf, obuf.len));
3009
3010 error_return:
3011         strbuf_release(&obuf);
3012         strbuf_release(&abuf);
3013         return ret;
3014 }
3015
3016 static int handle_modify_delete(struct merge_options *opt,
3017                                 const char *path,
3018                                 struct object_id *o_oid, int o_mode,
3019                                 struct object_id *a_oid, int a_mode,
3020                                 struct object_id *b_oid, int b_mode)
3021 {
3022         const char *modify_branch, *delete_branch;
3023         struct object_id *changed_oid;
3024         int changed_mode;
3025
3026         if (a_oid) {
3027                 modify_branch = opt->branch1;
3028                 delete_branch = opt->branch2;
3029                 changed_oid = a_oid;
3030                 changed_mode = a_mode;
3031         } else {
3032                 modify_branch = opt->branch2;
3033                 delete_branch = opt->branch1;
3034                 changed_oid = b_oid;
3035                 changed_mode = b_mode;
3036         }
3037
3038         return handle_change_delete(opt,
3039                                     path, NULL,
3040                                     o_oid, o_mode,
3041                                     changed_oid, changed_mode,
3042                                     modify_branch, delete_branch,
3043                                     _("modify"), _("modified"));
3044 }
3045
3046 static int handle_content_merge(struct merge_options *opt,
3047                                 const char *path,
3048                                 int is_dirty,
3049                                 struct object_id *o_oid, int o_mode,
3050                                 struct object_id *a_oid, int a_mode,
3051                                 struct object_id *b_oid, int b_mode,
3052                                 struct rename_conflict_info *ci)
3053 {
3054         const char *reason = _("content");
3055         const char *path1 = NULL, *path2 = NULL;
3056         struct merge_file_info mfi;
3057         struct diff_filespec one, a, b;
3058         unsigned df_conflict_remains = 0;
3059
3060         if (!o_oid) {
3061                 reason = _("add/add");
3062                 o_oid = (struct object_id *)&null_oid;
3063         }
3064         one.path = a.path = b.path = (char *)path;
3065         oidcpy(&one.oid, o_oid);
3066         one.mode = o_mode;
3067         oidcpy(&a.oid, a_oid);
3068         a.mode = a_mode;
3069         oidcpy(&b.oid, b_oid);
3070         b.mode = b_mode;
3071
3072         if (ci) {
3073                 struct diff_filepair *pair1 = ci->ren1->pair;
3074
3075                 path1 = (opt->branch1 == ci->branch1) ?
3076                         pair1->two->path : pair1->one->path;
3077                 /* If ci->ren2->pair != NULL, we are in
3078                  * RENAME_ONE_FILE_TO_ONE case.  Otherwise, we have a
3079                  * normal rename.
3080                  */
3081                 path2 = ((ci->ren2 && ci->ren2->pair) ||
3082                          opt->branch2 == ci->branch1) ?
3083                         pair1->two->path : pair1->one->path;
3084                 one.path = pair1->one->path;
3085                 a.path = (char *)path1;
3086                 b.path = (char *)path2;
3087
3088                 if (dir_in_way(opt->repo->index, path, !opt->call_depth,
3089                                S_ISGITLINK(pair1->two->mode)))
3090                         df_conflict_remains = 1;
3091         }
3092         if (merge_mode_and_contents(opt, &one, &a, &b, path,
3093                                     opt->branch1, opt->branch2,
3094                                     opt->call_depth * 2, &mfi))
3095                 return -1;
3096
3097         /*
3098          * We can skip updating the working tree file iff:
3099          *   a) The merge is clean
3100          *   b) The merge matches what was in HEAD (content, mode, pathname)
3101          *   c) The target path is usable (i.e. not involved in D/F conflict)
3102          */
3103         if (mfi.clean &&
3104             was_tracked_and_matches(opt, path, &mfi.oid, mfi.mode) &&
3105             !df_conflict_remains) {
3106                 int pos;
3107                 struct cache_entry *ce;
3108
3109                 output(opt, 3, _("Skipped %s (merged same as existing)"), path);
3110                 if (add_cacheinfo(opt, mfi.mode, &mfi.oid, path,
3111                                   0, (!opt->call_depth && !is_dirty), 0))
3112                         return -1;
3113                 /*
3114                  * However, add_cacheinfo() will delete the old cache entry
3115                  * and add a new one.  We need to copy over any skip_worktree
3116                  * flag to avoid making the file appear as if it were
3117                  * deleted by the user.
3118                  */
3119                 pos = index_name_pos(&opt->orig_index, path, strlen(path));
3120                 ce = opt->orig_index.cache[pos];
3121                 if (ce_skip_worktree(ce)) {
3122                         pos = index_name_pos(opt->repo->index, path, strlen(path));
3123                         ce = opt->repo->index->cache[pos];
3124                         ce->ce_flags |= CE_SKIP_WORKTREE;
3125                 }
3126                 return mfi.clean;
3127         }
3128
3129         if (!mfi.clean) {
3130                 if (S_ISGITLINK(mfi.mode))
3131                         reason = _("submodule");
3132                 output(opt, 1, _("CONFLICT (%s): Merge conflict in %s"),
3133                                 reason, path);
3134                 if (ci && !df_conflict_remains)
3135                         if (update_stages(opt, path, &one, &a, &b))
3136                                 return -1;
3137         }
3138
3139         if (df_conflict_remains || is_dirty) {
3140                 char *new_path;
3141                 if (opt->call_depth) {
3142                         remove_file_from_index(opt->repo->index, path);
3143                 } else {
3144                         if (!mfi.clean) {
3145                                 if (update_stages(opt, path, &one, &a, &b))
3146                                         return -1;
3147                         } else {
3148                                 int file_from_stage2 = was_tracked(opt, path);
3149                                 struct diff_filespec merged;
3150                                 oidcpy(&merged.oid, &mfi.oid);
3151                                 merged.mode = mfi.mode;
3152
3153                                 if (update_stages(opt, path, NULL,
3154                                                   file_from_stage2 ? &merged : NULL,
3155                                                   file_from_stage2 ? NULL : &merged))
3156                                         return -1;
3157                         }
3158
3159                 }
3160                 new_path = unique_path(opt, path, ci->branch1);
3161                 if (is_dirty) {
3162                         output(opt, 1, _("Refusing to lose dirty file at %s"),
3163                                path);
3164                 }
3165                 output(opt, 1, _("Adding as %s instead"), new_path);
3166                 if (update_file(opt, 0, &mfi.oid, mfi.mode, new_path)) {
3167                         free(new_path);
3168                         return -1;
3169                 }
3170                 free(new_path);
3171                 mfi.clean = 0;
3172         } else if (update_file(opt, mfi.clean, &mfi.oid, mfi.mode, path))
3173                 return -1;
3174         return !is_dirty && mfi.clean;
3175 }
3176
3177 static int handle_rename_normal(struct merge_options *opt,
3178                                 const char *path,
3179                                 struct object_id *o_oid, unsigned int o_mode,
3180                                 struct object_id *a_oid, unsigned int a_mode,
3181                                 struct object_id *b_oid, unsigned int b_mode,
3182                                 struct rename_conflict_info *ci)
3183 {
3184         /* Merge the content and write it out */
3185         return handle_content_merge(opt, path, was_dirty(opt, path),
3186                                     o_oid, o_mode, a_oid, a_mode, b_oid, b_mode,
3187                                     ci);
3188 }
3189
3190 /* Per entry merge function */
3191 static int process_entry(struct merge_options *opt,
3192                          const char *path, struct stage_data *entry)
3193 {
3194         int clean_merge = 1;
3195         int normalize = opt->renormalize;
3196         unsigned o_mode = entry->stages[1].mode;
3197         unsigned a_mode = entry->stages[2].mode;
3198         unsigned b_mode = entry->stages[3].mode;
3199         struct object_id *o_oid = stage_oid(&entry->stages[1].oid, o_mode);
3200         struct object_id *a_oid = stage_oid(&entry->stages[2].oid, a_mode);
3201         struct object_id *b_oid = stage_oid(&entry->stages[3].oid, b_mode);
3202
3203         entry->processed = 1;
3204         if (entry->rename_conflict_info) {
3205                 struct rename_conflict_info *ci = entry->rename_conflict_info;
3206                 switch (ci->rename_type) {
3207                 case RENAME_NORMAL:
3208                 case RENAME_ONE_FILE_TO_ONE:
3209                         clean_merge = handle_rename_normal(opt,
3210                                                            path,
3211                                                            o_oid, o_mode,
3212                                                            a_oid, a_mode,
3213                                                            b_oid, b_mode,
3214                                                            ci);
3215                         break;
3216                 case RENAME_VIA_DIR:
3217                         clean_merge = 1;
3218                         if (handle_rename_via_dir(opt, ci->ren1->pair, ci->branch1))
3219                                 clean_merge = -1;
3220                         break;
3221                 case RENAME_ADD:
3222                         /*
3223                          * Probably unclean merge, but if the renamed file
3224                          * merges cleanly and the result can then be
3225                          * two-way merged cleanly with the added file, I
3226                          * guess it's a clean merge?
3227                          */
3228                         clean_merge = handle_rename_add(opt, ci);
3229                         break;
3230                 case RENAME_DELETE:
3231                         clean_merge = 0;
3232                         if (handle_rename_delete(opt, ci->ren1->pair,
3233                                                  ci->branch1, ci->branch2))
3234                                 clean_merge = -1;
3235                         break;
3236                 case RENAME_ONE_FILE_TO_TWO:
3237                         clean_merge = 0;
3238                         if (handle_rename_rename_1to2(opt, ci))
3239                                 clean_merge = -1;
3240                         break;
3241                 case RENAME_TWO_FILES_TO_ONE:
3242                         /*
3243                          * Probably unclean merge, but if the two renamed
3244                          * files merge cleanly and the two resulting files
3245                          * can then be two-way merged cleanly, I guess it's
3246                          * a clean merge?
3247                          */
3248                         clean_merge = handle_rename_rename_2to1(opt, ci);
3249                         break;
3250                 default:
3251                         entry->processed = 0;
3252                         break;
3253                 }
3254         } else if (o_oid && (!a_oid || !b_oid)) {
3255                 /* Case A: Deleted in one */
3256                 if ((!a_oid && !b_oid) ||
3257                     (!b_oid && blob_unchanged(opt, o_oid, o_mode, a_oid, a_mode, normalize, path)) ||
3258                     (!a_oid && blob_unchanged(opt, o_oid, o_mode, b_oid, b_mode, normalize, path))) {
3259                         /* Deleted in both or deleted in one and
3260                          * unchanged in the other */
3261                         if (a_oid)
3262                                 output(opt, 2, _("Removing %s"), path);
3263                         /* do not touch working file if it did not exist */
3264                         remove_file(opt, 1, path, !a_oid);
3265                 } else {
3266                         /* Modify/delete; deleted side may have put a directory in the way */
3267                         clean_merge = 0;
3268                         if (handle_modify_delete(opt, path, o_oid, o_mode,
3269                                                  a_oid, a_mode, b_oid, b_mode))
3270                                 clean_merge = -1;
3271                 }
3272         } else if ((!o_oid && a_oid && !b_oid) ||
3273                    (!o_oid && !a_oid && b_oid)) {
3274                 /* Case B: Added in one. */
3275                 /* [nothing|directory] -> ([nothing|directory], file) */
3276
3277                 const char *add_branch;
3278                 const char *other_branch;
3279                 unsigned mode;
3280                 const struct object_id *oid;
3281                 const char *conf;
3282
3283                 if (a_oid) {
3284                         add_branch = opt->branch1;
3285                         other_branch = opt->branch2;
3286                         mode = a_mode;
3287                         oid = a_oid;
3288                         conf = _("file/directory");
3289                 } else {
3290                         add_branch = opt->branch2;
3291                         other_branch = opt->branch1;
3292                         mode = b_mode;
3293                         oid = b_oid;
3294                         conf = _("directory/file");
3295                 }
3296                 if (dir_in_way(opt->repo->index, path,
3297                                !opt->call_depth && !S_ISGITLINK(a_mode),
3298                                0)) {
3299                         char *new_path = unique_path(opt, path, add_branch);
3300                         clean_merge = 0;
3301                         output(opt, 1, _("CONFLICT (%s): There is a directory with name %s in %s. "
3302                                "Adding %s as %s"),
3303                                conf, path, other_branch, path, new_path);
3304                         if (update_file(opt, 0, oid, mode, new_path))
3305                                 clean_merge = -1;
3306                         else if (opt->call_depth)
3307                                 remove_file_from_index(opt->repo->index, path);
3308                         free(new_path);
3309                 } else {
3310                         output(opt, 2, _("Adding %s"), path);
3311                         /* do not overwrite file if already present */
3312                         if (update_file_flags(opt, oid, mode, path, 1, !a_oid))
3313                                 clean_merge = -1;
3314                 }
3315         } else if (a_oid && b_oid) {
3316                 if (!o_oid) {
3317                         /* Case C: Added in both (check for same permissions) */
3318                         output(opt, 1,
3319                                _("CONFLICT (add/add): Merge conflict in %s"),
3320                                path);
3321                         clean_merge = handle_file_collision(opt,
3322                                                             path, NULL, NULL,
3323                                                             opt->branch1,
3324                                                             opt->branch2,
3325                                                             a_oid, a_mode,
3326                                                             b_oid, b_mode);
3327                 } else {
3328                         /* case D: Modified in both, but differently. */
3329                         int is_dirty = 0; /* unpack_trees would have bailed if dirty */
3330                         clean_merge = handle_content_merge(opt, path,
3331                                                            is_dirty,
3332                                                            o_oid, o_mode,
3333                                                            a_oid, a_mode,
3334                                                            b_oid, b_mode,
3335                                                            NULL);
3336                 }
3337         } else if (!o_oid && !a_oid && !b_oid) {
3338                 /*
3339                  * this entry was deleted altogether. a_mode == 0 means
3340                  * we had that path and want to actively remove it.
3341                  */
3342                 remove_file(opt, 1, path, !a_mode);
3343         } else
3344                 BUG("fatal merge failure, shouldn't happen.");
3345
3346         return clean_merge;
3347 }
3348
3349 int merge_trees(struct merge_options *opt,
3350                 struct tree *head,
3351                 struct tree *merge,
3352                 struct tree *common,
3353                 struct tree **result)
3354 {
3355         struct index_state *istate = opt->repo->index;
3356         int code, clean;
3357         struct strbuf sb = STRBUF_INIT;
3358
3359         if (!opt->call_depth && repo_index_has_changes(opt->repo, head, &sb)) {
3360                 err(opt, _("Your local changes to the following files would be overwritten by merge:\n  %s"),
3361                     sb.buf);
3362                 return -1;
3363         }
3364
3365         if (opt->subtree_shift) {
3366                 merge = shift_tree_object(opt->repo, head, merge, opt->subtree_shift);
3367                 common = shift_tree_object(opt->repo, head, common, opt->subtree_shift);
3368         }
3369
3370         if (oid_eq(&common->object.oid, &merge->object.oid)) {
3371                 output(opt, 0, _("Already up to date!"));
3372                 *result = head;
3373                 return 1;
3374         }
3375
3376         code = unpack_trees_start(opt, common, head, merge);
3377
3378         if (code != 0) {
3379                 if (show(opt, 4) || opt->call_depth)
3380                         err(opt, _("merging of trees %s and %s failed"),
3381                             oid_to_hex(&head->object.oid),
3382                             oid_to_hex(&merge->object.oid));
3383                 unpack_trees_finish(opt);
3384                 return -1;
3385         }
3386
3387         if (unmerged_index(istate)) {
3388                 struct string_list *entries;
3389                 struct rename_info re_info;
3390                 int i;
3391                 /*
3392                  * Only need the hashmap while processing entries, so
3393                  * initialize it here and free it when we are done running
3394                  * through the entries. Keeping it in the merge_options as
3395                  * opposed to decaring a local hashmap is for convenience
3396                  * so that we don't have to pass it to around.
3397                  */
3398                 hashmap_init(&opt->current_file_dir_set, path_hashmap_cmp, NULL, 512);
3399                 get_files_dirs(opt, head);
3400                 get_files_dirs(opt, merge);
3401
3402                 entries = get_unmerged(opt->repo->index);
3403                 clean = detect_and_process_renames(opt, common, head, merge,
3404                                                    entries, &re_info);
3405                 record_df_conflict_files(opt, entries);
3406                 if (clean < 0)
3407                         goto cleanup;
3408                 for (i = entries->nr-1; 0 <= i; i--) {
3409                         const char *path = entries->items[i].string;
3410                         struct stage_data *e = entries->items[i].util;
3411                         if (!e->processed) {
3412                                 int ret = process_entry(opt, path, e);
3413                                 if (!ret)
3414                                         clean = 0;
3415                                 else if (ret < 0) {
3416                                         clean = ret;
3417                                         goto cleanup;
3418                                 }
3419                         }
3420                 }
3421                 for (i = 0; i < entries->nr; i++) {
3422                         struct stage_data *e = entries->items[i].util;
3423                         if (!e->processed)
3424                                 BUG("unprocessed path??? %s",
3425                                     entries->items[i].string);
3426                 }
3427
3428         cleanup:
3429                 final_cleanup_renames(&re_info);
3430
3431                 string_list_clear(entries, 1);
3432                 free(entries);
3433
3434                 hashmap_free(&opt->current_file_dir_set, 1);
3435
3436                 if (clean < 0) {
3437                         unpack_trees_finish(opt);
3438                         return clean;
3439                 }
3440         }
3441         else
3442                 clean = 1;
3443
3444         unpack_trees_finish(opt);
3445
3446         if (opt->call_depth && !(*result = write_tree_from_memory(opt)))
3447                 return -1;
3448
3449         return clean;
3450 }
3451
3452 static struct commit_list *reverse_commit_list(struct commit_list *list)
3453 {
3454         struct commit_list *next = NULL, *current, *backup;
3455         for (current = list; current; current = backup) {
3456                 backup = current->next;
3457                 current->next = next;
3458                 next = current;
3459         }
3460         return next;
3461 }
3462
3463 /*
3464  * Merge the commits h1 and h2, return the resulting virtual
3465  * commit object and a flag indicating the cleanness of the merge.
3466  */
3467 int merge_recursive(struct merge_options *opt,
3468                     struct commit *h1,
3469                     struct commit *h2,
3470                     struct commit_list *ca,
3471                     struct commit **result)
3472 {
3473         struct commit_list *iter;
3474         struct commit *merged_common_ancestors;
3475         struct tree *mrtree;
3476         int clean;
3477
3478         if (show(opt, 4)) {
3479                 output(opt, 4, _("Merging:"));
3480                 output_commit_title(opt, h1);
3481                 output_commit_title(opt, h2);
3482         }
3483
3484         if (!ca) {
3485                 ca = get_merge_bases(h1, h2);
3486                 ca = reverse_commit_list(ca);
3487         }
3488
3489         if (show(opt, 5)) {
3490                 unsigned cnt = commit_list_count(ca);
3491
3492                 output(opt, 5, Q_("found %u common ancestor:",
3493                                 "found %u common ancestors:", cnt), cnt);
3494                 for (iter = ca; iter; iter = iter->next)
3495                         output_commit_title(opt, iter->item);
3496         }
3497
3498         merged_common_ancestors = pop_commit(&ca);
3499         if (merged_common_ancestors == NULL) {
3500                 /* if there is no common ancestor, use an empty tree */
3501                 struct tree *tree;
3502
3503                 tree = lookup_tree(opt->repo, opt->repo->hash_algo->empty_tree);
3504                 merged_common_ancestors = make_virtual_commit(opt->repo, tree, "ancestor");
3505         }
3506
3507         for (iter = ca; iter; iter = iter->next) {
3508                 const char *saved_b1, *saved_b2;
3509                 opt->call_depth++;
3510                 /*
3511                  * When the merge fails, the result contains files
3512                  * with conflict markers. The cleanness flag is
3513                  * ignored (unless indicating an error), it was never
3514                  * actually used, as result of merge_trees has always
3515                  * overwritten it: the committed "conflicts" were
3516                  * already resolved.
3517                  */
3518                 discard_index(opt->repo->index);
3519                 saved_b1 = opt->branch1;
3520                 saved_b2 = opt->branch2;
3521                 opt->branch1 = "Temporary merge branch 1";
3522                 opt->branch2 = "Temporary merge branch 2";
3523                 if (merge_recursive(opt, merged_common_ancestors, iter->item,
3524                                     NULL, &merged_common_ancestors) < 0)
3525                         return -1;
3526                 opt->branch1 = saved_b1;
3527                 opt->branch2 = saved_b2;
3528                 opt->call_depth--;
3529
3530                 if (!merged_common_ancestors)
3531                         return err(opt, _("merge returned no commit"));
3532         }
3533
3534         discard_index(opt->repo->index);
3535         if (!opt->call_depth)
3536                 repo_read_index(opt->repo);
3537
3538         opt->ancestor = "merged common ancestors";
3539         clean = merge_trees(opt, get_commit_tree(h1), get_commit_tree(h2),
3540                             get_commit_tree(merged_common_ancestors),
3541                             &mrtree);
3542         if (clean < 0) {
3543                 flush_output(opt);
3544                 return clean;
3545         }
3546
3547         if (opt->call_depth) {
3548                 *result = make_virtual_commit(opt->repo, mrtree, "merged tree");
3549                 commit_list_insert(h1, &(*result)->parents);
3550                 commit_list_insert(h2, &(*result)->parents->next);
3551         }
3552         flush_output(opt);
3553         if (!opt->call_depth && opt->buffer_output < 2)
3554                 strbuf_release(&opt->obuf);
3555         if (show(opt, 2))
3556                 diff_warn_rename_limit("merge.renamelimit",
3557                                        opt->needed_rename_limit, 0);
3558         return clean;
3559 }
3560
3561 static struct commit *get_ref(struct repository *repo, const struct object_id *oid,
3562                               const char *name)
3563 {
3564         struct object *object;
3565
3566         object = deref_tag(repo, parse_object(repo, oid),
3567                            name, strlen(name));
3568         if (!object)
3569                 return NULL;
3570         if (object->type == OBJ_TREE)
3571                 return make_virtual_commit(repo, (struct tree*)object, name);
3572         if (object->type != OBJ_COMMIT)
3573                 return NULL;
3574         if (parse_commit((struct commit *)object))
3575                 return NULL;
3576         return (struct commit *)object;
3577 }
3578
3579 int merge_recursive_generic(struct merge_options *opt,
3580                             const struct object_id *head,
3581                             const struct object_id *merge,
3582                             int num_base_list,
3583                             const struct object_id **base_list,
3584                             struct commit **result)
3585 {
3586         int clean;
3587         struct lock_file lock = LOCK_INIT;
3588         struct commit *head_commit = get_ref(opt->repo, head, opt->branch1);
3589         struct commit *next_commit = get_ref(opt->repo, merge, opt->branch2);
3590         struct commit_list *ca = NULL;
3591
3592         if (base_list) {
3593                 int i;
3594                 for (i = 0; i < num_base_list; ++i) {
3595                         struct commit *base;
3596                         if (!(base = get_ref(opt->repo, base_list[i], oid_to_hex(base_list[i]))))
3597                                 return err(opt, _("Could not parse object '%s'"),
3598                                            oid_to_hex(base_list[i]));
3599                         commit_list_insert(base, &ca);
3600                 }
3601         }
3602
3603         repo_hold_locked_index(opt->repo, &lock, LOCK_DIE_ON_ERROR);
3604         clean = merge_recursive(opt, head_commit, next_commit, ca,
3605                                 result);
3606         if (clean < 0) {
3607                 rollback_lock_file(&lock);
3608                 return clean;
3609         }
3610
3611         if (write_locked_index(opt->repo->index, &lock,
3612                                COMMIT_LOCK | SKIP_IF_UNCHANGED))
3613                 return err(opt, _("Unable to write index."));
3614
3615         return clean ? 0 : 1;
3616 }
3617
3618 static void merge_recursive_config(struct merge_options *opt)
3619 {
3620         char *value = NULL;
3621         git_config_get_int("merge.verbosity", &opt->verbosity);
3622         git_config_get_int("diff.renamelimit", &opt->diff_rename_limit);
3623         git_config_get_int("merge.renamelimit", &opt->merge_rename_limit);
3624         if (!git_config_get_string("diff.renames", &value)) {
3625                 opt->diff_detect_rename = git_config_rename("diff.renames", value);
3626                 free(value);
3627         }
3628         if (!git_config_get_string("merge.renames", &value)) {
3629                 opt->merge_detect_rename = git_config_rename("merge.renames", value);
3630                 free(value);
3631         }
3632         git_config(git_xmerge_config, NULL);
3633 }
3634
3635 void init_merge_options(struct merge_options *opt,
3636                         struct repository *repo)
3637 {
3638         const char *merge_verbosity;
3639         memset(opt, 0, sizeof(struct merge_options));
3640         opt->repo = repo;
3641         opt->verbosity = 2;
3642         opt->buffer_output = 1;
3643         opt->diff_rename_limit = -1;
3644         opt->merge_rename_limit = -1;
3645         opt->renormalize = 0;
3646         opt->diff_detect_rename = -1;
3647         opt->merge_detect_rename = -1;
3648         opt->detect_directory_renames = 1;
3649         merge_recursive_config(opt);
3650         merge_verbosity = getenv("GIT_MERGE_VERBOSITY");
3651         if (merge_verbosity)
3652                 opt->verbosity = strtol(merge_verbosity, NULL, 10);
3653         if (opt->verbosity >= 5)
3654                 opt->buffer_output = 0;
3655         strbuf_init(&opt->obuf, 0);
3656         string_list_init(&opt->df_conflict_file_set, 1);
3657 }
3658
3659 int parse_merge_opt(struct merge_options *opt, const char *s)
3660 {
3661         const char *arg;
3662
3663         if (!s || !*s)
3664                 return -1;
3665         if (!strcmp(s, "ours"))
3666                 opt->recursive_variant = MERGE_RECURSIVE_OURS;
3667         else if (!strcmp(s, "theirs"))
3668                 opt->recursive_variant = MERGE_RECURSIVE_THEIRS;
3669         else if (!strcmp(s, "subtree"))
3670                 opt->subtree_shift = "";
3671         else if (skip_prefix(s, "subtree=", &arg))
3672                 opt->subtree_shift = arg;
3673         else if (!strcmp(s, "patience"))
3674                 opt->xdl_opts = DIFF_WITH_ALG(opt, PATIENCE_DIFF);
3675         else if (!strcmp(s, "histogram"))
3676                 opt->xdl_opts = DIFF_WITH_ALG(opt, HISTOGRAM_DIFF);
3677         else if (skip_prefix(s, "diff-algorithm=", &arg)) {
3678                 long value = parse_algorithm_value(arg);
3679                 if (value < 0)
3680                         return -1;
3681                 /* clear out previous settings */
3682                 DIFF_XDL_CLR(opt, NEED_MINIMAL);
3683                 opt->xdl_opts &= ~XDF_DIFF_ALGORITHM_MASK;
3684                 opt->xdl_opts |= value;
3685         }
3686         else if (!strcmp(s, "ignore-space-change"))
3687                 DIFF_XDL_SET(opt, IGNORE_WHITESPACE_CHANGE);
3688         else if (!strcmp(s, "ignore-all-space"))
3689                 DIFF_XDL_SET(opt, IGNORE_WHITESPACE);
3690         else if (!strcmp(s, "ignore-space-at-eol"))
3691                 DIFF_XDL_SET(opt, IGNORE_WHITESPACE_AT_EOL);
3692         else if (!strcmp(s, "ignore-cr-at-eol"))
3693                 DIFF_XDL_SET(opt, IGNORE_CR_AT_EOL);
3694         else if (!strcmp(s, "renormalize"))
3695                 opt->renormalize = 1;
3696         else if (!strcmp(s, "no-renormalize"))
3697                 opt->renormalize = 0;
3698         else if (!strcmp(s, "no-renames"))
3699                 opt->merge_detect_rename = 0;
3700         else if (!strcmp(s, "find-renames")) {
3701                 opt->merge_detect_rename = 1;
3702                 opt->rename_score = 0;
3703         }
3704         else if (skip_prefix(s, "find-renames=", &arg) ||
3705                  skip_prefix(s, "rename-threshold=", &arg)) {
3706                 if ((opt->rename_score = parse_rename_score(&arg)) == -1 || *arg != 0)
3707                         return -1;
3708                 opt->merge_detect_rename = 1;
3709         }
3710         /*
3711          * Please update $__git_merge_strategy_options in
3712          * git-completion.bash when you add new options
3713          */
3714         else
3715                 return -1;
3716         return 0;
3717 }