merge-recursive: avoid losing output and leaking memory holding that output
[git] / merge-recursive.c
1 /*
2  * Recursive Merge algorithm stolen from git-merge-recursive.py by
3  * Fredrik Kuivinen.
4  * The thieves were Alex Riesen and Johannes Schindelin, in June/July 2006
5  */
6 #include "cache.h"
7 #include "config.h"
8 #include "advice.h"
9 #include "lockfile.h"
10 #include "cache-tree.h"
11 #include "object-store.h"
12 #include "repository.h"
13 #include "commit.h"
14 #include "blob.h"
15 #include "builtin.h"
16 #include "tree-walk.h"
17 #include "diff.h"
18 #include "diffcore.h"
19 #include "tag.h"
20 #include "alloc.h"
21 #include "unpack-trees.h"
22 #include "string-list.h"
23 #include "xdiff-interface.h"
24 #include "ll-merge.h"
25 #include "attr.h"
26 #include "merge-recursive.h"
27 #include "dir.h"
28 #include "submodule.h"
29 #include "revision.h"
30 #include "commit-reach.h"
31
32 struct path_hashmap_entry {
33         struct hashmap_entry e;
34         char path[FLEX_ARRAY];
35 };
36
37 static int path_hashmap_cmp(const void *cmp_data,
38                             const void *entry,
39                             const void *entry_or_key,
40                             const void *keydata)
41 {
42         const struct path_hashmap_entry *a = entry;
43         const struct path_hashmap_entry *b = entry_or_key;
44         const char *key = keydata;
45
46         if (ignore_case)
47                 return strcasecmp(a->path, key ? key : b->path);
48         else
49                 return strcmp(a->path, key ? key : b->path);
50 }
51
52 static unsigned int path_hash(const char *path)
53 {
54         return ignore_case ? strihash(path) : strhash(path);
55 }
56
57 /*
58  * For dir_rename_entry, directory names are stored as a full path from the
59  * toplevel of the repository and do not include a trailing '/'.  Also:
60  *
61  *   dir:                original name of directory being renamed
62  *   non_unique_new_dir: if true, could not determine new_dir
63  *   new_dir:            final name of directory being renamed
64  *   possible_new_dirs:  temporary used to help determine new_dir; see comments
65  *                       in get_directory_renames() for details
66  */
67 struct dir_rename_entry {
68         struct hashmap_entry ent; /* must be the first member! */
69         char *dir;
70         unsigned non_unique_new_dir:1;
71         struct strbuf new_dir;
72         struct string_list possible_new_dirs;
73 };
74
75 static struct dir_rename_entry *dir_rename_find_entry(struct hashmap *hashmap,
76                                                       char *dir)
77 {
78         struct dir_rename_entry key;
79
80         if (dir == NULL)
81                 return NULL;
82         hashmap_entry_init(&key, strhash(dir));
83         key.dir = dir;
84         return hashmap_get(hashmap, &key, NULL);
85 }
86
87 static int dir_rename_cmp(const void *unused_cmp_data,
88                           const void *entry,
89                           const void *entry_or_key,
90                           const void *unused_keydata)
91 {
92         const struct dir_rename_entry *e1 = entry;
93         const struct dir_rename_entry *e2 = entry_or_key;
94
95         return strcmp(e1->dir, e2->dir);
96 }
97
98 static void dir_rename_init(struct hashmap *map)
99 {
100         hashmap_init(map, dir_rename_cmp, NULL, 0);
101 }
102
103 static void dir_rename_entry_init(struct dir_rename_entry *entry,
104                                   char *directory)
105 {
106         hashmap_entry_init(entry, strhash(directory));
107         entry->dir = directory;
108         entry->non_unique_new_dir = 0;
109         strbuf_init(&entry->new_dir, 0);
110         string_list_init(&entry->possible_new_dirs, 0);
111 }
112
113 struct collision_entry {
114         struct hashmap_entry ent; /* must be the first member! */
115         char *target_file;
116         struct string_list source_files;
117         unsigned reported_already:1;
118 };
119
120 static struct collision_entry *collision_find_entry(struct hashmap *hashmap,
121                                                     char *target_file)
122 {
123         struct collision_entry key;
124
125         hashmap_entry_init(&key, strhash(target_file));
126         key.target_file = target_file;
127         return hashmap_get(hashmap, &key, NULL);
128 }
129
130 static int collision_cmp(void *unused_cmp_data,
131                          const struct collision_entry *e1,
132                          const struct collision_entry *e2,
133                          const void *unused_keydata)
134 {
135         return strcmp(e1->target_file, e2->target_file);
136 }
137
138 static void collision_init(struct hashmap *map)
139 {
140         hashmap_init(map, (hashmap_cmp_fn) collision_cmp, NULL, 0);
141 }
142
143 static void flush_output(struct merge_options *opt)
144 {
145         if (opt->buffer_output < 2 && opt->obuf.len) {
146                 fputs(opt->obuf.buf, stdout);
147                 strbuf_reset(&opt->obuf);
148         }
149 }
150
151 static int err(struct merge_options *opt, const char *err, ...)
152 {
153         va_list params;
154
155         if (opt->buffer_output < 2)
156                 flush_output(opt);
157         else {
158                 strbuf_complete(&opt->obuf, '\n');
159                 strbuf_addstr(&opt->obuf, "error: ");
160         }
161         va_start(params, err);
162         strbuf_vaddf(&opt->obuf, err, params);
163         va_end(params);
164         if (opt->buffer_output > 1)
165                 strbuf_addch(&opt->obuf, '\n');
166         else {
167                 error("%s", opt->obuf.buf);
168                 strbuf_reset(&opt->obuf);
169         }
170
171         return -1;
172 }
173
174 static struct tree *shift_tree_object(struct repository *repo,
175                                       struct tree *one, struct tree *two,
176                                       const char *subtree_shift)
177 {
178         struct object_id shifted;
179
180         if (!*subtree_shift) {
181                 shift_tree(repo, &one->object.oid, &two->object.oid, &shifted, 0);
182         } else {
183                 shift_tree_by(repo, &one->object.oid, &two->object.oid, &shifted,
184                               subtree_shift);
185         }
186         if (oideq(&two->object.oid, &shifted))
187                 return two;
188         return lookup_tree(repo, &shifted);
189 }
190
191 static inline void set_commit_tree(struct commit *c, struct tree *t)
192 {
193         c->maybe_tree = t;
194 }
195
196 static struct commit *make_virtual_commit(struct repository *repo,
197                                           struct tree *tree,
198                                           const char *comment)
199 {
200         struct commit *commit = alloc_commit_node(repo);
201
202         set_merge_remote_desc(commit, comment, (struct object *)commit);
203         set_commit_tree(commit, tree);
204         commit->object.parsed = 1;
205         return commit;
206 }
207
208 /*
209  * Since we use get_tree_entry(), which does not put the read object into
210  * the object pool, we cannot rely on a == b.
211  */
212 static int oid_eq(const struct object_id *a, const struct object_id *b)
213 {
214         if (!a && !b)
215                 return 2;
216         return a && b && oideq(a, b);
217 }
218
219 enum rename_type {
220         RENAME_NORMAL = 0,
221         RENAME_VIA_DIR,
222         RENAME_ADD,
223         RENAME_DELETE,
224         RENAME_ONE_FILE_TO_ONE,
225         RENAME_ONE_FILE_TO_TWO,
226         RENAME_TWO_FILES_TO_ONE
227 };
228
229 /*
230  * Since we want to write the index eventually, we cannot reuse the index
231  * for these (temporary) data.
232  */
233 struct stage_data {
234         struct diff_filespec stages[4]; /* mostly for oid & mode; maybe path */
235         struct rename_conflict_info *rename_conflict_info;
236         unsigned processed:1;
237 };
238
239 struct rename {
240         unsigned processed:1;
241         struct diff_filepair *pair;
242         const char *branch; /* branch that the rename occurred on */
243         /*
244          * If directory rename detection affected this rename, what was its
245          * original type ('A' or 'R') and it's original destination before
246          * the directory rename (otherwise, '\0' and NULL for these two vars).
247          */
248         char dir_rename_original_type;
249         char *dir_rename_original_dest;
250         /*
251          * Purpose of src_entry and dst_entry:
252          *
253          * If 'before' is renamed to 'after' then src_entry will contain
254          * the versions of 'before' from the merge_base, HEAD, and MERGE in
255          * stages 1, 2, and 3; dst_entry will contain the respective
256          * versions of 'after' in corresponding locations.  Thus, we have a
257          * total of six modes and oids, though some will be null.  (Stage 0
258          * is ignored; we're interested in handling conflicts.)
259          *
260          * Since we don't turn on break-rewrites by default, neither
261          * src_entry nor dst_entry can have all three of their stages have
262          * non-null oids, meaning at most four of the six will be non-null.
263          * Also, since this is a rename, both src_entry and dst_entry will
264          * have at least one non-null oid, meaning at least two will be
265          * non-null.  Of the six oids, a typical rename will have three be
266          * non-null.  Only two implies a rename/delete, and four implies a
267          * rename/add.
268          */
269         struct stage_data *src_entry;
270         struct stage_data *dst_entry;
271 };
272
273 struct rename_conflict_info {
274         enum rename_type rename_type;
275         struct rename *ren1;
276         struct rename *ren2;
277 };
278
279 static inline void setup_rename_conflict_info(enum rename_type rename_type,
280                                               struct merge_options *opt,
281                                               struct rename *ren1,
282                                               struct rename *ren2)
283 {
284         struct rename_conflict_info *ci;
285
286         /*
287          * When we have two renames involved, it's easiest to get the
288          * correct things into stage 2 and 3, and to make sure that the
289          * content merge puts HEAD before the other branch if we just
290          * ensure that branch1 == opt->branch1.  So, simply flip arguments
291          * around if we don't have that.
292          */
293         if (ren2 && ren1->branch != opt->branch1) {
294                 setup_rename_conflict_info(rename_type, opt, ren2, ren1);
295                 return;
296         }
297
298         ci = xcalloc(1, sizeof(struct rename_conflict_info));
299         ci->rename_type = rename_type;
300         ci->ren1 = ren1;
301         ci->ren2 = ren2;
302
303         ci->ren1->dst_entry->processed = 0;
304         ci->ren1->dst_entry->rename_conflict_info = ci;
305         if (ren2) {
306                 ci->ren2->dst_entry->rename_conflict_info = ci;
307         }
308 }
309
310 static int show(struct merge_options *opt, int v)
311 {
312         return (!opt->call_depth && opt->verbosity >= v) || opt->verbosity >= 5;
313 }
314
315 __attribute__((format (printf, 3, 4)))
316 static void output(struct merge_options *opt, int v, const char *fmt, ...)
317 {
318         va_list ap;
319
320         if (!show(opt, v))
321                 return;
322
323         strbuf_addchars(&opt->obuf, ' ', opt->call_depth * 2);
324
325         va_start(ap, fmt);
326         strbuf_vaddf(&opt->obuf, fmt, ap);
327         va_end(ap);
328
329         strbuf_addch(&opt->obuf, '\n');
330         if (!opt->buffer_output)
331                 flush_output(opt);
332 }
333
334 static void output_commit_title(struct merge_options *opt, struct commit *commit)
335 {
336         struct merge_remote_desc *desc;
337
338         strbuf_addchars(&opt->obuf, ' ', opt->call_depth * 2);
339         desc = merge_remote_util(commit);
340         if (desc)
341                 strbuf_addf(&opt->obuf, "virtual %s\n", desc->name);
342         else {
343                 strbuf_add_unique_abbrev(&opt->obuf, &commit->object.oid,
344                                          DEFAULT_ABBREV);
345                 strbuf_addch(&opt->obuf, ' ');
346                 if (parse_commit(commit) != 0)
347                         strbuf_addstr(&opt->obuf, _("(bad commit)\n"));
348                 else {
349                         const char *title;
350                         const char *msg = get_commit_buffer(commit, NULL);
351                         int len = find_commit_subject(msg, &title);
352                         if (len)
353                                 strbuf_addf(&opt->obuf, "%.*s\n", len, title);
354                         unuse_commit_buffer(commit, msg);
355                 }
356         }
357         flush_output(opt);
358 }
359
360 static int add_cacheinfo(struct merge_options *opt,
361                          const struct diff_filespec *blob,
362                          const char *path, int stage, int refresh, int options)
363 {
364         struct index_state *istate = opt->repo->index;
365         struct cache_entry *ce;
366         int ret;
367
368         ce = make_cache_entry(istate, blob->mode, &blob->oid, path, stage, 0);
369         if (!ce)
370                 return err(opt, _("add_cacheinfo failed for path '%s'; merge aborting."), path);
371
372         ret = add_index_entry(istate, ce, options);
373         if (refresh) {
374                 struct cache_entry *nce;
375
376                 nce = refresh_cache_entry(istate, ce,
377                                           CE_MATCH_REFRESH | CE_MATCH_IGNORE_MISSING);
378                 if (!nce)
379                         return err(opt, _("add_cacheinfo failed to refresh for path '%s'; merge aborting."), path);
380                 if (nce != ce)
381                         ret = add_index_entry(istate, nce, options);
382         }
383         return ret;
384 }
385
386 static inline int merge_detect_rename(struct merge_options *opt)
387 {
388         return (opt->detect_renames >= 0) ? opt->detect_renames : 1;
389 }
390
391 static void init_tree_desc_from_tree(struct tree_desc *desc, struct tree *tree)
392 {
393         parse_tree(tree);
394         init_tree_desc(desc, tree->buffer, tree->size);
395 }
396
397 static int unpack_trees_start(struct merge_options *opt,
398                               struct tree *common,
399                               struct tree *head,
400                               struct tree *merge)
401 {
402         int rc;
403         struct tree_desc t[3];
404         struct index_state tmp_index = { NULL };
405
406         memset(&opt->unpack_opts, 0, sizeof(opt->unpack_opts));
407         if (opt->call_depth)
408                 opt->unpack_opts.index_only = 1;
409         else
410                 opt->unpack_opts.update = 1;
411         opt->unpack_opts.merge = 1;
412         opt->unpack_opts.head_idx = 2;
413         opt->unpack_opts.fn = threeway_merge;
414         opt->unpack_opts.src_index = opt->repo->index;
415         opt->unpack_opts.dst_index = &tmp_index;
416         opt->unpack_opts.aggressive = !merge_detect_rename(opt);
417         setup_unpack_trees_porcelain(&opt->unpack_opts, "merge");
418
419         init_tree_desc_from_tree(t+0, common);
420         init_tree_desc_from_tree(t+1, head);
421         init_tree_desc_from_tree(t+2, merge);
422
423         rc = unpack_trees(3, t, &opt->unpack_opts);
424         cache_tree_free(&opt->repo->index->cache_tree);
425
426         /*
427          * Update opt->repo->index to match the new results, AFTER saving a copy
428          * in opt->orig_index.  Update src_index to point to the saved copy.
429          * (verify_uptodate() checks src_index, and the original index is
430          * the one that had the necessary modification timestamps.)
431          */
432         opt->orig_index = *opt->repo->index;
433         *opt->repo->index = tmp_index;
434         opt->unpack_opts.src_index = &opt->orig_index;
435
436         return rc;
437 }
438
439 static void unpack_trees_finish(struct merge_options *opt)
440 {
441         discard_index(&opt->orig_index);
442         clear_unpack_trees_porcelain(&opt->unpack_opts);
443 }
444
445 static int save_files_dirs(const struct object_id *oid,
446                            struct strbuf *base, const char *path,
447                            unsigned int mode, int stage, void *context)
448 {
449         struct path_hashmap_entry *entry;
450         int baselen = base->len;
451         struct merge_options *opt = context;
452
453         strbuf_addstr(base, path);
454
455         FLEX_ALLOC_MEM(entry, path, base->buf, base->len);
456         hashmap_entry_init(entry, path_hash(entry->path));
457         hashmap_add(&opt->current_file_dir_set, entry);
458
459         strbuf_setlen(base, baselen);
460         return (S_ISDIR(mode) ? READ_TREE_RECURSIVE : 0);
461 }
462
463 static void get_files_dirs(struct merge_options *opt, struct tree *tree)
464 {
465         struct pathspec match_all;
466         memset(&match_all, 0, sizeof(match_all));
467         read_tree_recursive(opt->repo, tree, "", 0, 0,
468                             &match_all, save_files_dirs, opt);
469 }
470
471 static int get_tree_entry_if_blob(struct repository *r,
472                                   const struct object_id *tree,
473                                   const char *path,
474                                   struct diff_filespec *dfs)
475 {
476         int ret;
477
478         ret = get_tree_entry(r, tree, path, &dfs->oid, &dfs->mode);
479         if (S_ISDIR(dfs->mode)) {
480                 oidcpy(&dfs->oid, &null_oid);
481                 dfs->mode = 0;
482         }
483         return ret;
484 }
485
486 /*
487  * Returns an index_entry instance which doesn't have to correspond to
488  * a real cache entry in Git's index.
489  */
490 static struct stage_data *insert_stage_data(struct repository *r,
491                 const char *path,
492                 struct tree *o, struct tree *a, struct tree *b,
493                 struct string_list *entries)
494 {
495         struct string_list_item *item;
496         struct stage_data *e = xcalloc(1, sizeof(struct stage_data));
497         get_tree_entry_if_blob(r, &o->object.oid, path, &e->stages[1]);
498         get_tree_entry_if_blob(r, &a->object.oid, path, &e->stages[2]);
499         get_tree_entry_if_blob(r, &b->object.oid, path, &e->stages[3]);
500         item = string_list_insert(entries, path);
501         item->util = e;
502         return e;
503 }
504
505 /*
506  * Create a dictionary mapping file names to stage_data objects. The
507  * dictionary contains one entry for every path with a non-zero stage entry.
508  */
509 static struct string_list *get_unmerged(struct index_state *istate)
510 {
511         struct string_list *unmerged = xcalloc(1, sizeof(struct string_list));
512         int i;
513
514         unmerged->strdup_strings = 1;
515
516         for (i = 0; i < istate->cache_nr; i++) {
517                 struct string_list_item *item;
518                 struct stage_data *e;
519                 const struct cache_entry *ce = istate->cache[i];
520                 if (!ce_stage(ce))
521                         continue;
522
523                 item = string_list_lookup(unmerged, ce->name);
524                 if (!item) {
525                         item = string_list_insert(unmerged, ce->name);
526                         item->util = xcalloc(1, sizeof(struct stage_data));
527                 }
528                 e = item->util;
529                 e->stages[ce_stage(ce)].mode = ce->ce_mode;
530                 oidcpy(&e->stages[ce_stage(ce)].oid, &ce->oid);
531         }
532
533         return unmerged;
534 }
535
536 static int string_list_df_name_compare(const char *one, const char *two)
537 {
538         int onelen = strlen(one);
539         int twolen = strlen(two);
540         /*
541          * Here we only care that entries for D/F conflicts are
542          * adjacent, in particular with the file of the D/F conflict
543          * appearing before files below the corresponding directory.
544          * The order of the rest of the list is irrelevant for us.
545          *
546          * To achieve this, we sort with df_name_compare and provide
547          * the mode S_IFDIR so that D/F conflicts will sort correctly.
548          * We use the mode S_IFDIR for everything else for simplicity,
549          * since in other cases any changes in their order due to
550          * sorting cause no problems for us.
551          */
552         int cmp = df_name_compare(one, onelen, S_IFDIR,
553                                   two, twolen, S_IFDIR);
554         /*
555          * Now that 'foo' and 'foo/bar' compare equal, we have to make sure
556          * that 'foo' comes before 'foo/bar'.
557          */
558         if (cmp)
559                 return cmp;
560         return onelen - twolen;
561 }
562
563 static void record_df_conflict_files(struct merge_options *opt,
564                                      struct string_list *entries)
565 {
566         /* If there is a D/F conflict and the file for such a conflict
567          * currently exists in the working tree, we want to allow it to be
568          * removed to make room for the corresponding directory if needed.
569          * The files underneath the directories of such D/F conflicts will
570          * be processed before the corresponding file involved in the D/F
571          * conflict.  If the D/F directory ends up being removed by the
572          * merge, then we won't have to touch the D/F file.  If the D/F
573          * directory needs to be written to the working copy, then the D/F
574          * file will simply be removed (in make_room_for_path()) to make
575          * room for the necessary paths.  Note that if both the directory
576          * and the file need to be present, then the D/F file will be
577          * reinstated with a new unique name at the time it is processed.
578          */
579         struct string_list df_sorted_entries = STRING_LIST_INIT_NODUP;
580         const char *last_file = NULL;
581         int last_len = 0;
582         int i;
583
584         /*
585          * If we're merging merge-bases, we don't want to bother with
586          * any working directory changes.
587          */
588         if (opt->call_depth)
589                 return;
590
591         /* Ensure D/F conflicts are adjacent in the entries list. */
592         for (i = 0; i < entries->nr; i++) {
593                 struct string_list_item *next = &entries->items[i];
594                 string_list_append(&df_sorted_entries, next->string)->util =
595                                    next->util;
596         }
597         df_sorted_entries.cmp = string_list_df_name_compare;
598         string_list_sort(&df_sorted_entries);
599
600         string_list_clear(&opt->df_conflict_file_set, 1);
601         for (i = 0; i < df_sorted_entries.nr; i++) {
602                 const char *path = df_sorted_entries.items[i].string;
603                 int len = strlen(path);
604                 struct stage_data *e = df_sorted_entries.items[i].util;
605
606                 /*
607                  * Check if last_file & path correspond to a D/F conflict;
608                  * i.e. whether path is last_file+'/'+<something>.
609                  * If so, record that it's okay to remove last_file to make
610                  * room for path and friends if needed.
611                  */
612                 if (last_file &&
613                     len > last_len &&
614                     memcmp(path, last_file, last_len) == 0 &&
615                     path[last_len] == '/') {
616                         string_list_insert(&opt->df_conflict_file_set, last_file);
617                 }
618
619                 /*
620                  * Determine whether path could exist as a file in the
621                  * working directory as a possible D/F conflict.  This
622                  * will only occur when it exists in stage 2 as a
623                  * file.
624                  */
625                 if (S_ISREG(e->stages[2].mode) || S_ISLNK(e->stages[2].mode)) {
626                         last_file = path;
627                         last_len = len;
628                 } else {
629                         last_file = NULL;
630                 }
631         }
632         string_list_clear(&df_sorted_entries, 0);
633 }
634
635 static int update_stages(struct merge_options *opt, const char *path,
636                          const struct diff_filespec *o,
637                          const struct diff_filespec *a,
638                          const struct diff_filespec *b)
639 {
640
641         /*
642          * NOTE: It is usually a bad idea to call update_stages on a path
643          * before calling update_file on that same path, since it can
644          * sometimes lead to spurious "refusing to lose untracked file..."
645          * messages from update_file (via make_room_for path via
646          * would_lose_untracked).  Instead, reverse the order of the calls
647          * (executing update_file first and then update_stages).
648          */
649         int clear = 1;
650         int options = ADD_CACHE_OK_TO_ADD | ADD_CACHE_SKIP_DFCHECK;
651         if (clear)
652                 if (remove_file_from_index(opt->repo->index, path))
653                         return -1;
654         if (o)
655                 if (add_cacheinfo(opt, o, path, 1, 0, options))
656                         return -1;
657         if (a)
658                 if (add_cacheinfo(opt, a, path, 2, 0, options))
659                         return -1;
660         if (b)
661                 if (add_cacheinfo(opt, b, path, 3, 0, options))
662                         return -1;
663         return 0;
664 }
665
666 static void update_entry(struct stage_data *entry,
667                          struct diff_filespec *o,
668                          struct diff_filespec *a,
669                          struct diff_filespec *b)
670 {
671         entry->processed = 0;
672         entry->stages[1].mode = o->mode;
673         entry->stages[2].mode = a->mode;
674         entry->stages[3].mode = b->mode;
675         oidcpy(&entry->stages[1].oid, &o->oid);
676         oidcpy(&entry->stages[2].oid, &a->oid);
677         oidcpy(&entry->stages[3].oid, &b->oid);
678 }
679
680 static int remove_file(struct merge_options *opt, int clean,
681                        const char *path, int no_wd)
682 {
683         int update_cache = opt->call_depth || clean;
684         int update_working_directory = !opt->call_depth && !no_wd;
685
686         if (update_cache) {
687                 if (remove_file_from_index(opt->repo->index, path))
688                         return -1;
689         }
690         if (update_working_directory) {
691                 if (ignore_case) {
692                         struct cache_entry *ce;
693                         ce = index_file_exists(opt->repo->index, path, strlen(path),
694                                                ignore_case);
695                         if (ce && ce_stage(ce) == 0 && strcmp(path, ce->name))
696                                 return 0;
697                 }
698                 if (remove_path(path))
699                         return -1;
700         }
701         return 0;
702 }
703
704 /* add a string to a strbuf, but converting "/" to "_" */
705 static void add_flattened_path(struct strbuf *out, const char *s)
706 {
707         size_t i = out->len;
708         strbuf_addstr(out, s);
709         for (; i < out->len; i++)
710                 if (out->buf[i] == '/')
711                         out->buf[i] = '_';
712 }
713
714 static char *unique_path(struct merge_options *opt,
715                          const char *path,
716                          const char *branch)
717 {
718         struct path_hashmap_entry *entry;
719         struct strbuf newpath = STRBUF_INIT;
720         int suffix = 0;
721         size_t base_len;
722
723         strbuf_addf(&newpath, "%s~", path);
724         add_flattened_path(&newpath, branch);
725
726         base_len = newpath.len;
727         while (hashmap_get_from_hash(&opt->current_file_dir_set,
728                                      path_hash(newpath.buf), newpath.buf) ||
729                (!opt->call_depth && file_exists(newpath.buf))) {
730                 strbuf_setlen(&newpath, base_len);
731                 strbuf_addf(&newpath, "_%d", suffix++);
732         }
733
734         FLEX_ALLOC_MEM(entry, path, newpath.buf, newpath.len);
735         hashmap_entry_init(entry, path_hash(entry->path));
736         hashmap_add(&opt->current_file_dir_set, entry);
737         return strbuf_detach(&newpath, NULL);
738 }
739
740 /**
741  * Check whether a directory in the index is in the way of an incoming
742  * file.  Return 1 if so.  If check_working_copy is non-zero, also
743  * check the working directory.  If empty_ok is non-zero, also return
744  * 0 in the case where the working-tree dir exists but is empty.
745  */
746 static int dir_in_way(struct index_state *istate, const char *path,
747                       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 = index_name_pos(istate, dirpath.buf, dirpath.len);
757
758         if (pos < 0)
759                 pos = -1 - pos;
760         if (pos < istate->cache_nr &&
761             !strncmp(dirpath.buf, istate->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 /*
772  * Returns whether path was tracked in the index before the merge started,
773  * and its oid and mode match the specified values
774  */
775 static int was_tracked_and_matches(struct merge_options *opt, const char *path,
776                                    const struct diff_filespec *blob)
777 {
778         int pos = index_name_pos(&opt->orig_index, path, strlen(path));
779         struct cache_entry *ce;
780
781         if (0 > pos)
782                 /* we were not tracking this path before the merge */
783                 return 0;
784
785         /* See if the file we were tracking before matches */
786         ce = opt->orig_index.cache[pos];
787         return (oid_eq(&ce->oid, &blob->oid) && ce->ce_mode == blob->mode);
788 }
789
790 /*
791  * Returns whether path was tracked in the index before the merge started
792  */
793 static int was_tracked(struct merge_options *opt, const char *path)
794 {
795         int pos = index_name_pos(&opt->orig_index, path, strlen(path));
796
797         if (0 <= pos)
798                 /* we were tracking this path before the merge */
799                 return 1;
800
801         return 0;
802 }
803
804 static int would_lose_untracked(struct merge_options *opt, const char *path)
805 {
806         struct index_state *istate = opt->repo->index;
807
808         /*
809          * This may look like it can be simplified to:
810          *   return !was_tracked(opt, path) && file_exists(path)
811          * but it can't.  This function needs to know whether path was in
812          * the working tree due to EITHER having been tracked in the index
813          * before the merge OR having been put into the working copy and
814          * index by unpack_trees().  Due to that either-or requirement, we
815          * check the current index instead of the original one.
816          *
817          * Note that we do not need to worry about merge-recursive itself
818          * updating the index after unpack_trees() and before calling this
819          * function, because we strictly require all code paths in
820          * merge-recursive to update the working tree first and the index
821          * second.  Doing otherwise would break
822          * update_file()/would_lose_untracked(); see every comment in this
823          * file which mentions "update_stages".
824          */
825         int pos = index_name_pos(istate, path, strlen(path));
826
827         if (pos < 0)
828                 pos = -1 - pos;
829         while (pos < istate->cache_nr &&
830                !strcmp(path, istate->cache[pos]->name)) {
831                 /*
832                  * If stage #0, it is definitely tracked.
833                  * If it has stage #2 then it was tracked
834                  * before this merge started.  All other
835                  * cases the path was not tracked.
836                  */
837                 switch (ce_stage(istate->cache[pos])) {
838                 case 0:
839                 case 2:
840                         return 0;
841                 }
842                 pos++;
843         }
844         return file_exists(path);
845 }
846
847 static int was_dirty(struct merge_options *opt, const char *path)
848 {
849         struct cache_entry *ce;
850         int dirty = 1;
851
852         if (opt->call_depth || !was_tracked(opt, path))
853                 return !dirty;
854
855         ce = index_file_exists(opt->unpack_opts.src_index,
856                                path, strlen(path), ignore_case);
857         dirty = verify_uptodate(ce, &opt->unpack_opts) != 0;
858         return dirty;
859 }
860
861 static int make_room_for_path(struct merge_options *opt, const char *path)
862 {
863         int status, i;
864         const char *msg = _("failed to create path '%s'%s");
865
866         /* Unlink any D/F conflict files that are in the way */
867         for (i = 0; i < opt->df_conflict_file_set.nr; i++) {
868                 const char *df_path = opt->df_conflict_file_set.items[i].string;
869                 size_t pathlen = strlen(path);
870                 size_t df_pathlen = strlen(df_path);
871                 if (df_pathlen < pathlen &&
872                     path[df_pathlen] == '/' &&
873                     strncmp(path, df_path, df_pathlen) == 0) {
874                         output(opt, 3,
875                                _("Removing %s to make room for subdirectory\n"),
876                                df_path);
877                         unlink(df_path);
878                         unsorted_string_list_delete_item(&opt->df_conflict_file_set,
879                                                          i, 0);
880                         break;
881                 }
882         }
883
884         /* Make sure leading directories are created */
885         status = safe_create_leading_directories_const(path);
886         if (status) {
887                 if (status == SCLD_EXISTS)
888                         /* something else exists */
889                         return err(opt, msg, path, _(": perhaps a D/F conflict?"));
890                 return err(opt, msg, path, "");
891         }
892
893         /*
894          * Do not unlink a file in the work tree if we are not
895          * tracking it.
896          */
897         if (would_lose_untracked(opt, path))
898                 return err(opt, _("refusing to lose untracked file at '%s'"),
899                            path);
900
901         /* Successful unlink is good.. */
902         if (!unlink(path))
903                 return 0;
904         /* .. and so is no existing file */
905         if (errno == ENOENT)
906                 return 0;
907         /* .. but not some other error (who really cares what?) */
908         return err(opt, msg, path, _(": perhaps a D/F conflict?"));
909 }
910
911 static int update_file_flags(struct merge_options *opt,
912                              const struct diff_filespec *contents,
913                              const char *path,
914                              int update_cache,
915                              int update_wd)
916 {
917         int ret = 0;
918
919         if (opt->call_depth)
920                 update_wd = 0;
921
922         if (update_wd) {
923                 enum object_type type;
924                 void *buf;
925                 unsigned long size;
926
927                 if (S_ISGITLINK(contents->mode)) {
928                         /*
929                          * We may later decide to recursively descend into
930                          * the submodule directory and update its index
931                          * and/or work tree, but we do not do that now.
932                          */
933                         update_wd = 0;
934                         goto update_index;
935                 }
936
937                 buf = read_object_file(&contents->oid, &type, &size);
938                 if (!buf) {
939                         ret = err(opt, _("cannot read object %s '%s'"),
940                                   oid_to_hex(&contents->oid), path);
941                         goto free_buf;
942                 }
943                 if (type != OBJ_BLOB) {
944                         ret = err(opt, _("blob expected for %s '%s'"),
945                                   oid_to_hex(&contents->oid), path);
946                         goto free_buf;
947                 }
948                 if (S_ISREG(contents->mode)) {
949                         struct strbuf strbuf = STRBUF_INIT;
950                         if (convert_to_working_tree(opt->repo->index,
951                                                     path, buf, size, &strbuf)) {
952                                 free(buf);
953                                 size = strbuf.len;
954                                 buf = strbuf_detach(&strbuf, NULL);
955                         }
956                 }
957
958                 if (make_room_for_path(opt, path) < 0) {
959                         update_wd = 0;
960                         goto free_buf;
961                 }
962                 if (S_ISREG(contents->mode) ||
963                     (!has_symlinks && S_ISLNK(contents->mode))) {
964                         int fd;
965                         int mode = (contents->mode & 0100 ? 0777 : 0666);
966
967                         fd = open(path, O_WRONLY | O_TRUNC | O_CREAT, mode);
968                         if (fd < 0) {
969                                 ret = err(opt, _("failed to open '%s': %s"),
970                                           path, strerror(errno));
971                                 goto free_buf;
972                         }
973                         write_in_full(fd, buf, size);
974                         close(fd);
975                 } else if (S_ISLNK(contents->mode)) {
976                         char *lnk = xmemdupz(buf, size);
977                         safe_create_leading_directories_const(path);
978                         unlink(path);
979                         if (symlink(lnk, path))
980                                 ret = err(opt, _("failed to symlink '%s': %s"),
981                                           path, strerror(errno));
982                         free(lnk);
983                 } else
984                         ret = err(opt,
985                                   _("do not know what to do with %06o %s '%s'"),
986                                   contents->mode, oid_to_hex(&contents->oid), path);
987         free_buf:
988                 free(buf);
989         }
990 update_index:
991         if (!ret && update_cache)
992                 if (add_cacheinfo(opt, contents, path, 0, update_wd,
993                                   ADD_CACHE_OK_TO_ADD))
994                         return -1;
995         return ret;
996 }
997
998 static int update_file(struct merge_options *opt,
999                        int clean,
1000                        const struct diff_filespec *contents,
1001                        const char *path)
1002 {
1003         return update_file_flags(opt, contents, path,
1004                                  opt->call_depth || clean, !opt->call_depth);
1005 }
1006
1007 /* Low level file merging, update and removal */
1008
1009 struct merge_file_info {
1010         struct diff_filespec blob; /* mostly use oid & mode; sometimes path */
1011         unsigned clean:1,
1012                  merge:1;
1013 };
1014
1015 static int merge_3way(struct merge_options *opt,
1016                       mmbuffer_t *result_buf,
1017                       const struct diff_filespec *o,
1018                       const struct diff_filespec *a,
1019                       const struct diff_filespec *b,
1020                       const char *branch1,
1021                       const char *branch2,
1022                       const int extra_marker_size)
1023 {
1024         mmfile_t orig, src1, src2;
1025         struct ll_merge_options ll_opts = {0};
1026         char *base, *name1, *name2;
1027         int merge_status;
1028
1029         ll_opts.renormalize = opt->renormalize;
1030         ll_opts.extra_marker_size = extra_marker_size;
1031         ll_opts.xdl_opts = opt->xdl_opts;
1032
1033         if (opt->call_depth) {
1034                 ll_opts.virtual_ancestor = 1;
1035                 ll_opts.variant = 0;
1036         } else {
1037                 switch (opt->recursive_variant) {
1038                 case MERGE_RECURSIVE_OURS:
1039                         ll_opts.variant = XDL_MERGE_FAVOR_OURS;
1040                         break;
1041                 case MERGE_RECURSIVE_THEIRS:
1042                         ll_opts.variant = XDL_MERGE_FAVOR_THEIRS;
1043                         break;
1044                 default:
1045                         ll_opts.variant = 0;
1046                         break;
1047                 }
1048         }
1049
1050         assert(a->path && b->path && o->path && opt->ancestor);
1051         if (strcmp(a->path, b->path) || strcmp(a->path, o->path) != 0) {
1052                 base  = mkpathdup("%s:%s", opt->ancestor, o->path);
1053                 name1 = mkpathdup("%s:%s", branch1, a->path);
1054                 name2 = mkpathdup("%s:%s", branch2, b->path);
1055         } else {
1056                 base  = mkpathdup("%s", opt->ancestor);
1057                 name1 = mkpathdup("%s", branch1);
1058                 name2 = mkpathdup("%s", branch2);
1059         }
1060
1061         read_mmblob(&orig, &o->oid);
1062         read_mmblob(&src1, &a->oid);
1063         read_mmblob(&src2, &b->oid);
1064
1065         merge_status = ll_merge(result_buf, a->path, &orig, base,
1066                                 &src1, name1, &src2, name2,
1067                                 opt->repo->index, &ll_opts);
1068
1069         free(base);
1070         free(name1);
1071         free(name2);
1072         free(orig.ptr);
1073         free(src1.ptr);
1074         free(src2.ptr);
1075         return merge_status;
1076 }
1077
1078 static int find_first_merges(struct repository *repo,
1079                              struct object_array *result, const char *path,
1080                              struct commit *a, struct commit *b)
1081 {
1082         int i, j;
1083         struct object_array merges = OBJECT_ARRAY_INIT;
1084         struct commit *commit;
1085         int contains_another;
1086
1087         char merged_revision[GIT_MAX_HEXSZ + 2];
1088         const char *rev_args[] = { "rev-list", "--merges", "--ancestry-path",
1089                                    "--all", merged_revision, NULL };
1090         struct rev_info revs;
1091         struct setup_revision_opt rev_opts;
1092
1093         memset(result, 0, sizeof(struct object_array));
1094         memset(&rev_opts, 0, sizeof(rev_opts));
1095
1096         /* get all revisions that merge commit a */
1097         xsnprintf(merged_revision, sizeof(merged_revision), "^%s",
1098                   oid_to_hex(&a->object.oid));
1099         repo_init_revisions(repo, &revs, NULL);
1100         rev_opts.submodule = path;
1101         /* FIXME: can't handle linked worktrees in submodules yet */
1102         revs.single_worktree = path != NULL;
1103         setup_revisions(ARRAY_SIZE(rev_args)-1, rev_args, &revs, &rev_opts);
1104
1105         /* save all revisions from the above list that contain b */
1106         if (prepare_revision_walk(&revs))
1107                 die("revision walk setup failed");
1108         while ((commit = get_revision(&revs)) != NULL) {
1109                 struct object *o = &(commit->object);
1110                 if (in_merge_bases(b, commit))
1111                         add_object_array(o, NULL, &merges);
1112         }
1113         reset_revision_walk();
1114
1115         /* Now we've got all merges that contain a and b. Prune all
1116          * merges that contain another found merge and save them in
1117          * result.
1118          */
1119         for (i = 0; i < merges.nr; i++) {
1120                 struct commit *m1 = (struct commit *) merges.objects[i].item;
1121
1122                 contains_another = 0;
1123                 for (j = 0; j < merges.nr; j++) {
1124                         struct commit *m2 = (struct commit *) merges.objects[j].item;
1125                         if (i != j && in_merge_bases(m2, m1)) {
1126                                 contains_another = 1;
1127                                 break;
1128                         }
1129                 }
1130
1131                 if (!contains_another)
1132                         add_object_array(merges.objects[i].item, NULL, result);
1133         }
1134
1135         object_array_clear(&merges);
1136         return result->nr;
1137 }
1138
1139 static void print_commit(struct commit *commit)
1140 {
1141         struct strbuf sb = STRBUF_INIT;
1142         struct pretty_print_context ctx = {0};
1143         ctx.date_mode.type = DATE_NORMAL;
1144         format_commit_message(commit, " %h: %m %s", &sb, &ctx);
1145         fprintf(stderr, "%s\n", sb.buf);
1146         strbuf_release(&sb);
1147 }
1148
1149 static int is_valid(const struct diff_filespec *dfs)
1150 {
1151         return dfs->mode != 0 && !is_null_oid(&dfs->oid);
1152 }
1153
1154 static int merge_submodule(struct merge_options *opt,
1155                            struct object_id *result, const char *path,
1156                            const struct object_id *base, const struct object_id *a,
1157                            const struct object_id *b)
1158 {
1159         struct commit *commit_base, *commit_a, *commit_b;
1160         int parent_count;
1161         struct object_array merges;
1162
1163         int i;
1164         int search = !opt->call_depth;
1165
1166         /* store a in result in case we fail */
1167         oidcpy(result, a);
1168
1169         /* we can not handle deletion conflicts */
1170         if (is_null_oid(base))
1171                 return 0;
1172         if (is_null_oid(a))
1173                 return 0;
1174         if (is_null_oid(b))
1175                 return 0;
1176
1177         if (add_submodule_odb(path)) {
1178                 output(opt, 1, _("Failed to merge submodule %s (not checked out)"), path);
1179                 return 0;
1180         }
1181
1182         if (!(commit_base = lookup_commit_reference(opt->repo, base)) ||
1183             !(commit_a = lookup_commit_reference(opt->repo, a)) ||
1184             !(commit_b = lookup_commit_reference(opt->repo, b))) {
1185                 output(opt, 1, _("Failed to merge submodule %s (commits not present)"), path);
1186                 return 0;
1187         }
1188
1189         /* check whether both changes are forward */
1190         if (!in_merge_bases(commit_base, commit_a) ||
1191             !in_merge_bases(commit_base, commit_b)) {
1192                 output(opt, 1, _("Failed to merge submodule %s (commits don't follow merge-base)"), path);
1193                 return 0;
1194         }
1195
1196         /* Case #1: a is contained in b or vice versa */
1197         if (in_merge_bases(commit_a, commit_b)) {
1198                 oidcpy(result, b);
1199                 if (show(opt, 3)) {
1200                         output(opt, 3, _("Fast-forwarding submodule %s to the following commit:"), path);
1201                         output_commit_title(opt, commit_b);
1202                 } else if (show(opt, 2))
1203                         output(opt, 2, _("Fast-forwarding submodule %s"), path);
1204                 else
1205                         ; /* no output */
1206
1207                 return 1;
1208         }
1209         if (in_merge_bases(commit_b, commit_a)) {
1210                 oidcpy(result, a);
1211                 if (show(opt, 3)) {
1212                         output(opt, 3, _("Fast-forwarding submodule %s to the following commit:"), path);
1213                         output_commit_title(opt, commit_a);
1214                 } else if (show(opt, 2))
1215                         output(opt, 2, _("Fast-forwarding submodule %s"), path);
1216                 else
1217                         ; /* no output */
1218
1219                 return 1;
1220         }
1221
1222         /*
1223          * Case #2: There are one or more merges that contain a and b in
1224          * the submodule. If there is only one, then present it as a
1225          * suggestion to the user, but leave it marked unmerged so the
1226          * user needs to confirm the resolution.
1227          */
1228
1229         /* Skip the search if makes no sense to the calling context.  */
1230         if (!search)
1231                 return 0;
1232
1233         /* find commit which merges them */
1234         parent_count = find_first_merges(opt->repo, &merges, path,
1235                                          commit_a, commit_b);
1236         switch (parent_count) {
1237         case 0:
1238                 output(opt, 1, _("Failed to merge submodule %s (merge following commits not found)"), path);
1239                 break;
1240
1241         case 1:
1242                 output(opt, 1, _("Failed to merge submodule %s (not fast-forward)"), path);
1243                 output(opt, 2, _("Found a possible merge resolution for the submodule:\n"));
1244                 print_commit((struct commit *) merges.objects[0].item);
1245                 output(opt, 2, _(
1246                        "If this is correct simply add it to the index "
1247                        "for example\n"
1248                        "by using:\n\n"
1249                        "  git update-index --cacheinfo 160000 %s \"%s\"\n\n"
1250                        "which will accept this suggestion.\n"),
1251                        oid_to_hex(&merges.objects[0].item->oid), path);
1252                 break;
1253
1254         default:
1255                 output(opt, 1, _("Failed to merge submodule %s (multiple merges found)"), path);
1256                 for (i = 0; i < merges.nr; i++)
1257                         print_commit((struct commit *) merges.objects[i].item);
1258         }
1259
1260         object_array_clear(&merges);
1261         return 0;
1262 }
1263
1264 static int merge_mode_and_contents(struct merge_options *opt,
1265                                    const struct diff_filespec *o,
1266                                    const struct diff_filespec *a,
1267                                    const struct diff_filespec *b,
1268                                    const char *filename,
1269                                    const char *branch1,
1270                                    const char *branch2,
1271                                    const int extra_marker_size,
1272                                    struct merge_file_info *result)
1273 {
1274         if (opt->branch1 != branch1) {
1275                 /*
1276                  * It's weird getting a reverse merge with HEAD on the bottom
1277                  * side of the conflict markers and the other branch on the
1278                  * top.  Fix that.
1279                  */
1280                 return merge_mode_and_contents(opt, o, b, a,
1281                                                filename,
1282                                                branch2, branch1,
1283                                                extra_marker_size, result);
1284         }
1285
1286         result->merge = 0;
1287         result->clean = 1;
1288
1289         if ((S_IFMT & a->mode) != (S_IFMT & b->mode)) {
1290                 result->clean = 0;
1291                 if (S_ISREG(a->mode)) {
1292                         result->blob.mode = a->mode;
1293                         oidcpy(&result->blob.oid, &a->oid);
1294                 } else {
1295                         result->blob.mode = b->mode;
1296                         oidcpy(&result->blob.oid, &b->oid);
1297                 }
1298         } else {
1299                 if (!oid_eq(&a->oid, &o->oid) && !oid_eq(&b->oid, &o->oid))
1300                         result->merge = 1;
1301
1302                 /*
1303                  * Merge modes
1304                  */
1305                 if (a->mode == b->mode || a->mode == o->mode)
1306                         result->blob.mode = b->mode;
1307                 else {
1308                         result->blob.mode = a->mode;
1309                         if (b->mode != o->mode) {
1310                                 result->clean = 0;
1311                                 result->merge = 1;
1312                         }
1313                 }
1314
1315                 if (oid_eq(&a->oid, &b->oid) || oid_eq(&a->oid, &o->oid))
1316                         oidcpy(&result->blob.oid, &b->oid);
1317                 else if (oid_eq(&b->oid, &o->oid))
1318                         oidcpy(&result->blob.oid, &a->oid);
1319                 else if (S_ISREG(a->mode)) {
1320                         mmbuffer_t result_buf;
1321                         int ret = 0, merge_status;
1322
1323                         merge_status = merge_3way(opt, &result_buf, o, a, b,
1324                                                   branch1, branch2,
1325                                                   extra_marker_size);
1326
1327                         if ((merge_status < 0) || !result_buf.ptr)
1328                                 ret = err(opt, _("Failed to execute internal merge"));
1329
1330                         if (!ret &&
1331                             write_object_file(result_buf.ptr, result_buf.size,
1332                                               blob_type, &result->blob.oid))
1333                                 ret = err(opt, _("Unable to add %s to database"),
1334                                           a->path);
1335
1336                         free(result_buf.ptr);
1337                         if (ret)
1338                                 return ret;
1339                         result->clean = (merge_status == 0);
1340                 } else if (S_ISGITLINK(a->mode)) {
1341                         result->clean = merge_submodule(opt, &result->blob.oid,
1342                                                         o->path,
1343                                                         &o->oid,
1344                                                         &a->oid,
1345                                                         &b->oid);
1346                 } else if (S_ISLNK(a->mode)) {
1347                         switch (opt->recursive_variant) {
1348                         case MERGE_RECURSIVE_NORMAL:
1349                                 oidcpy(&result->blob.oid, &a->oid);
1350                                 if (!oid_eq(&a->oid, &b->oid))
1351                                         result->clean = 0;
1352                                 break;
1353                         case MERGE_RECURSIVE_OURS:
1354                                 oidcpy(&result->blob.oid, &a->oid);
1355                                 break;
1356                         case MERGE_RECURSIVE_THEIRS:
1357                                 oidcpy(&result->blob.oid, &b->oid);
1358                                 break;
1359                         }
1360                 } else
1361                         BUG("unsupported object type in the tree");
1362         }
1363
1364         if (result->merge)
1365                 output(opt, 2, _("Auto-merging %s"), filename);
1366
1367         return 0;
1368 }
1369
1370 static int handle_rename_via_dir(struct merge_options *opt,
1371                                  struct rename_conflict_info *ci)
1372 {
1373         /*
1374          * Handle file adds that need to be renamed due to directory rename
1375          * detection.  This differs from handle_rename_normal, because
1376          * there is no content merge to do; just move the file into the
1377          * desired final location.
1378          */
1379         const struct rename *ren = ci->ren1;
1380         const struct diff_filespec *dest = ren->pair->two;
1381         char *file_path = dest->path;
1382         int mark_conflicted = (opt->detect_directory_renames ==
1383                                MERGE_DIRECTORY_RENAMES_CONFLICT);
1384         assert(ren->dir_rename_original_dest);
1385
1386         if (!opt->call_depth && would_lose_untracked(opt, dest->path)) {
1387                 mark_conflicted = 1;
1388                 file_path = unique_path(opt, dest->path, ren->branch);
1389                 output(opt, 1, _("Error: Refusing to lose untracked file at %s; "
1390                                  "writing to %s instead."),
1391                        dest->path, file_path);
1392         }
1393
1394         if (mark_conflicted) {
1395                 /*
1396                  * Write the file in worktree at file_path.  In the index,
1397                  * only record the file at dest->path in the appropriate
1398                  * higher stage.
1399                  */
1400                 if (update_file(opt, 0, dest, file_path))
1401                         return -1;
1402                 if (file_path != dest->path)
1403                         free(file_path);
1404                 if (update_stages(opt, dest->path, NULL,
1405                                   ren->branch == opt->branch1 ? dest : NULL,
1406                                   ren->branch == opt->branch1 ? NULL : dest))
1407                         return -1;
1408                 return 0; /* not clean, but conflicted */
1409         } else {
1410                 /* Update dest->path both in index and in worktree */
1411                 if (update_file(opt, 1, dest, dest->path))
1412                         return -1;
1413                 return 1; /* clean */
1414         }
1415 }
1416
1417 static int handle_change_delete(struct merge_options *opt,
1418                                 const char *path, const char *old_path,
1419                                 const struct diff_filespec *o,
1420                                 const struct diff_filespec *changed,
1421                                 const char *change_branch,
1422                                 const char *delete_branch,
1423                                 const char *change, const char *change_past)
1424 {
1425         char *alt_path = NULL;
1426         const char *update_path = path;
1427         int ret = 0;
1428
1429         if (dir_in_way(opt->repo->index, path, !opt->call_depth, 0) ||
1430             (!opt->call_depth && would_lose_untracked(opt, path))) {
1431                 update_path = alt_path = unique_path(opt, path, change_branch);
1432         }
1433
1434         if (opt->call_depth) {
1435                 /*
1436                  * We cannot arbitrarily accept either a_sha or b_sha as
1437                  * correct; since there is no true "middle point" between
1438                  * them, simply reuse the base version for virtual merge base.
1439                  */
1440                 ret = remove_file_from_index(opt->repo->index, path);
1441                 if (!ret)
1442                         ret = update_file(opt, 0, o, update_path);
1443         } else {
1444                 /*
1445                  * Despite the four nearly duplicate messages and argument
1446                  * lists below and the ugliness of the nested if-statements,
1447                  * having complete messages makes the job easier for
1448                  * translators.
1449                  *
1450                  * The slight variance among the cases is due to the fact
1451                  * that:
1452                  *   1) directory/file conflicts (in effect if
1453                  *      !alt_path) could cause us to need to write the
1454                  *      file to a different path.
1455                  *   2) renames (in effect if !old_path) could mean that
1456                  *      there are two names for the path that the user
1457                  *      may know the file by.
1458                  */
1459                 if (!alt_path) {
1460                         if (!old_path) {
1461                                 output(opt, 1, _("CONFLICT (%s/delete): %s deleted in %s "
1462                                        "and %s in %s. Version %s of %s left in tree."),
1463                                        change, path, delete_branch, change_past,
1464                                        change_branch, change_branch, path);
1465                         } else {
1466                                 output(opt, 1, _("CONFLICT (%s/delete): %s deleted in %s "
1467                                        "and %s to %s in %s. Version %s of %s left in tree."),
1468                                        change, old_path, delete_branch, change_past, path,
1469                                        change_branch, change_branch, path);
1470                         }
1471                 } else {
1472                         if (!old_path) {
1473                                 output(opt, 1, _("CONFLICT (%s/delete): %s deleted in %s "
1474                                        "and %s in %s. Version %s of %s left in tree at %s."),
1475                                        change, path, delete_branch, change_past,
1476                                        change_branch, change_branch, path, alt_path);
1477                         } else {
1478                                 output(opt, 1, _("CONFLICT (%s/delete): %s deleted in %s "
1479                                        "and %s to %s in %s. Version %s of %s left in tree at %s."),
1480                                        change, old_path, delete_branch, change_past, path,
1481                                        change_branch, change_branch, path, alt_path);
1482                         }
1483                 }
1484                 /*
1485                  * No need to call update_file() on path when change_branch ==
1486                  * opt->branch1 && !alt_path, since that would needlessly touch
1487                  * path.  We could call update_file_flags() with update_cache=0
1488                  * and update_wd=0, but that's a no-op.
1489                  */
1490                 if (change_branch != opt->branch1 || alt_path)
1491                         ret = update_file(opt, 0, changed, update_path);
1492         }
1493         free(alt_path);
1494
1495         return ret;
1496 }
1497
1498 static int handle_rename_delete(struct merge_options *opt,
1499                                 struct rename_conflict_info *ci)
1500 {
1501         const struct rename *ren = ci->ren1;
1502         const struct diff_filespec *orig = ren->pair->one;
1503         const struct diff_filespec *dest = ren->pair->two;
1504         const char *rename_branch = ren->branch;
1505         const char *delete_branch = (opt->branch1 == ren->branch ?
1506                                      opt->branch2 : opt->branch1);
1507
1508         if (handle_change_delete(opt,
1509                                  opt->call_depth ? orig->path : dest->path,
1510                                  opt->call_depth ? NULL : orig->path,
1511                                  orig, dest,
1512                                  rename_branch, delete_branch,
1513                                  _("rename"), _("renamed")))
1514                 return -1;
1515
1516         if (opt->call_depth)
1517                 return remove_file_from_index(opt->repo->index, dest->path);
1518         else
1519                 return update_stages(opt, dest->path, NULL,
1520                                      rename_branch == opt->branch1 ? dest : NULL,
1521                                      rename_branch == opt->branch1 ? NULL : dest);
1522 }
1523
1524 static int handle_file_collision(struct merge_options *opt,
1525                                  const char *collide_path,
1526                                  const char *prev_path1,
1527                                  const char *prev_path2,
1528                                  const char *branch1, const char *branch2,
1529                                  struct diff_filespec *a,
1530                                  struct diff_filespec *b)
1531 {
1532         struct merge_file_info mfi;
1533         struct diff_filespec null;
1534         char *alt_path = NULL;
1535         const char *update_path = collide_path;
1536
1537         /*
1538          * It's easiest to get the correct things into stage 2 and 3, and
1539          * to make sure that the content merge puts HEAD before the other
1540          * branch if we just ensure that branch1 == opt->branch1.  So, simply
1541          * flip arguments around if we don't have that.
1542          */
1543         if (branch1 != opt->branch1) {
1544                 return handle_file_collision(opt, collide_path,
1545                                              prev_path2, prev_path1,
1546                                              branch2, branch1,
1547                                              b, a);
1548         }
1549
1550         /*
1551          * In the recursive case, we just opt to undo renames
1552          */
1553         if (opt->call_depth && (prev_path1 || prev_path2)) {
1554                 /* Put first file (a->oid, a->mode) in its original spot */
1555                 if (prev_path1) {
1556                         if (update_file(opt, 1, a, prev_path1))
1557                                 return -1;
1558                 } else {
1559                         if (update_file(opt, 1, a, collide_path))
1560                                 return -1;
1561                 }
1562
1563                 /* Put second file (b->oid, b->mode) in its original spot */
1564                 if (prev_path2) {
1565                         if (update_file(opt, 1, b, prev_path2))
1566                                 return -1;
1567                 } else {
1568                         if (update_file(opt, 1, b, collide_path))
1569                                 return -1;
1570                 }
1571
1572                 /* Don't leave something at collision path if unrenaming both */
1573                 if (prev_path1 && prev_path2)
1574                         remove_file(opt, 1, collide_path, 0);
1575
1576                 return 0;
1577         }
1578
1579         /* Remove rename sources if rename/add or rename/rename(2to1) */
1580         if (prev_path1)
1581                 remove_file(opt, 1, prev_path1,
1582                             opt->call_depth || would_lose_untracked(opt, prev_path1));
1583         if (prev_path2)
1584                 remove_file(opt, 1, prev_path2,
1585                             opt->call_depth || would_lose_untracked(opt, prev_path2));
1586
1587         /*
1588          * Remove the collision path, if it wouldn't cause dirty contents
1589          * or an untracked file to get lost.  We'll either overwrite with
1590          * merged contents, or just write out to differently named files.
1591          */
1592         if (was_dirty(opt, collide_path)) {
1593                 output(opt, 1, _("Refusing to lose dirty file at %s"),
1594                        collide_path);
1595                 update_path = alt_path = unique_path(opt, collide_path, "merged");
1596         } else if (would_lose_untracked(opt, collide_path)) {
1597                 /*
1598                  * Only way we get here is if both renames were from
1599                  * a directory rename AND user had an untracked file
1600                  * at the location where both files end up after the
1601                  * two directory renames.  See testcase 10d of t6043.
1602                  */
1603                 output(opt, 1, _("Refusing to lose untracked file at "
1604                                "%s, even though it's in the way."),
1605                        collide_path);
1606                 update_path = alt_path = unique_path(opt, collide_path, "merged");
1607         } else {
1608                 /*
1609                  * FIXME: It's possible that the two files are identical
1610                  * and that the current working copy happens to match, in
1611                  * which case we are unnecessarily touching the working
1612                  * tree file.  It's not a likely enough scenario that I
1613                  * want to code up the checks for it and a better fix is
1614                  * available if we restructure how unpack_trees() and
1615                  * merge-recursive interoperate anyway, so punting for
1616                  * now...
1617                  */
1618                 remove_file(opt, 0, collide_path, 0);
1619         }
1620
1621         /* Store things in diff_filespecs for functions that need it */
1622         null.path = (char *)collide_path;
1623         oidcpy(&null.oid, &null_oid);
1624         null.mode = 0;
1625
1626         if (merge_mode_and_contents(opt, &null, a, b, collide_path,
1627                                     branch1, branch2, opt->call_depth * 2, &mfi))
1628                 return -1;
1629         mfi.clean &= !alt_path;
1630         if (update_file(opt, mfi.clean, &mfi.blob, update_path))
1631                 return -1;
1632         if (!mfi.clean && !opt->call_depth &&
1633             update_stages(opt, collide_path, NULL, a, b))
1634                 return -1;
1635         free(alt_path);
1636         /*
1637          * FIXME: If both a & b both started with conflicts (only possible
1638          * if they came from a rename/rename(2to1)), but had IDENTICAL
1639          * contents including those conflicts, then in the next line we claim
1640          * it was clean.  If someone cares about this case, we should have the
1641          * caller notify us if we started with conflicts.
1642          */
1643         return mfi.clean;
1644 }
1645
1646 static int handle_rename_add(struct merge_options *opt,
1647                              struct rename_conflict_info *ci)
1648 {
1649         /* a was renamed to c, and a separate c was added. */
1650         struct diff_filespec *a = ci->ren1->pair->one;
1651         struct diff_filespec *c = ci->ren1->pair->two;
1652         char *path = c->path;
1653         char *prev_path_desc;
1654         struct merge_file_info mfi;
1655
1656         const char *rename_branch = ci->ren1->branch;
1657         const char *add_branch = (opt->branch1 == rename_branch ?
1658                                   opt->branch2 : opt->branch1);
1659         int other_stage = (ci->ren1->branch == opt->branch1 ? 3 : 2);
1660
1661         output(opt, 1, _("CONFLICT (rename/add): "
1662                "Rename %s->%s in %s.  Added %s in %s"),
1663                a->path, c->path, rename_branch,
1664                c->path, add_branch);
1665
1666         prev_path_desc = xstrfmt("version of %s from %s", path, a->path);
1667         ci->ren1->src_entry->stages[other_stage].path = a->path;
1668         if (merge_mode_and_contents(opt, a, c,
1669                                     &ci->ren1->src_entry->stages[other_stage],
1670                                     prev_path_desc,
1671                                     opt->branch1, opt->branch2,
1672                                     1 + opt->call_depth * 2, &mfi))
1673                 return -1;
1674         free(prev_path_desc);
1675
1676         ci->ren1->dst_entry->stages[other_stage].path = mfi.blob.path = c->path;
1677         return handle_file_collision(opt,
1678                                      c->path, a->path, NULL,
1679                                      rename_branch, add_branch,
1680                                      &mfi.blob,
1681                                      &ci->ren1->dst_entry->stages[other_stage]);
1682 }
1683
1684 static char *find_path_for_conflict(struct merge_options *opt,
1685                                     const char *path,
1686                                     const char *branch1,
1687                                     const char *branch2)
1688 {
1689         char *new_path = NULL;
1690         if (dir_in_way(opt->repo->index, path, !opt->call_depth, 0)) {
1691                 new_path = unique_path(opt, path, branch1);
1692                 output(opt, 1, _("%s is a directory in %s adding "
1693                                "as %s instead"),
1694                        path, branch2, new_path);
1695         } else if (would_lose_untracked(opt, path)) {
1696                 new_path = unique_path(opt, path, branch1);
1697                 output(opt, 1, _("Refusing to lose untracked file"
1698                                " at %s; adding as %s instead"),
1699                        path, new_path);
1700         }
1701
1702         return new_path;
1703 }
1704
1705 static int handle_rename_rename_1to2(struct merge_options *opt,
1706                                      struct rename_conflict_info *ci)
1707 {
1708         /* One file was renamed in both branches, but to different names. */
1709         struct merge_file_info mfi;
1710         struct diff_filespec *add;
1711         struct diff_filespec *o = ci->ren1->pair->one;
1712         struct diff_filespec *a = ci->ren1->pair->two;
1713         struct diff_filespec *b = ci->ren2->pair->two;
1714         char *path_desc;
1715
1716         output(opt, 1, _("CONFLICT (rename/rename): "
1717                "Rename \"%s\"->\"%s\" in branch \"%s\" "
1718                "rename \"%s\"->\"%s\" in \"%s\"%s"),
1719                o->path, a->path, ci->ren1->branch,
1720                o->path, b->path, ci->ren2->branch,
1721                opt->call_depth ? _(" (left unresolved)") : "");
1722
1723         path_desc = xstrfmt("%s and %s, both renamed from %s",
1724                             a->path, b->path, o->path);
1725         if (merge_mode_and_contents(opt, o, a, b, path_desc,
1726                                     ci->ren1->branch, ci->ren2->branch,
1727                                     opt->call_depth * 2, &mfi))
1728                 return -1;
1729         free(path_desc);
1730
1731         if (opt->call_depth) {
1732                 /*
1733                  * FIXME: For rename/add-source conflicts (if we could detect
1734                  * such), this is wrong.  We should instead find a unique
1735                  * pathname and then either rename the add-source file to that
1736                  * unique path, or use that unique path instead of src here.
1737                  */
1738                 if (update_file(opt, 0, &mfi.blob, o->path))
1739                         return -1;
1740
1741                 /*
1742                  * Above, we put the merged content at the merge-base's
1743                  * path.  Now we usually need to delete both a->path and
1744                  * b->path.  However, the rename on each side of the merge
1745                  * could also be involved in a rename/add conflict.  In
1746                  * such cases, we should keep the added file around,
1747                  * resolving the conflict at that path in its favor.
1748                  */
1749                 add = &ci->ren1->dst_entry->stages[2 ^ 1];
1750                 if (is_valid(add)) {
1751                         if (update_file(opt, 0, add, a->path))
1752                                 return -1;
1753                 }
1754                 else
1755                         remove_file_from_index(opt->repo->index, a->path);
1756                 add = &ci->ren2->dst_entry->stages[3 ^ 1];
1757                 if (is_valid(add)) {
1758                         if (update_file(opt, 0, add, b->path))
1759                                 return -1;
1760                 }
1761                 else
1762                         remove_file_from_index(opt->repo->index, b->path);
1763         } else {
1764                 /*
1765                  * For each destination path, we need to see if there is a
1766                  * rename/add collision.  If not, we can write the file out
1767                  * to the specified location.
1768                  */
1769                 add = &ci->ren1->dst_entry->stages[2 ^ 1];
1770                 if (is_valid(add)) {
1771                         add->path = mfi.blob.path = a->path;
1772                         if (handle_file_collision(opt, a->path,
1773                                                   NULL, NULL,
1774                                                   ci->ren1->branch,
1775                                                   ci->ren2->branch,
1776                                                   &mfi.blob, add) < 0)
1777                                 return -1;
1778                 } else {
1779                         char *new_path = find_path_for_conflict(opt, a->path,
1780                                                                 ci->ren1->branch,
1781                                                                 ci->ren2->branch);
1782                         if (update_file(opt, 0, &mfi.blob,
1783                                         new_path ? new_path : a->path))
1784                                 return -1;
1785                         free(new_path);
1786                         if (update_stages(opt, a->path, NULL, a, NULL))
1787                                 return -1;
1788                 }
1789
1790                 add = &ci->ren2->dst_entry->stages[3 ^ 1];
1791                 if (is_valid(add)) {
1792                         add->path = mfi.blob.path = b->path;
1793                         if (handle_file_collision(opt, b->path,
1794                                                   NULL, NULL,
1795                                                   ci->ren1->branch,
1796                                                   ci->ren2->branch,
1797                                                   add, &mfi.blob) < 0)
1798                                 return -1;
1799                 } else {
1800                         char *new_path = find_path_for_conflict(opt, b->path,
1801                                                                 ci->ren2->branch,
1802                                                                 ci->ren1->branch);
1803                         if (update_file(opt, 0, &mfi.blob,
1804                                         new_path ? new_path : b->path))
1805                                 return -1;
1806                         free(new_path);
1807                         if (update_stages(opt, b->path, NULL, NULL, b))
1808                                 return -1;
1809                 }
1810         }
1811
1812         return 0;
1813 }
1814
1815 static int handle_rename_rename_2to1(struct merge_options *opt,
1816                                      struct rename_conflict_info *ci)
1817 {
1818         /* Two files, a & b, were renamed to the same thing, c. */
1819         struct diff_filespec *a = ci->ren1->pair->one;
1820         struct diff_filespec *b = ci->ren2->pair->one;
1821         struct diff_filespec *c1 = ci->ren1->pair->two;
1822         struct diff_filespec *c2 = ci->ren2->pair->two;
1823         char *path = c1->path; /* == c2->path */
1824         char *path_side_1_desc;
1825         char *path_side_2_desc;
1826         struct merge_file_info mfi_c1;
1827         struct merge_file_info mfi_c2;
1828         int ostage1, ostage2;
1829
1830         output(opt, 1, _("CONFLICT (rename/rename): "
1831                "Rename %s->%s in %s. "
1832                "Rename %s->%s in %s"),
1833                a->path, c1->path, ci->ren1->branch,
1834                b->path, c2->path, ci->ren2->branch);
1835
1836         path_side_1_desc = xstrfmt("version of %s from %s", path, a->path);
1837         path_side_2_desc = xstrfmt("version of %s from %s", path, b->path);
1838         ostage1 = ci->ren1->branch == opt->branch1 ? 3 : 2;
1839         ostage2 = ostage1 ^ 1;
1840         ci->ren1->src_entry->stages[ostage1].path = a->path;
1841         ci->ren2->src_entry->stages[ostage2].path = b->path;
1842         if (merge_mode_and_contents(opt, a, c1,
1843                                     &ci->ren1->src_entry->stages[ostage1],
1844                                     path_side_1_desc,
1845                                     opt->branch1, opt->branch2,
1846                                     1 + opt->call_depth * 2, &mfi_c1) ||
1847             merge_mode_and_contents(opt, b,
1848                                     &ci->ren2->src_entry->stages[ostage2],
1849                                     c2, path_side_2_desc,
1850                                     opt->branch1, opt->branch2,
1851                                     1 + opt->call_depth * 2, &mfi_c2))
1852                 return -1;
1853         free(path_side_1_desc);
1854         free(path_side_2_desc);
1855         mfi_c1.blob.path = path;
1856         mfi_c2.blob.path = path;
1857
1858         return handle_file_collision(opt, path, a->path, b->path,
1859                                      ci->ren1->branch, ci->ren2->branch,
1860                                      &mfi_c1.blob, &mfi_c2.blob);
1861 }
1862
1863 /*
1864  * Get the diff_filepairs changed between o_tree and tree.
1865  */
1866 static struct diff_queue_struct *get_diffpairs(struct merge_options *opt,
1867                                                struct tree *o_tree,
1868                                                struct tree *tree)
1869 {
1870         struct diff_queue_struct *ret;
1871         struct diff_options opts;
1872
1873         repo_diff_setup(opt->repo, &opts);
1874         opts.flags.recursive = 1;
1875         opts.flags.rename_empty = 0;
1876         opts.detect_rename = merge_detect_rename(opt);
1877         /*
1878          * We do not have logic to handle the detection of copies.  In
1879          * fact, it may not even make sense to add such logic: would we
1880          * really want a change to a base file to be propagated through
1881          * multiple other files by a merge?
1882          */
1883         if (opts.detect_rename > DIFF_DETECT_RENAME)
1884                 opts.detect_rename = DIFF_DETECT_RENAME;
1885         opts.rename_limit = (opt->rename_limit >= 0) ? opt->rename_limit : 1000;
1886         opts.rename_score = opt->rename_score;
1887         opts.show_rename_progress = opt->show_rename_progress;
1888         opts.output_format = DIFF_FORMAT_NO_OUTPUT;
1889         diff_setup_done(&opts);
1890         diff_tree_oid(&o_tree->object.oid, &tree->object.oid, "", &opts);
1891         diffcore_std(&opts);
1892         if (opts.needed_rename_limit > opt->needed_rename_limit)
1893                 opt->needed_rename_limit = opts.needed_rename_limit;
1894
1895         ret = xmalloc(sizeof(*ret));
1896         *ret = diff_queued_diff;
1897
1898         opts.output_format = DIFF_FORMAT_NO_OUTPUT;
1899         diff_queued_diff.nr = 0;
1900         diff_queued_diff.queue = NULL;
1901         diff_flush(&opts);
1902         return ret;
1903 }
1904
1905 static int tree_has_path(struct repository *r, struct tree *tree,
1906                          const char *path)
1907 {
1908         struct object_id hashy;
1909         unsigned short mode_o;
1910
1911         return !get_tree_entry(r,
1912                                &tree->object.oid, path,
1913                                &hashy, &mode_o);
1914 }
1915
1916 /*
1917  * Return a new string that replaces the beginning portion (which matches
1918  * entry->dir), with entry->new_dir.  In perl-speak:
1919  *   new_path_name = (old_path =~ s/entry->dir/entry->new_dir/);
1920  * NOTE:
1921  *   Caller must ensure that old_path starts with entry->dir + '/'.
1922  */
1923 static char *apply_dir_rename(struct dir_rename_entry *entry,
1924                               const char *old_path)
1925 {
1926         struct strbuf new_path = STRBUF_INIT;
1927         int oldlen, newlen;
1928
1929         if (entry->non_unique_new_dir)
1930                 return NULL;
1931
1932         oldlen = strlen(entry->dir);
1933         newlen = entry->new_dir.len + (strlen(old_path) - oldlen) + 1;
1934         strbuf_grow(&new_path, newlen);
1935         strbuf_addbuf(&new_path, &entry->new_dir);
1936         strbuf_addstr(&new_path, &old_path[oldlen]);
1937
1938         return strbuf_detach(&new_path, NULL);
1939 }
1940
1941 static void get_renamed_dir_portion(const char *old_path, const char *new_path,
1942                                     char **old_dir, char **new_dir)
1943 {
1944         char *end_of_old, *end_of_new;
1945         int old_len, new_len;
1946
1947         *old_dir = NULL;
1948         *new_dir = NULL;
1949
1950         /*
1951          * For
1952          *    "a/b/c/d/e/foo.c" -> "a/b/some/thing/else/e/foo.c"
1953          * the "e/foo.c" part is the same, we just want to know that
1954          *    "a/b/c/d" was renamed to "a/b/some/thing/else"
1955          * so, for this example, this function returns "a/b/c/d" in
1956          * *old_dir and "a/b/some/thing/else" in *new_dir.
1957          *
1958          * Also, if the basename of the file changed, we don't care.  We
1959          * want to know which portion of the directory, if any, changed.
1960          */
1961         end_of_old = strrchr(old_path, '/');
1962         end_of_new = strrchr(new_path, '/');
1963
1964         if (end_of_old == NULL || end_of_new == NULL)
1965                 return;
1966         while (*--end_of_new == *--end_of_old &&
1967                end_of_old != old_path &&
1968                end_of_new != new_path)
1969                 ; /* Do nothing; all in the while loop */
1970         /*
1971          * We've found the first non-matching character in the directory
1972          * paths.  That means the current directory we were comparing
1973          * represents the rename.  Move end_of_old and end_of_new back
1974          * to the full directory name.
1975          */
1976         if (*end_of_old == '/')
1977                 end_of_old++;
1978         if (*end_of_old != '/')
1979                 end_of_new++;
1980         end_of_old = strchr(end_of_old, '/');
1981         end_of_new = strchr(end_of_new, '/');
1982
1983         /*
1984          * It may have been the case that old_path and new_path were the same
1985          * directory all along.  Don't claim a rename if they're the same.
1986          */
1987         old_len = end_of_old - old_path;
1988         new_len = end_of_new - new_path;
1989
1990         if (old_len != new_len || strncmp(old_path, new_path, old_len)) {
1991                 *old_dir = xstrndup(old_path, old_len);
1992                 *new_dir = xstrndup(new_path, new_len);
1993         }
1994 }
1995
1996 static void remove_hashmap_entries(struct hashmap *dir_renames,
1997                                    struct string_list *items_to_remove)
1998 {
1999         int i;
2000         struct dir_rename_entry *entry;
2001
2002         for (i = 0; i < items_to_remove->nr; i++) {
2003                 entry = items_to_remove->items[i].util;
2004                 hashmap_remove(dir_renames, entry, NULL);
2005         }
2006         string_list_clear(items_to_remove, 0);
2007 }
2008
2009 /*
2010  * See if there is a directory rename for path, and if there are any file
2011  * level conflicts for the renamed location.  If there is a rename and
2012  * there are no conflicts, return the new name.  Otherwise, return NULL.
2013  */
2014 static char *handle_path_level_conflicts(struct merge_options *opt,
2015                                          const char *path,
2016                                          struct dir_rename_entry *entry,
2017                                          struct hashmap *collisions,
2018                                          struct tree *tree)
2019 {
2020         char *new_path = NULL;
2021         struct collision_entry *collision_ent;
2022         int clean = 1;
2023         struct strbuf collision_paths = STRBUF_INIT;
2024
2025         /*
2026          * entry has the mapping of old directory name to new directory name
2027          * that we want to apply to path.
2028          */
2029         new_path = apply_dir_rename(entry, path);
2030
2031         if (!new_path) {
2032                 /* This should only happen when entry->non_unique_new_dir set */
2033                 if (!entry->non_unique_new_dir)
2034                         BUG("entry->non_unqiue_dir not set and !new_path");
2035                 output(opt, 1, _("CONFLICT (directory rename split): "
2036                                "Unclear where to place %s because directory "
2037                                "%s was renamed to multiple other directories, "
2038                                "with no destination getting a majority of the "
2039                                "files."),
2040                        path, entry->dir);
2041                 clean = 0;
2042                 return NULL;
2043         }
2044
2045         /*
2046          * The caller needs to have ensured that it has pre-populated
2047          * collisions with all paths that map to new_path.  Do a quick check
2048          * to ensure that's the case.
2049          */
2050         collision_ent = collision_find_entry(collisions, new_path);
2051         if (collision_ent == NULL)
2052                 BUG("collision_ent is NULL");
2053
2054         /*
2055          * Check for one-sided add/add/.../add conflicts, i.e.
2056          * where implicit renames from the other side doing
2057          * directory rename(s) can affect this side of history
2058          * to put multiple paths into the same location.  Warn
2059          * and bail on directory renames for such paths.
2060          */
2061         if (collision_ent->reported_already) {
2062                 clean = 0;
2063         } else if (tree_has_path(opt->repo, tree, new_path)) {
2064                 collision_ent->reported_already = 1;
2065                 strbuf_add_separated_string_list(&collision_paths, ", ",
2066                                                  &collision_ent->source_files);
2067                 output(opt, 1, _("CONFLICT (implicit dir rename): Existing "
2068                                "file/dir at %s in the way of implicit "
2069                                "directory rename(s) putting the following "
2070                                "path(s) there: %s."),
2071                        new_path, collision_paths.buf);
2072                 clean = 0;
2073         } else if (collision_ent->source_files.nr > 1) {
2074                 collision_ent->reported_already = 1;
2075                 strbuf_add_separated_string_list(&collision_paths, ", ",
2076                                                  &collision_ent->source_files);
2077                 output(opt, 1, _("CONFLICT (implicit dir rename): Cannot map "
2078                                "more than one path to %s; implicit directory "
2079                                "renames tried to put these paths there: %s"),
2080                        new_path, collision_paths.buf);
2081                 clean = 0;
2082         }
2083
2084         /* Free memory we no longer need */
2085         strbuf_release(&collision_paths);
2086         if (!clean && new_path) {
2087                 free(new_path);
2088                 return NULL;
2089         }
2090
2091         return new_path;
2092 }
2093
2094 /*
2095  * There are a couple things we want to do at the directory level:
2096  *   1. Check for both sides renaming to the same thing, in order to avoid
2097  *      implicit renaming of files that should be left in place.  (See
2098  *      testcase 6b in t6043 for details.)
2099  *   2. Prune directory renames if there are still files left in the
2100  *      the original directory.  These represent a partial directory rename,
2101  *      i.e. a rename where only some of the files within the directory
2102  *      were renamed elsewhere.  (Technically, this could be done earlier
2103  *      in get_directory_renames(), except that would prevent us from
2104  *      doing the previous check and thus failing testcase 6b.)
2105  *   3. Check for rename/rename(1to2) conflicts (at the directory level).
2106  *      In the future, we could potentially record this info as well and
2107  *      omit reporting rename/rename(1to2) conflicts for each path within
2108  *      the affected directories, thus cleaning up the merge output.
2109  *   NOTE: We do NOT check for rename/rename(2to1) conflicts at the
2110  *         directory level, because merging directories is fine.  If it
2111  *         causes conflicts for files within those merged directories, then
2112  *         that should be detected at the individual path level.
2113  */
2114 static void handle_directory_level_conflicts(struct merge_options *opt,
2115                                              struct hashmap *dir_re_head,
2116                                              struct tree *head,
2117                                              struct hashmap *dir_re_merge,
2118                                              struct tree *merge)
2119 {
2120         struct hashmap_iter iter;
2121         struct dir_rename_entry *head_ent;
2122         struct dir_rename_entry *merge_ent;
2123
2124         struct string_list remove_from_head = STRING_LIST_INIT_NODUP;
2125         struct string_list remove_from_merge = STRING_LIST_INIT_NODUP;
2126
2127         hashmap_iter_init(dir_re_head, &iter);
2128         while ((head_ent = hashmap_iter_next(&iter))) {
2129                 merge_ent = dir_rename_find_entry(dir_re_merge, head_ent->dir);
2130                 if (merge_ent &&
2131                     !head_ent->non_unique_new_dir &&
2132                     !merge_ent->non_unique_new_dir &&
2133                     !strbuf_cmp(&head_ent->new_dir, &merge_ent->new_dir)) {
2134                         /* 1. Renamed identically; remove it from both sides */
2135                         string_list_append(&remove_from_head,
2136                                            head_ent->dir)->util = head_ent;
2137                         strbuf_release(&head_ent->new_dir);
2138                         string_list_append(&remove_from_merge,
2139                                            merge_ent->dir)->util = merge_ent;
2140                         strbuf_release(&merge_ent->new_dir);
2141                 } else if (tree_has_path(opt->repo, head, head_ent->dir)) {
2142                         /* 2. This wasn't a directory rename after all */
2143                         string_list_append(&remove_from_head,
2144                                            head_ent->dir)->util = head_ent;
2145                         strbuf_release(&head_ent->new_dir);
2146                 }
2147         }
2148
2149         remove_hashmap_entries(dir_re_head, &remove_from_head);
2150         remove_hashmap_entries(dir_re_merge, &remove_from_merge);
2151
2152         hashmap_iter_init(dir_re_merge, &iter);
2153         while ((merge_ent = hashmap_iter_next(&iter))) {
2154                 head_ent = dir_rename_find_entry(dir_re_head, merge_ent->dir);
2155                 if (tree_has_path(opt->repo, merge, merge_ent->dir)) {
2156                         /* 2. This wasn't a directory rename after all */
2157                         string_list_append(&remove_from_merge,
2158                                            merge_ent->dir)->util = merge_ent;
2159                 } else if (head_ent &&
2160                            !head_ent->non_unique_new_dir &&
2161                            !merge_ent->non_unique_new_dir) {
2162                         /* 3. rename/rename(1to2) */
2163                         /*
2164                          * We can assume it's not rename/rename(1to1) because
2165                          * that was case (1), already checked above.  So we
2166                          * know that head_ent->new_dir and merge_ent->new_dir
2167                          * are different strings.
2168                          */
2169                         output(opt, 1, _("CONFLICT (rename/rename): "
2170                                        "Rename directory %s->%s in %s. "
2171                                        "Rename directory %s->%s in %s"),
2172                                head_ent->dir, head_ent->new_dir.buf, opt->branch1,
2173                                head_ent->dir, merge_ent->new_dir.buf, opt->branch2);
2174                         string_list_append(&remove_from_head,
2175                                            head_ent->dir)->util = head_ent;
2176                         strbuf_release(&head_ent->new_dir);
2177                         string_list_append(&remove_from_merge,
2178                                            merge_ent->dir)->util = merge_ent;
2179                         strbuf_release(&merge_ent->new_dir);
2180                 }
2181         }
2182
2183         remove_hashmap_entries(dir_re_head, &remove_from_head);
2184         remove_hashmap_entries(dir_re_merge, &remove_from_merge);
2185 }
2186
2187 static struct hashmap *get_directory_renames(struct diff_queue_struct *pairs)
2188 {
2189         struct hashmap *dir_renames;
2190         struct hashmap_iter iter;
2191         struct dir_rename_entry *entry;
2192         int i;
2193
2194         /*
2195          * Typically, we think of a directory rename as all files from a
2196          * certain directory being moved to a target directory.  However,
2197          * what if someone first moved two files from the original
2198          * directory in one commit, and then renamed the directory
2199          * somewhere else in a later commit?  At merge time, we just know
2200          * that files from the original directory went to two different
2201          * places, and that the bulk of them ended up in the same place.
2202          * We want each directory rename to represent where the bulk of the
2203          * files from that directory end up; this function exists to find
2204          * where the bulk of the files went.
2205          *
2206          * The first loop below simply iterates through the list of file
2207          * renames, finding out how often each directory rename pair
2208          * possibility occurs.
2209          */
2210         dir_renames = xmalloc(sizeof(*dir_renames));
2211         dir_rename_init(dir_renames);
2212         for (i = 0; i < pairs->nr; ++i) {
2213                 struct string_list_item *item;
2214                 int *count;
2215                 struct diff_filepair *pair = pairs->queue[i];
2216                 char *old_dir, *new_dir;
2217
2218                 /* File not part of directory rename if it wasn't renamed */
2219                 if (pair->status != 'R')
2220                         continue;
2221
2222                 get_renamed_dir_portion(pair->one->path, pair->two->path,
2223                                         &old_dir,        &new_dir);
2224                 if (!old_dir)
2225                         /* Directory didn't change at all; ignore this one. */
2226                         continue;
2227
2228                 entry = dir_rename_find_entry(dir_renames, old_dir);
2229                 if (!entry) {
2230                         entry = xmalloc(sizeof(*entry));
2231                         dir_rename_entry_init(entry, old_dir);
2232                         hashmap_put(dir_renames, entry);
2233                 } else {
2234                         free(old_dir);
2235                 }
2236                 item = string_list_lookup(&entry->possible_new_dirs, new_dir);
2237                 if (!item) {
2238                         item = string_list_insert(&entry->possible_new_dirs,
2239                                                   new_dir);
2240                         item->util = xcalloc(1, sizeof(int));
2241                 } else {
2242                         free(new_dir);
2243                 }
2244                 count = item->util;
2245                 *count += 1;
2246         }
2247
2248         /*
2249          * For each directory with files moved out of it, we find out which
2250          * target directory received the most files so we can declare it to
2251          * be the "winning" target location for the directory rename.  This
2252          * winner gets recorded in new_dir.  If there is no winner
2253          * (multiple target directories received the same number of files),
2254          * we set non_unique_new_dir.  Once we've determined the winner (or
2255          * that there is no winner), we no longer need possible_new_dirs.
2256          */
2257         hashmap_iter_init(dir_renames, &iter);
2258         while ((entry = hashmap_iter_next(&iter))) {
2259                 int max = 0;
2260                 int bad_max = 0;
2261                 char *best = NULL;
2262
2263                 for (i = 0; i < entry->possible_new_dirs.nr; i++) {
2264                         int *count = entry->possible_new_dirs.items[i].util;
2265
2266                         if (*count == max)
2267                                 bad_max = max;
2268                         else if (*count > max) {
2269                                 max = *count;
2270                                 best = entry->possible_new_dirs.items[i].string;
2271                         }
2272                 }
2273                 if (bad_max == max)
2274                         entry->non_unique_new_dir = 1;
2275                 else {
2276                         assert(entry->new_dir.len == 0);
2277                         strbuf_addstr(&entry->new_dir, best);
2278                 }
2279                 /*
2280                  * The relevant directory sub-portion of the original full
2281                  * filepaths were xstrndup'ed before inserting into
2282                  * possible_new_dirs, and instead of manually iterating the
2283                  * list and free'ing each, just lie and tell
2284                  * possible_new_dirs that it did the strdup'ing so that it
2285                  * will free them for us.
2286                  */
2287                 entry->possible_new_dirs.strdup_strings = 1;
2288                 string_list_clear(&entry->possible_new_dirs, 1);
2289         }
2290
2291         return dir_renames;
2292 }
2293
2294 static struct dir_rename_entry *check_dir_renamed(const char *path,
2295                                                   struct hashmap *dir_renames)
2296 {
2297         char *temp = xstrdup(path);
2298         char *end;
2299         struct dir_rename_entry *entry = NULL;
2300
2301         while ((end = strrchr(temp, '/'))) {
2302                 *end = '\0';
2303                 entry = dir_rename_find_entry(dir_renames, temp);
2304                 if (entry)
2305                         break;
2306         }
2307         free(temp);
2308         return entry;
2309 }
2310
2311 static void compute_collisions(struct hashmap *collisions,
2312                                struct hashmap *dir_renames,
2313                                struct diff_queue_struct *pairs)
2314 {
2315         int i;
2316
2317         /*
2318          * Multiple files can be mapped to the same path due to directory
2319          * renames done by the other side of history.  Since that other
2320          * side of history could have merged multiple directories into one,
2321          * if our side of history added the same file basename to each of
2322          * those directories, then all N of them would get implicitly
2323          * renamed by the directory rename detection into the same path,
2324          * and we'd get an add/add/.../add conflict, and all those adds
2325          * from *this* side of history.  This is not representable in the
2326          * index, and users aren't going to easily be able to make sense of
2327          * it.  So we need to provide a good warning about what's
2328          * happening, and fall back to no-directory-rename detection
2329          * behavior for those paths.
2330          *
2331          * See testcases 9e and all of section 5 from t6043 for examples.
2332          */
2333         collision_init(collisions);
2334
2335         for (i = 0; i < pairs->nr; ++i) {
2336                 struct dir_rename_entry *dir_rename_ent;
2337                 struct collision_entry *collision_ent;
2338                 char *new_path;
2339                 struct diff_filepair *pair = pairs->queue[i];
2340
2341                 if (pair->status != 'A' && pair->status != 'R')
2342                         continue;
2343                 dir_rename_ent = check_dir_renamed(pair->two->path,
2344                                                    dir_renames);
2345                 if (!dir_rename_ent)
2346                         continue;
2347
2348                 new_path = apply_dir_rename(dir_rename_ent, pair->two->path);
2349                 if (!new_path)
2350                         /*
2351                          * dir_rename_ent->non_unique_new_path is true, which
2352                          * means there is no directory rename for us to use,
2353                          * which means it won't cause us any additional
2354                          * collisions.
2355                          */
2356                         continue;
2357                 collision_ent = collision_find_entry(collisions, new_path);
2358                 if (!collision_ent) {
2359                         collision_ent = xcalloc(1,
2360                                                 sizeof(struct collision_entry));
2361                         hashmap_entry_init(collision_ent, strhash(new_path));
2362                         hashmap_put(collisions, collision_ent);
2363                         collision_ent->target_file = new_path;
2364                 } else {
2365                         free(new_path);
2366                 }
2367                 string_list_insert(&collision_ent->source_files,
2368                                    pair->two->path);
2369         }
2370 }
2371
2372 static char *check_for_directory_rename(struct merge_options *opt,
2373                                         const char *path,
2374                                         struct tree *tree,
2375                                         struct hashmap *dir_renames,
2376                                         struct hashmap *dir_rename_exclusions,
2377                                         struct hashmap *collisions,
2378                                         int *clean_merge)
2379 {
2380         char *new_path = NULL;
2381         struct dir_rename_entry *entry = check_dir_renamed(path, dir_renames);
2382         struct dir_rename_entry *oentry = NULL;
2383
2384         if (!entry)
2385                 return new_path;
2386
2387         /*
2388          * This next part is a little weird.  We do not want to do an
2389          * implicit rename into a directory we renamed on our side, because
2390          * that will result in a spurious rename/rename(1to2) conflict.  An
2391          * example:
2392          *   Base commit: dumbdir/afile, otherdir/bfile
2393          *   Side 1:      smrtdir/afile, otherdir/bfile
2394          *   Side 2:      dumbdir/afile, dumbdir/bfile
2395          * Here, while working on Side 1, we could notice that otherdir was
2396          * renamed/merged to dumbdir, and change the diff_filepair for
2397          * otherdir/bfile into a rename into dumbdir/bfile.  However, Side
2398          * 2 will notice the rename from dumbdir to smrtdir, and do the
2399          * transitive rename to move it from dumbdir/bfile to
2400          * smrtdir/bfile.  That gives us bfile in dumbdir vs being in
2401          * smrtdir, a rename/rename(1to2) conflict.  We really just want
2402          * the file to end up in smrtdir.  And the way to achieve that is
2403          * to not let Side1 do the rename to dumbdir, since we know that is
2404          * the source of one of our directory renames.
2405          *
2406          * That's why oentry and dir_rename_exclusions is here.
2407          *
2408          * As it turns out, this also prevents N-way transient rename
2409          * confusion; See testcases 9c and 9d of t6043.
2410          */
2411         oentry = dir_rename_find_entry(dir_rename_exclusions, entry->new_dir.buf);
2412         if (oentry) {
2413                 output(opt, 1, _("WARNING: Avoiding applying %s -> %s rename "
2414                                "to %s, because %s itself was renamed."),
2415                        entry->dir, entry->new_dir.buf, path, entry->new_dir.buf);
2416         } else {
2417                 new_path = handle_path_level_conflicts(opt, path, entry,
2418                                                        collisions, tree);
2419                 *clean_merge &= (new_path != NULL);
2420         }
2421
2422         return new_path;
2423 }
2424
2425 static void apply_directory_rename_modifications(struct merge_options *opt,
2426                                                  struct diff_filepair *pair,
2427                                                  char *new_path,
2428                                                  struct rename *re,
2429                                                  struct tree *tree,
2430                                                  struct tree *o_tree,
2431                                                  struct tree *a_tree,
2432                                                  struct tree *b_tree,
2433                                                  struct string_list *entries)
2434 {
2435         struct string_list_item *item;
2436         int stage = (tree == a_tree ? 2 : 3);
2437         int update_wd;
2438
2439         /*
2440          * In all cases where we can do directory rename detection,
2441          * unpack_trees() will have read pair->two->path into the
2442          * index and the working copy.  We need to remove it so that
2443          * we can instead place it at new_path.  It is guaranteed to
2444          * not be untracked (unpack_trees() would have errored out
2445          * saying the file would have been overwritten), but it might
2446          * be dirty, though.
2447          */
2448         update_wd = !was_dirty(opt, pair->two->path);
2449         if (!update_wd)
2450                 output(opt, 1, _("Refusing to lose dirty file at %s"),
2451                        pair->two->path);
2452         remove_file(opt, 1, pair->two->path, !update_wd);
2453
2454         /* Find or create a new re->dst_entry */
2455         item = string_list_lookup(entries, new_path);
2456         if (item) {
2457                 /*
2458                  * Since we're renaming on this side of history, and it's
2459                  * due to a directory rename on the other side of history
2460                  * (which we only allow when the directory in question no
2461                  * longer exists on the other side of history), the
2462                  * original entry for re->dst_entry is no longer
2463                  * necessary...
2464                  */
2465                 re->dst_entry->processed = 1;
2466
2467                 /*
2468                  * ...because we'll be using this new one.
2469                  */
2470                 re->dst_entry = item->util;
2471         } else {
2472                 /*
2473                  * re->dst_entry is for the before-dir-rename path, and we
2474                  * need it to hold information for the after-dir-rename
2475                  * path.  Before creating a new entry, we need to mark the
2476                  * old one as unnecessary (...unless it is shared by
2477                  * src_entry, i.e. this didn't use to be a rename, in which
2478                  * case we can just allow the normal processing to happen
2479                  * for it).
2480                  */
2481                 if (pair->status == 'R')
2482                         re->dst_entry->processed = 1;
2483
2484                 re->dst_entry = insert_stage_data(opt->repo, new_path,
2485                                                   o_tree, a_tree, b_tree,
2486                                                   entries);
2487                 item = string_list_insert(entries, new_path);
2488                 item->util = re->dst_entry;
2489         }
2490
2491         /*
2492          * Update the stage_data with the information about the path we are
2493          * moving into place.  That slot will be empty and available for us
2494          * to write to because of the collision checks in
2495          * handle_path_level_conflicts().  In other words,
2496          * re->dst_entry->stages[stage].oid will be the null_oid, so it's
2497          * open for us to write to.
2498          *
2499          * It may be tempting to actually update the index at this point as
2500          * well, using update_stages_for_stage_data(), but as per the big
2501          * "NOTE" in update_stages(), doing so will modify the current
2502          * in-memory index which will break calls to would_lose_untracked()
2503          * that we need to make.  Instead, we need to just make sure that
2504          * the various handle_rename_*() functions update the index
2505          * explicitly rather than relying on unpack_trees() to have done it.
2506          */
2507         get_tree_entry(opt->repo,
2508                        &tree->object.oid,
2509                        pair->two->path,
2510                        &re->dst_entry->stages[stage].oid,
2511                        &re->dst_entry->stages[stage].mode);
2512
2513         /*
2514          * Record the original change status (or 'type' of change).  If it
2515          * was originally an add ('A'), this lets us differentiate later
2516          * between a RENAME_DELETE conflict and RENAME_VIA_DIR (they
2517          * otherwise look the same).  If it was originally a rename ('R'),
2518          * this lets us remember and report accurately about the transitive
2519          * renaming that occurred via the directory rename detection.  Also,
2520          * record the original destination name.
2521          */
2522         re->dir_rename_original_type = pair->status;
2523         re->dir_rename_original_dest = pair->two->path;
2524
2525         /*
2526          * We don't actually look at pair->status again, but it seems
2527          * pedagogically correct to adjust it.
2528          */
2529         pair->status = 'R';
2530
2531         /*
2532          * Finally, record the new location.
2533          */
2534         pair->two->path = new_path;
2535 }
2536
2537 /*
2538  * Get information of all renames which occurred in 'pairs', making use of
2539  * any implicit directory renames inferred from the other side of history.
2540  * We need the three trees in the merge ('o_tree', 'a_tree' and 'b_tree')
2541  * to be able to associate the correct cache entries with the rename
2542  * information; tree is always equal to either a_tree or b_tree.
2543  */
2544 static struct string_list *get_renames(struct merge_options *opt,
2545                                        const char *branch,
2546                                        struct diff_queue_struct *pairs,
2547                                        struct hashmap *dir_renames,
2548                                        struct hashmap *dir_rename_exclusions,
2549                                        struct tree *tree,
2550                                        struct tree *o_tree,
2551                                        struct tree *a_tree,
2552                                        struct tree *b_tree,
2553                                        struct string_list *entries,
2554                                        int *clean_merge)
2555 {
2556         int i;
2557         struct hashmap collisions;
2558         struct hashmap_iter iter;
2559         struct collision_entry *e;
2560         struct string_list *renames;
2561
2562         compute_collisions(&collisions, dir_renames, pairs);
2563         renames = xcalloc(1, sizeof(struct string_list));
2564
2565         for (i = 0; i < pairs->nr; ++i) {
2566                 struct string_list_item *item;
2567                 struct rename *re;
2568                 struct diff_filepair *pair = pairs->queue[i];
2569                 char *new_path; /* non-NULL only with directory renames */
2570
2571                 if (pair->status != 'A' && pair->status != 'R') {
2572                         diff_free_filepair(pair);
2573                         continue;
2574                 }
2575                 new_path = check_for_directory_rename(opt, pair->two->path, tree,
2576                                                       dir_renames,
2577                                                       dir_rename_exclusions,
2578                                                       &collisions,
2579                                                       clean_merge);
2580                 if (pair->status != 'R' && !new_path) {
2581                         diff_free_filepair(pair);
2582                         continue;
2583                 }
2584
2585                 re = xmalloc(sizeof(*re));
2586                 re->processed = 0;
2587                 re->pair = pair;
2588                 re->branch = branch;
2589                 re->dir_rename_original_type = '\0';
2590                 re->dir_rename_original_dest = NULL;
2591                 item = string_list_lookup(entries, re->pair->one->path);
2592                 if (!item)
2593                         re->src_entry = insert_stage_data(opt->repo,
2594                                         re->pair->one->path,
2595                                         o_tree, a_tree, b_tree, entries);
2596                 else
2597                         re->src_entry = item->util;
2598
2599                 item = string_list_lookup(entries, re->pair->two->path);
2600                 if (!item)
2601                         re->dst_entry = insert_stage_data(opt->repo,
2602                                         re->pair->two->path,
2603                                         o_tree, a_tree, b_tree, entries);
2604                 else
2605                         re->dst_entry = item->util;
2606                 item = string_list_insert(renames, pair->one->path);
2607                 item->util = re;
2608                 if (new_path)
2609                         apply_directory_rename_modifications(opt, pair, new_path,
2610                                                              re, tree, o_tree,
2611                                                              a_tree, b_tree,
2612                                                              entries);
2613         }
2614
2615         hashmap_iter_init(&collisions, &iter);
2616         while ((e = hashmap_iter_next(&iter))) {
2617                 free(e->target_file);
2618                 string_list_clear(&e->source_files, 0);
2619         }
2620         hashmap_free(&collisions, 1);
2621         return renames;
2622 }
2623
2624 static int process_renames(struct merge_options *opt,
2625                            struct string_list *a_renames,
2626                            struct string_list *b_renames)
2627 {
2628         int clean_merge = 1, i, j;
2629         struct string_list a_by_dst = STRING_LIST_INIT_NODUP;
2630         struct string_list b_by_dst = STRING_LIST_INIT_NODUP;
2631         const struct rename *sre;
2632
2633         for (i = 0; i < a_renames->nr; i++) {
2634                 sre = a_renames->items[i].util;
2635                 string_list_insert(&a_by_dst, sre->pair->two->path)->util
2636                         = (void *)sre;
2637         }
2638         for (i = 0; i < b_renames->nr; i++) {
2639                 sre = b_renames->items[i].util;
2640                 string_list_insert(&b_by_dst, sre->pair->two->path)->util
2641                         = (void *)sre;
2642         }
2643
2644         for (i = 0, j = 0; i < a_renames->nr || j < b_renames->nr;) {
2645                 struct string_list *renames1, *renames2Dst;
2646                 struct rename *ren1 = NULL, *ren2 = NULL;
2647                 const char *ren1_src, *ren1_dst;
2648                 struct string_list_item *lookup;
2649
2650                 if (i >= a_renames->nr) {
2651                         ren2 = b_renames->items[j++].util;
2652                 } else if (j >= b_renames->nr) {
2653                         ren1 = a_renames->items[i++].util;
2654                 } else {
2655                         int compare = strcmp(a_renames->items[i].string,
2656                                              b_renames->items[j].string);
2657                         if (compare <= 0)
2658                                 ren1 = a_renames->items[i++].util;
2659                         if (compare >= 0)
2660                                 ren2 = b_renames->items[j++].util;
2661                 }
2662
2663                 /* TODO: refactor, so that 1/2 are not needed */
2664                 if (ren1) {
2665                         renames1 = a_renames;
2666                         renames2Dst = &b_by_dst;
2667                 } else {
2668                         renames1 = b_renames;
2669                         renames2Dst = &a_by_dst;
2670                         SWAP(ren2, ren1);
2671                 }
2672
2673                 if (ren1->processed)
2674                         continue;
2675                 ren1->processed = 1;
2676                 ren1->dst_entry->processed = 1;
2677                 /* BUG: We should only mark src_entry as processed if we
2678                  * are not dealing with a rename + add-source case.
2679                  */
2680                 ren1->src_entry->processed = 1;
2681
2682                 ren1_src = ren1->pair->one->path;
2683                 ren1_dst = ren1->pair->two->path;
2684
2685                 if (ren2) {
2686                         /* One file renamed on both sides */
2687                         const char *ren2_src = ren2->pair->one->path;
2688                         const char *ren2_dst = ren2->pair->two->path;
2689                         enum rename_type rename_type;
2690                         if (strcmp(ren1_src, ren2_src) != 0)
2691                                 BUG("ren1_src != ren2_src");
2692                         ren2->dst_entry->processed = 1;
2693                         ren2->processed = 1;
2694                         if (strcmp(ren1_dst, ren2_dst) != 0) {
2695                                 rename_type = RENAME_ONE_FILE_TO_TWO;
2696                                 clean_merge = 0;
2697                         } else {
2698                                 rename_type = RENAME_ONE_FILE_TO_ONE;
2699                                 /* BUG: We should only remove ren1_src in
2700                                  * the base stage (think of rename +
2701                                  * add-source cases).
2702                                  */
2703                                 remove_file(opt, 1, ren1_src, 1);
2704                                 update_entry(ren1->dst_entry,
2705                                              ren1->pair->one,
2706                                              ren1->pair->two,
2707                                              ren2->pair->two);
2708                         }
2709                         setup_rename_conflict_info(rename_type, opt, ren1, ren2);
2710                 } else if ((lookup = string_list_lookup(renames2Dst, ren1_dst))) {
2711                         /* Two different files renamed to the same thing */
2712                         char *ren2_dst;
2713                         ren2 = lookup->util;
2714                         ren2_dst = ren2->pair->two->path;
2715                         if (strcmp(ren1_dst, ren2_dst) != 0)
2716                                 BUG("ren1_dst != ren2_dst");
2717
2718                         clean_merge = 0;
2719                         ren2->processed = 1;
2720                         /*
2721                          * BUG: We should only mark src_entry as processed
2722                          * if we are not dealing with a rename + add-source
2723                          * case.
2724                          */
2725                         ren2->src_entry->processed = 1;
2726
2727                         setup_rename_conflict_info(RENAME_TWO_FILES_TO_ONE,
2728                                                    opt, ren1, ren2);
2729                 } else {
2730                         /* Renamed in 1, maybe changed in 2 */
2731                         /* we only use sha1 and mode of these */
2732                         struct diff_filespec src_other, dst_other;
2733                         int try_merge;
2734
2735                         /*
2736                          * unpack_trees loads entries from common-commit
2737                          * into stage 1, from head-commit into stage 2, and
2738                          * from merge-commit into stage 3.  We keep track
2739                          * of which side corresponds to the rename.
2740                          */
2741                         int renamed_stage = a_renames == renames1 ? 2 : 3;
2742                         int other_stage =   a_renames == renames1 ? 3 : 2;
2743
2744                         /* BUG: We should only remove ren1_src in the base
2745                          * stage and in other_stage (think of rename +
2746                          * add-source case).
2747                          */
2748                         remove_file(opt, 1, ren1_src,
2749                                     renamed_stage == 2 || !was_tracked(opt, ren1_src));
2750
2751                         oidcpy(&src_other.oid,
2752                                &ren1->src_entry->stages[other_stage].oid);
2753                         src_other.mode = ren1->src_entry->stages[other_stage].mode;
2754                         oidcpy(&dst_other.oid,
2755                                &ren1->dst_entry->stages[other_stage].oid);
2756                         dst_other.mode = ren1->dst_entry->stages[other_stage].mode;
2757                         try_merge = 0;
2758
2759                         if (oid_eq(&src_other.oid, &null_oid) &&
2760                             ren1->dir_rename_original_type == 'A') {
2761                                 setup_rename_conflict_info(RENAME_VIA_DIR,
2762                                                            opt, ren1, NULL);
2763                         } else if (oid_eq(&src_other.oid, &null_oid)) {
2764                                 setup_rename_conflict_info(RENAME_DELETE,
2765                                                            opt, ren1, NULL);
2766                         } else if ((dst_other.mode == ren1->pair->two->mode) &&
2767                                    oid_eq(&dst_other.oid, &ren1->pair->two->oid)) {
2768                                 /*
2769                                  * Added file on the other side identical to
2770                                  * the file being renamed: clean merge.
2771                                  * Also, there is no need to overwrite the
2772                                  * file already in the working copy, so call
2773                                  * update_file_flags() instead of
2774                                  * update_file().
2775                                  */
2776                                 if (update_file_flags(opt,
2777                                                       ren1->pair->two,
2778                                                       ren1_dst,
2779                                                       1, /* update_cache */
2780                                                       0  /* update_wd    */))
2781                                         clean_merge = -1;
2782                         } else if (!oid_eq(&dst_other.oid, &null_oid)) {
2783                                 /*
2784                                  * Probably not a clean merge, but it's
2785                                  * premature to set clean_merge to 0 here,
2786                                  * because if the rename merges cleanly and
2787                                  * the merge exactly matches the newly added
2788                                  * file, then the merge will be clean.
2789                                  */
2790                                 setup_rename_conflict_info(RENAME_ADD,
2791                                                            opt, ren1, NULL);
2792                         } else
2793                                 try_merge = 1;
2794
2795                         if (clean_merge < 0)
2796                                 goto cleanup_and_return;
2797                         if (try_merge) {
2798                                 struct diff_filespec *o, *a, *b;
2799                                 src_other.path = (char *)ren1_src;
2800
2801                                 o = ren1->pair->one;
2802                                 if (a_renames == renames1) {
2803                                         a = ren1->pair->two;
2804                                         b = &src_other;
2805                                 } else {
2806                                         b = ren1->pair->two;
2807                                         a = &src_other;
2808                                 }
2809                                 update_entry(ren1->dst_entry, o, a, b);
2810                                 setup_rename_conflict_info(RENAME_NORMAL,
2811                                                            opt, ren1, NULL);
2812                         }
2813                 }
2814         }
2815 cleanup_and_return:
2816         string_list_clear(&a_by_dst, 0);
2817         string_list_clear(&b_by_dst, 0);
2818
2819         return clean_merge;
2820 }
2821
2822 struct rename_info {
2823         struct string_list *head_renames;
2824         struct string_list *merge_renames;
2825 };
2826
2827 static void initial_cleanup_rename(struct diff_queue_struct *pairs,
2828                                    struct hashmap *dir_renames)
2829 {
2830         struct hashmap_iter iter;
2831         struct dir_rename_entry *e;
2832
2833         hashmap_iter_init(dir_renames, &iter);
2834         while ((e = hashmap_iter_next(&iter))) {
2835                 free(e->dir);
2836                 strbuf_release(&e->new_dir);
2837                 /* possible_new_dirs already cleared in get_directory_renames */
2838         }
2839         hashmap_free(dir_renames, 1);
2840         free(dir_renames);
2841
2842         free(pairs->queue);
2843         free(pairs);
2844 }
2845
2846 static int detect_and_process_renames(struct merge_options *opt,
2847                                       struct tree *common,
2848                                       struct tree *head,
2849                                       struct tree *merge,
2850                                       struct string_list *entries,
2851                                       struct rename_info *ri)
2852 {
2853         struct diff_queue_struct *head_pairs, *merge_pairs;
2854         struct hashmap *dir_re_head, *dir_re_merge;
2855         int clean = 1;
2856
2857         ri->head_renames = NULL;
2858         ri->merge_renames = NULL;
2859
2860         if (!merge_detect_rename(opt))
2861                 return 1;
2862
2863         head_pairs = get_diffpairs(opt, common, head);
2864         merge_pairs = get_diffpairs(opt, common, merge);
2865
2866         if ((opt->detect_directory_renames == MERGE_DIRECTORY_RENAMES_TRUE) ||
2867             (opt->detect_directory_renames == MERGE_DIRECTORY_RENAMES_CONFLICT &&
2868              !opt->call_depth)) {
2869                 dir_re_head = get_directory_renames(head_pairs);
2870                 dir_re_merge = get_directory_renames(merge_pairs);
2871
2872                 handle_directory_level_conflicts(opt,
2873                                                  dir_re_head, head,
2874                                                  dir_re_merge, merge);
2875         } else {
2876                 dir_re_head  = xmalloc(sizeof(*dir_re_head));
2877                 dir_re_merge = xmalloc(sizeof(*dir_re_merge));
2878                 dir_rename_init(dir_re_head);
2879                 dir_rename_init(dir_re_merge);
2880         }
2881
2882         ri->head_renames  = get_renames(opt, opt->branch1, head_pairs,
2883                                         dir_re_merge, dir_re_head, head,
2884                                         common, head, merge, entries,
2885                                         &clean);
2886         if (clean < 0)
2887                 goto cleanup;
2888         ri->merge_renames = get_renames(opt, opt->branch2, merge_pairs,
2889                                         dir_re_head, dir_re_merge, merge,
2890                                         common, head, merge, entries,
2891                                         &clean);
2892         if (clean < 0)
2893                 goto cleanup;
2894         clean &= process_renames(opt, ri->head_renames, ri->merge_renames);
2895
2896 cleanup:
2897         /*
2898          * Some cleanup is deferred until cleanup_renames() because the
2899          * data structures are still needed and referenced in
2900          * process_entry().  But there are a few things we can free now.
2901          */
2902         initial_cleanup_rename(head_pairs, dir_re_head);
2903         initial_cleanup_rename(merge_pairs, dir_re_merge);
2904
2905         return clean;
2906 }
2907
2908 static void final_cleanup_rename(struct string_list *rename)
2909 {
2910         const struct rename *re;
2911         int i;
2912
2913         if (rename == NULL)
2914                 return;
2915
2916         for (i = 0; i < rename->nr; i++) {
2917                 re = rename->items[i].util;
2918                 diff_free_filepair(re->pair);
2919         }
2920         string_list_clear(rename, 1);
2921         free(rename);
2922 }
2923
2924 static void final_cleanup_renames(struct rename_info *re_info)
2925 {
2926         final_cleanup_rename(re_info->head_renames);
2927         final_cleanup_rename(re_info->merge_renames);
2928 }
2929
2930 static int read_oid_strbuf(struct merge_options *opt,
2931                            const struct object_id *oid,
2932                            struct strbuf *dst)
2933 {
2934         void *buf;
2935         enum object_type type;
2936         unsigned long size;
2937         buf = read_object_file(oid, &type, &size);
2938         if (!buf)
2939                 return err(opt, _("cannot read object %s"), oid_to_hex(oid));
2940         if (type != OBJ_BLOB) {
2941                 free(buf);
2942                 return err(opt, _("object %s is not a blob"), oid_to_hex(oid));
2943         }
2944         strbuf_attach(dst, buf, size, size + 1);
2945         return 0;
2946 }
2947
2948 static int blob_unchanged(struct merge_options *opt,
2949                           const struct diff_filespec *o,
2950                           const struct diff_filespec *a,
2951                           int renormalize, const char *path)
2952 {
2953         struct strbuf obuf = STRBUF_INIT;
2954         struct strbuf abuf = STRBUF_INIT;
2955         int ret = 0; /* assume changed for safety */
2956         const struct index_state *idx = opt->repo->index;
2957
2958         if (a->mode != o->mode)
2959                 return 0;
2960         if (oid_eq(&o->oid, &a->oid))
2961                 return 1;
2962         if (!renormalize)
2963                 return 0;
2964
2965         if (read_oid_strbuf(opt, &o->oid, &obuf) ||
2966             read_oid_strbuf(opt, &a->oid, &abuf))
2967                 goto error_return;
2968         /*
2969          * Note: binary | is used so that both renormalizations are
2970          * performed.  Comparison can be skipped if both files are
2971          * unchanged since their sha1s have already been compared.
2972          */
2973         if (renormalize_buffer(idx, path, obuf.buf, obuf.len, &obuf) |
2974             renormalize_buffer(idx, path, abuf.buf, abuf.len, &abuf))
2975                 ret = (obuf.len == abuf.len && !memcmp(obuf.buf, abuf.buf, obuf.len));
2976
2977 error_return:
2978         strbuf_release(&obuf);
2979         strbuf_release(&abuf);
2980         return ret;
2981 }
2982
2983 static int handle_modify_delete(struct merge_options *opt,
2984                                 const char *path,
2985                                 const struct diff_filespec *o,
2986                                 const struct diff_filespec *a,
2987                                 const struct diff_filespec *b)
2988 {
2989         const char *modify_branch, *delete_branch;
2990         const struct diff_filespec *changed;
2991
2992         if (is_valid(a)) {
2993                 modify_branch = opt->branch1;
2994                 delete_branch = opt->branch2;
2995                 changed = a;
2996         } else {
2997                 modify_branch = opt->branch2;
2998                 delete_branch = opt->branch1;
2999                 changed = b;
3000         }
3001
3002         return handle_change_delete(opt,
3003                                     path, NULL,
3004                                     o, changed,
3005                                     modify_branch, delete_branch,
3006                                     _("modify"), _("modified"));
3007 }
3008
3009 static int handle_content_merge(struct merge_file_info *mfi,
3010                                 struct merge_options *opt,
3011                                 const char *path,
3012                                 int is_dirty,
3013                                 const struct diff_filespec *o,
3014                                 const struct diff_filespec *a,
3015                                 const struct diff_filespec *b,
3016                                 struct rename_conflict_info *ci)
3017 {
3018         const char *reason = _("content");
3019         unsigned df_conflict_remains = 0;
3020
3021         if (!is_valid(o))
3022                 reason = _("add/add");
3023
3024         assert(o->path && a->path && b->path);
3025         if (ci && dir_in_way(opt->repo->index, path, !opt->call_depth,
3026                              S_ISGITLINK(ci->ren1->pair->two->mode)))
3027                 df_conflict_remains = 1;
3028
3029         if (merge_mode_and_contents(opt, o, a, b, path,
3030                                     opt->branch1, opt->branch2,
3031                                     opt->call_depth * 2, mfi))
3032                 return -1;
3033
3034         /*
3035          * We can skip updating the working tree file iff:
3036          *   a) The merge is clean
3037          *   b) The merge matches what was in HEAD (content, mode, pathname)
3038          *   c) The target path is usable (i.e. not involved in D/F conflict)
3039          */
3040         if (mfi->clean && was_tracked_and_matches(opt, path, &mfi->blob) &&
3041             !df_conflict_remains) {
3042                 int pos;
3043                 struct cache_entry *ce;
3044
3045                 output(opt, 3, _("Skipped %s (merged same as existing)"), path);
3046                 if (add_cacheinfo(opt, &mfi->blob, path,
3047                                   0, (!opt->call_depth && !is_dirty), 0))
3048                         return -1;
3049                 /*
3050                  * However, add_cacheinfo() will delete the old cache entry
3051                  * and add a new one.  We need to copy over any skip_worktree
3052                  * flag to avoid making the file appear as if it were
3053                  * deleted by the user.
3054                  */
3055                 pos = index_name_pos(&opt->orig_index, path, strlen(path));
3056                 ce = opt->orig_index.cache[pos];
3057                 if (ce_skip_worktree(ce)) {
3058                         pos = index_name_pos(opt->repo->index, path, strlen(path));
3059                         ce = opt->repo->index->cache[pos];
3060                         ce->ce_flags |= CE_SKIP_WORKTREE;
3061                 }
3062                 return mfi->clean;
3063         }
3064
3065         if (!mfi->clean) {
3066                 if (S_ISGITLINK(mfi->blob.mode))
3067                         reason = _("submodule");
3068                 output(opt, 1, _("CONFLICT (%s): Merge conflict in %s"),
3069                                 reason, path);
3070                 if (ci && !df_conflict_remains)
3071                         if (update_stages(opt, path, o, a, b))
3072                                 return -1;
3073         }
3074
3075         if (df_conflict_remains || is_dirty) {
3076                 char *new_path;
3077                 if (opt->call_depth) {
3078                         remove_file_from_index(opt->repo->index, path);
3079                 } else {
3080                         if (!mfi->clean) {
3081                                 if (update_stages(opt, path, o, a, b))
3082                                         return -1;
3083                         } else {
3084                                 int file_from_stage2 = was_tracked(opt, path);
3085
3086                                 if (update_stages(opt, path, NULL,
3087                                                   file_from_stage2 ? &mfi->blob : NULL,
3088                                                   file_from_stage2 ? NULL : &mfi->blob))
3089                                         return -1;
3090                         }
3091
3092                 }
3093                 new_path = unique_path(opt, path, ci->ren1->branch);
3094                 if (is_dirty) {
3095                         output(opt, 1, _("Refusing to lose dirty file at %s"),
3096                                path);
3097                 }
3098                 output(opt, 1, _("Adding as %s instead"), new_path);
3099                 if (update_file(opt, 0, &mfi->blob, new_path)) {
3100                         free(new_path);
3101                         return -1;
3102                 }
3103                 free(new_path);
3104                 mfi->clean = 0;
3105         } else if (update_file(opt, mfi->clean, &mfi->blob, path))
3106                 return -1;
3107         return !is_dirty && mfi->clean;
3108 }
3109
3110 static int handle_rename_normal(struct merge_options *opt,
3111                                 const char *path,
3112                                 const struct diff_filespec *o,
3113                                 const struct diff_filespec *a,
3114                                 const struct diff_filespec *b,
3115                                 struct rename_conflict_info *ci)
3116 {
3117         struct rename *ren = ci->ren1;
3118         struct merge_file_info mfi;
3119         int clean;
3120         int side = (ren->branch == opt->branch1 ? 2 : 3);
3121
3122         /* Merge the content and write it out */
3123         clean = handle_content_merge(&mfi, opt, path, was_dirty(opt, path),
3124                                      o, a, b, ci);
3125
3126         if (clean &&
3127             opt->detect_directory_renames == MERGE_DIRECTORY_RENAMES_CONFLICT &&
3128             ren->dir_rename_original_dest) {
3129                 if (update_stages(opt, path,
3130                                   NULL,
3131                                   side == 2 ? &mfi.blob : NULL,
3132                                   side == 2 ? NULL : &mfi.blob))
3133                         return -1;
3134                 clean = 0; /* not clean, but conflicted */
3135         }
3136         return clean;
3137 }
3138
3139 static void dir_rename_warning(const char *msg,
3140                                int is_add,
3141                                int clean,
3142                                struct merge_options *opt,
3143                                struct rename *ren)
3144 {
3145         const char *other_branch;
3146         other_branch = (ren->branch == opt->branch1 ?
3147                         opt->branch2 : opt->branch1);
3148         if (is_add) {
3149                 output(opt, clean ? 2 : 1, msg,
3150                        ren->pair->one->path, ren->branch,
3151                        other_branch, ren->pair->two->path);
3152                 return;
3153         }
3154         output(opt, clean ? 2 : 1, msg,
3155                ren->pair->one->path, ren->dir_rename_original_dest, ren->branch,
3156                other_branch, ren->pair->two->path);
3157 }
3158 static int warn_about_dir_renamed_entries(struct merge_options *opt,
3159                                           struct rename *ren)
3160 {
3161         const char *msg;
3162         int clean = 1, is_add;
3163
3164         if (!ren)
3165                 return clean;
3166
3167         /* Return early if ren was not affected/created by a directory rename */
3168         if (!ren->dir_rename_original_dest)
3169                 return clean;
3170
3171         /* Sanity checks */
3172         assert(opt->detect_directory_renames > MERGE_DIRECTORY_RENAMES_NONE);
3173         assert(ren->dir_rename_original_type == 'A' ||
3174                ren->dir_rename_original_type == 'R');
3175
3176         /* Check whether to treat directory renames as a conflict */
3177         clean = (opt->detect_directory_renames == MERGE_DIRECTORY_RENAMES_TRUE);
3178
3179         is_add = (ren->dir_rename_original_type == 'A');
3180         if (ren->dir_rename_original_type == 'A' && clean) {
3181                 msg = _("Path updated: %s added in %s inside a "
3182                         "directory that was renamed in %s; moving it to %s.");
3183         } else if (ren->dir_rename_original_type == 'A' && !clean) {
3184                 msg = _("CONFLICT (file location): %s added in %s "
3185                         "inside a directory that was renamed in %s, "
3186                         "suggesting it should perhaps be moved to %s.");
3187         } else if (ren->dir_rename_original_type == 'R' && clean) {
3188                 msg = _("Path updated: %s renamed to %s in %s, inside a "
3189                         "directory that was renamed in %s; moving it to %s.");
3190         } else if (ren->dir_rename_original_type == 'R' && !clean) {
3191                 msg = _("CONFLICT (file location): %s renamed to %s in %s, "
3192                         "inside a directory that was renamed in %s, "
3193                         "suggesting it should perhaps be moved to %s.");
3194         } else {
3195                 BUG("Impossible dir_rename_original_type/clean combination");
3196         }
3197         dir_rename_warning(msg, is_add, clean, opt, ren);
3198
3199         return clean;
3200 }
3201
3202 /* Per entry merge function */
3203 static int process_entry(struct merge_options *opt,
3204                          const char *path, struct stage_data *entry)
3205 {
3206         int clean_merge = 1;
3207         int normalize = opt->renormalize;
3208
3209         struct diff_filespec *o = &entry->stages[1];
3210         struct diff_filespec *a = &entry->stages[2];
3211         struct diff_filespec *b = &entry->stages[3];
3212         int o_valid = is_valid(o);
3213         int a_valid = is_valid(a);
3214         int b_valid = is_valid(b);
3215         o->path = a->path = b->path = (char*)path;
3216
3217         entry->processed = 1;
3218         if (entry->rename_conflict_info) {
3219                 struct rename_conflict_info *ci = entry->rename_conflict_info;
3220                 struct diff_filespec *temp;
3221                 int path_clean;
3222
3223                 path_clean = warn_about_dir_renamed_entries(opt, ci->ren1);
3224                 path_clean &= warn_about_dir_renamed_entries(opt, ci->ren2);
3225
3226                 /*
3227                  * For cases with a single rename, {o,a,b}->path have all been
3228                  * set to the rename target path; we need to set two of these
3229                  * back to the rename source.
3230                  * For rename/rename conflicts, we'll manually fix paths below.
3231                  */
3232                 temp = (opt->branch1 == ci->ren1->branch) ? b : a;
3233                 o->path = temp->path = ci->ren1->pair->one->path;
3234                 if (ci->ren2) {
3235                         assert(opt->branch1 == ci->ren1->branch);
3236                 }
3237
3238                 switch (ci->rename_type) {
3239                 case RENAME_NORMAL:
3240                 case RENAME_ONE_FILE_TO_ONE:
3241                         clean_merge = handle_rename_normal(opt, path, o, a, b,
3242                                                            ci);
3243                         break;
3244                 case RENAME_VIA_DIR:
3245                         clean_merge = handle_rename_via_dir(opt, ci);
3246                         break;
3247                 case RENAME_ADD:
3248                         /*
3249                          * Probably unclean merge, but if the renamed file
3250                          * merges cleanly and the result can then be
3251                          * two-way merged cleanly with the added file, I
3252                          * guess it's a clean merge?
3253                          */
3254                         clean_merge = handle_rename_add(opt, ci);
3255                         break;
3256                 case RENAME_DELETE:
3257                         clean_merge = 0;
3258                         if (handle_rename_delete(opt, ci))
3259                                 clean_merge = -1;
3260                         break;
3261                 case RENAME_ONE_FILE_TO_TWO:
3262                         /*
3263                          * Manually fix up paths; note:
3264                          * ren[12]->pair->one->path are equal.
3265                          */
3266                         o->path = ci->ren1->pair->one->path;
3267                         a->path = ci->ren1->pair->two->path;
3268                         b->path = ci->ren2->pair->two->path;
3269
3270                         clean_merge = 0;
3271                         if (handle_rename_rename_1to2(opt, ci))
3272                                 clean_merge = -1;
3273                         break;
3274                 case RENAME_TWO_FILES_TO_ONE:
3275                         /*
3276                          * Manually fix up paths; note,
3277                          * ren[12]->pair->two->path are actually equal.
3278                          */
3279                         o->path = NULL;
3280                         a->path = ci->ren1->pair->two->path;
3281                         b->path = ci->ren2->pair->two->path;
3282
3283                         /*
3284                          * Probably unclean merge, but if the two renamed
3285                          * files merge cleanly and the two resulting files
3286                          * can then be two-way merged cleanly, I guess it's
3287                          * a clean merge?
3288                          */
3289                         clean_merge = handle_rename_rename_2to1(opt, ci);
3290                         break;
3291                 default:
3292                         entry->processed = 0;
3293                         break;
3294                 }
3295                 if (path_clean < clean_merge)
3296                         clean_merge = path_clean;
3297         } else if (o_valid && (!a_valid || !b_valid)) {
3298                 /* Case A: Deleted in one */
3299                 if ((!a_valid && !b_valid) ||
3300                     (!b_valid && blob_unchanged(opt, o, a, normalize, path)) ||
3301                     (!a_valid && blob_unchanged(opt, o, b, normalize, path))) {
3302                         /* Deleted in both or deleted in one and
3303                          * unchanged in the other */
3304                         if (a_valid)
3305                                 output(opt, 2, _("Removing %s"), path);
3306                         /* do not touch working file if it did not exist */
3307                         remove_file(opt, 1, path, !a_valid);
3308                 } else {
3309                         /* Modify/delete; deleted side may have put a directory in the way */
3310                         clean_merge = 0;
3311                         if (handle_modify_delete(opt, path, o, a, b))
3312                                 clean_merge = -1;
3313                 }
3314         } else if ((!o_valid && a_valid && !b_valid) ||
3315                    (!o_valid && !a_valid && b_valid)) {
3316                 /* Case B: Added in one. */
3317                 /* [nothing|directory] -> ([nothing|directory], file) */
3318
3319                 const char *add_branch;
3320                 const char *other_branch;
3321                 const char *conf;
3322                 const struct diff_filespec *contents;
3323
3324                 if (a_valid) {
3325                         add_branch = opt->branch1;
3326                         other_branch = opt->branch2;
3327                         contents = a;
3328                         conf = _("file/directory");
3329                 } else {
3330                         add_branch = opt->branch2;
3331                         other_branch = opt->branch1;
3332                         contents = b;
3333                         conf = _("directory/file");
3334                 }
3335                 if (dir_in_way(opt->repo->index, path,
3336                                !opt->call_depth && !S_ISGITLINK(a->mode),
3337                                0)) {
3338                         char *new_path = unique_path(opt, path, add_branch);
3339                         clean_merge = 0;
3340                         output(opt, 1, _("CONFLICT (%s): There is a directory with name %s in %s. "
3341                                "Adding %s as %s"),
3342                                conf, path, other_branch, path, new_path);
3343                         if (update_file(opt, 0, contents, new_path))
3344                                 clean_merge = -1;
3345                         else if (opt->call_depth)
3346                                 remove_file_from_index(opt->repo->index, path);
3347                         free(new_path);
3348                 } else {
3349                         output(opt, 2, _("Adding %s"), path);
3350                         /* do not overwrite file if already present */
3351                         if (update_file_flags(opt, contents, path, 1, !a_valid))
3352                                 clean_merge = -1;
3353                 }
3354         } else if (a_valid && b_valid) {
3355                 if (!o_valid) {
3356                         /* Case C: Added in both (check for same permissions) */
3357                         output(opt, 1,
3358                                _("CONFLICT (add/add): Merge conflict in %s"),
3359                                path);
3360                         clean_merge = handle_file_collision(opt,
3361                                                             path, NULL, NULL,
3362                                                             opt->branch1,
3363                                                             opt->branch2,
3364                                                             a, b);
3365                 } else {
3366                         /* case D: Modified in both, but differently. */
3367                         struct merge_file_info mfi;
3368                         int is_dirty = 0; /* unpack_trees would have bailed if dirty */
3369                         clean_merge = handle_content_merge(&mfi, opt, path,
3370                                                            is_dirty,
3371                                                            o, a, b, NULL);
3372                 }
3373         } else if (!o_valid && !a_valid && !b_valid) {
3374                 /*
3375                  * this entry was deleted altogether. a_mode == 0 means
3376                  * we had that path and want to actively remove it.
3377                  */
3378                 remove_file(opt, 1, path, !a->mode);
3379         } else
3380                 BUG("fatal merge failure, shouldn't happen.");
3381
3382         return clean_merge;
3383 }
3384
3385 static int merge_trees_internal(struct merge_options *opt,
3386                                 struct tree *head,
3387                                 struct tree *merge,
3388                                 struct tree *merge_base,
3389                                 struct tree **result)
3390 {
3391         struct index_state *istate = opt->repo->index;
3392         int code, clean;
3393
3394         if (opt->subtree_shift) {
3395                 merge = shift_tree_object(opt->repo, head, merge,
3396                                           opt->subtree_shift);
3397                 merge_base = shift_tree_object(opt->repo, head, merge_base,
3398                                                opt->subtree_shift);
3399         }
3400
3401         if (oid_eq(&merge_base->object.oid, &merge->object.oid)) {
3402                 output(opt, 0, _("Already up to date!"));
3403                 *result = head;
3404                 return 1;
3405         }
3406
3407         code = unpack_trees_start(opt, merge_base, head, merge);
3408
3409         if (code != 0) {
3410                 if (show(opt, 4) || opt->call_depth)
3411                         err(opt, _("merging of trees %s and %s failed"),
3412                             oid_to_hex(&head->object.oid),
3413                             oid_to_hex(&merge->object.oid));
3414                 unpack_trees_finish(opt);
3415                 return -1;
3416         }
3417
3418         if (unmerged_index(istate)) {
3419                 struct string_list *entries;
3420                 struct rename_info re_info;
3421                 int i;
3422                 /*
3423                  * Only need the hashmap while processing entries, so
3424                  * initialize it here and free it when we are done running
3425                  * through the entries. Keeping it in the merge_options as
3426                  * opposed to decaring a local hashmap is for convenience
3427                  * so that we don't have to pass it to around.
3428                  */
3429                 hashmap_init(&opt->current_file_dir_set, path_hashmap_cmp,
3430                              NULL, 512);
3431                 get_files_dirs(opt, head);
3432                 get_files_dirs(opt, merge);
3433
3434                 entries = get_unmerged(opt->repo->index);
3435                 clean = detect_and_process_renames(opt, merge_base, head, merge,
3436                                                    entries, &re_info);
3437                 record_df_conflict_files(opt, entries);
3438                 if (clean < 0)
3439                         goto cleanup;
3440                 for (i = entries->nr-1; 0 <= i; i--) {
3441                         const char *path = entries->items[i].string;
3442                         struct stage_data *e = entries->items[i].util;
3443                         if (!e->processed) {
3444                                 int ret = process_entry(opt, path, e);
3445                                 if (!ret)
3446                                         clean = 0;
3447                                 else if (ret < 0) {
3448                                         clean = ret;
3449                                         goto cleanup;
3450                                 }
3451                         }
3452                 }
3453                 for (i = 0; i < entries->nr; i++) {
3454                         struct stage_data *e = entries->items[i].util;
3455                         if (!e->processed)
3456                                 BUG("unprocessed path??? %s",
3457                                     entries->items[i].string);
3458                 }
3459
3460         cleanup:
3461                 final_cleanup_renames(&re_info);
3462
3463                 string_list_clear(entries, 1);
3464                 free(entries);
3465
3466                 hashmap_free(&opt->current_file_dir_set, 1);
3467
3468                 if (clean < 0) {
3469                         unpack_trees_finish(opt);
3470                         return clean;
3471                 }
3472         }
3473         else
3474                 clean = 1;
3475
3476         unpack_trees_finish(opt);
3477
3478         if (opt->call_depth &&
3479             !(*result = write_in_core_index_as_tree(opt->repo)))
3480                 return -1;
3481
3482         return clean;
3483 }
3484
3485 static struct commit_list *reverse_commit_list(struct commit_list *list)
3486 {
3487         struct commit_list *next = NULL, *current, *backup;
3488         for (current = list; current; current = backup) {
3489                 backup = current->next;
3490                 current->next = next;
3491                 next = current;
3492         }
3493         return next;
3494 }
3495
3496 /*
3497  * Merge the commits h1 and h2, return the resulting virtual
3498  * commit object and a flag indicating the cleanness of the merge.
3499  */
3500 static int merge_recursive_internal(struct merge_options *opt,
3501                                     struct commit *h1,
3502                                     struct commit *h2,
3503                                     struct commit_list *merge_bases,
3504                                     struct commit **result)
3505 {
3506         struct commit_list *iter;
3507         struct commit *merged_merge_bases;
3508         struct tree *result_tree;
3509         int clean;
3510         const char *ancestor_name;
3511         struct strbuf merge_base_abbrev = STRBUF_INIT;
3512
3513         if (show(opt, 4)) {
3514                 output(opt, 4, _("Merging:"));
3515                 output_commit_title(opt, h1);
3516                 output_commit_title(opt, h2);
3517         }
3518
3519         if (!merge_bases) {
3520                 merge_bases = get_merge_bases(h1, h2);
3521                 merge_bases = reverse_commit_list(merge_bases);
3522         }
3523
3524         if (show(opt, 5)) {
3525                 unsigned cnt = commit_list_count(merge_bases);
3526
3527                 output(opt, 5, Q_("found %u common ancestor:",
3528                                 "found %u common ancestors:", cnt), cnt);
3529                 for (iter = merge_bases; iter; iter = iter->next)
3530                         output_commit_title(opt, iter->item);
3531         }
3532
3533         merged_merge_bases = pop_commit(&merge_bases);
3534         if (merged_merge_bases == NULL) {
3535                 /* if there is no common ancestor, use an empty tree */
3536                 struct tree *tree;
3537
3538                 tree = lookup_tree(opt->repo, opt->repo->hash_algo->empty_tree);
3539                 merged_merge_bases = make_virtual_commit(opt->repo, tree,
3540                                                          "ancestor");
3541                 ancestor_name = "empty tree";
3542         } else if (merge_bases) {
3543                 ancestor_name = "merged common ancestors";
3544         } else {
3545                 strbuf_add_unique_abbrev(&merge_base_abbrev,
3546                                          &merged_merge_bases->object.oid,
3547                                          DEFAULT_ABBREV);
3548                 ancestor_name = merge_base_abbrev.buf;
3549         }
3550
3551         for (iter = merge_bases; iter; iter = iter->next) {
3552                 const char *saved_b1, *saved_b2;
3553                 opt->call_depth++;
3554                 /*
3555                  * When the merge fails, the result contains files
3556                  * with conflict markers. The cleanness flag is
3557                  * ignored (unless indicating an error), it was never
3558                  * actually used, as result of merge_trees has always
3559                  * overwritten it: the committed "conflicts" were
3560                  * already resolved.
3561                  */
3562                 discard_index(opt->repo->index);
3563                 saved_b1 = opt->branch1;
3564                 saved_b2 = opt->branch2;
3565                 opt->branch1 = "Temporary merge branch 1";
3566                 opt->branch2 = "Temporary merge branch 2";
3567                 if (merge_recursive_internal(opt, merged_merge_bases, iter->item,
3568                                              NULL, &merged_merge_bases) < 0)
3569                         return -1;
3570                 opt->branch1 = saved_b1;
3571                 opt->branch2 = saved_b2;
3572                 opt->call_depth--;
3573
3574                 if (!merged_merge_bases)
3575                         return err(opt, _("merge returned no commit"));
3576         }
3577
3578         discard_index(opt->repo->index);
3579         if (!opt->call_depth)
3580                 repo_read_index(opt->repo);
3581
3582         opt->ancestor = ancestor_name;
3583         clean = merge_trees_internal(opt,
3584                                      repo_get_commit_tree(opt->repo, h1),
3585                                      repo_get_commit_tree(opt->repo, h2),
3586                                      repo_get_commit_tree(opt->repo,
3587                                                           merged_merge_bases),
3588                                      &result_tree);
3589         strbuf_release(&merge_base_abbrev);
3590         if (clean < 0) {
3591                 flush_output(opt);
3592                 return clean;
3593         }
3594
3595         if (opt->call_depth) {
3596                 *result = make_virtual_commit(opt->repo, result_tree,
3597                                               "merged tree");
3598                 commit_list_insert(h1, &(*result)->parents);
3599                 commit_list_insert(h2, &(*result)->parents->next);
3600         }
3601         return clean;
3602 }
3603
3604 static int merge_start(struct merge_options *opt, struct tree *head)
3605 {
3606         struct strbuf sb = STRBUF_INIT;
3607
3608         if (repo_index_has_changes(opt->repo, head, &sb)) {
3609                 err(opt, _("Your local changes to the following files would be overwritten by merge:\n  %s"),
3610                     sb.buf);
3611                 strbuf_release(&sb);
3612                 return -1;
3613         }
3614
3615         return 0;
3616 }
3617
3618 static void merge_finalize(struct merge_options *opt)
3619 {
3620         flush_output(opt);
3621         if (!opt->call_depth && opt->buffer_output < 2)
3622                 strbuf_release(&opt->obuf);
3623         if (show(opt, 2))
3624                 diff_warn_rename_limit("merge.renamelimit",
3625                                        opt->needed_rename_limit, 0);
3626 }
3627
3628 int merge_trees(struct merge_options *opt,
3629                 struct tree *head,
3630                 struct tree *merge,
3631                 struct tree *merge_base)
3632 {
3633         int clean;
3634         struct tree *ignored;
3635
3636         assert(opt->ancestor != NULL);
3637
3638         if (merge_start(opt, head))
3639                 return -1;
3640         clean = merge_trees_internal(opt, head, merge, merge_base, &ignored);
3641         merge_finalize(opt);
3642
3643         return clean;
3644 }
3645
3646 int merge_recursive(struct merge_options *opt,
3647                     struct commit *h1,
3648                     struct commit *h2,
3649                     struct commit_list *merge_bases,
3650                     struct commit **result)
3651 {
3652         int clean;
3653
3654         assert(opt->ancestor == NULL);
3655
3656         if (merge_start(opt, repo_get_commit_tree(opt->repo, h1)))
3657                 return -1;
3658         clean = merge_recursive_internal(opt, h1, h2, merge_bases, result);
3659         merge_finalize(opt);
3660
3661         return clean;
3662 }
3663
3664 static struct commit *get_ref(struct repository *repo,
3665                               const struct object_id *oid,
3666                               const char *name)
3667 {
3668         struct object *object;
3669
3670         object = deref_tag(repo, parse_object(repo, oid),
3671                            name, strlen(name));
3672         if (!object)
3673                 return NULL;
3674         if (object->type == OBJ_TREE)
3675                 return make_virtual_commit(repo, (struct tree*)object, name);
3676         if (object->type != OBJ_COMMIT)
3677                 return NULL;
3678         if (parse_commit((struct commit *)object))
3679                 return NULL;
3680         return (struct commit *)object;
3681 }
3682
3683 int merge_recursive_generic(struct merge_options *opt,
3684                             const struct object_id *head,
3685                             const struct object_id *merge,
3686                             int num_merge_bases,
3687                             const struct object_id **merge_bases,
3688                             struct commit **result)
3689 {
3690         int clean;
3691         struct lock_file lock = LOCK_INIT;
3692         struct commit *head_commit = get_ref(opt->repo, head, opt->branch1);
3693         struct commit *next_commit = get_ref(opt->repo, merge, opt->branch2);
3694         struct commit_list *ca = NULL;
3695
3696         if (merge_bases) {
3697                 int i;
3698                 for (i = 0; i < num_merge_bases; ++i) {
3699                         struct commit *base;
3700                         if (!(base = get_ref(opt->repo, merge_bases[i],
3701                                              oid_to_hex(merge_bases[i]))))
3702                                 return err(opt, _("Could not parse object '%s'"),
3703                                            oid_to_hex(merge_bases[i]));
3704                         commit_list_insert(base, &ca);
3705                 }
3706         }
3707
3708         repo_hold_locked_index(opt->repo, &lock, LOCK_DIE_ON_ERROR);
3709         clean = merge_recursive(opt, head_commit, next_commit, ca,
3710                                 result);
3711         if (clean < 0) {
3712                 rollback_lock_file(&lock);
3713                 return clean;
3714         }
3715
3716         if (write_locked_index(opt->repo->index, &lock,
3717                                COMMIT_LOCK | SKIP_IF_UNCHANGED))
3718                 return err(opt, _("Unable to write index."));
3719
3720         return clean ? 0 : 1;
3721 }
3722
3723 static void merge_recursive_config(struct merge_options *opt)
3724 {
3725         char *value = NULL;
3726         git_config_get_int("merge.verbosity", &opt->verbosity);
3727         git_config_get_int("diff.renamelimit", &opt->rename_limit);
3728         git_config_get_int("merge.renamelimit", &opt->rename_limit);
3729         if (!git_config_get_string("diff.renames", &value)) {
3730                 opt->detect_renames = git_config_rename("diff.renames", value);
3731                 free(value);
3732         }
3733         if (!git_config_get_string("merge.renames", &value)) {
3734                 opt->detect_renames = git_config_rename("merge.renames", value);
3735                 free(value);
3736         }
3737         if (!git_config_get_string("merge.directoryrenames", &value)) {
3738                 int boolval = git_parse_maybe_bool(value);
3739                 if (0 <= boolval) {
3740                         opt->detect_directory_renames = boolval ?
3741                                 MERGE_DIRECTORY_RENAMES_TRUE :
3742                                 MERGE_DIRECTORY_RENAMES_NONE;
3743                 } else if (!strcasecmp(value, "conflict")) {
3744                         opt->detect_directory_renames =
3745                                 MERGE_DIRECTORY_RENAMES_CONFLICT;
3746                 } /* avoid erroring on values from future versions of git */
3747                 free(value);
3748         }
3749         git_config(git_xmerge_config, NULL);
3750 }
3751
3752 void init_merge_options(struct merge_options *opt,
3753                         struct repository *repo)
3754 {
3755         const char *merge_verbosity;
3756         memset(opt, 0, sizeof(struct merge_options));
3757
3758         opt->repo = repo;
3759
3760         opt->detect_renames = -1;
3761         opt->detect_directory_renames = MERGE_DIRECTORY_RENAMES_CONFLICT;
3762         opt->rename_limit = -1;
3763
3764         opt->verbosity = 2;
3765         opt->buffer_output = 1;
3766         strbuf_init(&opt->obuf, 0);
3767
3768         opt->renormalize = 0;
3769
3770         string_list_init(&opt->df_conflict_file_set, 1);
3771
3772         merge_recursive_config(opt);
3773         merge_verbosity = getenv("GIT_MERGE_VERBOSITY");
3774         if (merge_verbosity)
3775                 opt->verbosity = strtol(merge_verbosity, NULL, 10);
3776         if (opt->verbosity >= 5)
3777                 opt->buffer_output = 0;
3778 }
3779
3780 int parse_merge_opt(struct merge_options *opt, const char *s)
3781 {
3782         const char *arg;
3783
3784         if (!s || !*s)
3785                 return -1;
3786         if (!strcmp(s, "ours"))
3787                 opt->recursive_variant = MERGE_RECURSIVE_OURS;
3788         else if (!strcmp(s, "theirs"))
3789                 opt->recursive_variant = MERGE_RECURSIVE_THEIRS;
3790         else if (!strcmp(s, "subtree"))
3791                 opt->subtree_shift = "";
3792         else if (skip_prefix(s, "subtree=", &arg))
3793                 opt->subtree_shift = arg;
3794         else if (!strcmp(s, "patience"))
3795                 opt->xdl_opts = DIFF_WITH_ALG(opt, PATIENCE_DIFF);
3796         else if (!strcmp(s, "histogram"))
3797                 opt->xdl_opts = DIFF_WITH_ALG(opt, HISTOGRAM_DIFF);
3798         else if (skip_prefix(s, "diff-algorithm=", &arg)) {
3799                 long value = parse_algorithm_value(arg);
3800                 if (value < 0)
3801                         return -1;
3802                 /* clear out previous settings */
3803                 DIFF_XDL_CLR(opt, NEED_MINIMAL);
3804                 opt->xdl_opts &= ~XDF_DIFF_ALGORITHM_MASK;
3805                 opt->xdl_opts |= value;
3806         }
3807         else if (!strcmp(s, "ignore-space-change"))
3808                 DIFF_XDL_SET(opt, IGNORE_WHITESPACE_CHANGE);
3809         else if (!strcmp(s, "ignore-all-space"))
3810                 DIFF_XDL_SET(opt, IGNORE_WHITESPACE);
3811         else if (!strcmp(s, "ignore-space-at-eol"))
3812                 DIFF_XDL_SET(opt, IGNORE_WHITESPACE_AT_EOL);
3813         else if (!strcmp(s, "ignore-cr-at-eol"))
3814                 DIFF_XDL_SET(opt, IGNORE_CR_AT_EOL);
3815         else if (!strcmp(s, "renormalize"))
3816                 opt->renormalize = 1;
3817         else if (!strcmp(s, "no-renormalize"))
3818                 opt->renormalize = 0;
3819         else if (!strcmp(s, "no-renames"))
3820                 opt->detect_renames = 0;
3821         else if (!strcmp(s, "find-renames")) {
3822                 opt->detect_renames = 1;
3823                 opt->rename_score = 0;
3824         }
3825         else if (skip_prefix(s, "find-renames=", &arg) ||
3826                  skip_prefix(s, "rename-threshold=", &arg)) {
3827                 if ((opt->rename_score = parse_rename_score(&arg)) == -1 || *arg != 0)
3828                         return -1;
3829                 opt->detect_renames = 1;
3830         }
3831         /*
3832          * Please update $__git_merge_strategy_options in
3833          * git-completion.bash when you add new options
3834          */
3835         else
3836                 return -1;
3837         return 0;
3838 }