merge-recursive: apply necessary modifications for directory renames
[git] / merge-recursive.c
1 /*
2  * Recursive Merge algorithm stolen from git-merge-recursive.py by
3  * Fredrik Kuivinen.
4  * The thieves were Alex Riesen and Johannes Schindelin, in June/July 2006
5  */
6 #include "cache.h"
7 #include "config.h"
8 #include "advice.h"
9 #include "lockfile.h"
10 #include "cache-tree.h"
11 #include "commit.h"
12 #include "blob.h"
13 #include "builtin.h"
14 #include "tree-walk.h"
15 #include "diff.h"
16 #include "diffcore.h"
17 #include "tag.h"
18 #include "unpack-trees.h"
19 #include "string-list.h"
20 #include "xdiff-interface.h"
21 #include "ll-merge.h"
22 #include "attr.h"
23 #include "merge-recursive.h"
24 #include "dir.h"
25 #include "submodule.h"
26
27 struct path_hashmap_entry {
28         struct hashmap_entry e;
29         char path[FLEX_ARRAY];
30 };
31
32 static int path_hashmap_cmp(const void *cmp_data,
33                             const void *entry,
34                             const void *entry_or_key,
35                             const void *keydata)
36 {
37         const struct path_hashmap_entry *a = entry;
38         const struct path_hashmap_entry *b = entry_or_key;
39         const char *key = keydata;
40
41         if (ignore_case)
42                 return strcasecmp(a->path, key ? key : b->path);
43         else
44                 return strcmp(a->path, key ? key : b->path);
45 }
46
47 static unsigned int path_hash(const char *path)
48 {
49         return ignore_case ? strihash(path) : strhash(path);
50 }
51
52 static struct dir_rename_entry *dir_rename_find_entry(struct hashmap *hashmap,
53                                                       char *dir)
54 {
55         struct dir_rename_entry key;
56
57         if (dir == NULL)
58                 return NULL;
59         hashmap_entry_init(&key, strhash(dir));
60         key.dir = dir;
61         return hashmap_get(hashmap, &key, NULL);
62 }
63
64 static int dir_rename_cmp(const void *unused_cmp_data,
65                           const void *entry,
66                           const void *entry_or_key,
67                           const void *unused_keydata)
68 {
69         const struct dir_rename_entry *e1 = entry;
70         const struct dir_rename_entry *e2 = entry_or_key;
71
72         return strcmp(e1->dir, e2->dir);
73 }
74
75 static void dir_rename_init(struct hashmap *map)
76 {
77         hashmap_init(map, dir_rename_cmp, NULL, 0);
78 }
79
80 static void dir_rename_entry_init(struct dir_rename_entry *entry,
81                                   char *directory)
82 {
83         hashmap_entry_init(entry, strhash(directory));
84         entry->dir = directory;
85         entry->non_unique_new_dir = 0;
86         strbuf_init(&entry->new_dir, 0);
87         string_list_init(&entry->possible_new_dirs, 0);
88 }
89
90 static struct collision_entry *collision_find_entry(struct hashmap *hashmap,
91                                                     char *target_file)
92 {
93         struct collision_entry key;
94
95         hashmap_entry_init(&key, strhash(target_file));
96         key.target_file = target_file;
97         return hashmap_get(hashmap, &key, NULL);
98 }
99
100 static int collision_cmp(void *unused_cmp_data,
101                          const struct collision_entry *e1,
102                          const struct collision_entry *e2,
103                          const void *unused_keydata)
104 {
105         return strcmp(e1->target_file, e2->target_file);
106 }
107
108 static void collision_init(struct hashmap *map)
109 {
110         hashmap_init(map, (hashmap_cmp_fn) collision_cmp, NULL, 0);
111 }
112
113 static void flush_output(struct merge_options *o)
114 {
115         if (o->buffer_output < 2 && o->obuf.len) {
116                 fputs(o->obuf.buf, stdout);
117                 strbuf_reset(&o->obuf);
118         }
119 }
120
121 static int err(struct merge_options *o, const char *err, ...)
122 {
123         va_list params;
124
125         if (o->buffer_output < 2)
126                 flush_output(o);
127         else {
128                 strbuf_complete(&o->obuf, '\n');
129                 strbuf_addstr(&o->obuf, "error: ");
130         }
131         va_start(params, err);
132         strbuf_vaddf(&o->obuf, err, params);
133         va_end(params);
134         if (o->buffer_output > 1)
135                 strbuf_addch(&o->obuf, '\n');
136         else {
137                 error("%s", o->obuf.buf);
138                 strbuf_reset(&o->obuf);
139         }
140
141         return -1;
142 }
143
144 static struct tree *shift_tree_object(struct tree *one, struct tree *two,
145                                       const char *subtree_shift)
146 {
147         struct object_id shifted;
148
149         if (!*subtree_shift) {
150                 shift_tree(&one->object.oid, &two->object.oid, &shifted, 0);
151         } else {
152                 shift_tree_by(&one->object.oid, &two->object.oid, &shifted,
153                               subtree_shift);
154         }
155         if (!oidcmp(&two->object.oid, &shifted))
156                 return two;
157         return lookup_tree(&shifted);
158 }
159
160 static struct commit *make_virtual_commit(struct tree *tree, const char *comment)
161 {
162         struct commit *commit = alloc_commit_node();
163
164         set_merge_remote_desc(commit, comment, (struct object *)commit);
165         commit->tree = tree;
166         commit->object.parsed = 1;
167         return commit;
168 }
169
170 /*
171  * Since we use get_tree_entry(), which does not put the read object into
172  * the object pool, we cannot rely on a == b.
173  */
174 static int oid_eq(const struct object_id *a, const struct object_id *b)
175 {
176         if (!a && !b)
177                 return 2;
178         return a && b && oidcmp(a, b) == 0;
179 }
180
181 enum rename_type {
182         RENAME_NORMAL = 0,
183         RENAME_DIR,
184         RENAME_DELETE,
185         RENAME_ONE_FILE_TO_ONE,
186         RENAME_ONE_FILE_TO_TWO,
187         RENAME_TWO_FILES_TO_ONE
188 };
189
190 struct rename_conflict_info {
191         enum rename_type rename_type;
192         struct diff_filepair *pair1;
193         struct diff_filepair *pair2;
194         const char *branch1;
195         const char *branch2;
196         struct stage_data *dst_entry1;
197         struct stage_data *dst_entry2;
198         struct diff_filespec ren1_other;
199         struct diff_filespec ren2_other;
200 };
201
202 /*
203  * Since we want to write the index eventually, we cannot reuse the index
204  * for these (temporary) data.
205  */
206 struct stage_data {
207         struct {
208                 unsigned mode;
209                 struct object_id oid;
210         } stages[4];
211         struct rename_conflict_info *rename_conflict_info;
212         unsigned processed:1;
213 };
214
215 static inline void setup_rename_conflict_info(enum rename_type rename_type,
216                                               struct diff_filepair *pair1,
217                                               struct diff_filepair *pair2,
218                                               const char *branch1,
219                                               const char *branch2,
220                                               struct stage_data *dst_entry1,
221                                               struct stage_data *dst_entry2,
222                                               struct merge_options *o,
223                                               struct stage_data *src_entry1,
224                                               struct stage_data *src_entry2)
225 {
226         struct rename_conflict_info *ci = xcalloc(1, sizeof(struct rename_conflict_info));
227         ci->rename_type = rename_type;
228         ci->pair1 = pair1;
229         ci->branch1 = branch1;
230         ci->branch2 = branch2;
231
232         ci->dst_entry1 = dst_entry1;
233         dst_entry1->rename_conflict_info = ci;
234         dst_entry1->processed = 0;
235
236         assert(!pair2 == !dst_entry2);
237         if (dst_entry2) {
238                 ci->dst_entry2 = dst_entry2;
239                 ci->pair2 = pair2;
240                 dst_entry2->rename_conflict_info = ci;
241         }
242
243         if (rename_type == RENAME_TWO_FILES_TO_ONE) {
244                 /*
245                  * For each rename, there could have been
246                  * modifications on the side of history where that
247                  * file was not renamed.
248                  */
249                 int ostage1 = o->branch1 == branch1 ? 3 : 2;
250                 int ostage2 = ostage1 ^ 1;
251
252                 ci->ren1_other.path = pair1->one->path;
253                 oidcpy(&ci->ren1_other.oid, &src_entry1->stages[ostage1].oid);
254                 ci->ren1_other.mode = src_entry1->stages[ostage1].mode;
255
256                 ci->ren2_other.path = pair2->one->path;
257                 oidcpy(&ci->ren2_other.oid, &src_entry2->stages[ostage2].oid);
258                 ci->ren2_other.mode = src_entry2->stages[ostage2].mode;
259         }
260 }
261
262 static int show(struct merge_options *o, int v)
263 {
264         return (!o->call_depth && o->verbosity >= v) || o->verbosity >= 5;
265 }
266
267 __attribute__((format (printf, 3, 4)))
268 static void output(struct merge_options *o, int v, const char *fmt, ...)
269 {
270         va_list ap;
271
272         if (!show(o, v))
273                 return;
274
275         strbuf_addchars(&o->obuf, ' ', o->call_depth * 2);
276
277         va_start(ap, fmt);
278         strbuf_vaddf(&o->obuf, fmt, ap);
279         va_end(ap);
280
281         strbuf_addch(&o->obuf, '\n');
282         if (!o->buffer_output)
283                 flush_output(o);
284 }
285
286 static void output_commit_title(struct merge_options *o, struct commit *commit)
287 {
288         strbuf_addchars(&o->obuf, ' ', o->call_depth * 2);
289         if (commit->util)
290                 strbuf_addf(&o->obuf, "virtual %s\n",
291                         merge_remote_util(commit)->name);
292         else {
293                 strbuf_add_unique_abbrev(&o->obuf, commit->object.oid.hash,
294                                          DEFAULT_ABBREV);
295                 strbuf_addch(&o->obuf, ' ');
296                 if (parse_commit(commit) != 0)
297                         strbuf_addstr(&o->obuf, _("(bad commit)\n"));
298                 else {
299                         const char *title;
300                         const char *msg = get_commit_buffer(commit, NULL);
301                         int len = find_commit_subject(msg, &title);
302                         if (len)
303                                 strbuf_addf(&o->obuf, "%.*s\n", len, title);
304                         unuse_commit_buffer(commit, msg);
305                 }
306         }
307         flush_output(o);
308 }
309
310 static int add_cacheinfo(struct merge_options *o,
311                 unsigned int mode, const struct object_id *oid,
312                 const char *path, int stage, int refresh, int options)
313 {
314         struct cache_entry *ce;
315         int ret;
316
317         ce = make_cache_entry(mode, oid ? oid->hash : null_sha1, path, stage, 0);
318         if (!ce)
319                 return err(o, _("addinfo_cache failed for path '%s'"), path);
320
321         ret = add_cache_entry(ce, options);
322         if (refresh) {
323                 struct cache_entry *nce;
324
325                 nce = refresh_cache_entry(ce, CE_MATCH_REFRESH | CE_MATCH_IGNORE_MISSING);
326                 if (!nce)
327                         return err(o, _("addinfo_cache failed for path '%s'"), path);
328                 if (nce != ce)
329                         ret = add_cache_entry(nce, options);
330         }
331         return ret;
332 }
333
334 static void init_tree_desc_from_tree(struct tree_desc *desc, struct tree *tree)
335 {
336         parse_tree(tree);
337         init_tree_desc(desc, tree->buffer, tree->size);
338 }
339
340 static int git_merge_trees(int index_only,
341                            struct tree *common,
342                            struct tree *head,
343                            struct tree *merge)
344 {
345         int rc;
346         struct tree_desc t[3];
347         struct unpack_trees_options opts;
348
349         memset(&opts, 0, sizeof(opts));
350         if (index_only)
351                 opts.index_only = 1;
352         else
353                 opts.update = 1;
354         opts.merge = 1;
355         opts.head_idx = 2;
356         opts.fn = threeway_merge;
357         opts.src_index = &the_index;
358         opts.dst_index = &the_index;
359         setup_unpack_trees_porcelain(&opts, "merge");
360
361         init_tree_desc_from_tree(t+0, common);
362         init_tree_desc_from_tree(t+1, head);
363         init_tree_desc_from_tree(t+2, merge);
364
365         rc = unpack_trees(3, t, &opts);
366         cache_tree_free(&active_cache_tree);
367         return rc;
368 }
369
370 struct tree *write_tree_from_memory(struct merge_options *o)
371 {
372         struct tree *result = NULL;
373
374         if (unmerged_cache()) {
375                 int i;
376                 fprintf(stderr, "BUG: There are unmerged index entries:\n");
377                 for (i = 0; i < active_nr; i++) {
378                         const struct cache_entry *ce = active_cache[i];
379                         if (ce_stage(ce))
380                                 fprintf(stderr, "BUG: %d %.*s\n", ce_stage(ce),
381                                         (int)ce_namelen(ce), ce->name);
382                 }
383                 die("BUG: unmerged index entries in merge-recursive.c");
384         }
385
386         if (!active_cache_tree)
387                 active_cache_tree = cache_tree();
388
389         if (!cache_tree_fully_valid(active_cache_tree) &&
390             cache_tree_update(&the_index, 0) < 0) {
391                 err(o, _("error building trees"));
392                 return NULL;
393         }
394
395         result = lookup_tree(&active_cache_tree->oid);
396
397         return result;
398 }
399
400 static int save_files_dirs(const unsigned char *sha1,
401                 struct strbuf *base, const char *path,
402                 unsigned int mode, int stage, void *context)
403 {
404         struct path_hashmap_entry *entry;
405         int baselen = base->len;
406         struct merge_options *o = context;
407
408         strbuf_addstr(base, path);
409
410         FLEX_ALLOC_MEM(entry, path, base->buf, base->len);
411         hashmap_entry_init(entry, path_hash(entry->path));
412         hashmap_add(&o->current_file_dir_set, entry);
413
414         strbuf_setlen(base, baselen);
415         return (S_ISDIR(mode) ? READ_TREE_RECURSIVE : 0);
416 }
417
418 static void get_files_dirs(struct merge_options *o, struct tree *tree)
419 {
420         struct pathspec match_all;
421         memset(&match_all, 0, sizeof(match_all));
422         read_tree_recursive(tree, "", 0, 0, &match_all, save_files_dirs, o);
423 }
424
425 static int get_tree_entry_if_blob(const unsigned char *tree,
426                                   const char *path,
427                                   unsigned char *hashy,
428                                   unsigned int *mode_o)
429 {
430         int ret;
431
432         ret = get_tree_entry(tree, path, hashy, mode_o);
433         if (S_ISDIR(*mode_o)) {
434                 hashcpy(hashy, null_sha1);
435                 *mode_o = 0;
436         }
437         return ret;
438 }
439
440 /*
441  * Returns an index_entry instance which doesn't have to correspond to
442  * a real cache entry in Git's index.
443  */
444 static struct stage_data *insert_stage_data(const char *path,
445                 struct tree *o, struct tree *a, struct tree *b,
446                 struct string_list *entries)
447 {
448         struct string_list_item *item;
449         struct stage_data *e = xcalloc(1, sizeof(struct stage_data));
450         get_tree_entry_if_blob(o->object.oid.hash, path,
451                                e->stages[1].oid.hash, &e->stages[1].mode);
452         get_tree_entry_if_blob(a->object.oid.hash, path,
453                                e->stages[2].oid.hash, &e->stages[2].mode);
454         get_tree_entry_if_blob(b->object.oid.hash, path,
455                                e->stages[3].oid.hash, &e->stages[3].mode);
456         item = string_list_insert(entries, path);
457         item->util = e;
458         return e;
459 }
460
461 /*
462  * Create a dictionary mapping file names to stage_data objects. The
463  * dictionary contains one entry for every path with a non-zero stage entry.
464  */
465 static struct string_list *get_unmerged(void)
466 {
467         struct string_list *unmerged = xcalloc(1, sizeof(struct string_list));
468         int i;
469
470         unmerged->strdup_strings = 1;
471
472         for (i = 0; i < active_nr; i++) {
473                 struct string_list_item *item;
474                 struct stage_data *e;
475                 const struct cache_entry *ce = active_cache[i];
476                 if (!ce_stage(ce))
477                         continue;
478
479                 item = string_list_lookup(unmerged, ce->name);
480                 if (!item) {
481                         item = string_list_insert(unmerged, ce->name);
482                         item->util = xcalloc(1, sizeof(struct stage_data));
483                 }
484                 e = item->util;
485                 e->stages[ce_stage(ce)].mode = ce->ce_mode;
486                 oidcpy(&e->stages[ce_stage(ce)].oid, &ce->oid);
487         }
488
489         return unmerged;
490 }
491
492 static int string_list_df_name_compare(const char *one, const char *two)
493 {
494         int onelen = strlen(one);
495         int twolen = strlen(two);
496         /*
497          * Here we only care that entries for D/F conflicts are
498          * adjacent, in particular with the file of the D/F conflict
499          * appearing before files below the corresponding directory.
500          * The order of the rest of the list is irrelevant for us.
501          *
502          * To achieve this, we sort with df_name_compare and provide
503          * the mode S_IFDIR so that D/F conflicts will sort correctly.
504          * We use the mode S_IFDIR for everything else for simplicity,
505          * since in other cases any changes in their order due to
506          * sorting cause no problems for us.
507          */
508         int cmp = df_name_compare(one, onelen, S_IFDIR,
509                                   two, twolen, S_IFDIR);
510         /*
511          * Now that 'foo' and 'foo/bar' compare equal, we have to make sure
512          * that 'foo' comes before 'foo/bar'.
513          */
514         if (cmp)
515                 return cmp;
516         return onelen - twolen;
517 }
518
519 static void record_df_conflict_files(struct merge_options *o,
520                                      struct string_list *entries)
521 {
522         /* If there is a D/F conflict and the file for such a conflict
523          * currently exist in the working tree, we want to allow it to be
524          * removed to make room for the corresponding directory if needed.
525          * The files underneath the directories of such D/F conflicts will
526          * be processed before the corresponding file involved in the D/F
527          * conflict.  If the D/F directory ends up being removed by the
528          * merge, then we won't have to touch the D/F file.  If the D/F
529          * directory needs to be written to the working copy, then the D/F
530          * file will simply be removed (in make_room_for_path()) to make
531          * room for the necessary paths.  Note that if both the directory
532          * and the file need to be present, then the D/F file will be
533          * reinstated with a new unique name at the time it is processed.
534          */
535         struct string_list df_sorted_entries = STRING_LIST_INIT_NODUP;
536         const char *last_file = NULL;
537         int last_len = 0;
538         int i;
539
540         /*
541          * If we're merging merge-bases, we don't want to bother with
542          * any working directory changes.
543          */
544         if (o->call_depth)
545                 return;
546
547         /* Ensure D/F conflicts are adjacent in the entries list. */
548         for (i = 0; i < entries->nr; i++) {
549                 struct string_list_item *next = &entries->items[i];
550                 string_list_append(&df_sorted_entries, next->string)->util =
551                                    next->util;
552         }
553         df_sorted_entries.cmp = string_list_df_name_compare;
554         string_list_sort(&df_sorted_entries);
555
556         string_list_clear(&o->df_conflict_file_set, 1);
557         for (i = 0; i < df_sorted_entries.nr; i++) {
558                 const char *path = df_sorted_entries.items[i].string;
559                 int len = strlen(path);
560                 struct stage_data *e = df_sorted_entries.items[i].util;
561
562                 /*
563                  * Check if last_file & path correspond to a D/F conflict;
564                  * i.e. whether path is last_file+'/'+<something>.
565                  * If so, record that it's okay to remove last_file to make
566                  * room for path and friends if needed.
567                  */
568                 if (last_file &&
569                     len > last_len &&
570                     memcmp(path, last_file, last_len) == 0 &&
571                     path[last_len] == '/') {
572                         string_list_insert(&o->df_conflict_file_set, last_file);
573                 }
574
575                 /*
576                  * Determine whether path could exist as a file in the
577                  * working directory as a possible D/F conflict.  This
578                  * will only occur when it exists in stage 2 as a
579                  * file.
580                  */
581                 if (S_ISREG(e->stages[2].mode) || S_ISLNK(e->stages[2].mode)) {
582                         last_file = path;
583                         last_len = len;
584                 } else {
585                         last_file = NULL;
586                 }
587         }
588         string_list_clear(&df_sorted_entries, 0);
589 }
590
591 struct rename {
592         struct diff_filepair *pair;
593         /*
594          * Purpose of src_entry and dst_entry:
595          *
596          * If 'before' is renamed to 'after' then src_entry will contain
597          * the versions of 'before' from the merge_base, HEAD, and MERGE in
598          * stages 1, 2, and 3; dst_entry will contain the respective
599          * versions of 'after' in corresponding locations.  Thus, we have a
600          * total of six modes and oids, though some will be null.  (Stage 0
601          * is ignored; we're interested in handling conflicts.)
602          *
603          * Since we don't turn on break-rewrites by default, neither
604          * src_entry nor dst_entry can have all three of their stages have
605          * non-null oids, meaning at most four of the six will be non-null.
606          * Also, since this is a rename, both src_entry and dst_entry will
607          * have at least one non-null oid, meaning at least two will be
608          * non-null.  Of the six oids, a typical rename will have three be
609          * non-null.  Only two implies a rename/delete, and four implies a
610          * rename/add.
611          */
612         struct stage_data *src_entry;
613         struct stage_data *dst_entry;
614         unsigned add_turned_into_rename:1;
615         unsigned processed:1;
616 };
617
618 static int update_stages(struct merge_options *opt, const char *path,
619                          const struct diff_filespec *o,
620                          const struct diff_filespec *a,
621                          const struct diff_filespec *b)
622 {
623
624         /*
625          * NOTE: It is usually a bad idea to call update_stages on a path
626          * before calling update_file on that same path, since it can
627          * sometimes lead to spurious "refusing to lose untracked file..."
628          * messages from update_file (via make_room_for path via
629          * would_lose_untracked).  Instead, reverse the order of the calls
630          * (executing update_file first and then update_stages).
631          */
632         int clear = 1;
633         int options = ADD_CACHE_OK_TO_ADD | ADD_CACHE_SKIP_DFCHECK;
634         if (clear)
635                 if (remove_file_from_cache(path))
636                         return -1;
637         if (o)
638                 if (add_cacheinfo(opt, o->mode, &o->oid, path, 1, 0, options))
639                         return -1;
640         if (a)
641                 if (add_cacheinfo(opt, a->mode, &a->oid, path, 2, 0, options))
642                         return -1;
643         if (b)
644                 if (add_cacheinfo(opt, b->mode, &b->oid, path, 3, 0, options))
645                         return -1;
646         return 0;
647 }
648
649 static int update_stages_for_stage_data(struct merge_options *opt,
650                                         const char *path,
651                                         const struct stage_data *stage_data)
652 {
653         struct diff_filespec o, a, b;
654
655         o.mode = stage_data->stages[1].mode;
656         oidcpy(&o.oid, &stage_data->stages[1].oid);
657
658         a.mode = stage_data->stages[2].mode;
659         oidcpy(&a.oid, &stage_data->stages[2].oid);
660
661         b.mode = stage_data->stages[3].mode;
662         oidcpy(&b.oid, &stage_data->stages[3].oid);
663
664         return update_stages(opt, path,
665                              is_null_oid(&o.oid) ? NULL : &o,
666                              is_null_oid(&a.oid) ? NULL : &a,
667                              is_null_oid(&b.oid) ? NULL : &b);
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 *o, int clean,
685                        const char *path, int no_wd)
686 {
687         int update_cache = o->call_depth || clean;
688         int update_working_directory = !o->call_depth && !no_wd;
689
690         if (update_cache) {
691                 if (remove_file_from_cache(path))
692                         return -1;
693         }
694         if (update_working_directory) {
695                 if (ignore_case) {
696                         struct cache_entry *ce;
697                         ce = cache_file_exists(path, strlen(path), ignore_case);
698                         if (ce && ce_stage(ce) == 0 && strcmp(path, ce->name))
699                                 return 0;
700                 }
701                 if (remove_path(path))
702                         return -1;
703         }
704         return 0;
705 }
706
707 /* add a string to a strbuf, but converting "/" to "_" */
708 static void add_flattened_path(struct strbuf *out, const char *s)
709 {
710         size_t i = out->len;
711         strbuf_addstr(out, s);
712         for (; i < out->len; i++)
713                 if (out->buf[i] == '/')
714                         out->buf[i] = '_';
715 }
716
717 static char *unique_path(struct merge_options *o, const char *path, const char *branch)
718 {
719         struct path_hashmap_entry *entry;
720         struct strbuf newpath = STRBUF_INIT;
721         int suffix = 0;
722         size_t base_len;
723
724         strbuf_addf(&newpath, "%s~", path);
725         add_flattened_path(&newpath, branch);
726
727         base_len = newpath.len;
728         while (hashmap_get_from_hash(&o->current_file_dir_set,
729                                      path_hash(newpath.buf), newpath.buf) ||
730                (!o->call_depth && file_exists(newpath.buf))) {
731                 strbuf_setlen(&newpath, base_len);
732                 strbuf_addf(&newpath, "_%d", suffix++);
733         }
734
735         FLEX_ALLOC_MEM(entry, path, newpath.buf, newpath.len);
736         hashmap_entry_init(entry, path_hash(entry->path));
737         hashmap_add(&o->current_file_dir_set, entry);
738         return strbuf_detach(&newpath, NULL);
739 }
740
741 /**
742  * Check whether a directory in the index is in the way of an incoming
743  * file.  Return 1 if so.  If check_working_copy is non-zero, also
744  * check the working directory.  If empty_ok is non-zero, also return
745  * 0 in the case where the working-tree dir exists but is empty.
746  */
747 static int dir_in_way(const char *path, int check_working_copy, int empty_ok)
748 {
749         int pos;
750         struct strbuf dirpath = STRBUF_INIT;
751         struct stat st;
752
753         strbuf_addstr(&dirpath, path);
754         strbuf_addch(&dirpath, '/');
755
756         pos = cache_name_pos(dirpath.buf, dirpath.len);
757
758         if (pos < 0)
759                 pos = -1 - pos;
760         if (pos < active_nr &&
761             !strncmp(dirpath.buf, active_cache[pos]->name, dirpath.len)) {
762                 strbuf_release(&dirpath);
763                 return 1;
764         }
765
766         strbuf_release(&dirpath);
767         return check_working_copy && !lstat(path, &st) && S_ISDIR(st.st_mode) &&
768                 !(empty_ok && is_empty_dir(path));
769 }
770
771 static int was_tracked(const char *path)
772 {
773         int pos = cache_name_pos(path, strlen(path));
774
775         if (0 <= pos)
776                 /* we have been tracking this path */
777                 return 1;
778
779         /*
780          * Look for an unmerged entry for the path,
781          * specifically stage #2, which would indicate
782          * that "our" side before the merge started
783          * had the path tracked (and resulted in a conflict).
784          */
785         for (pos = -1 - pos;
786              pos < active_nr && !strcmp(path, active_cache[pos]->name);
787              pos++)
788                 if (ce_stage(active_cache[pos]) == 2)
789                         return 1;
790         return 0;
791 }
792
793 static int would_lose_untracked(const char *path)
794 {
795         return !was_tracked(path) && file_exists(path);
796 }
797
798 static int make_room_for_path(struct merge_options *o, const char *path)
799 {
800         int status, i;
801         const char *msg = _("failed to create path '%s'%s");
802
803         /* Unlink any D/F conflict files that are in the way */
804         for (i = 0; i < o->df_conflict_file_set.nr; i++) {
805                 const char *df_path = o->df_conflict_file_set.items[i].string;
806                 size_t pathlen = strlen(path);
807                 size_t df_pathlen = strlen(df_path);
808                 if (df_pathlen < pathlen &&
809                     path[df_pathlen] == '/' &&
810                     strncmp(path, df_path, df_pathlen) == 0) {
811                         output(o, 3,
812                                _("Removing %s to make room for subdirectory\n"),
813                                df_path);
814                         unlink(df_path);
815                         unsorted_string_list_delete_item(&o->df_conflict_file_set,
816                                                          i, 0);
817                         break;
818                 }
819         }
820
821         /* Make sure leading directories are created */
822         status = safe_create_leading_directories_const(path);
823         if (status) {
824                 if (status == SCLD_EXISTS)
825                         /* something else exists */
826                         return err(o, msg, path, _(": perhaps a D/F conflict?"));
827                 return err(o, msg, path, "");
828         }
829
830         /*
831          * Do not unlink a file in the work tree if we are not
832          * tracking it.
833          */
834         if (would_lose_untracked(path))
835                 return err(o, _("refusing to lose untracked file at '%s'"),
836                              path);
837
838         /* Successful unlink is good.. */
839         if (!unlink(path))
840                 return 0;
841         /* .. and so is no existing file */
842         if (errno == ENOENT)
843                 return 0;
844         /* .. but not some other error (who really cares what?) */
845         return err(o, msg, path, _(": perhaps a D/F conflict?"));
846 }
847
848 static int update_file_flags(struct merge_options *o,
849                              const struct object_id *oid,
850                              unsigned mode,
851                              const char *path,
852                              int update_cache,
853                              int update_wd)
854 {
855         int ret = 0;
856
857         if (o->call_depth)
858                 update_wd = 0;
859
860         if (update_wd) {
861                 enum object_type type;
862                 void *buf;
863                 unsigned long size;
864
865                 if (S_ISGITLINK(mode)) {
866                         /*
867                          * We may later decide to recursively descend into
868                          * the submodule directory and update its index
869                          * and/or work tree, but we do not do that now.
870                          */
871                         update_wd = 0;
872                         goto update_index;
873                 }
874
875                 buf = read_sha1_file(oid->hash, &type, &size);
876                 if (!buf)
877                         return err(o, _("cannot read object %s '%s'"), oid_to_hex(oid), path);
878                 if (type != OBJ_BLOB) {
879                         ret = err(o, _("blob expected for %s '%s'"), oid_to_hex(oid), path);
880                         goto free_buf;
881                 }
882                 if (S_ISREG(mode)) {
883                         struct strbuf strbuf = STRBUF_INIT;
884                         if (convert_to_working_tree(path, buf, size, &strbuf)) {
885                                 free(buf);
886                                 size = strbuf.len;
887                                 buf = strbuf_detach(&strbuf, NULL);
888                         }
889                 }
890
891                 if (make_room_for_path(o, path) < 0) {
892                         update_wd = 0;
893                         goto free_buf;
894                 }
895                 if (S_ISREG(mode) || (!has_symlinks && S_ISLNK(mode))) {
896                         int fd;
897                         if (mode & 0100)
898                                 mode = 0777;
899                         else
900                                 mode = 0666;
901                         fd = open(path, O_WRONLY | O_TRUNC | O_CREAT, mode);
902                         if (fd < 0) {
903                                 ret = err(o, _("failed to open '%s': %s"),
904                                           path, strerror(errno));
905                                 goto free_buf;
906                         }
907                         write_in_full(fd, buf, size);
908                         close(fd);
909                 } else if (S_ISLNK(mode)) {
910                         char *lnk = xmemdupz(buf, size);
911                         safe_create_leading_directories_const(path);
912                         unlink(path);
913                         if (symlink(lnk, path))
914                                 ret = err(o, _("failed to symlink '%s': %s"),
915                                         path, strerror(errno));
916                         free(lnk);
917                 } else
918                         ret = err(o,
919                                   _("do not know what to do with %06o %s '%s'"),
920                                   mode, oid_to_hex(oid), path);
921  free_buf:
922                 free(buf);
923         }
924  update_index:
925         if (!ret && update_cache)
926                 add_cacheinfo(o, mode, oid, path, 0, update_wd, ADD_CACHE_OK_TO_ADD);
927         return ret;
928 }
929
930 static int update_file(struct merge_options *o,
931                        int clean,
932                        const struct object_id *oid,
933                        unsigned mode,
934                        const char *path)
935 {
936         return update_file_flags(o, oid, mode, path, o->call_depth || clean, !o->call_depth);
937 }
938
939 /* Low level file merging, update and removal */
940
941 struct merge_file_info {
942         struct object_id oid;
943         unsigned mode;
944         unsigned clean:1,
945                  merge:1;
946 };
947
948 static int merge_3way(struct merge_options *o,
949                       mmbuffer_t *result_buf,
950                       const struct diff_filespec *one,
951                       const struct diff_filespec *a,
952                       const struct diff_filespec *b,
953                       const char *branch1,
954                       const char *branch2)
955 {
956         mmfile_t orig, src1, src2;
957         struct ll_merge_options ll_opts = {0};
958         char *base_name, *name1, *name2;
959         int merge_status;
960
961         ll_opts.renormalize = o->renormalize;
962         ll_opts.xdl_opts = o->xdl_opts;
963
964         if (o->call_depth) {
965                 ll_opts.virtual_ancestor = 1;
966                 ll_opts.variant = 0;
967         } else {
968                 switch (o->recursive_variant) {
969                 case MERGE_RECURSIVE_OURS:
970                         ll_opts.variant = XDL_MERGE_FAVOR_OURS;
971                         break;
972                 case MERGE_RECURSIVE_THEIRS:
973                         ll_opts.variant = XDL_MERGE_FAVOR_THEIRS;
974                         break;
975                 default:
976                         ll_opts.variant = 0;
977                         break;
978                 }
979         }
980
981         if (strcmp(a->path, b->path) ||
982             (o->ancestor != NULL && strcmp(a->path, one->path) != 0)) {
983                 base_name = o->ancestor == NULL ? NULL :
984                         mkpathdup("%s:%s", o->ancestor, one->path);
985                 name1 = mkpathdup("%s:%s", branch1, a->path);
986                 name2 = mkpathdup("%s:%s", branch2, b->path);
987         } else {
988                 base_name = o->ancestor == NULL ? NULL :
989                         mkpathdup("%s", o->ancestor);
990                 name1 = mkpathdup("%s", branch1);
991                 name2 = mkpathdup("%s", branch2);
992         }
993
994         read_mmblob(&orig, &one->oid);
995         read_mmblob(&src1, &a->oid);
996         read_mmblob(&src2, &b->oid);
997
998         merge_status = ll_merge(result_buf, a->path, &orig, base_name,
999                                 &src1, name1, &src2, name2, &ll_opts);
1000
1001         free(base_name);
1002         free(name1);
1003         free(name2);
1004         free(orig.ptr);
1005         free(src1.ptr);
1006         free(src2.ptr);
1007         return merge_status;
1008 }
1009
1010 static int merge_file_1(struct merge_options *o,
1011                                            const struct diff_filespec *one,
1012                                            const struct diff_filespec *a,
1013                                            const struct diff_filespec *b,
1014                                            const char *branch1,
1015                                            const char *branch2,
1016                                            struct merge_file_info *result)
1017 {
1018         result->merge = 0;
1019         result->clean = 1;
1020
1021         if ((S_IFMT & a->mode) != (S_IFMT & b->mode)) {
1022                 result->clean = 0;
1023                 if (S_ISREG(a->mode)) {
1024                         result->mode = a->mode;
1025                         oidcpy(&result->oid, &a->oid);
1026                 } else {
1027                         result->mode = b->mode;
1028                         oidcpy(&result->oid, &b->oid);
1029                 }
1030         } else {
1031                 if (!oid_eq(&a->oid, &one->oid) && !oid_eq(&b->oid, &one->oid))
1032                         result->merge = 1;
1033
1034                 /*
1035                  * Merge modes
1036                  */
1037                 if (a->mode == b->mode || a->mode == one->mode)
1038                         result->mode = b->mode;
1039                 else {
1040                         result->mode = a->mode;
1041                         if (b->mode != one->mode) {
1042                                 result->clean = 0;
1043                                 result->merge = 1;
1044                         }
1045                 }
1046
1047                 if (oid_eq(&a->oid, &b->oid) || oid_eq(&a->oid, &one->oid))
1048                         oidcpy(&result->oid, &b->oid);
1049                 else if (oid_eq(&b->oid, &one->oid))
1050                         oidcpy(&result->oid, &a->oid);
1051                 else if (S_ISREG(a->mode)) {
1052                         mmbuffer_t result_buf;
1053                         int ret = 0, merge_status;
1054
1055                         merge_status = merge_3way(o, &result_buf, one, a, b,
1056                                                   branch1, branch2);
1057
1058                         if ((merge_status < 0) || !result_buf.ptr)
1059                                 ret = err(o, _("Failed to execute internal merge"));
1060
1061                         if (!ret && write_sha1_file(result_buf.ptr, result_buf.size,
1062                                                     blob_type, result->oid.hash))
1063                                 ret = err(o, _("Unable to add %s to database"),
1064                                           a->path);
1065
1066                         free(result_buf.ptr);
1067                         if (ret)
1068                                 return ret;
1069                         result->clean = (merge_status == 0);
1070                 } else if (S_ISGITLINK(a->mode)) {
1071                         result->clean = merge_submodule(&result->oid,
1072                                                        one->path,
1073                                                        &one->oid,
1074                                                        &a->oid,
1075                                                        &b->oid,
1076                                                        !o->call_depth);
1077                 } else if (S_ISLNK(a->mode)) {
1078                         oidcpy(&result->oid, &a->oid);
1079
1080                         if (!oid_eq(&a->oid, &b->oid))
1081                                 result->clean = 0;
1082                 } else
1083                         die("BUG: unsupported object type in the tree");
1084         }
1085
1086         return 0;
1087 }
1088
1089 static int merge_file_special_markers(struct merge_options *o,
1090                            const struct diff_filespec *one,
1091                            const struct diff_filespec *a,
1092                            const struct diff_filespec *b,
1093                            const char *branch1,
1094                            const char *filename1,
1095                            const char *branch2,
1096                            const char *filename2,
1097                            struct merge_file_info *mfi)
1098 {
1099         char *side1 = NULL;
1100         char *side2 = NULL;
1101         int ret;
1102
1103         if (filename1)
1104                 side1 = xstrfmt("%s:%s", branch1, filename1);
1105         if (filename2)
1106                 side2 = xstrfmt("%s:%s", branch2, filename2);
1107
1108         ret = merge_file_1(o, one, a, b,
1109                            side1 ? side1 : branch1,
1110                            side2 ? side2 : branch2, mfi);
1111         free(side1);
1112         free(side2);
1113         return ret;
1114 }
1115
1116 static int merge_file_one(struct merge_options *o,
1117                                          const char *path,
1118                                          const struct object_id *o_oid, int o_mode,
1119                                          const struct object_id *a_oid, int a_mode,
1120                                          const struct object_id *b_oid, int b_mode,
1121                                          const char *branch1,
1122                                          const char *branch2,
1123                                          struct merge_file_info *mfi)
1124 {
1125         struct diff_filespec one, a, b;
1126
1127         one.path = a.path = b.path = (char *)path;
1128         oidcpy(&one.oid, o_oid);
1129         one.mode = o_mode;
1130         oidcpy(&a.oid, a_oid);
1131         a.mode = a_mode;
1132         oidcpy(&b.oid, b_oid);
1133         b.mode = b_mode;
1134         return merge_file_1(o, &one, &a, &b, branch1, branch2, mfi);
1135 }
1136
1137 static int conflict_rename_dir(struct merge_options *o,
1138                                struct diff_filepair *pair,
1139                                const char *rename_branch,
1140                                const char *other_branch)
1141 {
1142         const struct diff_filespec *dest = pair->two;
1143
1144         if (update_file(o, 1, &dest->oid, dest->mode, dest->path))
1145                 return -1;
1146         return 0;
1147 }
1148
1149 static int handle_change_delete(struct merge_options *o,
1150                                  const char *path, const char *old_path,
1151                                  const struct object_id *o_oid, int o_mode,
1152                                  const struct object_id *changed_oid,
1153                                  int changed_mode,
1154                                  const char *change_branch,
1155                                  const char *delete_branch,
1156                                  const char *change, const char *change_past)
1157 {
1158         char *alt_path = NULL;
1159         const char *update_path = path;
1160         int ret = 0;
1161
1162         if (dir_in_way(path, !o->call_depth, 0)) {
1163                 update_path = alt_path = unique_path(o, path, change_branch);
1164         }
1165
1166         if (o->call_depth) {
1167                 /*
1168                  * We cannot arbitrarily accept either a_sha or b_sha as
1169                  * correct; since there is no true "middle point" between
1170                  * them, simply reuse the base version for virtual merge base.
1171                  */
1172                 ret = remove_file_from_cache(path);
1173                 if (!ret)
1174                         ret = update_file(o, 0, o_oid, o_mode, update_path);
1175         } else {
1176                 if (!alt_path) {
1177                         if (!old_path) {
1178                                 output(o, 1, _("CONFLICT (%s/delete): %s deleted in %s "
1179                                        "and %s in %s. Version %s of %s left in tree."),
1180                                        change, path, delete_branch, change_past,
1181                                        change_branch, change_branch, path);
1182                         } else {
1183                                 output(o, 1, _("CONFLICT (%s/delete): %s deleted in %s "
1184                                        "and %s to %s in %s. Version %s of %s left in tree."),
1185                                        change, old_path, delete_branch, change_past, path,
1186                                        change_branch, change_branch, path);
1187                         }
1188                 } else {
1189                         if (!old_path) {
1190                                 output(o, 1, _("CONFLICT (%s/delete): %s deleted in %s "
1191                                        "and %s in %s. Version %s of %s left in tree at %s."),
1192                                        change, path, delete_branch, change_past,
1193                                        change_branch, change_branch, path, alt_path);
1194                         } else {
1195                                 output(o, 1, _("CONFLICT (%s/delete): %s deleted in %s "
1196                                        "and %s to %s in %s. Version %s of %s left in tree at %s."),
1197                                        change, old_path, delete_branch, change_past, path,
1198                                        change_branch, change_branch, path, alt_path);
1199                         }
1200                 }
1201                 /*
1202                  * No need to call update_file() on path when change_branch ==
1203                  * o->branch1 && !alt_path, since that would needlessly touch
1204                  * path.  We could call update_file_flags() with update_cache=0
1205                  * and update_wd=0, but that's a no-op.
1206                  */
1207                 if (change_branch != o->branch1 || alt_path)
1208                         ret = update_file(o, 0, changed_oid, changed_mode, update_path);
1209         }
1210         free(alt_path);
1211
1212         return ret;
1213 }
1214
1215 static int conflict_rename_delete(struct merge_options *o,
1216                                    struct diff_filepair *pair,
1217                                    const char *rename_branch,
1218                                    const char *delete_branch)
1219 {
1220         const struct diff_filespec *orig = pair->one;
1221         const struct diff_filespec *dest = pair->two;
1222
1223         if (handle_change_delete(o,
1224                                  o->call_depth ? orig->path : dest->path,
1225                                  o->call_depth ? NULL : orig->path,
1226                                  &orig->oid, orig->mode,
1227                                  &dest->oid, dest->mode,
1228                                  rename_branch, delete_branch,
1229                                  _("rename"), _("renamed")))
1230                 return -1;
1231
1232         if (o->call_depth)
1233                 return remove_file_from_cache(dest->path);
1234         else
1235                 return update_stages(o, dest->path, NULL,
1236                                      rename_branch == o->branch1 ? dest : NULL,
1237                                      rename_branch == o->branch1 ? NULL : dest);
1238 }
1239
1240 static struct diff_filespec *filespec_from_entry(struct diff_filespec *target,
1241                                                  struct stage_data *entry,
1242                                                  int stage)
1243 {
1244         struct object_id *oid = &entry->stages[stage].oid;
1245         unsigned mode = entry->stages[stage].mode;
1246         if (mode == 0 || is_null_oid(oid))
1247                 return NULL;
1248         oidcpy(&target->oid, oid);
1249         target->mode = mode;
1250         return target;
1251 }
1252
1253 static int handle_file(struct merge_options *o,
1254                         struct diff_filespec *rename,
1255                         int stage,
1256                         struct rename_conflict_info *ci)
1257 {
1258         char *dst_name = rename->path;
1259         struct stage_data *dst_entry;
1260         const char *cur_branch, *other_branch;
1261         struct diff_filespec other;
1262         struct diff_filespec *add;
1263         int ret;
1264
1265         if (stage == 2) {
1266                 dst_entry = ci->dst_entry1;
1267                 cur_branch = ci->branch1;
1268                 other_branch = ci->branch2;
1269         } else {
1270                 dst_entry = ci->dst_entry2;
1271                 cur_branch = ci->branch2;
1272                 other_branch = ci->branch1;
1273         }
1274
1275         add = filespec_from_entry(&other, dst_entry, stage ^ 1);
1276         if (add) {
1277                 char *add_name = unique_path(o, rename->path, other_branch);
1278                 if (update_file(o, 0, &add->oid, add->mode, add_name))
1279                         return -1;
1280
1281                 remove_file(o, 0, rename->path, 0);
1282                 dst_name = unique_path(o, rename->path, cur_branch);
1283         } else {
1284                 if (dir_in_way(rename->path, !o->call_depth, 0)) {
1285                         dst_name = unique_path(o, rename->path, cur_branch);
1286                         output(o, 1, _("%s is a directory in %s adding as %s instead"),
1287                                rename->path, other_branch, dst_name);
1288                 }
1289         }
1290         if ((ret = update_file(o, 0, &rename->oid, rename->mode, dst_name)))
1291                 ; /* fall through, do allow dst_name to be released */
1292         else if (stage == 2)
1293                 ret = update_stages(o, rename->path, NULL, rename, add);
1294         else
1295                 ret = update_stages(o, rename->path, NULL, add, rename);
1296
1297         if (dst_name != rename->path)
1298                 free(dst_name);
1299
1300         return ret;
1301 }
1302
1303 static int conflict_rename_rename_1to2(struct merge_options *o,
1304                                         struct rename_conflict_info *ci)
1305 {
1306         /* One file was renamed in both branches, but to different names. */
1307         struct diff_filespec *one = ci->pair1->one;
1308         struct diff_filespec *a = ci->pair1->two;
1309         struct diff_filespec *b = ci->pair2->two;
1310
1311         output(o, 1, _("CONFLICT (rename/rename): "
1312                "Rename \"%s\"->\"%s\" in branch \"%s\" "
1313                "rename \"%s\"->\"%s\" in \"%s\"%s"),
1314                one->path, a->path, ci->branch1,
1315                one->path, b->path, ci->branch2,
1316                o->call_depth ? _(" (left unresolved)") : "");
1317         if (o->call_depth) {
1318                 struct merge_file_info mfi;
1319                 struct diff_filespec other;
1320                 struct diff_filespec *add;
1321                 if (merge_file_one(o, one->path,
1322                                  &one->oid, one->mode,
1323                                  &a->oid, a->mode,
1324                                  &b->oid, b->mode,
1325                                  ci->branch1, ci->branch2, &mfi))
1326                         return -1;
1327
1328                 /*
1329                  * FIXME: For rename/add-source conflicts (if we could detect
1330                  * such), this is wrong.  We should instead find a unique
1331                  * pathname and then either rename the add-source file to that
1332                  * unique path, or use that unique path instead of src here.
1333                  */
1334                 if (update_file(o, 0, &mfi.oid, mfi.mode, one->path))
1335                         return -1;
1336
1337                 /*
1338                  * Above, we put the merged content at the merge-base's
1339                  * path.  Now we usually need to delete both a->path and
1340                  * b->path.  However, the rename on each side of the merge
1341                  * could also be involved in a rename/add conflict.  In
1342                  * such cases, we should keep the added file around,
1343                  * resolving the conflict at that path in its favor.
1344                  */
1345                 add = filespec_from_entry(&other, ci->dst_entry1, 2 ^ 1);
1346                 if (add) {
1347                         if (update_file(o, 0, &add->oid, add->mode, a->path))
1348                                 return -1;
1349                 }
1350                 else
1351                         remove_file_from_cache(a->path);
1352                 add = filespec_from_entry(&other, ci->dst_entry2, 3 ^ 1);
1353                 if (add) {
1354                         if (update_file(o, 0, &add->oid, add->mode, b->path))
1355                                 return -1;
1356                 }
1357                 else
1358                         remove_file_from_cache(b->path);
1359         } else if (handle_file(o, a, 2, ci) || handle_file(o, b, 3, ci))
1360                 return -1;
1361
1362         return 0;
1363 }
1364
1365 static int conflict_rename_rename_2to1(struct merge_options *o,
1366                                         struct rename_conflict_info *ci)
1367 {
1368         /* Two files, a & b, were renamed to the same thing, c. */
1369         struct diff_filespec *a = ci->pair1->one;
1370         struct diff_filespec *b = ci->pair2->one;
1371         struct diff_filespec *c1 = ci->pair1->two;
1372         struct diff_filespec *c2 = ci->pair2->two;
1373         char *path = c1->path; /* == c2->path */
1374         struct merge_file_info mfi_c1;
1375         struct merge_file_info mfi_c2;
1376         int ret;
1377
1378         output(o, 1, _("CONFLICT (rename/rename): "
1379                "Rename %s->%s in %s. "
1380                "Rename %s->%s in %s"),
1381                a->path, c1->path, ci->branch1,
1382                b->path, c2->path, ci->branch2);
1383
1384         remove_file(o, 1, a->path, o->call_depth || would_lose_untracked(a->path));
1385         remove_file(o, 1, b->path, o->call_depth || would_lose_untracked(b->path));
1386
1387         if (merge_file_special_markers(o, a, c1, &ci->ren1_other,
1388                                        o->branch1, c1->path,
1389                                        o->branch2, ci->ren1_other.path, &mfi_c1) ||
1390             merge_file_special_markers(o, b, &ci->ren2_other, c2,
1391                                        o->branch1, ci->ren2_other.path,
1392                                        o->branch2, c2->path, &mfi_c2))
1393                 return -1;
1394
1395         if (o->call_depth) {
1396                 /*
1397                  * If mfi_c1.clean && mfi_c2.clean, then it might make
1398                  * sense to do a two-way merge of those results.  But, I
1399                  * think in all cases, it makes sense to have the virtual
1400                  * merge base just undo the renames; they can be detected
1401                  * again later for the non-recursive merge.
1402                  */
1403                 remove_file(o, 0, path, 0);
1404                 ret = update_file(o, 0, &mfi_c1.oid, mfi_c1.mode, a->path);
1405                 if (!ret)
1406                         ret = update_file(o, 0, &mfi_c2.oid, mfi_c2.mode,
1407                                           b->path);
1408         } else {
1409                 char *new_path1 = unique_path(o, path, ci->branch1);
1410                 char *new_path2 = unique_path(o, path, ci->branch2);
1411                 output(o, 1, _("Renaming %s to %s and %s to %s instead"),
1412                        a->path, new_path1, b->path, new_path2);
1413                 remove_file(o, 0, path, 0);
1414                 ret = update_file(o, 0, &mfi_c1.oid, mfi_c1.mode, new_path1);
1415                 if (!ret)
1416                         ret = update_file(o, 0, &mfi_c2.oid, mfi_c2.mode,
1417                                           new_path2);
1418                 /*
1419                  * unpack_trees() actually populates the index for us for
1420                  * "normal" rename/rename(2to1) situtations so that the
1421                  * correct entries are at the higher stages, which would
1422                  * make the call below to update_stages_for_stage_data
1423                  * unnecessary.  However, if either of the renames came
1424                  * from a directory rename, then unpack_trees() will not
1425                  * have gotten the right data loaded into the index, so we
1426                  * need to do so now.  (While it'd be tempting to move this
1427                  * call to update_stages_for_stage_data() to
1428                  * apply_directory_rename_modifications(), that would break
1429                  * our intermediate calls to would_lose_untracked() since
1430                  * those rely on the current in-memory index.  See also the
1431                  * big "NOTE" in update_stages()).
1432                  */
1433                 if (update_stages_for_stage_data(o, path, ci->dst_entry1))
1434                         ret = -1;
1435
1436                 free(new_path2);
1437                 free(new_path1);
1438         }
1439
1440         return ret;
1441 }
1442
1443 /*
1444  * Get the diff_filepairs changed between o_tree and tree.
1445  */
1446 static struct diff_queue_struct *get_diffpairs(struct merge_options *o,
1447                                                struct tree *o_tree,
1448                                                struct tree *tree)
1449 {
1450         struct diff_queue_struct *ret;
1451         struct diff_options opts;
1452
1453         diff_setup(&opts);
1454         opts.flags.recursive = 1;
1455         opts.flags.rename_empty = 0;
1456         opts.detect_rename = DIFF_DETECT_RENAME;
1457         opts.rename_limit = o->merge_rename_limit >= 0 ? o->merge_rename_limit :
1458                             o->diff_rename_limit >= 0 ? o->diff_rename_limit :
1459                             1000;
1460         opts.rename_score = o->rename_score;
1461         opts.show_rename_progress = o->show_rename_progress;
1462         opts.output_format = DIFF_FORMAT_NO_OUTPUT;
1463         diff_setup_done(&opts);
1464         diff_tree_oid(&o_tree->object.oid, &tree->object.oid, "", &opts);
1465         diffcore_std(&opts);
1466         if (opts.needed_rename_limit > o->needed_rename_limit)
1467                 o->needed_rename_limit = opts.needed_rename_limit;
1468
1469         ret = xmalloc(sizeof(*ret));
1470         *ret = diff_queued_diff;
1471
1472         opts.output_format = DIFF_FORMAT_NO_OUTPUT;
1473         diff_queued_diff.nr = 0;
1474         diff_queued_diff.queue = NULL;
1475         diff_flush(&opts);
1476         return ret;
1477 }
1478
1479 static int tree_has_path(struct tree *tree, const char *path)
1480 {
1481         unsigned char hashy[GIT_MAX_RAWSZ];
1482         unsigned int mode_o;
1483
1484         return !get_tree_entry(tree->object.oid.hash, path,
1485                                hashy, &mode_o);
1486 }
1487
1488 /*
1489  * Return a new string that replaces the beginning portion (which matches
1490  * entry->dir), with entry->new_dir.  In perl-speak:
1491  *   new_path_name = (old_path =~ s/entry->dir/entry->new_dir/);
1492  * NOTE:
1493  *   Caller must ensure that old_path starts with entry->dir + '/'.
1494  */
1495 static char *apply_dir_rename(struct dir_rename_entry *entry,
1496                               const char *old_path)
1497 {
1498         struct strbuf new_path = STRBUF_INIT;
1499         int oldlen, newlen;
1500
1501         if (entry->non_unique_new_dir)
1502                 return NULL;
1503
1504         oldlen = strlen(entry->dir);
1505         newlen = entry->new_dir.len + (strlen(old_path) - oldlen) + 1;
1506         strbuf_grow(&new_path, newlen);
1507         strbuf_addbuf(&new_path, &entry->new_dir);
1508         strbuf_addstr(&new_path, &old_path[oldlen]);
1509
1510         return strbuf_detach(&new_path, NULL);
1511 }
1512
1513 static void get_renamed_dir_portion(const char *old_path, const char *new_path,
1514                                     char **old_dir, char **new_dir)
1515 {
1516         char *end_of_old, *end_of_new;
1517         int old_len, new_len;
1518
1519         *old_dir = NULL;
1520         *new_dir = NULL;
1521
1522         /*
1523          * For
1524          *    "a/b/c/d/e/foo.c" -> "a/b/some/thing/else/e/foo.c"
1525          * the "e/foo.c" part is the same, we just want to know that
1526          *    "a/b/c/d" was renamed to "a/b/some/thing/else"
1527          * so, for this example, this function returns "a/b/c/d" in
1528          * *old_dir and "a/b/some/thing/else" in *new_dir.
1529          *
1530          * Also, if the basename of the file changed, we don't care.  We
1531          * want to know which portion of the directory, if any, changed.
1532          */
1533         end_of_old = strrchr(old_path, '/');
1534         end_of_new = strrchr(new_path, '/');
1535
1536         if (end_of_old == NULL || end_of_new == NULL)
1537                 return;
1538         while (*--end_of_new == *--end_of_old &&
1539                end_of_old != old_path &&
1540                end_of_new != new_path)
1541                 ; /* Do nothing; all in the while loop */
1542         /*
1543          * We've found the first non-matching character in the directory
1544          * paths.  That means the current directory we were comparing
1545          * represents the rename.  Move end_of_old and end_of_new back
1546          * to the full directory name.
1547          */
1548         if (*end_of_old == '/')
1549                 end_of_old++;
1550         if (*end_of_old != '/')
1551                 end_of_new++;
1552         end_of_old = strchr(end_of_old, '/');
1553         end_of_new = strchr(end_of_new, '/');
1554
1555         /*
1556          * It may have been the case that old_path and new_path were the same
1557          * directory all along.  Don't claim a rename if they're the same.
1558          */
1559         old_len = end_of_old - old_path;
1560         new_len = end_of_new - new_path;
1561
1562         if (old_len != new_len || strncmp(old_path, new_path, old_len)) {
1563                 *old_dir = xstrndup(old_path, old_len);
1564                 *new_dir = xstrndup(new_path, new_len);
1565         }
1566 }
1567
1568 static void remove_hashmap_entries(struct hashmap *dir_renames,
1569                                    struct string_list *items_to_remove)
1570 {
1571         int i;
1572         struct dir_rename_entry *entry;
1573
1574         for (i = 0; i < items_to_remove->nr; i++) {
1575                 entry = items_to_remove->items[i].util;
1576                 hashmap_remove(dir_renames, entry, NULL);
1577         }
1578         string_list_clear(items_to_remove, 0);
1579 }
1580
1581 /*
1582  * See if there is a directory rename for path, and if there are any file
1583  * level conflicts for the renamed location.  If there is a rename and
1584  * there are no conflicts, return the new name.  Otherwise, return NULL.
1585  */
1586 static char *handle_path_level_conflicts(struct merge_options *o,
1587                                          const char *path,
1588                                          struct dir_rename_entry *entry,
1589                                          struct hashmap *collisions,
1590                                          struct tree *tree)
1591 {
1592         char *new_path = NULL;
1593         struct collision_entry *collision_ent;
1594         int clean = 1;
1595         struct strbuf collision_paths = STRBUF_INIT;
1596
1597         /*
1598          * entry has the mapping of old directory name to new directory name
1599          * that we want to apply to path.
1600          */
1601         new_path = apply_dir_rename(entry, path);
1602
1603         if (!new_path) {
1604                 /* This should only happen when entry->non_unique_new_dir set */
1605                 if (!entry->non_unique_new_dir)
1606                         BUG("entry->non_unqiue_dir not set and !new_path");
1607                 output(o, 1, _("CONFLICT (directory rename split): "
1608                                "Unclear where to place %s because directory "
1609                                "%s was renamed to multiple other directories, "
1610                                "with no destination getting a majority of the "
1611                                "files."),
1612                        path, entry->dir);
1613                 clean = 0;
1614                 return NULL;
1615         }
1616
1617         /*
1618          * The caller needs to have ensured that it has pre-populated
1619          * collisions with all paths that map to new_path.  Do a quick check
1620          * to ensure that's the case.
1621          */
1622         collision_ent = collision_find_entry(collisions, new_path);
1623         if (collision_ent == NULL)
1624                 BUG("collision_ent is NULL");
1625
1626         /*
1627          * Check for one-sided add/add/.../add conflicts, i.e.
1628          * where implicit renames from the other side doing
1629          * directory rename(s) can affect this side of history
1630          * to put multiple paths into the same location.  Warn
1631          * and bail on directory renames for such paths.
1632          */
1633         if (collision_ent->reported_already) {
1634                 clean = 0;
1635         } else if (tree_has_path(tree, new_path)) {
1636                 collision_ent->reported_already = 1;
1637                 strbuf_add_separated_string_list(&collision_paths, ", ",
1638                                                  &collision_ent->source_files);
1639                 output(o, 1, _("CONFLICT (implicit dir rename): Existing "
1640                                "file/dir at %s in the way of implicit "
1641                                "directory rename(s) putting the following "
1642                                "path(s) there: %s."),
1643                        new_path, collision_paths.buf);
1644                 clean = 0;
1645         } else if (collision_ent->source_files.nr > 1) {
1646                 collision_ent->reported_already = 1;
1647                 strbuf_add_separated_string_list(&collision_paths, ", ",
1648                                                  &collision_ent->source_files);
1649                 output(o, 1, _("CONFLICT (implicit dir rename): Cannot map "
1650                                "more than one path to %s; implicit directory "
1651                                "renames tried to put these paths there: %s"),
1652                        new_path, collision_paths.buf);
1653                 clean = 0;
1654         }
1655
1656         /* Free memory we no longer need */
1657         strbuf_release(&collision_paths);
1658         if (!clean && new_path) {
1659                 free(new_path);
1660                 return NULL;
1661         }
1662
1663         return new_path;
1664 }
1665
1666 /*
1667  * There are a couple things we want to do at the directory level:
1668  *   1. Check for both sides renaming to the same thing, in order to avoid
1669  *      implicit renaming of files that should be left in place.  (See
1670  *      testcase 6b in t6043 for details.)
1671  *   2. Prune directory renames if there are still files left in the
1672  *      the original directory.  These represent a partial directory rename,
1673  *      i.e. a rename where only some of the files within the directory
1674  *      were renamed elsewhere.  (Technically, this could be done earlier
1675  *      in get_directory_renames(), except that would prevent us from
1676  *      doing the previous check and thus failing testcase 6b.)
1677  *   3. Check for rename/rename(1to2) conflicts (at the directory level).
1678  *      In the future, we could potentially record this info as well and
1679  *      omit reporting rename/rename(1to2) conflicts for each path within
1680  *      the affected directories, thus cleaning up the merge output.
1681  *   NOTE: We do NOT check for rename/rename(2to1) conflicts at the
1682  *         directory level, because merging directories is fine.  If it
1683  *         causes conflicts for files within those merged directories, then
1684  *         that should be detected at the individual path level.
1685  */
1686 static void handle_directory_level_conflicts(struct merge_options *o,
1687                                              struct hashmap *dir_re_head,
1688                                              struct tree *head,
1689                                              struct hashmap *dir_re_merge,
1690                                              struct tree *merge)
1691 {
1692         struct hashmap_iter iter;
1693         struct dir_rename_entry *head_ent;
1694         struct dir_rename_entry *merge_ent;
1695
1696         struct string_list remove_from_head = STRING_LIST_INIT_NODUP;
1697         struct string_list remove_from_merge = STRING_LIST_INIT_NODUP;
1698
1699         hashmap_iter_init(dir_re_head, &iter);
1700         while ((head_ent = hashmap_iter_next(&iter))) {
1701                 merge_ent = dir_rename_find_entry(dir_re_merge, head_ent->dir);
1702                 if (merge_ent &&
1703                     !head_ent->non_unique_new_dir &&
1704                     !merge_ent->non_unique_new_dir &&
1705                     !strbuf_cmp(&head_ent->new_dir, &merge_ent->new_dir)) {
1706                         /* 1. Renamed identically; remove it from both sides */
1707                         string_list_append(&remove_from_head,
1708                                            head_ent->dir)->util = head_ent;
1709                         strbuf_release(&head_ent->new_dir);
1710                         string_list_append(&remove_from_merge,
1711                                            merge_ent->dir)->util = merge_ent;
1712                         strbuf_release(&merge_ent->new_dir);
1713                 } else if (tree_has_path(head, head_ent->dir)) {
1714                         /* 2. This wasn't a directory rename after all */
1715                         string_list_append(&remove_from_head,
1716                                            head_ent->dir)->util = head_ent;
1717                         strbuf_release(&head_ent->new_dir);
1718                 }
1719         }
1720
1721         remove_hashmap_entries(dir_re_head, &remove_from_head);
1722         remove_hashmap_entries(dir_re_merge, &remove_from_merge);
1723
1724         hashmap_iter_init(dir_re_merge, &iter);
1725         while ((merge_ent = hashmap_iter_next(&iter))) {
1726                 head_ent = dir_rename_find_entry(dir_re_head, merge_ent->dir);
1727                 if (tree_has_path(merge, merge_ent->dir)) {
1728                         /* 2. This wasn't a directory rename after all */
1729                         string_list_append(&remove_from_merge,
1730                                            merge_ent->dir)->util = merge_ent;
1731                 } else if (head_ent &&
1732                            !head_ent->non_unique_new_dir &&
1733                            !merge_ent->non_unique_new_dir) {
1734                         /* 3. rename/rename(1to2) */
1735                         /*
1736                          * We can assume it's not rename/rename(1to1) because
1737                          * that was case (1), already checked above.  So we
1738                          * know that head_ent->new_dir and merge_ent->new_dir
1739                          * are different strings.
1740                          */
1741                         output(o, 1, _("CONFLICT (rename/rename): "
1742                                        "Rename directory %s->%s in %s. "
1743                                        "Rename directory %s->%s in %s"),
1744                                head_ent->dir, head_ent->new_dir.buf, o->branch1,
1745                                head_ent->dir, merge_ent->new_dir.buf, o->branch2);
1746                         string_list_append(&remove_from_head,
1747                                            head_ent->dir)->util = head_ent;
1748                         strbuf_release(&head_ent->new_dir);
1749                         string_list_append(&remove_from_merge,
1750                                            merge_ent->dir)->util = merge_ent;
1751                         strbuf_release(&merge_ent->new_dir);
1752                 }
1753         }
1754
1755         remove_hashmap_entries(dir_re_head, &remove_from_head);
1756         remove_hashmap_entries(dir_re_merge, &remove_from_merge);
1757 }
1758
1759 static struct hashmap *get_directory_renames(struct diff_queue_struct *pairs,
1760                                              struct tree *tree)
1761 {
1762         struct hashmap *dir_renames;
1763         struct hashmap_iter iter;
1764         struct dir_rename_entry *entry;
1765         int i;
1766
1767         /*
1768          * Typically, we think of a directory rename as all files from a
1769          * certain directory being moved to a target directory.  However,
1770          * what if someone first moved two files from the original
1771          * directory in one commit, and then renamed the directory
1772          * somewhere else in a later commit?  At merge time, we just know
1773          * that files from the original directory went to two different
1774          * places, and that the bulk of them ended up in the same place.
1775          * We want each directory rename to represent where the bulk of the
1776          * files from that directory end up; this function exists to find
1777          * where the bulk of the files went.
1778          *
1779          * The first loop below simply iterates through the list of file
1780          * renames, finding out how often each directory rename pair
1781          * possibility occurs.
1782          */
1783         dir_renames = xmalloc(sizeof(struct hashmap));
1784         dir_rename_init(dir_renames);
1785         for (i = 0; i < pairs->nr; ++i) {
1786                 struct string_list_item *item;
1787                 int *count;
1788                 struct diff_filepair *pair = pairs->queue[i];
1789                 char *old_dir, *new_dir;
1790
1791                 /* File not part of directory rename if it wasn't renamed */
1792                 if (pair->status != 'R')
1793                         continue;
1794
1795                 get_renamed_dir_portion(pair->one->path, pair->two->path,
1796                                         &old_dir,        &new_dir);
1797                 if (!old_dir)
1798                         /* Directory didn't change at all; ignore this one. */
1799                         continue;
1800
1801                 entry = dir_rename_find_entry(dir_renames, old_dir);
1802                 if (!entry) {
1803                         entry = xmalloc(sizeof(struct dir_rename_entry));
1804                         dir_rename_entry_init(entry, old_dir);
1805                         hashmap_put(dir_renames, entry);
1806                 } else {
1807                         free(old_dir);
1808                 }
1809                 item = string_list_lookup(&entry->possible_new_dirs, new_dir);
1810                 if (!item) {
1811                         item = string_list_insert(&entry->possible_new_dirs,
1812                                                   new_dir);
1813                         item->util = xcalloc(1, sizeof(int));
1814                 } else {
1815                         free(new_dir);
1816                 }
1817                 count = item->util;
1818                 *count += 1;
1819         }
1820
1821         /*
1822          * For each directory with files moved out of it, we find out which
1823          * target directory received the most files so we can declare it to
1824          * be the "winning" target location for the directory rename.  This
1825          * winner gets recorded in new_dir.  If there is no winner
1826          * (multiple target directories received the same number of files),
1827          * we set non_unique_new_dir.  Once we've determined the winner (or
1828          * that there is no winner), we no longer need possible_new_dirs.
1829          */
1830         hashmap_iter_init(dir_renames, &iter);
1831         while ((entry = hashmap_iter_next(&iter))) {
1832                 int max = 0;
1833                 int bad_max = 0;
1834                 char *best = NULL;
1835
1836                 for (i = 0; i < entry->possible_new_dirs.nr; i++) {
1837                         int *count = entry->possible_new_dirs.items[i].util;
1838
1839                         if (*count == max)
1840                                 bad_max = max;
1841                         else if (*count > max) {
1842                                 max = *count;
1843                                 best = entry->possible_new_dirs.items[i].string;
1844                         }
1845                 }
1846                 if (bad_max == max)
1847                         entry->non_unique_new_dir = 1;
1848                 else {
1849                         assert(entry->new_dir.len == 0);
1850                         strbuf_addstr(&entry->new_dir, best);
1851                 }
1852                 /*
1853                  * The relevant directory sub-portion of the original full
1854                  * filepaths were xstrndup'ed before inserting into
1855                  * possible_new_dirs, and instead of manually iterating the
1856                  * list and free'ing each, just lie and tell
1857                  * possible_new_dirs that it did the strdup'ing so that it
1858                  * will free them for us.
1859                  */
1860                 entry->possible_new_dirs.strdup_strings = 1;
1861                 string_list_clear(&entry->possible_new_dirs, 1);
1862         }
1863
1864         return dir_renames;
1865 }
1866
1867 static struct dir_rename_entry *check_dir_renamed(const char *path,
1868                                                   struct hashmap *dir_renames)
1869 {
1870         char temp[PATH_MAX];
1871         char *end;
1872         struct dir_rename_entry *entry;
1873
1874         strcpy(temp, path);
1875         while ((end = strrchr(temp, '/'))) {
1876                 *end = '\0';
1877                 entry = dir_rename_find_entry(dir_renames, temp);
1878                 if (entry)
1879                         return entry;
1880         }
1881         return NULL;
1882 }
1883
1884 static void compute_collisions(struct hashmap *collisions,
1885                                struct hashmap *dir_renames,
1886                                struct diff_queue_struct *pairs)
1887 {
1888         int i;
1889
1890         /*
1891          * Multiple files can be mapped to the same path due to directory
1892          * renames done by the other side of history.  Since that other
1893          * side of history could have merged multiple directories into one,
1894          * if our side of history added the same file basename to each of
1895          * those directories, then all N of them would get implicitly
1896          * renamed by the directory rename detection into the same path,
1897          * and we'd get an add/add/.../add conflict, and all those adds
1898          * from *this* side of history.  This is not representable in the
1899          * index, and users aren't going to easily be able to make sense of
1900          * it.  So we need to provide a good warning about what's
1901          * happening, and fall back to no-directory-rename detection
1902          * behavior for those paths.
1903          *
1904          * See testcases 9e and all of section 5 from t6043 for examples.
1905          */
1906         collision_init(collisions);
1907
1908         for (i = 0; i < pairs->nr; ++i) {
1909                 struct dir_rename_entry *dir_rename_ent;
1910                 struct collision_entry *collision_ent;
1911                 char *new_path;
1912                 struct diff_filepair *pair = pairs->queue[i];
1913
1914                 if (pair->status == 'D')
1915                         continue;
1916                 dir_rename_ent = check_dir_renamed(pair->two->path,
1917                                                    dir_renames);
1918                 if (!dir_rename_ent)
1919                         continue;
1920
1921                 new_path = apply_dir_rename(dir_rename_ent, pair->two->path);
1922                 if (!new_path)
1923                         /*
1924                          * dir_rename_ent->non_unique_new_path is true, which
1925                          * means there is no directory rename for us to use,
1926                          * which means it won't cause us any additional
1927                          * collisions.
1928                          */
1929                         continue;
1930                 collision_ent = collision_find_entry(collisions, new_path);
1931                 if (!collision_ent) {
1932                         collision_ent = xcalloc(1,
1933                                                 sizeof(struct collision_entry));
1934                         hashmap_entry_init(collision_ent, strhash(new_path));
1935                         hashmap_put(collisions, collision_ent);
1936                         collision_ent->target_file = new_path;
1937                 } else {
1938                         free(new_path);
1939                 }
1940                 string_list_insert(&collision_ent->source_files,
1941                                    pair->two->path);
1942         }
1943 }
1944
1945 static char *check_for_directory_rename(struct merge_options *o,
1946                                         const char *path,
1947                                         struct tree *tree,
1948                                         struct hashmap *dir_renames,
1949                                         struct hashmap *dir_rename_exclusions,
1950                                         struct hashmap *collisions,
1951                                         int *clean_merge)
1952 {
1953         char *new_path = NULL;
1954         struct dir_rename_entry *entry = check_dir_renamed(path, dir_renames);
1955         struct dir_rename_entry *oentry = NULL;
1956
1957         if (!entry)
1958                 return new_path;
1959
1960         /*
1961          * This next part is a little weird.  We do not want to do an
1962          * implicit rename into a directory we renamed on our side, because
1963          * that will result in a spurious rename/rename(1to2) conflict.  An
1964          * example:
1965          *   Base commit: dumbdir/afile, otherdir/bfile
1966          *   Side 1:      smrtdir/afile, otherdir/bfile
1967          *   Side 2:      dumbdir/afile, dumbdir/bfile
1968          * Here, while working on Side 1, we could notice that otherdir was
1969          * renamed/merged to dumbdir, and change the diff_filepair for
1970          * otherdir/bfile into a rename into dumbdir/bfile.  However, Side
1971          * 2 will notice the rename from dumbdir to smrtdir, and do the
1972          * transitive rename to move it from dumbdir/bfile to
1973          * smrtdir/bfile.  That gives us bfile in dumbdir vs being in
1974          * smrtdir, a rename/rename(1to2) conflict.  We really just want
1975          * the file to end up in smrtdir.  And the way to achieve that is
1976          * to not let Side1 do the rename to dumbdir, since we know that is
1977          * the source of one of our directory renames.
1978          *
1979          * That's why oentry and dir_rename_exclusions is here.
1980          *
1981          * As it turns out, this also prevents N-way transient rename
1982          * confusion; See testcases 9c and 9d of t6043.
1983          */
1984         oentry = dir_rename_find_entry(dir_rename_exclusions, entry->new_dir.buf);
1985         if (oentry) {
1986                 output(o, 1, _("WARNING: Avoiding applying %s -> %s rename "
1987                                "to %s, because %s itself was renamed."),
1988                        entry->dir, entry->new_dir.buf, path, entry->new_dir.buf);
1989         } else {
1990                 new_path = handle_path_level_conflicts(o, path, entry,
1991                                                        collisions, tree);
1992                 *clean_merge &= (new_path != NULL);
1993         }
1994
1995         return new_path;
1996 }
1997
1998 static void apply_directory_rename_modifications(struct merge_options *o,
1999                                                  struct diff_filepair *pair,
2000                                                  char *new_path,
2001                                                  struct rename *re,
2002                                                  struct tree *tree,
2003                                                  struct tree *o_tree,
2004                                                  struct tree *a_tree,
2005                                                  struct tree *b_tree,
2006                                                  struct string_list *entries,
2007                                                  int *clean)
2008 {
2009         struct string_list_item *item;
2010         int stage = (tree == a_tree ? 2 : 3);
2011
2012         /*
2013          * In all cases where we can do directory rename detection,
2014          * unpack_trees() will have read pair->two->path into the
2015          * index and the working copy.  We need to remove it so that
2016          * we can instead place it at new_path.  It is guaranteed to
2017          * not be untracked (unpack_trees() would have errored out
2018          * saying the file would have been overwritten), but it might
2019          * be dirty, though.
2020          */
2021         remove_file(o, 1, pair->two->path, 0 /* no_wd */);
2022
2023         /* Find or create a new re->dst_entry */
2024         item = string_list_lookup(entries, new_path);
2025         if (item) {
2026                 /*
2027                  * Since we're renaming on this side of history, and it's
2028                  * due to a directory rename on the other side of history
2029                  * (which we only allow when the directory in question no
2030                  * longer exists on the other side of history), the
2031                  * original entry for re->dst_entry is no longer
2032                  * necessary...
2033                  */
2034                 re->dst_entry->processed = 1;
2035
2036                 /*
2037                  * ...because we'll be using this new one.
2038                  */
2039                 re->dst_entry = item->util;
2040         } else {
2041                 /*
2042                  * re->dst_entry is for the before-dir-rename path, and we
2043                  * need it to hold information for the after-dir-rename
2044                  * path.  Before creating a new entry, we need to mark the
2045                  * old one as unnecessary (...unless it is shared by
2046                  * src_entry, i.e. this didn't use to be a rename, in which
2047                  * case we can just allow the normal processing to happen
2048                  * for it).
2049                  */
2050                 if (pair->status == 'R')
2051                         re->dst_entry->processed = 1;
2052
2053                 re->dst_entry = insert_stage_data(new_path,
2054                                                   o_tree, a_tree, b_tree,
2055                                                   entries);
2056                 item = string_list_insert(entries, new_path);
2057                 item->util = re->dst_entry;
2058         }
2059
2060         /*
2061          * Update the stage_data with the information about the path we are
2062          * moving into place.  That slot will be empty and available for us
2063          * to write to because of the collision checks in
2064          * handle_path_level_conflicts().  In other words,
2065          * re->dst_entry->stages[stage].oid will be the null_oid, so it's
2066          * open for us to write to.
2067          *
2068          * It may be tempting to actually update the index at this point as
2069          * well, using update_stages_for_stage_data(), but as per the big
2070          * "NOTE" in update_stages(), doing so will modify the current
2071          * in-memory index which will break calls to would_lose_untracked()
2072          * that we need to make.  Instead, we need to just make sure that
2073          * the various conflict_rename_*() functions update the index
2074          * explicitly rather than relying on unpack_trees() to have done it.
2075          */
2076         get_tree_entry(tree->object.oid.hash,
2077                        pair->two->path,
2078                        re->dst_entry->stages[stage].oid.hash,
2079                        &re->dst_entry->stages[stage].mode);
2080
2081         /* Update pair status */
2082         if (pair->status == 'A') {
2083                 /*
2084                  * Recording rename information for this add makes it look
2085                  * like a rename/delete conflict.  Make sure we can
2086                  * correctly handle this as an add that was moved to a new
2087                  * directory instead of reporting a rename/delete conflict.
2088                  */
2089                 re->add_turned_into_rename = 1;
2090         }
2091         /*
2092          * We don't actually look at pair->status again, but it seems
2093          * pedagogically correct to adjust it.
2094          */
2095         pair->status = 'R';
2096
2097         /*
2098          * Finally, record the new location.
2099          */
2100         pair->two->path = new_path;
2101 }
2102
2103 /*
2104  * Get information of all renames which occurred in 'pairs', making use of
2105  * any implicit directory renames inferred from the other side of history.
2106  * We need the three trees in the merge ('o_tree', 'a_tree' and 'b_tree')
2107  * to be able to associate the correct cache entries with the rename
2108  * information; tree is always equal to either a_tree or b_tree.
2109  */
2110 static struct string_list *get_renames(struct merge_options *o,
2111                                        struct diff_queue_struct *pairs,
2112                                        struct hashmap *dir_renames,
2113                                        struct hashmap *dir_rename_exclusions,
2114                                        struct tree *tree,
2115                                        struct tree *o_tree,
2116                                        struct tree *a_tree,
2117                                        struct tree *b_tree,
2118                                        struct string_list *entries,
2119                                        int *clean_merge)
2120 {
2121         int i;
2122         struct hashmap collisions;
2123         struct hashmap_iter iter;
2124         struct collision_entry *e;
2125         struct string_list *renames;
2126
2127         compute_collisions(&collisions, dir_renames, pairs);
2128         renames = xcalloc(1, sizeof(struct string_list));
2129
2130         for (i = 0; i < pairs->nr; ++i) {
2131                 struct string_list_item *item;
2132                 struct rename *re;
2133                 struct diff_filepair *pair = pairs->queue[i];
2134                 char *new_path; /* non-NULL only with directory renames */
2135
2136                 if (pair->status == 'D') {
2137                         diff_free_filepair(pair);
2138                         continue;
2139                 }
2140                 new_path = check_for_directory_rename(o, pair->two->path, tree,
2141                                                       dir_renames,
2142                                                       dir_rename_exclusions,
2143                                                       &collisions,
2144                                                       clean_merge);
2145                 if (pair->status != 'R' && !new_path) {
2146                         diff_free_filepair(pair);
2147                         continue;
2148                 }
2149
2150                 re = xmalloc(sizeof(*re));
2151                 re->processed = 0;
2152                 re->add_turned_into_rename = 0;
2153                 re->pair = pair;
2154                 item = string_list_lookup(entries, re->pair->one->path);
2155                 if (!item)
2156                         re->src_entry = insert_stage_data(re->pair->one->path,
2157                                         o_tree, a_tree, b_tree, entries);
2158                 else
2159                         re->src_entry = item->util;
2160
2161                 item = string_list_lookup(entries, re->pair->two->path);
2162                 if (!item)
2163                         re->dst_entry = insert_stage_data(re->pair->two->path,
2164                                         o_tree, a_tree, b_tree, entries);
2165                 else
2166                         re->dst_entry = item->util;
2167                 item = string_list_insert(renames, pair->one->path);
2168                 item->util = re;
2169                 if (new_path)
2170                         apply_directory_rename_modifications(o, pair, new_path,
2171                                                              re, tree, o_tree,
2172                                                              a_tree, b_tree,
2173                                                              entries,
2174                                                              clean_merge);
2175         }
2176
2177         hashmap_iter_init(&collisions, &iter);
2178         while ((e = hashmap_iter_next(&iter))) {
2179                 free(e->target_file);
2180                 string_list_clear(&e->source_files, 0);
2181         }
2182         hashmap_free(&collisions, 1);
2183         return renames;
2184 }
2185
2186 static int process_renames(struct merge_options *o,
2187                            struct string_list *a_renames,
2188                            struct string_list *b_renames)
2189 {
2190         int clean_merge = 1, i, j;
2191         struct string_list a_by_dst = STRING_LIST_INIT_NODUP;
2192         struct string_list b_by_dst = STRING_LIST_INIT_NODUP;
2193         const struct rename *sre;
2194
2195         for (i = 0; i < a_renames->nr; i++) {
2196                 sre = a_renames->items[i].util;
2197                 string_list_insert(&a_by_dst, sre->pair->two->path)->util
2198                         = (void *)sre;
2199         }
2200         for (i = 0; i < b_renames->nr; i++) {
2201                 sre = b_renames->items[i].util;
2202                 string_list_insert(&b_by_dst, sre->pair->two->path)->util
2203                         = (void *)sre;
2204         }
2205
2206         for (i = 0, j = 0; i < a_renames->nr || j < b_renames->nr;) {
2207                 struct string_list *renames1, *renames2Dst;
2208                 struct rename *ren1 = NULL, *ren2 = NULL;
2209                 const char *branch1, *branch2;
2210                 const char *ren1_src, *ren1_dst;
2211                 struct string_list_item *lookup;
2212
2213                 if (i >= a_renames->nr) {
2214                         ren2 = b_renames->items[j++].util;
2215                 } else if (j >= b_renames->nr) {
2216                         ren1 = a_renames->items[i++].util;
2217                 } else {
2218                         int compare = strcmp(a_renames->items[i].string,
2219                                              b_renames->items[j].string);
2220                         if (compare <= 0)
2221                                 ren1 = a_renames->items[i++].util;
2222                         if (compare >= 0)
2223                                 ren2 = b_renames->items[j++].util;
2224                 }
2225
2226                 /* TODO: refactor, so that 1/2 are not needed */
2227                 if (ren1) {
2228                         renames1 = a_renames;
2229                         renames2Dst = &b_by_dst;
2230                         branch1 = o->branch1;
2231                         branch2 = o->branch2;
2232                 } else {
2233                         renames1 = b_renames;
2234                         renames2Dst = &a_by_dst;
2235                         branch1 = o->branch2;
2236                         branch2 = o->branch1;
2237                         SWAP(ren2, ren1);
2238                 }
2239
2240                 if (ren1->processed)
2241                         continue;
2242                 ren1->processed = 1;
2243                 ren1->dst_entry->processed = 1;
2244                 /* BUG: We should only mark src_entry as processed if we
2245                  * are not dealing with a rename + add-source case.
2246                  */
2247                 ren1->src_entry->processed = 1;
2248
2249                 ren1_src = ren1->pair->one->path;
2250                 ren1_dst = ren1->pair->two->path;
2251
2252                 if (ren2) {
2253                         /* One file renamed on both sides */
2254                         const char *ren2_src = ren2->pair->one->path;
2255                         const char *ren2_dst = ren2->pair->two->path;
2256                         enum rename_type rename_type;
2257                         if (strcmp(ren1_src, ren2_src) != 0)
2258                                 die("BUG: ren1_src != ren2_src");
2259                         ren2->dst_entry->processed = 1;
2260                         ren2->processed = 1;
2261                         if (strcmp(ren1_dst, ren2_dst) != 0) {
2262                                 rename_type = RENAME_ONE_FILE_TO_TWO;
2263                                 clean_merge = 0;
2264                         } else {
2265                                 rename_type = RENAME_ONE_FILE_TO_ONE;
2266                                 /* BUG: We should only remove ren1_src in
2267                                  * the base stage (think of rename +
2268                                  * add-source cases).
2269                                  */
2270                                 remove_file(o, 1, ren1_src, 1);
2271                                 update_entry(ren1->dst_entry,
2272                                              ren1->pair->one,
2273                                              ren1->pair->two,
2274                                              ren2->pair->two);
2275                         }
2276                         setup_rename_conflict_info(rename_type,
2277                                                    ren1->pair,
2278                                                    ren2->pair,
2279                                                    branch1,
2280                                                    branch2,
2281                                                    ren1->dst_entry,
2282                                                    ren2->dst_entry,
2283                                                    o,
2284                                                    NULL,
2285                                                    NULL);
2286                 } else if ((lookup = string_list_lookup(renames2Dst, ren1_dst))) {
2287                         /* Two different files renamed to the same thing */
2288                         char *ren2_dst;
2289                         ren2 = lookup->util;
2290                         ren2_dst = ren2->pair->two->path;
2291                         if (strcmp(ren1_dst, ren2_dst) != 0)
2292                                 die("BUG: ren1_dst != ren2_dst");
2293
2294                         clean_merge = 0;
2295                         ren2->processed = 1;
2296                         /*
2297                          * BUG: We should only mark src_entry as processed
2298                          * if we are not dealing with a rename + add-source
2299                          * case.
2300                          */
2301                         ren2->src_entry->processed = 1;
2302
2303                         setup_rename_conflict_info(RENAME_TWO_FILES_TO_ONE,
2304                                                    ren1->pair,
2305                                                    ren2->pair,
2306                                                    branch1,
2307                                                    branch2,
2308                                                    ren1->dst_entry,
2309                                                    ren2->dst_entry,
2310                                                    o,
2311                                                    ren1->src_entry,
2312                                                    ren2->src_entry);
2313
2314                 } else {
2315                         /* Renamed in 1, maybe changed in 2 */
2316                         /* we only use sha1 and mode of these */
2317                         struct diff_filespec src_other, dst_other;
2318                         int try_merge;
2319
2320                         /*
2321                          * unpack_trees loads entries from common-commit
2322                          * into stage 1, from head-commit into stage 2, and
2323                          * from merge-commit into stage 3.  We keep track
2324                          * of which side corresponds to the rename.
2325                          */
2326                         int renamed_stage = a_renames == renames1 ? 2 : 3;
2327                         int other_stage =   a_renames == renames1 ? 3 : 2;
2328
2329                         /* BUG: We should only remove ren1_src in the base
2330                          * stage and in other_stage (think of rename +
2331                          * add-source case).
2332                          */
2333                         remove_file(o, 1, ren1_src,
2334                                     renamed_stage == 2 || !was_tracked(ren1_src));
2335
2336                         oidcpy(&src_other.oid,
2337                                &ren1->src_entry->stages[other_stage].oid);
2338                         src_other.mode = ren1->src_entry->stages[other_stage].mode;
2339                         oidcpy(&dst_other.oid,
2340                                &ren1->dst_entry->stages[other_stage].oid);
2341                         dst_other.mode = ren1->dst_entry->stages[other_stage].mode;
2342                         try_merge = 0;
2343
2344                         if (oid_eq(&src_other.oid, &null_oid) &&
2345                             ren1->add_turned_into_rename) {
2346                                 setup_rename_conflict_info(RENAME_DIR,
2347                                                            ren1->pair,
2348                                                            NULL,
2349                                                            branch1,
2350                                                            branch2,
2351                                                            ren1->dst_entry,
2352                                                            NULL,
2353                                                            o,
2354                                                            NULL,
2355                                                            NULL);
2356                         } else if (oid_eq(&src_other.oid, &null_oid)) {
2357                                 setup_rename_conflict_info(RENAME_DELETE,
2358                                                            ren1->pair,
2359                                                            NULL,
2360                                                            branch1,
2361                                                            branch2,
2362                                                            ren1->dst_entry,
2363                                                            NULL,
2364                                                            o,
2365                                                            NULL,
2366                                                            NULL);
2367                         } else if ((dst_other.mode == ren1->pair->two->mode) &&
2368                                    oid_eq(&dst_other.oid, &ren1->pair->two->oid)) {
2369                                 /*
2370                                  * Added file on the other side identical to
2371                                  * the file being renamed: clean merge.
2372                                  * Also, there is no need to overwrite the
2373                                  * file already in the working copy, so call
2374                                  * update_file_flags() instead of
2375                                  * update_file().
2376                                  */
2377                                 if (update_file_flags(o,
2378                                                       &ren1->pair->two->oid,
2379                                                       ren1->pair->two->mode,
2380                                                       ren1_dst,
2381                                                       1, /* update_cache */
2382                                                       0  /* update_wd    */))
2383                                         clean_merge = -1;
2384                         } else if (!oid_eq(&dst_other.oid, &null_oid)) {
2385                                 clean_merge = 0;
2386                                 try_merge = 1;
2387                                 output(o, 1, _("CONFLICT (rename/add): Rename %s->%s in %s. "
2388                                        "%s added in %s"),
2389                                        ren1_src, ren1_dst, branch1,
2390                                        ren1_dst, branch2);
2391                                 if (o->call_depth) {
2392                                         struct merge_file_info mfi;
2393                                         if (merge_file_one(o, ren1_dst, &null_oid, 0,
2394                                                            &ren1->pair->two->oid,
2395                                                            ren1->pair->two->mode,
2396                                                            &dst_other.oid,
2397                                                            dst_other.mode,
2398                                                            branch1, branch2, &mfi)) {
2399                                                 clean_merge = -1;
2400                                                 goto cleanup_and_return;
2401                                         }
2402                                         output(o, 1, _("Adding merged %s"), ren1_dst);
2403                                         if (update_file(o, 0, &mfi.oid,
2404                                                         mfi.mode, ren1_dst))
2405                                                 clean_merge = -1;
2406                                         try_merge = 0;
2407                                 } else {
2408                                         char *new_path = unique_path(o, ren1_dst, branch2);
2409                                         output(o, 1, _("Adding as %s instead"), new_path);
2410                                         if (update_file(o, 0, &dst_other.oid,
2411                                                         dst_other.mode, new_path))
2412                                                 clean_merge = -1;
2413                                         free(new_path);
2414                                 }
2415                         } else
2416                                 try_merge = 1;
2417
2418                         if (clean_merge < 0)
2419                                 goto cleanup_and_return;
2420                         if (try_merge) {
2421                                 struct diff_filespec *one, *a, *b;
2422                                 src_other.path = (char *)ren1_src;
2423
2424                                 one = ren1->pair->one;
2425                                 if (a_renames == renames1) {
2426                                         a = ren1->pair->two;
2427                                         b = &src_other;
2428                                 } else {
2429                                         b = ren1->pair->two;
2430                                         a = &src_other;
2431                                 }
2432                                 update_entry(ren1->dst_entry, one, a, b);
2433                                 setup_rename_conflict_info(RENAME_NORMAL,
2434                                                            ren1->pair,
2435                                                            NULL,
2436                                                            branch1,
2437                                                            NULL,
2438                                                            ren1->dst_entry,
2439                                                            NULL,
2440                                                            o,
2441                                                            NULL,
2442                                                            NULL);
2443                         }
2444                 }
2445         }
2446 cleanup_and_return:
2447         string_list_clear(&a_by_dst, 0);
2448         string_list_clear(&b_by_dst, 0);
2449
2450         return clean_merge;
2451 }
2452
2453 struct rename_info {
2454         struct string_list *head_renames;
2455         struct string_list *merge_renames;
2456 };
2457
2458 static void initial_cleanup_rename(struct diff_queue_struct *pairs,
2459                                    struct hashmap *dir_renames)
2460 {
2461         struct hashmap_iter iter;
2462         struct dir_rename_entry *e;
2463
2464         hashmap_iter_init(dir_renames, &iter);
2465         while ((e = hashmap_iter_next(&iter))) {
2466                 free(e->dir);
2467                 strbuf_release(&e->new_dir);
2468                 /* possible_new_dirs already cleared in get_directory_renames */
2469         }
2470         hashmap_free(dir_renames, 1);
2471         free(dir_renames);
2472
2473         free(pairs->queue);
2474         free(pairs);
2475 }
2476
2477 static int handle_renames(struct merge_options *o,
2478                           struct tree *common,
2479                           struct tree *head,
2480                           struct tree *merge,
2481                           struct string_list *entries,
2482                           struct rename_info *ri)
2483 {
2484         struct diff_queue_struct *head_pairs, *merge_pairs;
2485         struct hashmap *dir_re_head, *dir_re_merge;
2486         int clean = 1;
2487
2488         ri->head_renames = NULL;
2489         ri->merge_renames = NULL;
2490
2491         if (!o->detect_rename)
2492                 return 1;
2493
2494         head_pairs = get_diffpairs(o, common, head);
2495         merge_pairs = get_diffpairs(o, common, merge);
2496
2497         dir_re_head = get_directory_renames(head_pairs, head);
2498         dir_re_merge = get_directory_renames(merge_pairs, merge);
2499
2500         handle_directory_level_conflicts(o,
2501                                          dir_re_head, head,
2502                                          dir_re_merge, merge);
2503
2504         ri->head_renames  = get_renames(o, head_pairs,
2505                                         dir_re_merge, dir_re_head, head,
2506                                         common, head, merge, entries,
2507                                         &clean);
2508         if (clean < 0)
2509                 goto cleanup;
2510         ri->merge_renames = get_renames(o, merge_pairs,
2511                                         dir_re_head, dir_re_merge, merge,
2512                                         common, head, merge, entries,
2513                                         &clean);
2514         if (clean < 0)
2515                 goto cleanup;
2516         clean &= process_renames(o, ri->head_renames, ri->merge_renames);
2517
2518 cleanup:
2519         /*
2520          * Some cleanup is deferred until cleanup_renames() because the
2521          * data structures are still needed and referenced in
2522          * process_entry().  But there are a few things we can free now.
2523          */
2524         initial_cleanup_rename(head_pairs, dir_re_head);
2525         initial_cleanup_rename(merge_pairs, dir_re_merge);
2526
2527         return clean;
2528 }
2529
2530 static void final_cleanup_rename(struct string_list *rename)
2531 {
2532         const struct rename *re;
2533         int i;
2534
2535         if (rename == NULL)
2536                 return;
2537
2538         for (i = 0; i < rename->nr; i++) {
2539                 re = rename->items[i].util;
2540                 diff_free_filepair(re->pair);
2541         }
2542         string_list_clear(rename, 1);
2543         free(rename);
2544 }
2545
2546 static void final_cleanup_renames(struct rename_info *re_info)
2547 {
2548         final_cleanup_rename(re_info->head_renames);
2549         final_cleanup_rename(re_info->merge_renames);
2550 }
2551
2552 static struct object_id *stage_oid(const struct object_id *oid, unsigned mode)
2553 {
2554         return (is_null_oid(oid) || mode == 0) ? NULL: (struct object_id *)oid;
2555 }
2556
2557 static int read_oid_strbuf(struct merge_options *o,
2558         const struct object_id *oid, struct strbuf *dst)
2559 {
2560         void *buf;
2561         enum object_type type;
2562         unsigned long size;
2563         buf = read_sha1_file(oid->hash, &type, &size);
2564         if (!buf)
2565                 return err(o, _("cannot read object %s"), oid_to_hex(oid));
2566         if (type != OBJ_BLOB) {
2567                 free(buf);
2568                 return err(o, _("object %s is not a blob"), oid_to_hex(oid));
2569         }
2570         strbuf_attach(dst, buf, size, size + 1);
2571         return 0;
2572 }
2573
2574 static int blob_unchanged(struct merge_options *opt,
2575                           const struct object_id *o_oid,
2576                           unsigned o_mode,
2577                           const struct object_id *a_oid,
2578                           unsigned a_mode,
2579                           int renormalize, const char *path)
2580 {
2581         struct strbuf o = STRBUF_INIT;
2582         struct strbuf a = STRBUF_INIT;
2583         int ret = 0; /* assume changed for safety */
2584
2585         if (a_mode != o_mode)
2586                 return 0;
2587         if (oid_eq(o_oid, a_oid))
2588                 return 1;
2589         if (!renormalize)
2590                 return 0;
2591
2592         assert(o_oid && a_oid);
2593         if (read_oid_strbuf(opt, o_oid, &o) || read_oid_strbuf(opt, a_oid, &a))
2594                 goto error_return;
2595         /*
2596          * Note: binary | is used so that both renormalizations are
2597          * performed.  Comparison can be skipped if both files are
2598          * unchanged since their sha1s have already been compared.
2599          */
2600         if (renormalize_buffer(&the_index, path, o.buf, o.len, &o) |
2601             renormalize_buffer(&the_index, path, a.buf, a.len, &a))
2602                 ret = (o.len == a.len && !memcmp(o.buf, a.buf, o.len));
2603
2604 error_return:
2605         strbuf_release(&o);
2606         strbuf_release(&a);
2607         return ret;
2608 }
2609
2610 static int handle_modify_delete(struct merge_options *o,
2611                                  const char *path,
2612                                  struct object_id *o_oid, int o_mode,
2613                                  struct object_id *a_oid, int a_mode,
2614                                  struct object_id *b_oid, int b_mode)
2615 {
2616         const char *modify_branch, *delete_branch;
2617         struct object_id *changed_oid;
2618         int changed_mode;
2619
2620         if (a_oid) {
2621                 modify_branch = o->branch1;
2622                 delete_branch = o->branch2;
2623                 changed_oid = a_oid;
2624                 changed_mode = a_mode;
2625         } else {
2626                 modify_branch = o->branch2;
2627                 delete_branch = o->branch1;
2628                 changed_oid = b_oid;
2629                 changed_mode = b_mode;
2630         }
2631
2632         return handle_change_delete(o,
2633                                     path, NULL,
2634                                     o_oid, o_mode,
2635                                     changed_oid, changed_mode,
2636                                     modify_branch, delete_branch,
2637                                     _("modify"), _("modified"));
2638 }
2639
2640 static int merge_content(struct merge_options *o,
2641                          const char *path,
2642                          struct object_id *o_oid, int o_mode,
2643                          struct object_id *a_oid, int a_mode,
2644                          struct object_id *b_oid, int b_mode,
2645                          struct rename_conflict_info *rename_conflict_info)
2646 {
2647         const char *reason = _("content");
2648         const char *path1 = NULL, *path2 = NULL;
2649         struct merge_file_info mfi;
2650         struct diff_filespec one, a, b;
2651         unsigned df_conflict_remains = 0;
2652
2653         if (!o_oid) {
2654                 reason = _("add/add");
2655                 o_oid = (struct object_id *)&null_oid;
2656         }
2657         one.path = a.path = b.path = (char *)path;
2658         oidcpy(&one.oid, o_oid);
2659         one.mode = o_mode;
2660         oidcpy(&a.oid, a_oid);
2661         a.mode = a_mode;
2662         oidcpy(&b.oid, b_oid);
2663         b.mode = b_mode;
2664
2665         if (rename_conflict_info) {
2666                 struct diff_filepair *pair1 = rename_conflict_info->pair1;
2667
2668                 path1 = (o->branch1 == rename_conflict_info->branch1) ?
2669                         pair1->two->path : pair1->one->path;
2670                 /* If rename_conflict_info->pair2 != NULL, we are in
2671                  * RENAME_ONE_FILE_TO_ONE case.  Otherwise, we have a
2672                  * normal rename.
2673                  */
2674                 path2 = (rename_conflict_info->pair2 ||
2675                          o->branch2 == rename_conflict_info->branch1) ?
2676                         pair1->two->path : pair1->one->path;
2677
2678                 if (dir_in_way(path, !o->call_depth,
2679                                S_ISGITLINK(pair1->two->mode)))
2680                         df_conflict_remains = 1;
2681         }
2682         if (merge_file_special_markers(o, &one, &a, &b,
2683                                        o->branch1, path1,
2684                                        o->branch2, path2, &mfi))
2685                 return -1;
2686
2687         if (mfi.clean && !df_conflict_remains &&
2688             oid_eq(&mfi.oid, a_oid) && mfi.mode == a_mode) {
2689                 int path_renamed_outside_HEAD;
2690                 output(o, 3, _("Skipped %s (merged same as existing)"), path);
2691                 /*
2692                  * The content merge resulted in the same file contents we
2693                  * already had.  We can return early if those file contents
2694                  * are recorded at the correct path (which may not be true
2695                  * if the merge involves a rename).
2696                  */
2697                 path_renamed_outside_HEAD = !path2 || !strcmp(path, path2);
2698                 if (!path_renamed_outside_HEAD) {
2699                         add_cacheinfo(o, mfi.mode, &mfi.oid, path,
2700                                       0, (!o->call_depth), 0);
2701                         return mfi.clean;
2702                 }
2703         } else
2704                 output(o, 2, _("Auto-merging %s"), path);
2705
2706         if (!mfi.clean) {
2707                 if (S_ISGITLINK(mfi.mode))
2708                         reason = _("submodule");
2709                 output(o, 1, _("CONFLICT (%s): Merge conflict in %s"),
2710                                 reason, path);
2711                 if (rename_conflict_info && !df_conflict_remains)
2712                         if (update_stages(o, path, &one, &a, &b))
2713                                 return -1;
2714         }
2715
2716         if (df_conflict_remains) {
2717                 char *new_path;
2718                 if (o->call_depth) {
2719                         remove_file_from_cache(path);
2720                 } else {
2721                         if (!mfi.clean) {
2722                                 if (update_stages(o, path, &one, &a, &b))
2723                                         return -1;
2724                         } else {
2725                                 int file_from_stage2 = was_tracked(path);
2726                                 struct diff_filespec merged;
2727                                 oidcpy(&merged.oid, &mfi.oid);
2728                                 merged.mode = mfi.mode;
2729
2730                                 if (update_stages(o, path, NULL,
2731                                                   file_from_stage2 ? &merged : NULL,
2732                                                   file_from_stage2 ? NULL : &merged))
2733                                         return -1;
2734                         }
2735
2736                 }
2737                 new_path = unique_path(o, path, rename_conflict_info->branch1);
2738                 output(o, 1, _("Adding as %s instead"), new_path);
2739                 if (update_file(o, 0, &mfi.oid, mfi.mode, new_path)) {
2740                         free(new_path);
2741                         return -1;
2742                 }
2743                 free(new_path);
2744                 mfi.clean = 0;
2745         } else if (update_file(o, mfi.clean, &mfi.oid, mfi.mode, path))
2746                 return -1;
2747         return mfi.clean;
2748 }
2749
2750 /* Per entry merge function */
2751 static int process_entry(struct merge_options *o,
2752                          const char *path, struct stage_data *entry)
2753 {
2754         int clean_merge = 1;
2755         int normalize = o->renormalize;
2756         unsigned o_mode = entry->stages[1].mode;
2757         unsigned a_mode = entry->stages[2].mode;
2758         unsigned b_mode = entry->stages[3].mode;
2759         struct object_id *o_oid = stage_oid(&entry->stages[1].oid, o_mode);
2760         struct object_id *a_oid = stage_oid(&entry->stages[2].oid, a_mode);
2761         struct object_id *b_oid = stage_oid(&entry->stages[3].oid, b_mode);
2762
2763         entry->processed = 1;
2764         if (entry->rename_conflict_info) {
2765                 struct rename_conflict_info *conflict_info = entry->rename_conflict_info;
2766                 switch (conflict_info->rename_type) {
2767                 case RENAME_NORMAL:
2768                 case RENAME_ONE_FILE_TO_ONE:
2769                         clean_merge = merge_content(o, path,
2770                                                     o_oid, o_mode, a_oid, a_mode, b_oid, b_mode,
2771                                                     conflict_info);
2772                         break;
2773                 case RENAME_DIR:
2774                         clean_merge = 1;
2775                         if (conflict_rename_dir(o,
2776                                                 conflict_info->pair1,
2777                                                 conflict_info->branch1,
2778                                                 conflict_info->branch2))
2779                                 clean_merge = -1;
2780                         break;
2781                 case RENAME_DELETE:
2782                         clean_merge = 0;
2783                         if (conflict_rename_delete(o,
2784                                                    conflict_info->pair1,
2785                                                    conflict_info->branch1,
2786                                                    conflict_info->branch2))
2787                                 clean_merge = -1;
2788                         break;
2789                 case RENAME_ONE_FILE_TO_TWO:
2790                         clean_merge = 0;
2791                         if (conflict_rename_rename_1to2(o, conflict_info))
2792                                 clean_merge = -1;
2793                         break;
2794                 case RENAME_TWO_FILES_TO_ONE:
2795                         clean_merge = 0;
2796                         if (conflict_rename_rename_2to1(o, conflict_info))
2797                                 clean_merge = -1;
2798                         break;
2799                 default:
2800                         entry->processed = 0;
2801                         break;
2802                 }
2803         } else if (o_oid && (!a_oid || !b_oid)) {
2804                 /* Case A: Deleted in one */
2805                 if ((!a_oid && !b_oid) ||
2806                     (!b_oid && blob_unchanged(o, o_oid, o_mode, a_oid, a_mode, normalize, path)) ||
2807                     (!a_oid && blob_unchanged(o, o_oid, o_mode, b_oid, b_mode, normalize, path))) {
2808                         /* Deleted in both or deleted in one and
2809                          * unchanged in the other */
2810                         if (a_oid)
2811                                 output(o, 2, _("Removing %s"), path);
2812                         /* do not touch working file if it did not exist */
2813                         remove_file(o, 1, path, !a_oid);
2814                 } else {
2815                         /* Modify/delete; deleted side may have put a directory in the way */
2816                         clean_merge = 0;
2817                         if (handle_modify_delete(o, path, o_oid, o_mode,
2818                                                  a_oid, a_mode, b_oid, b_mode))
2819                                 clean_merge = -1;
2820                 }
2821         } else if ((!o_oid && a_oid && !b_oid) ||
2822                    (!o_oid && !a_oid && b_oid)) {
2823                 /* Case B: Added in one. */
2824                 /* [nothing|directory] -> ([nothing|directory], file) */
2825
2826                 const char *add_branch;
2827                 const char *other_branch;
2828                 unsigned mode;
2829                 const struct object_id *oid;
2830                 const char *conf;
2831
2832                 if (a_oid) {
2833                         add_branch = o->branch1;
2834                         other_branch = o->branch2;
2835                         mode = a_mode;
2836                         oid = a_oid;
2837                         conf = _("file/directory");
2838                 } else {
2839                         add_branch = o->branch2;
2840                         other_branch = o->branch1;
2841                         mode = b_mode;
2842                         oid = b_oid;
2843                         conf = _("directory/file");
2844                 }
2845                 if (dir_in_way(path,
2846                                !o->call_depth && !S_ISGITLINK(a_mode),
2847                                0)) {
2848                         char *new_path = unique_path(o, path, add_branch);
2849                         clean_merge = 0;
2850                         output(o, 1, _("CONFLICT (%s): There is a directory with name %s in %s. "
2851                                "Adding %s as %s"),
2852                                conf, path, other_branch, path, new_path);
2853                         if (update_file(o, 0, oid, mode, new_path))
2854                                 clean_merge = -1;
2855                         else if (o->call_depth)
2856                                 remove_file_from_cache(path);
2857                         free(new_path);
2858                 } else {
2859                         output(o, 2, _("Adding %s"), path);
2860                         /* do not overwrite file if already present */
2861                         if (update_file_flags(o, oid, mode, path, 1, !a_oid))
2862                                 clean_merge = -1;
2863                 }
2864         } else if (a_oid && b_oid) {
2865                 /* Case C: Added in both (check for same permissions) and */
2866                 /* case D: Modified in both, but differently. */
2867                 clean_merge = merge_content(o, path,
2868                                             o_oid, o_mode, a_oid, a_mode, b_oid, b_mode,
2869                                             NULL);
2870         } else if (!o_oid && !a_oid && !b_oid) {
2871                 /*
2872                  * this entry was deleted altogether. a_mode == 0 means
2873                  * we had that path and want to actively remove it.
2874                  */
2875                 remove_file(o, 1, path, !a_mode);
2876         } else
2877                 die("BUG: fatal merge failure, shouldn't happen.");
2878
2879         return clean_merge;
2880 }
2881
2882 int merge_trees(struct merge_options *o,
2883                 struct tree *head,
2884                 struct tree *merge,
2885                 struct tree *common,
2886                 struct tree **result)
2887 {
2888         int code, clean;
2889
2890         if (o->subtree_shift) {
2891                 merge = shift_tree_object(head, merge, o->subtree_shift);
2892                 common = shift_tree_object(head, common, o->subtree_shift);
2893         }
2894
2895         if (oid_eq(&common->object.oid, &merge->object.oid)) {
2896                 struct strbuf sb = STRBUF_INIT;
2897
2898                 if (!o->call_depth && index_has_changes(&sb)) {
2899                         err(o, _("Dirty index: cannot merge (dirty: %s)"),
2900                             sb.buf);
2901                         return 0;
2902                 }
2903                 output(o, 0, _("Already up to date!"));
2904                 *result = head;
2905                 return 1;
2906         }
2907
2908         code = git_merge_trees(o->call_depth, common, head, merge);
2909
2910         if (code != 0) {
2911                 if (show(o, 4) || o->call_depth)
2912                         err(o, _("merging of trees %s and %s failed"),
2913                             oid_to_hex(&head->object.oid),
2914                             oid_to_hex(&merge->object.oid));
2915                 return -1;
2916         }
2917
2918         if (unmerged_cache()) {
2919                 struct string_list *entries;
2920                 struct rename_info re_info;
2921                 int i;
2922                 /*
2923                  * Only need the hashmap while processing entries, so
2924                  * initialize it here and free it when we are done running
2925                  * through the entries. Keeping it in the merge_options as
2926                  * opposed to decaring a local hashmap is for convenience
2927                  * so that we don't have to pass it to around.
2928                  */
2929                 hashmap_init(&o->current_file_dir_set, path_hashmap_cmp, NULL, 512);
2930                 get_files_dirs(o, head);
2931                 get_files_dirs(o, merge);
2932
2933                 entries = get_unmerged();
2934                 clean = handle_renames(o, common, head, merge, entries,
2935                                        &re_info);
2936                 record_df_conflict_files(o, entries);
2937                 if (clean < 0)
2938                         goto cleanup;
2939                 for (i = entries->nr-1; 0 <= i; i--) {
2940                         const char *path = entries->items[i].string;
2941                         struct stage_data *e = entries->items[i].util;
2942                         if (!e->processed) {
2943                                 int ret = process_entry(o, path, e);
2944                                 if (!ret)
2945                                         clean = 0;
2946                                 else if (ret < 0) {
2947                                         clean = ret;
2948                                         goto cleanup;
2949                                 }
2950                         }
2951                 }
2952                 for (i = 0; i < entries->nr; i++) {
2953                         struct stage_data *e = entries->items[i].util;
2954                         if (!e->processed)
2955                                 die("BUG: unprocessed path??? %s",
2956                                     entries->items[i].string);
2957                 }
2958
2959 cleanup:
2960                 final_cleanup_renames(&re_info);
2961
2962                 string_list_clear(entries, 1);
2963                 free(entries);
2964
2965                 hashmap_free(&o->current_file_dir_set, 1);
2966
2967                 if (clean < 0)
2968                         return clean;
2969         }
2970         else
2971                 clean = 1;
2972
2973         if (o->call_depth && !(*result = write_tree_from_memory(o)))
2974                 return -1;
2975
2976         return clean;
2977 }
2978
2979 static struct commit_list *reverse_commit_list(struct commit_list *list)
2980 {
2981         struct commit_list *next = NULL, *current, *backup;
2982         for (current = list; current; current = backup) {
2983                 backup = current->next;
2984                 current->next = next;
2985                 next = current;
2986         }
2987         return next;
2988 }
2989
2990 /*
2991  * Merge the commits h1 and h2, return the resulting virtual
2992  * commit object and a flag indicating the cleanness of the merge.
2993  */
2994 int merge_recursive(struct merge_options *o,
2995                     struct commit *h1,
2996                     struct commit *h2,
2997                     struct commit_list *ca,
2998                     struct commit **result)
2999 {
3000         struct commit_list *iter;
3001         struct commit *merged_common_ancestors;
3002         struct tree *mrtree = mrtree;
3003         int clean;
3004
3005         if (show(o, 4)) {
3006                 output(o, 4, _("Merging:"));
3007                 output_commit_title(o, h1);
3008                 output_commit_title(o, h2);
3009         }
3010
3011         if (!ca) {
3012                 ca = get_merge_bases(h1, h2);
3013                 ca = reverse_commit_list(ca);
3014         }
3015
3016         if (show(o, 5)) {
3017                 unsigned cnt = commit_list_count(ca);
3018
3019                 output(o, 5, Q_("found %u common ancestor:",
3020                                 "found %u common ancestors:", cnt), cnt);
3021                 for (iter = ca; iter; iter = iter->next)
3022                         output_commit_title(o, iter->item);
3023         }
3024
3025         merged_common_ancestors = pop_commit(&ca);
3026         if (merged_common_ancestors == NULL) {
3027                 /* if there is no common ancestor, use an empty tree */
3028                 struct tree *tree;
3029
3030                 tree = lookup_tree(the_hash_algo->empty_tree);
3031                 merged_common_ancestors = make_virtual_commit(tree, "ancestor");
3032         }
3033
3034         for (iter = ca; iter; iter = iter->next) {
3035                 const char *saved_b1, *saved_b2;
3036                 o->call_depth++;
3037                 /*
3038                  * When the merge fails, the result contains files
3039                  * with conflict markers. The cleanness flag is
3040                  * ignored (unless indicating an error), it was never
3041                  * actually used, as result of merge_trees has always
3042                  * overwritten it: the committed "conflicts" were
3043                  * already resolved.
3044                  */
3045                 discard_cache();
3046                 saved_b1 = o->branch1;
3047                 saved_b2 = o->branch2;
3048                 o->branch1 = "Temporary merge branch 1";
3049                 o->branch2 = "Temporary merge branch 2";
3050                 if (merge_recursive(o, merged_common_ancestors, iter->item,
3051                                     NULL, &merged_common_ancestors) < 0)
3052                         return -1;
3053                 o->branch1 = saved_b1;
3054                 o->branch2 = saved_b2;
3055                 o->call_depth--;
3056
3057                 if (!merged_common_ancestors)
3058                         return err(o, _("merge returned no commit"));
3059         }
3060
3061         discard_cache();
3062         if (!o->call_depth)
3063                 read_cache();
3064
3065         o->ancestor = "merged common ancestors";
3066         clean = merge_trees(o, h1->tree, h2->tree, merged_common_ancestors->tree,
3067                             &mrtree);
3068         if (clean < 0) {
3069                 flush_output(o);
3070                 return clean;
3071         }
3072
3073         if (o->call_depth) {
3074                 *result = make_virtual_commit(mrtree, "merged tree");
3075                 commit_list_insert(h1, &(*result)->parents);
3076                 commit_list_insert(h2, &(*result)->parents->next);
3077         }
3078         flush_output(o);
3079         if (!o->call_depth && o->buffer_output < 2)
3080                 strbuf_release(&o->obuf);
3081         if (show(o, 2))
3082                 diff_warn_rename_limit("merge.renamelimit",
3083                                        o->needed_rename_limit, 0);
3084         return clean;
3085 }
3086
3087 static struct commit *get_ref(const struct object_id *oid, const char *name)
3088 {
3089         struct object *object;
3090
3091         object = deref_tag(parse_object(oid), name, strlen(name));
3092         if (!object)
3093                 return NULL;
3094         if (object->type == OBJ_TREE)
3095                 return make_virtual_commit((struct tree*)object, name);
3096         if (object->type != OBJ_COMMIT)
3097                 return NULL;
3098         if (parse_commit((struct commit *)object))
3099                 return NULL;
3100         return (struct commit *)object;
3101 }
3102
3103 int merge_recursive_generic(struct merge_options *o,
3104                             const struct object_id *head,
3105                             const struct object_id *merge,
3106                             int num_base_list,
3107                             const struct object_id **base_list,
3108                             struct commit **result)
3109 {
3110         int clean;
3111         struct lock_file lock = LOCK_INIT;
3112         struct commit *head_commit = get_ref(head, o->branch1);
3113         struct commit *next_commit = get_ref(merge, o->branch2);
3114         struct commit_list *ca = NULL;
3115
3116         if (base_list) {
3117                 int i;
3118                 for (i = 0; i < num_base_list; ++i) {
3119                         struct commit *base;
3120                         if (!(base = get_ref(base_list[i], oid_to_hex(base_list[i]))))
3121                                 return err(o, _("Could not parse object '%s'"),
3122                                         oid_to_hex(base_list[i]));
3123                         commit_list_insert(base, &ca);
3124                 }
3125         }
3126
3127         hold_locked_index(&lock, LOCK_DIE_ON_ERROR);
3128         clean = merge_recursive(o, head_commit, next_commit, ca,
3129                         result);
3130         if (clean < 0)
3131                 return clean;
3132
3133         if (active_cache_changed &&
3134             write_locked_index(&the_index, &lock, COMMIT_LOCK))
3135                 return err(o, _("Unable to write index."));
3136
3137         return clean ? 0 : 1;
3138 }
3139
3140 static void merge_recursive_config(struct merge_options *o)
3141 {
3142         git_config_get_int("merge.verbosity", &o->verbosity);
3143         git_config_get_int("diff.renamelimit", &o->diff_rename_limit);
3144         git_config_get_int("merge.renamelimit", &o->merge_rename_limit);
3145         git_config(git_xmerge_config, NULL);
3146 }
3147
3148 void init_merge_options(struct merge_options *o)
3149 {
3150         const char *merge_verbosity;
3151         memset(o, 0, sizeof(struct merge_options));
3152         o->verbosity = 2;
3153         o->buffer_output = 1;
3154         o->diff_rename_limit = -1;
3155         o->merge_rename_limit = -1;
3156         o->renormalize = 0;
3157         o->detect_rename = 1;
3158         merge_recursive_config(o);
3159         merge_verbosity = getenv("GIT_MERGE_VERBOSITY");
3160         if (merge_verbosity)
3161                 o->verbosity = strtol(merge_verbosity, NULL, 10);
3162         if (o->verbosity >= 5)
3163                 o->buffer_output = 0;
3164         strbuf_init(&o->obuf, 0);
3165         string_list_init(&o->df_conflict_file_set, 1);
3166 }
3167
3168 int parse_merge_opt(struct merge_options *o, const char *s)
3169 {
3170         const char *arg;
3171
3172         if (!s || !*s)
3173                 return -1;
3174         if (!strcmp(s, "ours"))
3175                 o->recursive_variant = MERGE_RECURSIVE_OURS;
3176         else if (!strcmp(s, "theirs"))
3177                 o->recursive_variant = MERGE_RECURSIVE_THEIRS;
3178         else if (!strcmp(s, "subtree"))
3179                 o->subtree_shift = "";
3180         else if (skip_prefix(s, "subtree=", &arg))
3181                 o->subtree_shift = arg;
3182         else if (!strcmp(s, "patience"))
3183                 o->xdl_opts = DIFF_WITH_ALG(o, PATIENCE_DIFF);
3184         else if (!strcmp(s, "histogram"))
3185                 o->xdl_opts = DIFF_WITH_ALG(o, HISTOGRAM_DIFF);
3186         else if (skip_prefix(s, "diff-algorithm=", &arg)) {
3187                 long value = parse_algorithm_value(arg);
3188                 if (value < 0)
3189                         return -1;
3190                 /* clear out previous settings */
3191                 DIFF_XDL_CLR(o, NEED_MINIMAL);
3192                 o->xdl_opts &= ~XDF_DIFF_ALGORITHM_MASK;
3193                 o->xdl_opts |= value;
3194         }
3195         else if (!strcmp(s, "ignore-space-change"))
3196                 DIFF_XDL_SET(o, IGNORE_WHITESPACE_CHANGE);
3197         else if (!strcmp(s, "ignore-all-space"))
3198                 DIFF_XDL_SET(o, IGNORE_WHITESPACE);
3199         else if (!strcmp(s, "ignore-space-at-eol"))
3200                 DIFF_XDL_SET(o, IGNORE_WHITESPACE_AT_EOL);
3201         else if (!strcmp(s, "ignore-cr-at-eol"))
3202                 DIFF_XDL_SET(o, IGNORE_CR_AT_EOL);
3203         else if (!strcmp(s, "renormalize"))
3204                 o->renormalize = 1;
3205         else if (!strcmp(s, "no-renormalize"))
3206                 o->renormalize = 0;
3207         else if (!strcmp(s, "no-renames"))
3208                 o->detect_rename = 0;
3209         else if (!strcmp(s, "find-renames")) {
3210                 o->detect_rename = 1;
3211                 o->rename_score = 0;
3212         }
3213         else if (skip_prefix(s, "find-renames=", &arg) ||
3214                  skip_prefix(s, "rename-threshold=", &arg)) {
3215                 if ((o->rename_score = parse_rename_score(&arg)) == -1 || *arg != 0)
3216                         return -1;
3217                 o->detect_rename = 1;
3218         }
3219         else
3220                 return -1;
3221         return 0;
3222 }