merge-recursive: move some definitions around to clean up the header
[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->merge_detect_rename >= 0 ? opt->merge_detect_rename :
389                 opt->diff_detect_rename >= 0 ? opt->diff_detect_rename : 1;
390 }
391
392 static void init_tree_desc_from_tree(struct tree_desc *desc, struct tree *tree)
393 {
394         parse_tree(tree);
395         init_tree_desc(desc, tree->buffer, tree->size);
396 }
397
398 static int unpack_trees_start(struct merge_options *opt,
399                               struct tree *common,
400                               struct tree *head,
401                               struct tree *merge)
402 {
403         int rc;
404         struct tree_desc t[3];
405         struct index_state tmp_index = { NULL };
406
407         memset(&opt->unpack_opts, 0, sizeof(opt->unpack_opts));
408         if (opt->call_depth)
409                 opt->unpack_opts.index_only = 1;
410         else
411                 opt->unpack_opts.update = 1;
412         opt->unpack_opts.merge = 1;
413         opt->unpack_opts.head_idx = 2;
414         opt->unpack_opts.fn = threeway_merge;
415         opt->unpack_opts.src_index = opt->repo->index;
416         opt->unpack_opts.dst_index = &tmp_index;
417         opt->unpack_opts.aggressive = !merge_detect_rename(opt);
418         setup_unpack_trees_porcelain(&opt->unpack_opts, "merge");
419
420         init_tree_desc_from_tree(t+0, common);
421         init_tree_desc_from_tree(t+1, head);
422         init_tree_desc_from_tree(t+2, merge);
423
424         rc = unpack_trees(3, t, &opt->unpack_opts);
425         cache_tree_free(&opt->repo->index->cache_tree);
426
427         /*
428          * Update opt->repo->index to match the new results, AFTER saving a copy
429          * in opt->orig_index.  Update src_index to point to the saved copy.
430          * (verify_uptodate() checks src_index, and the original index is
431          * the one that had the necessary modification timestamps.)
432          */
433         opt->orig_index = *opt->repo->index;
434         *opt->repo->index = tmp_index;
435         opt->unpack_opts.src_index = &opt->orig_index;
436
437         return rc;
438 }
439
440 static void unpack_trees_finish(struct merge_options *opt)
441 {
442         discard_index(&opt->orig_index);
443         clear_unpack_trees_porcelain(&opt->unpack_opts);
444 }
445
446 static int save_files_dirs(const struct object_id *oid,
447                            struct strbuf *base, const char *path,
448                            unsigned int mode, int stage, void *context)
449 {
450         struct path_hashmap_entry *entry;
451         int baselen = base->len;
452         struct merge_options *opt = context;
453
454         strbuf_addstr(base, path);
455
456         FLEX_ALLOC_MEM(entry, path, base->buf, base->len);
457         hashmap_entry_init(entry, path_hash(entry->path));
458         hashmap_add(&opt->current_file_dir_set, entry);
459
460         strbuf_setlen(base, baselen);
461         return (S_ISDIR(mode) ? READ_TREE_RECURSIVE : 0);
462 }
463
464 static void get_files_dirs(struct merge_options *opt, struct tree *tree)
465 {
466         struct pathspec match_all;
467         memset(&match_all, 0, sizeof(match_all));
468         read_tree_recursive(opt->repo, tree, "", 0, 0,
469                             &match_all, save_files_dirs, opt);
470 }
471
472 static int get_tree_entry_if_blob(struct repository *r,
473                                   const struct object_id *tree,
474                                   const char *path,
475                                   struct diff_filespec *dfs)
476 {
477         int ret;
478
479         ret = get_tree_entry(r, tree, path, &dfs->oid, &dfs->mode);
480         if (S_ISDIR(dfs->mode)) {
481                 oidcpy(&dfs->oid, &null_oid);
482                 dfs->mode = 0;
483         }
484         return ret;
485 }
486
487 /*
488  * Returns an index_entry instance which doesn't have to correspond to
489  * a real cache entry in Git's index.
490  */
491 static struct stage_data *insert_stage_data(struct repository *r,
492                 const char *path,
493                 struct tree *o, struct tree *a, struct tree *b,
494                 struct string_list *entries)
495 {
496         struct string_list_item *item;
497         struct stage_data *e = xcalloc(1, sizeof(struct stage_data));
498         get_tree_entry_if_blob(r, &o->object.oid, path, &e->stages[1]);
499         get_tree_entry_if_blob(r, &a->object.oid, path, &e->stages[2]);
500         get_tree_entry_if_blob(r, &b->object.oid, path, &e->stages[3]);
501         item = string_list_insert(entries, path);
502         item->util = e;
503         return e;
504 }
505
506 /*
507  * Create a dictionary mapping file names to stage_data objects. The
508  * dictionary contains one entry for every path with a non-zero stage entry.
509  */
510 static struct string_list *get_unmerged(struct index_state *istate)
511 {
512         struct string_list *unmerged = xcalloc(1, sizeof(struct string_list));
513         int i;
514
515         unmerged->strdup_strings = 1;
516
517         for (i = 0; i < istate->cache_nr; i++) {
518                 struct string_list_item *item;
519                 struct stage_data *e;
520                 const struct cache_entry *ce = istate->cache[i];
521                 if (!ce_stage(ce))
522                         continue;
523
524                 item = string_list_lookup(unmerged, ce->name);
525                 if (!item) {
526                         item = string_list_insert(unmerged, ce->name);
527                         item->util = xcalloc(1, sizeof(struct stage_data));
528                 }
529                 e = item->util;
530                 e->stages[ce_stage(ce)].mode = ce->ce_mode;
531                 oidcpy(&e->stages[ce_stage(ce)].oid, &ce->oid);
532         }
533
534         return unmerged;
535 }
536
537 static int string_list_df_name_compare(const char *one, const char *two)
538 {
539         int onelen = strlen(one);
540         int twolen = strlen(two);
541         /*
542          * Here we only care that entries for D/F conflicts are
543          * adjacent, in particular with the file of the D/F conflict
544          * appearing before files below the corresponding directory.
545          * The order of the rest of the list is irrelevant for us.
546          *
547          * To achieve this, we sort with df_name_compare and provide
548          * the mode S_IFDIR so that D/F conflicts will sort correctly.
549          * We use the mode S_IFDIR for everything else for simplicity,
550          * since in other cases any changes in their order due to
551          * sorting cause no problems for us.
552          */
553         int cmp = df_name_compare(one, onelen, S_IFDIR,
554                                   two, twolen, S_IFDIR);
555         /*
556          * Now that 'foo' and 'foo/bar' compare equal, we have to make sure
557          * that 'foo' comes before 'foo/bar'.
558          */
559         if (cmp)
560                 return cmp;
561         return onelen - twolen;
562 }
563
564 static void record_df_conflict_files(struct merge_options *opt,
565                                      struct string_list *entries)
566 {
567         /* If there is a D/F conflict and the file for such a conflict
568          * currently exists in the working tree, we want to allow it to be
569          * removed to make room for the corresponding directory if needed.
570          * The files underneath the directories of such D/F conflicts will
571          * be processed before the corresponding file involved in the D/F
572          * conflict.  If the D/F directory ends up being removed by the
573          * merge, then we won't have to touch the D/F file.  If the D/F
574          * directory needs to be written to the working copy, then the D/F
575          * file will simply be removed (in make_room_for_path()) to make
576          * room for the necessary paths.  Note that if both the directory
577          * and the file need to be present, then the D/F file will be
578          * reinstated with a new unique name at the time it is processed.
579          */
580         struct string_list df_sorted_entries = STRING_LIST_INIT_NODUP;
581         const char *last_file = NULL;
582         int last_len = 0;
583         int i;
584
585         /*
586          * If we're merging merge-bases, we don't want to bother with
587          * any working directory changes.
588          */
589         if (opt->call_depth)
590                 return;
591
592         /* Ensure D/F conflicts are adjacent in the entries list. */
593         for (i = 0; i < entries->nr; i++) {
594                 struct string_list_item *next = &entries->items[i];
595                 string_list_append(&df_sorted_entries, next->string)->util =
596                                    next->util;
597         }
598         df_sorted_entries.cmp = string_list_df_name_compare;
599         string_list_sort(&df_sorted_entries);
600
601         string_list_clear(&opt->df_conflict_file_set, 1);
602         for (i = 0; i < df_sorted_entries.nr; i++) {
603                 const char *path = df_sorted_entries.items[i].string;
604                 int len = strlen(path);
605                 struct stage_data *e = df_sorted_entries.items[i].util;
606
607                 /*
608                  * Check if last_file & path correspond to a D/F conflict;
609                  * i.e. whether path is last_file+'/'+<something>.
610                  * If so, record that it's okay to remove last_file to make
611                  * room for path and friends if needed.
612                  */
613                 if (last_file &&
614                     len > last_len &&
615                     memcmp(path, last_file, last_len) == 0 &&
616                     path[last_len] == '/') {
617                         string_list_insert(&opt->df_conflict_file_set, last_file);
618                 }
619
620                 /*
621                  * Determine whether path could exist as a file in the
622                  * working directory as a possible D/F conflict.  This
623                  * will only occur when it exists in stage 2 as a
624                  * file.
625                  */
626                 if (S_ISREG(e->stages[2].mode) || S_ISLNK(e->stages[2].mode)) {
627                         last_file = path;
628                         last_len = len;
629                 } else {
630                         last_file = NULL;
631                 }
632         }
633         string_list_clear(&df_sorted_entries, 0);
634 }
635
636 static int update_stages(struct merge_options *opt, const char *path,
637                          const struct diff_filespec *o,
638                          const struct diff_filespec *a,
639                          const struct diff_filespec *b)
640 {
641
642         /*
643          * NOTE: It is usually a bad idea to call update_stages on a path
644          * before calling update_file on that same path, since it can
645          * sometimes lead to spurious "refusing to lose untracked file..."
646          * messages from update_file (via make_room_for path via
647          * would_lose_untracked).  Instead, reverse the order of the calls
648          * (executing update_file first and then update_stages).
649          */
650         int clear = 1;
651         int options = ADD_CACHE_OK_TO_ADD | ADD_CACHE_SKIP_DFCHECK;
652         if (clear)
653                 if (remove_file_from_index(opt->repo->index, path))
654                         return -1;
655         if (o)
656                 if (add_cacheinfo(opt, o, path, 1, 0, options))
657                         return -1;
658         if (a)
659                 if (add_cacheinfo(opt, a, path, 2, 0, options))
660                         return -1;
661         if (b)
662                 if (add_cacheinfo(opt, b, path, 3, 0, options))
663                         return -1;
664         return 0;
665 }
666
667 static void update_entry(struct stage_data *entry,
668                          struct diff_filespec *o,
669                          struct diff_filespec *a,
670                          struct diff_filespec *b)
671 {
672         entry->processed = 0;
673         entry->stages[1].mode = o->mode;
674         entry->stages[2].mode = a->mode;
675         entry->stages[3].mode = b->mode;
676         oidcpy(&entry->stages[1].oid, &o->oid);
677         oidcpy(&entry->stages[2].oid, &a->oid);
678         oidcpy(&entry->stages[3].oid, &b->oid);
679 }
680
681 static int remove_file(struct merge_options *opt, int clean,
682                        const char *path, int no_wd)
683 {
684         int update_cache = opt->call_depth || clean;
685         int update_working_directory = !opt->call_depth && !no_wd;
686
687         if (update_cache) {
688                 if (remove_file_from_index(opt->repo->index, path))
689                         return -1;
690         }
691         if (update_working_directory) {
692                 if (ignore_case) {
693                         struct cache_entry *ce;
694                         ce = index_file_exists(opt->repo->index, path, strlen(path),
695                                                ignore_case);
696                         if (ce && ce_stage(ce) == 0 && strcmp(path, ce->name))
697                                 return 0;
698                 }
699                 if (remove_path(path))
700                         return -1;
701         }
702         return 0;
703 }
704
705 /* add a string to a strbuf, but converting "/" to "_" */
706 static void add_flattened_path(struct strbuf *out, const char *s)
707 {
708         size_t i = out->len;
709         strbuf_addstr(out, s);
710         for (; i < out->len; i++)
711                 if (out->buf[i] == '/')
712                         out->buf[i] = '_';
713 }
714
715 static char *unique_path(struct merge_options *opt,
716                          const char *path,
717                          const char *branch)
718 {
719         struct path_hashmap_entry *entry;
720         struct strbuf newpath = STRBUF_INIT;
721         int suffix = 0;
722         size_t base_len;
723
724         strbuf_addf(&newpath, "%s~", path);
725         add_flattened_path(&newpath, branch);
726
727         base_len = newpath.len;
728         while (hashmap_get_from_hash(&opt->current_file_dir_set,
729                                      path_hash(newpath.buf), newpath.buf) ||
730                (!opt->call_depth && file_exists(newpath.buf))) {
731                 strbuf_setlen(&newpath, base_len);
732                 strbuf_addf(&newpath, "_%d", suffix++);
733         }
734
735         FLEX_ALLOC_MEM(entry, path, newpath.buf, newpath.len);
736         hashmap_entry_init(entry, path_hash(entry->path));
737         hashmap_add(&opt->current_file_dir_set, entry);
738         return strbuf_detach(&newpath, NULL);
739 }
740
741 /**
742  * Check whether a directory in the index is in the way of an incoming
743  * file.  Return 1 if so.  If check_working_copy is non-zero, also
744  * check the working directory.  If empty_ok is non-zero, also return
745  * 0 in the case where the working-tree dir exists but is empty.
746  */
747 static int dir_in_way(struct index_state *istate, const char *path,
748                       int check_working_copy, int empty_ok)
749 {
750         int pos;
751         struct strbuf dirpath = STRBUF_INIT;
752         struct stat st;
753
754         strbuf_addstr(&dirpath, path);
755         strbuf_addch(&dirpath, '/');
756
757         pos = index_name_pos(istate, dirpath.buf, dirpath.len);
758
759         if (pos < 0)
760                 pos = -1 - pos;
761         if (pos < istate->cache_nr &&
762             !strncmp(dirpath.buf, istate->cache[pos]->name, dirpath.len)) {
763                 strbuf_release(&dirpath);
764                 return 1;
765         }
766
767         strbuf_release(&dirpath);
768         return check_working_copy && !lstat(path, &st) && S_ISDIR(st.st_mode) &&
769                 !(empty_ok && is_empty_dir(path));
770 }
771
772 /*
773  * Returns whether path was tracked in the index before the merge started,
774  * and its oid and mode match the specified values
775  */
776 static int was_tracked_and_matches(struct merge_options *opt, const char *path,
777                                    const struct diff_filespec *blob)
778 {
779         int pos = index_name_pos(&opt->orig_index, path, strlen(path));
780         struct cache_entry *ce;
781
782         if (0 > pos)
783                 /* we were not tracking this path before the merge */
784                 return 0;
785
786         /* See if the file we were tracking before matches */
787         ce = opt->orig_index.cache[pos];
788         return (oid_eq(&ce->oid, &blob->oid) && ce->ce_mode == blob->mode);
789 }
790
791 /*
792  * Returns whether path was tracked in the index before the merge started
793  */
794 static int was_tracked(struct merge_options *opt, const char *path)
795 {
796         int pos = index_name_pos(&opt->orig_index, path, strlen(path));
797
798         if (0 <= pos)
799                 /* we were tracking this path before the merge */
800                 return 1;
801
802         return 0;
803 }
804
805 static int would_lose_untracked(struct merge_options *opt, const char *path)
806 {
807         struct index_state *istate = opt->repo->index;
808
809         /*
810          * This may look like it can be simplified to:
811          *   return !was_tracked(opt, path) && file_exists(path)
812          * but it can't.  This function needs to know whether path was in
813          * the working tree due to EITHER having been tracked in the index
814          * before the merge OR having been put into the working copy and
815          * index by unpack_trees().  Due to that either-or requirement, we
816          * check the current index instead of the original one.
817          *
818          * Note that we do not need to worry about merge-recursive itself
819          * updating the index after unpack_trees() and before calling this
820          * function, because we strictly require all code paths in
821          * merge-recursive to update the working tree first and the index
822          * second.  Doing otherwise would break
823          * update_file()/would_lose_untracked(); see every comment in this
824          * file which mentions "update_stages".
825          */
826         int pos = index_name_pos(istate, path, strlen(path));
827
828         if (pos < 0)
829                 pos = -1 - pos;
830         while (pos < istate->cache_nr &&
831                !strcmp(path, istate->cache[pos]->name)) {
832                 /*
833                  * If stage #0, it is definitely tracked.
834                  * If it has stage #2 then it was tracked
835                  * before this merge started.  All other
836                  * cases the path was not tracked.
837                  */
838                 switch (ce_stage(istate->cache[pos])) {
839                 case 0:
840                 case 2:
841                         return 0;
842                 }
843                 pos++;
844         }
845         return file_exists(path);
846 }
847
848 static int was_dirty(struct merge_options *opt, const char *path)
849 {
850         struct cache_entry *ce;
851         int dirty = 1;
852
853         if (opt->call_depth || !was_tracked(opt, path))
854                 return !dirty;
855
856         ce = index_file_exists(opt->unpack_opts.src_index,
857                                path, strlen(path), ignore_case);
858         dirty = verify_uptodate(ce, &opt->unpack_opts) != 0;
859         return dirty;
860 }
861
862 static int make_room_for_path(struct merge_options *opt, const char *path)
863 {
864         int status, i;
865         const char *msg = _("failed to create path '%s'%s");
866
867         /* Unlink any D/F conflict files that are in the way */
868         for (i = 0; i < opt->df_conflict_file_set.nr; i++) {
869                 const char *df_path = opt->df_conflict_file_set.items[i].string;
870                 size_t pathlen = strlen(path);
871                 size_t df_pathlen = strlen(df_path);
872                 if (df_pathlen < pathlen &&
873                     path[df_pathlen] == '/' &&
874                     strncmp(path, df_path, df_pathlen) == 0) {
875                         output(opt, 3,
876                                _("Removing %s to make room for subdirectory\n"),
877                                df_path);
878                         unlink(df_path);
879                         unsorted_string_list_delete_item(&opt->df_conflict_file_set,
880                                                          i, 0);
881                         break;
882                 }
883         }
884
885         /* Make sure leading directories are created */
886         status = safe_create_leading_directories_const(path);
887         if (status) {
888                 if (status == SCLD_EXISTS)
889                         /* something else exists */
890                         return err(opt, msg, path, _(": perhaps a D/F conflict?"));
891                 return err(opt, msg, path, "");
892         }
893
894         /*
895          * Do not unlink a file in the work tree if we are not
896          * tracking it.
897          */
898         if (would_lose_untracked(opt, path))
899                 return err(opt, _("refusing to lose untracked file at '%s'"),
900                            path);
901
902         /* Successful unlink is good.. */
903         if (!unlink(path))
904                 return 0;
905         /* .. and so is no existing file */
906         if (errno == ENOENT)
907                 return 0;
908         /* .. but not some other error (who really cares what?) */
909         return err(opt, msg, path, _(": perhaps a D/F conflict?"));
910 }
911
912 static int update_file_flags(struct merge_options *opt,
913                              const struct diff_filespec *contents,
914                              const char *path,
915                              int update_cache,
916                              int update_wd)
917 {
918         int ret = 0;
919
920         if (opt->call_depth)
921                 update_wd = 0;
922
923         if (update_wd) {
924                 enum object_type type;
925                 void *buf;
926                 unsigned long size;
927
928                 if (S_ISGITLINK(contents->mode)) {
929                         /*
930                          * We may later decide to recursively descend into
931                          * the submodule directory and update its index
932                          * and/or work tree, but we do not do that now.
933                          */
934                         update_wd = 0;
935                         goto update_index;
936                 }
937
938                 buf = read_object_file(&contents->oid, &type, &size);
939                 if (!buf) {
940                         ret = err(opt, _("cannot read object %s '%s'"),
941                                   oid_to_hex(&contents->oid), path);
942                         goto free_buf;
943                 }
944                 if (type != OBJ_BLOB) {
945                         ret = err(opt, _("blob expected for %s '%s'"),
946                                   oid_to_hex(&contents->oid), path);
947                         goto free_buf;
948                 }
949                 if (S_ISREG(contents->mode)) {
950                         struct strbuf strbuf = STRBUF_INIT;
951                         if (convert_to_working_tree(opt->repo->index,
952                                                     path, buf, size, &strbuf)) {
953                                 free(buf);
954                                 size = strbuf.len;
955                                 buf = strbuf_detach(&strbuf, NULL);
956                         }
957                 }
958
959                 if (make_room_for_path(opt, path) < 0) {
960                         update_wd = 0;
961                         goto free_buf;
962                 }
963                 if (S_ISREG(contents->mode) ||
964                     (!has_symlinks && S_ISLNK(contents->mode))) {
965                         int fd;
966                         int mode = (contents->mode & 0100 ? 0777 : 0666);
967
968                         fd = open(path, O_WRONLY | O_TRUNC | O_CREAT, mode);
969                         if (fd < 0) {
970                                 ret = err(opt, _("failed to open '%s': %s"),
971                                           path, strerror(errno));
972                                 goto free_buf;
973                         }
974                         write_in_full(fd, buf, size);
975                         close(fd);
976                 } else if (S_ISLNK(contents->mode)) {
977                         char *lnk = xmemdupz(buf, size);
978                         safe_create_leading_directories_const(path);
979                         unlink(path);
980                         if (symlink(lnk, path))
981                                 ret = err(opt, _("failed to symlink '%s': %s"),
982                                           path, strerror(errno));
983                         free(lnk);
984                 } else
985                         ret = err(opt,
986                                   _("do not know what to do with %06o %s '%s'"),
987                                   contents->mode, oid_to_hex(&contents->oid), path);
988         free_buf:
989                 free(buf);
990         }
991 update_index:
992         if (!ret && update_cache)
993                 if (add_cacheinfo(opt, contents, path, 0, update_wd,
994                                   ADD_CACHE_OK_TO_ADD))
995                         return -1;
996         return ret;
997 }
998
999 static int update_file(struct merge_options *opt,
1000                        int clean,
1001                        const struct diff_filespec *contents,
1002                        const char *path)
1003 {
1004         return update_file_flags(opt, contents, path,
1005                                  opt->call_depth || clean, !opt->call_depth);
1006 }
1007
1008 /* Low level file merging, update and removal */
1009
1010 struct merge_file_info {
1011         struct diff_filespec blob; /* mostly use oid & mode; sometimes path */
1012         unsigned clean:1,
1013                  merge:1;
1014 };
1015
1016 static int merge_3way(struct merge_options *opt,
1017                       mmbuffer_t *result_buf,
1018                       const struct diff_filespec *o,
1019                       const struct diff_filespec *a,
1020                       const struct diff_filespec *b,
1021                       const char *branch1,
1022                       const char *branch2,
1023                       const int extra_marker_size)
1024 {
1025         mmfile_t orig, src1, src2;
1026         struct ll_merge_options ll_opts = {0};
1027         char *base, *name1, *name2;
1028         int merge_status;
1029
1030         ll_opts.renormalize = opt->renormalize;
1031         ll_opts.extra_marker_size = extra_marker_size;
1032         ll_opts.xdl_opts = opt->xdl_opts;
1033
1034         if (opt->call_depth) {
1035                 ll_opts.virtual_ancestor = 1;
1036                 ll_opts.variant = 0;
1037         } else {
1038                 switch (opt->recursive_variant) {
1039                 case MERGE_RECURSIVE_OURS:
1040                         ll_opts.variant = XDL_MERGE_FAVOR_OURS;
1041                         break;
1042                 case MERGE_RECURSIVE_THEIRS:
1043                         ll_opts.variant = XDL_MERGE_FAVOR_THEIRS;
1044                         break;
1045                 default:
1046                         ll_opts.variant = 0;
1047                         break;
1048                 }
1049         }
1050
1051         assert(a->path && b->path && o->path && opt->ancestor);
1052         if (strcmp(a->path, b->path) || strcmp(a->path, o->path) != 0) {
1053                 base  = mkpathdup("%s:%s", opt->ancestor, o->path);
1054                 name1 = mkpathdup("%s:%s", branch1, a->path);
1055                 name2 = mkpathdup("%s:%s", branch2, b->path);
1056         } else {
1057                 base  = mkpathdup("%s", opt->ancestor);
1058                 name1 = mkpathdup("%s", branch1);
1059                 name2 = mkpathdup("%s", branch2);
1060         }
1061
1062         read_mmblob(&orig, &o->oid);
1063         read_mmblob(&src1, &a->oid);
1064         read_mmblob(&src2, &b->oid);
1065
1066         merge_status = ll_merge(result_buf, a->path, &orig, base,
1067                                 &src1, name1, &src2, name2,
1068                                 opt->repo->index, &ll_opts);
1069
1070         free(base);
1071         free(name1);
1072         free(name2);
1073         free(orig.ptr);
1074         free(src1.ptr);
1075         free(src2.ptr);
1076         return merge_status;
1077 }
1078
1079 static int find_first_merges(struct repository *repo,
1080                              struct object_array *result, const char *path,
1081                              struct commit *a, struct commit *b)
1082 {
1083         int i, j;
1084         struct object_array merges = OBJECT_ARRAY_INIT;
1085         struct commit *commit;
1086         int contains_another;
1087
1088         char merged_revision[GIT_MAX_HEXSZ + 2];
1089         const char *rev_args[] = { "rev-list", "--merges", "--ancestry-path",
1090                                    "--all", merged_revision, NULL };
1091         struct rev_info revs;
1092         struct setup_revision_opt rev_opts;
1093
1094         memset(result, 0, sizeof(struct object_array));
1095         memset(&rev_opts, 0, sizeof(rev_opts));
1096
1097         /* get all revisions that merge commit a */
1098         xsnprintf(merged_revision, sizeof(merged_revision), "^%s",
1099                   oid_to_hex(&a->object.oid));
1100         repo_init_revisions(repo, &revs, NULL);
1101         rev_opts.submodule = path;
1102         /* FIXME: can't handle linked worktrees in submodules yet */
1103         revs.single_worktree = path != NULL;
1104         setup_revisions(ARRAY_SIZE(rev_args)-1, rev_args, &revs, &rev_opts);
1105
1106         /* save all revisions from the above list that contain b */
1107         if (prepare_revision_walk(&revs))
1108                 die("revision walk setup failed");
1109         while ((commit = get_revision(&revs)) != NULL) {
1110                 struct object *o = &(commit->object);
1111                 if (in_merge_bases(b, commit))
1112                         add_object_array(o, NULL, &merges);
1113         }
1114         reset_revision_walk();
1115
1116         /* Now we've got all merges that contain a and b. Prune all
1117          * merges that contain another found merge and save them in
1118          * result.
1119          */
1120         for (i = 0; i < merges.nr; i++) {
1121                 struct commit *m1 = (struct commit *) merges.objects[i].item;
1122
1123                 contains_another = 0;
1124                 for (j = 0; j < merges.nr; j++) {
1125                         struct commit *m2 = (struct commit *) merges.objects[j].item;
1126                         if (i != j && in_merge_bases(m2, m1)) {
1127                                 contains_another = 1;
1128                                 break;
1129                         }
1130                 }
1131
1132                 if (!contains_another)
1133                         add_object_array(merges.objects[i].item, NULL, result);
1134         }
1135
1136         object_array_clear(&merges);
1137         return result->nr;
1138 }
1139
1140 static void print_commit(struct commit *commit)
1141 {
1142         struct strbuf sb = STRBUF_INIT;
1143         struct pretty_print_context ctx = {0};
1144         ctx.date_mode.type = DATE_NORMAL;
1145         format_commit_message(commit, " %h: %m %s", &sb, &ctx);
1146         fprintf(stderr, "%s\n", sb.buf);
1147         strbuf_release(&sb);
1148 }
1149
1150 static int is_valid(const struct diff_filespec *dfs)
1151 {
1152         return dfs->mode != 0 && !is_null_oid(&dfs->oid);
1153 }
1154
1155 static int merge_submodule(struct merge_options *opt,
1156                            struct object_id *result, const char *path,
1157                            const struct object_id *base, const struct object_id *a,
1158                            const struct object_id *b)
1159 {
1160         struct commit *commit_base, *commit_a, *commit_b;
1161         int parent_count;
1162         struct object_array merges;
1163
1164         int i;
1165         int search = !opt->call_depth;
1166
1167         /* store a in result in case we fail */
1168         oidcpy(result, a);
1169
1170         /* we can not handle deletion conflicts */
1171         if (is_null_oid(base))
1172                 return 0;
1173         if (is_null_oid(a))
1174                 return 0;
1175         if (is_null_oid(b))
1176                 return 0;
1177
1178         if (add_submodule_odb(path)) {
1179                 output(opt, 1, _("Failed to merge submodule %s (not checked out)"), path);
1180                 return 0;
1181         }
1182
1183         if (!(commit_base = lookup_commit_reference(opt->repo, base)) ||
1184             !(commit_a = lookup_commit_reference(opt->repo, a)) ||
1185             !(commit_b = lookup_commit_reference(opt->repo, b))) {
1186                 output(opt, 1, _("Failed to merge submodule %s (commits not present)"), path);
1187                 return 0;
1188         }
1189
1190         /* check whether both changes are forward */
1191         if (!in_merge_bases(commit_base, commit_a) ||
1192             !in_merge_bases(commit_base, commit_b)) {
1193                 output(opt, 1, _("Failed to merge submodule %s (commits don't follow merge-base)"), path);
1194                 return 0;
1195         }
1196
1197         /* Case #1: a is contained in b or vice versa */
1198         if (in_merge_bases(commit_a, commit_b)) {
1199                 oidcpy(result, b);
1200                 if (show(opt, 3)) {
1201                         output(opt, 3, _("Fast-forwarding submodule %s to the following commit:"), path);
1202                         output_commit_title(opt, commit_b);
1203                 } else if (show(opt, 2))
1204                         output(opt, 2, _("Fast-forwarding submodule %s"), path);
1205                 else
1206                         ; /* no output */
1207
1208                 return 1;
1209         }
1210         if (in_merge_bases(commit_b, commit_a)) {
1211                 oidcpy(result, a);
1212                 if (show(opt, 3)) {
1213                         output(opt, 3, _("Fast-forwarding submodule %s to the following commit:"), path);
1214                         output_commit_title(opt, commit_a);
1215                 } else if (show(opt, 2))
1216                         output(opt, 2, _("Fast-forwarding submodule %s"), path);
1217                 else
1218                         ; /* no output */
1219
1220                 return 1;
1221         }
1222
1223         /*
1224          * Case #2: There are one or more merges that contain a and b in
1225          * the submodule. If there is only one, then present it as a
1226          * suggestion to the user, but leave it marked unmerged so the
1227          * user needs to confirm the resolution.
1228          */
1229
1230         /* Skip the search if makes no sense to the calling context.  */
1231         if (!search)
1232                 return 0;
1233
1234         /* find commit which merges them */
1235         parent_count = find_first_merges(opt->repo, &merges, path,
1236                                          commit_a, commit_b);
1237         switch (parent_count) {
1238         case 0:
1239                 output(opt, 1, _("Failed to merge submodule %s (merge following commits not found)"), path);
1240                 break;
1241
1242         case 1:
1243                 output(opt, 1, _("Failed to merge submodule %s (not fast-forward)"), path);
1244                 output(opt, 2, _("Found a possible merge resolution for the submodule:\n"));
1245                 print_commit((struct commit *) merges.objects[0].item);
1246                 output(opt, 2, _(
1247                        "If this is correct simply add it to the index "
1248                        "for example\n"
1249                        "by using:\n\n"
1250                        "  git update-index --cacheinfo 160000 %s \"%s\"\n\n"
1251                        "which will accept this suggestion.\n"),
1252                        oid_to_hex(&merges.objects[0].item->oid), path);
1253                 break;
1254
1255         default:
1256                 output(opt, 1, _("Failed to merge submodule %s (multiple merges found)"), path);
1257                 for (i = 0; i < merges.nr; i++)
1258                         print_commit((struct commit *) merges.objects[i].item);
1259         }
1260
1261         object_array_clear(&merges);
1262         return 0;
1263 }
1264
1265 static int merge_mode_and_contents(struct merge_options *opt,
1266                                    const struct diff_filespec *o,
1267                                    const struct diff_filespec *a,
1268                                    const struct diff_filespec *b,
1269                                    const char *filename,
1270                                    const char *branch1,
1271                                    const char *branch2,
1272                                    const int extra_marker_size,
1273                                    struct merge_file_info *result)
1274 {
1275         if (opt->branch1 != branch1) {
1276                 /*
1277                  * It's weird getting a reverse merge with HEAD on the bottom
1278                  * side of the conflict markers and the other branch on the
1279                  * top.  Fix that.
1280                  */
1281                 return merge_mode_and_contents(opt, o, b, a,
1282                                                filename,
1283                                                branch2, branch1,
1284                                                extra_marker_size, result);
1285         }
1286
1287         result->merge = 0;
1288         result->clean = 1;
1289
1290         if ((S_IFMT & a->mode) != (S_IFMT & b->mode)) {
1291                 result->clean = 0;
1292                 if (S_ISREG(a->mode)) {
1293                         result->blob.mode = a->mode;
1294                         oidcpy(&result->blob.oid, &a->oid);
1295                 } else {
1296                         result->blob.mode = b->mode;
1297                         oidcpy(&result->blob.oid, &b->oid);
1298                 }
1299         } else {
1300                 if (!oid_eq(&a->oid, &o->oid) && !oid_eq(&b->oid, &o->oid))
1301                         result->merge = 1;
1302
1303                 /*
1304                  * Merge modes
1305                  */
1306                 if (a->mode == b->mode || a->mode == o->mode)
1307                         result->blob.mode = b->mode;
1308                 else {
1309                         result->blob.mode = a->mode;
1310                         if (b->mode != o->mode) {
1311                                 result->clean = 0;
1312                                 result->merge = 1;
1313                         }
1314                 }
1315
1316                 if (oid_eq(&a->oid, &b->oid) || oid_eq(&a->oid, &o->oid))
1317                         oidcpy(&result->blob.oid, &b->oid);
1318                 else if (oid_eq(&b->oid, &o->oid))
1319                         oidcpy(&result->blob.oid, &a->oid);
1320                 else if (S_ISREG(a->mode)) {
1321                         mmbuffer_t result_buf;
1322                         int ret = 0, merge_status;
1323
1324                         merge_status = merge_3way(opt, &result_buf, o, a, b,
1325                                                   branch1, branch2,
1326                                                   extra_marker_size);
1327
1328                         if ((merge_status < 0) || !result_buf.ptr)
1329                                 ret = err(opt, _("Failed to execute internal merge"));
1330
1331                         if (!ret &&
1332                             write_object_file(result_buf.ptr, result_buf.size,
1333                                               blob_type, &result->blob.oid))
1334                                 ret = err(opt, _("Unable to add %s to database"),
1335                                           a->path);
1336
1337                         free(result_buf.ptr);
1338                         if (ret)
1339                                 return ret;
1340                         result->clean = (merge_status == 0);
1341                 } else if (S_ISGITLINK(a->mode)) {
1342                         result->clean = merge_submodule(opt, &result->blob.oid,
1343                                                         o->path,
1344                                                         &o->oid,
1345                                                         &a->oid,
1346                                                         &b->oid);
1347                 } else if (S_ISLNK(a->mode)) {
1348                         switch (opt->recursive_variant) {
1349                         case MERGE_RECURSIVE_NORMAL:
1350                                 oidcpy(&result->blob.oid, &a->oid);
1351                                 if (!oid_eq(&a->oid, &b->oid))
1352                                         result->clean = 0;
1353                                 break;
1354                         case MERGE_RECURSIVE_OURS:
1355                                 oidcpy(&result->blob.oid, &a->oid);
1356                                 break;
1357                         case MERGE_RECURSIVE_THEIRS:
1358                                 oidcpy(&result->blob.oid, &b->oid);
1359                                 break;
1360                         }
1361                 } else
1362                         BUG("unsupported object type in the tree");
1363         }
1364
1365         if (result->merge)
1366                 output(opt, 2, _("Auto-merging %s"), filename);
1367
1368         return 0;
1369 }
1370
1371 static int handle_rename_via_dir(struct merge_options *opt,
1372                                  struct rename_conflict_info *ci)
1373 {
1374         /*
1375          * Handle file adds that need to be renamed due to directory rename
1376          * detection.  This differs from handle_rename_normal, because
1377          * there is no content merge to do; just move the file into the
1378          * desired final location.
1379          */
1380         const struct rename *ren = ci->ren1;
1381         const struct diff_filespec *dest = ren->pair->two;
1382         char *file_path = dest->path;
1383         int mark_conflicted = (opt->detect_directory_renames ==
1384                                MERGE_DIRECTORY_RENAMES_CONFLICT);
1385         assert(ren->dir_rename_original_dest);
1386
1387         if (!opt->call_depth && would_lose_untracked(opt, dest->path)) {
1388                 mark_conflicted = 1;
1389                 file_path = unique_path(opt, dest->path, ren->branch);
1390                 output(opt, 1, _("Error: Refusing to lose untracked file at %s; "
1391                                  "writing to %s instead."),
1392                        dest->path, file_path);
1393         }
1394
1395         if (mark_conflicted) {
1396                 /*
1397                  * Write the file in worktree at file_path.  In the index,
1398                  * only record the file at dest->path in the appropriate
1399                  * higher stage.
1400                  */
1401                 if (update_file(opt, 0, dest, file_path))
1402                         return -1;
1403                 if (file_path != dest->path)
1404                         free(file_path);
1405                 if (update_stages(opt, dest->path, NULL,
1406                                   ren->branch == opt->branch1 ? dest : NULL,
1407                                   ren->branch == opt->branch1 ? NULL : dest))
1408                         return -1;
1409                 return 0; /* not clean, but conflicted */
1410         } else {
1411                 /* Update dest->path both in index and in worktree */
1412                 if (update_file(opt, 1, dest, dest->path))
1413                         return -1;
1414                 return 1; /* clean */
1415         }
1416 }
1417
1418 static int handle_change_delete(struct merge_options *opt,
1419                                 const char *path, const char *old_path,
1420                                 const struct diff_filespec *o,
1421                                 const struct diff_filespec *changed,
1422                                 const char *change_branch,
1423                                 const char *delete_branch,
1424                                 const char *change, const char *change_past)
1425 {
1426         char *alt_path = NULL;
1427         const char *update_path = path;
1428         int ret = 0;
1429
1430         if (dir_in_way(opt->repo->index, path, !opt->call_depth, 0) ||
1431             (!opt->call_depth && would_lose_untracked(opt, path))) {
1432                 update_path = alt_path = unique_path(opt, path, change_branch);
1433         }
1434
1435         if (opt->call_depth) {
1436                 /*
1437                  * We cannot arbitrarily accept either a_sha or b_sha as
1438                  * correct; since there is no true "middle point" between
1439                  * them, simply reuse the base version for virtual merge base.
1440                  */
1441                 ret = remove_file_from_index(opt->repo->index, path);
1442                 if (!ret)
1443                         ret = update_file(opt, 0, o, update_path);
1444         } else {
1445                 /*
1446                  * Despite the four nearly duplicate messages and argument
1447                  * lists below and the ugliness of the nested if-statements,
1448                  * having complete messages makes the job easier for
1449                  * translators.
1450                  *
1451                  * The slight variance among the cases is due to the fact
1452                  * that:
1453                  *   1) directory/file conflicts (in effect if
1454                  *      !alt_path) could cause us to need to write the
1455                  *      file to a different path.
1456                  *   2) renames (in effect if !old_path) could mean that
1457                  *      there are two names for the path that the user
1458                  *      may know the file by.
1459                  */
1460                 if (!alt_path) {
1461                         if (!old_path) {
1462                                 output(opt, 1, _("CONFLICT (%s/delete): %s deleted in %s "
1463                                        "and %s in %s. Version %s of %s left in tree."),
1464                                        change, path, delete_branch, change_past,
1465                                        change_branch, change_branch, path);
1466                         } else {
1467                                 output(opt, 1, _("CONFLICT (%s/delete): %s deleted in %s "
1468                                        "and %s to %s in %s. Version %s of %s left in tree."),
1469                                        change, old_path, delete_branch, change_past, path,
1470                                        change_branch, change_branch, path);
1471                         }
1472                 } else {
1473                         if (!old_path) {
1474                                 output(opt, 1, _("CONFLICT (%s/delete): %s deleted in %s "
1475                                        "and %s in %s. Version %s of %s left in tree at %s."),
1476                                        change, path, delete_branch, change_past,
1477                                        change_branch, change_branch, path, alt_path);
1478                         } else {
1479                                 output(opt, 1, _("CONFLICT (%s/delete): %s deleted in %s "
1480                                        "and %s to %s in %s. Version %s of %s left in tree at %s."),
1481                                        change, old_path, delete_branch, change_past, path,
1482                                        change_branch, change_branch, path, alt_path);
1483                         }
1484                 }
1485                 /*
1486                  * No need to call update_file() on path when change_branch ==
1487                  * opt->branch1 && !alt_path, since that would needlessly touch
1488                  * path.  We could call update_file_flags() with update_cache=0
1489                  * and update_wd=0, but that's a no-op.
1490                  */
1491                 if (change_branch != opt->branch1 || alt_path)
1492                         ret = update_file(opt, 0, changed, update_path);
1493         }
1494         free(alt_path);
1495
1496         return ret;
1497 }
1498
1499 static int handle_rename_delete(struct merge_options *opt,
1500                                 struct rename_conflict_info *ci)
1501 {
1502         const struct rename *ren = ci->ren1;
1503         const struct diff_filespec *orig = ren->pair->one;
1504         const struct diff_filespec *dest = ren->pair->two;
1505         const char *rename_branch = ren->branch;
1506         const char *delete_branch = (opt->branch1 == ren->branch ?
1507                                      opt->branch2 : opt->branch1);
1508
1509         if (handle_change_delete(opt,
1510                                  opt->call_depth ? orig->path : dest->path,
1511                                  opt->call_depth ? NULL : orig->path,
1512                                  orig, dest,
1513                                  rename_branch, delete_branch,
1514                                  _("rename"), _("renamed")))
1515                 return -1;
1516
1517         if (opt->call_depth)
1518                 return remove_file_from_index(opt->repo->index, dest->path);
1519         else
1520                 return update_stages(opt, dest->path, NULL,
1521                                      rename_branch == opt->branch1 ? dest : NULL,
1522                                      rename_branch == opt->branch1 ? NULL : dest);
1523 }
1524
1525 static int handle_file_collision(struct merge_options *opt,
1526                                  const char *collide_path,
1527                                  const char *prev_path1,
1528                                  const char *prev_path2,
1529                                  const char *branch1, const char *branch2,
1530                                  struct diff_filespec *a,
1531                                  struct diff_filespec *b)
1532 {
1533         struct merge_file_info mfi;
1534         struct diff_filespec null;
1535         char *alt_path = NULL;
1536         const char *update_path = collide_path;
1537
1538         /*
1539          * It's easiest to get the correct things into stage 2 and 3, and
1540          * to make sure that the content merge puts HEAD before the other
1541          * branch if we just ensure that branch1 == opt->branch1.  So, simply
1542          * flip arguments around if we don't have that.
1543          */
1544         if (branch1 != opt->branch1) {
1545                 return handle_file_collision(opt, collide_path,
1546                                              prev_path2, prev_path1,
1547                                              branch2, branch1,
1548                                              b, a);
1549         }
1550
1551         /*
1552          * In the recursive case, we just opt to undo renames
1553          */
1554         if (opt->call_depth && (prev_path1 || prev_path2)) {
1555                 /* Put first file (a->oid, a->mode) in its original spot */
1556                 if (prev_path1) {
1557                         if (update_file(opt, 1, a, prev_path1))
1558                                 return -1;
1559                 } else {
1560                         if (update_file(opt, 1, a, collide_path))
1561                                 return -1;
1562                 }
1563
1564                 /* Put second file (b->oid, b->mode) in its original spot */
1565                 if (prev_path2) {
1566                         if (update_file(opt, 1, b, prev_path2))
1567                                 return -1;
1568                 } else {
1569                         if (update_file(opt, 1, b, collide_path))
1570                                 return -1;
1571                 }
1572
1573                 /* Don't leave something at collision path if unrenaming both */
1574                 if (prev_path1 && prev_path2)
1575                         remove_file(opt, 1, collide_path, 0);
1576
1577                 return 0;
1578         }
1579
1580         /* Remove rename sources if rename/add or rename/rename(2to1) */
1581         if (prev_path1)
1582                 remove_file(opt, 1, prev_path1,
1583                             opt->call_depth || would_lose_untracked(opt, prev_path1));
1584         if (prev_path2)
1585                 remove_file(opt, 1, prev_path2,
1586                             opt->call_depth || would_lose_untracked(opt, prev_path2));
1587
1588         /*
1589          * Remove the collision path, if it wouldn't cause dirty contents
1590          * or an untracked file to get lost.  We'll either overwrite with
1591          * merged contents, or just write out to differently named files.
1592          */
1593         if (was_dirty(opt, collide_path)) {
1594                 output(opt, 1, _("Refusing to lose dirty file at %s"),
1595                        collide_path);
1596                 update_path = alt_path = unique_path(opt, collide_path, "merged");
1597         } else if (would_lose_untracked(opt, collide_path)) {
1598                 /*
1599                  * Only way we get here is if both renames were from
1600                  * a directory rename AND user had an untracked file
1601                  * at the location where both files end up after the
1602                  * two directory renames.  See testcase 10d of t6043.
1603                  */
1604                 output(opt, 1, _("Refusing to lose untracked file at "
1605                                "%s, even though it's in the way."),
1606                        collide_path);
1607                 update_path = alt_path = unique_path(opt, collide_path, "merged");
1608         } else {
1609                 /*
1610                  * FIXME: It's possible that the two files are identical
1611                  * and that the current working copy happens to match, in
1612                  * which case we are unnecessarily touching the working
1613                  * tree file.  It's not a likely enough scenario that I
1614                  * want to code up the checks for it and a better fix is
1615                  * available if we restructure how unpack_trees() and
1616                  * merge-recursive interoperate anyway, so punting for
1617                  * now...
1618                  */
1619                 remove_file(opt, 0, collide_path, 0);
1620         }
1621
1622         /* Store things in diff_filespecs for functions that need it */
1623         null.path = (char *)collide_path;
1624         oidcpy(&null.oid, &null_oid);
1625         null.mode = 0;
1626
1627         if (merge_mode_and_contents(opt, &null, a, b, collide_path,
1628                                     branch1, branch2, opt->call_depth * 2, &mfi))
1629                 return -1;
1630         mfi.clean &= !alt_path;
1631         if (update_file(opt, mfi.clean, &mfi.blob, update_path))
1632                 return -1;
1633         if (!mfi.clean && !opt->call_depth &&
1634             update_stages(opt, collide_path, NULL, a, b))
1635                 return -1;
1636         free(alt_path);
1637         /*
1638          * FIXME: If both a & b both started with conflicts (only possible
1639          * if they came from a rename/rename(2to1)), but had IDENTICAL
1640          * contents including those conflicts, then in the next line we claim
1641          * it was clean.  If someone cares about this case, we should have the
1642          * caller notify us if we started with conflicts.
1643          */
1644         return mfi.clean;
1645 }
1646
1647 static int handle_rename_add(struct merge_options *opt,
1648                              struct rename_conflict_info *ci)
1649 {
1650         /* a was renamed to c, and a separate c was added. */
1651         struct diff_filespec *a = ci->ren1->pair->one;
1652         struct diff_filespec *c = ci->ren1->pair->two;
1653         char *path = c->path;
1654         char *prev_path_desc;
1655         struct merge_file_info mfi;
1656
1657         const char *rename_branch = ci->ren1->branch;
1658         const char *add_branch = (opt->branch1 == rename_branch ?
1659                                   opt->branch2 : opt->branch1);
1660         int other_stage = (ci->ren1->branch == opt->branch1 ? 3 : 2);
1661
1662         output(opt, 1, _("CONFLICT (rename/add): "
1663                "Rename %s->%s in %s.  Added %s in %s"),
1664                a->path, c->path, rename_branch,
1665                c->path, add_branch);
1666
1667         prev_path_desc = xstrfmt("version of %s from %s", path, a->path);
1668         ci->ren1->src_entry->stages[other_stage].path = a->path;
1669         if (merge_mode_and_contents(opt, a, c,
1670                                     &ci->ren1->src_entry->stages[other_stage],
1671                                     prev_path_desc,
1672                                     opt->branch1, opt->branch2,
1673                                     1 + opt->call_depth * 2, &mfi))
1674                 return -1;
1675         free(prev_path_desc);
1676
1677         ci->ren1->dst_entry->stages[other_stage].path = mfi.blob.path = c->path;
1678         return handle_file_collision(opt,
1679                                      c->path, a->path, NULL,
1680                                      rename_branch, add_branch,
1681                                      &mfi.blob,
1682                                      &ci->ren1->dst_entry->stages[other_stage]);
1683 }
1684
1685 static char *find_path_for_conflict(struct merge_options *opt,
1686                                     const char *path,
1687                                     const char *branch1,
1688                                     const char *branch2)
1689 {
1690         char *new_path = NULL;
1691         if (dir_in_way(opt->repo->index, path, !opt->call_depth, 0)) {
1692                 new_path = unique_path(opt, path, branch1);
1693                 output(opt, 1, _("%s is a directory in %s adding "
1694                                "as %s instead"),
1695                        path, branch2, new_path);
1696         } else if (would_lose_untracked(opt, path)) {
1697                 new_path = unique_path(opt, path, branch1);
1698                 output(opt, 1, _("Refusing to lose untracked file"
1699                                " at %s; adding as %s instead"),
1700                        path, new_path);
1701         }
1702
1703         return new_path;
1704 }
1705
1706 static int handle_rename_rename_1to2(struct merge_options *opt,
1707                                      struct rename_conflict_info *ci)
1708 {
1709         /* One file was renamed in both branches, but to different names. */
1710         struct merge_file_info mfi;
1711         struct diff_filespec *add;
1712         struct diff_filespec *o = ci->ren1->pair->one;
1713         struct diff_filespec *a = ci->ren1->pair->two;
1714         struct diff_filespec *b = ci->ren2->pair->two;
1715         char *path_desc;
1716
1717         output(opt, 1, _("CONFLICT (rename/rename): "
1718                "Rename \"%s\"->\"%s\" in branch \"%s\" "
1719                "rename \"%s\"->\"%s\" in \"%s\"%s"),
1720                o->path, a->path, ci->ren1->branch,
1721                o->path, b->path, ci->ren2->branch,
1722                opt->call_depth ? _(" (left unresolved)") : "");
1723
1724         path_desc = xstrfmt("%s and %s, both renamed from %s",
1725                             a->path, b->path, o->path);
1726         if (merge_mode_and_contents(opt, o, a, b, path_desc,
1727                                     ci->ren1->branch, ci->ren2->branch,
1728                                     opt->call_depth * 2, &mfi))
1729                 return -1;
1730         free(path_desc);
1731
1732         if (opt->call_depth) {
1733                 /*
1734                  * FIXME: For rename/add-source conflicts (if we could detect
1735                  * such), this is wrong.  We should instead find a unique
1736                  * pathname and then either rename the add-source file to that
1737                  * unique path, or use that unique path instead of src here.
1738                  */
1739                 if (update_file(opt, 0, &mfi.blob, o->path))
1740                         return -1;
1741
1742                 /*
1743                  * Above, we put the merged content at the merge-base's
1744                  * path.  Now we usually need to delete both a->path and
1745                  * b->path.  However, the rename on each side of the merge
1746                  * could also be involved in a rename/add conflict.  In
1747                  * such cases, we should keep the added file around,
1748                  * resolving the conflict at that path in its favor.
1749                  */
1750                 add = &ci->ren1->dst_entry->stages[2 ^ 1];
1751                 if (is_valid(add)) {
1752                         if (update_file(opt, 0, add, a->path))
1753                                 return -1;
1754                 }
1755                 else
1756                         remove_file_from_index(opt->repo->index, a->path);
1757                 add = &ci->ren2->dst_entry->stages[3 ^ 1];
1758                 if (is_valid(add)) {
1759                         if (update_file(opt, 0, add, b->path))
1760                                 return -1;
1761                 }
1762                 else
1763                         remove_file_from_index(opt->repo->index, b->path);
1764         } else {
1765                 /*
1766                  * For each destination path, we need to see if there is a
1767                  * rename/add collision.  If not, we can write the file out
1768                  * to the specified location.
1769                  */
1770                 add = &ci->ren1->dst_entry->stages[2 ^ 1];
1771                 if (is_valid(add)) {
1772                         add->path = mfi.blob.path = a->path;
1773                         if (handle_file_collision(opt, a->path,
1774                                                   NULL, NULL,
1775                                                   ci->ren1->branch,
1776                                                   ci->ren2->branch,
1777                                                   &mfi.blob, add) < 0)
1778                                 return -1;
1779                 } else {
1780                         char *new_path = find_path_for_conflict(opt, a->path,
1781                                                                 ci->ren1->branch,
1782                                                                 ci->ren2->branch);
1783                         if (update_file(opt, 0, &mfi.blob,
1784                                         new_path ? new_path : a->path))
1785                                 return -1;
1786                         free(new_path);
1787                         if (update_stages(opt, a->path, NULL, a, NULL))
1788                                 return -1;
1789                 }
1790
1791                 add = &ci->ren2->dst_entry->stages[3 ^ 1];
1792                 if (is_valid(add)) {
1793                         add->path = mfi.blob.path = b->path;
1794                         if (handle_file_collision(opt, b->path,
1795                                                   NULL, NULL,
1796                                                   ci->ren1->branch,
1797                                                   ci->ren2->branch,
1798                                                   add, &mfi.blob) < 0)
1799                                 return -1;
1800                 } else {
1801                         char *new_path = find_path_for_conflict(opt, b->path,
1802                                                                 ci->ren2->branch,
1803                                                                 ci->ren1->branch);
1804                         if (update_file(opt, 0, &mfi.blob,
1805                                         new_path ? new_path : b->path))
1806                                 return -1;
1807                         free(new_path);
1808                         if (update_stages(opt, b->path, NULL, NULL, b))
1809                                 return -1;
1810                 }
1811         }
1812
1813         return 0;
1814 }
1815
1816 static int handle_rename_rename_2to1(struct merge_options *opt,
1817                                      struct rename_conflict_info *ci)
1818 {
1819         /* Two files, a & b, were renamed to the same thing, c. */
1820         struct diff_filespec *a = ci->ren1->pair->one;
1821         struct diff_filespec *b = ci->ren2->pair->one;
1822         struct diff_filespec *c1 = ci->ren1->pair->two;
1823         struct diff_filespec *c2 = ci->ren2->pair->two;
1824         char *path = c1->path; /* == c2->path */
1825         char *path_side_1_desc;
1826         char *path_side_2_desc;
1827         struct merge_file_info mfi_c1;
1828         struct merge_file_info mfi_c2;
1829         int ostage1, ostage2;
1830
1831         output(opt, 1, _("CONFLICT (rename/rename): "
1832                "Rename %s->%s in %s. "
1833                "Rename %s->%s in %s"),
1834                a->path, c1->path, ci->ren1->branch,
1835                b->path, c2->path, ci->ren2->branch);
1836
1837         path_side_1_desc = xstrfmt("version of %s from %s", path, a->path);
1838         path_side_2_desc = xstrfmt("version of %s from %s", path, b->path);
1839         ostage1 = ci->ren1->branch == opt->branch1 ? 3 : 2;
1840         ostage2 = ostage1 ^ 1;
1841         ci->ren1->src_entry->stages[ostage1].path = a->path;
1842         ci->ren2->src_entry->stages[ostage2].path = b->path;
1843         if (merge_mode_and_contents(opt, a, c1,
1844                                     &ci->ren1->src_entry->stages[ostage1],
1845                                     path_side_1_desc,
1846                                     opt->branch1, opt->branch2,
1847                                     1 + opt->call_depth * 2, &mfi_c1) ||
1848             merge_mode_and_contents(opt, b,
1849                                     &ci->ren2->src_entry->stages[ostage2],
1850                                     c2, path_side_2_desc,
1851                                     opt->branch1, opt->branch2,
1852                                     1 + opt->call_depth * 2, &mfi_c2))
1853                 return -1;
1854         free(path_side_1_desc);
1855         free(path_side_2_desc);
1856         mfi_c1.blob.path = path;
1857         mfi_c2.blob.path = path;
1858
1859         return handle_file_collision(opt, path, a->path, b->path,
1860                                      ci->ren1->branch, ci->ren2->branch,
1861                                      &mfi_c1.blob, &mfi_c2.blob);
1862 }
1863
1864 /*
1865  * Get the diff_filepairs changed between o_tree and tree.
1866  */
1867 static struct diff_queue_struct *get_diffpairs(struct merge_options *opt,
1868                                                struct tree *o_tree,
1869                                                struct tree *tree)
1870 {
1871         struct diff_queue_struct *ret;
1872         struct diff_options opts;
1873
1874         repo_diff_setup(opt->repo, &opts);
1875         opts.flags.recursive = 1;
1876         opts.flags.rename_empty = 0;
1877         opts.detect_rename = merge_detect_rename(opt);
1878         /*
1879          * We do not have logic to handle the detection of copies.  In
1880          * fact, it may not even make sense to add such logic: would we
1881          * really want a change to a base file to be propagated through
1882          * multiple other files by a merge?
1883          */
1884         if (opts.detect_rename > DIFF_DETECT_RENAME)
1885                 opts.detect_rename = DIFF_DETECT_RENAME;
1886         opts.rename_limit = opt->merge_rename_limit >= 0 ? opt->merge_rename_limit :
1887                             opt->diff_rename_limit >= 0 ? opt->diff_rename_limit :
1888                             1000;
1889         opts.rename_score = opt->rename_score;
1890         opts.show_rename_progress = opt->show_rename_progress;
1891         opts.output_format = DIFF_FORMAT_NO_OUTPUT;
1892         diff_setup_done(&opts);
1893         diff_tree_oid(&o_tree->object.oid, &tree->object.oid, "", &opts);
1894         diffcore_std(&opts);
1895         if (opts.needed_rename_limit > opt->needed_rename_limit)
1896                 opt->needed_rename_limit = opts.needed_rename_limit;
1897
1898         ret = xmalloc(sizeof(*ret));
1899         *ret = diff_queued_diff;
1900
1901         opts.output_format = DIFF_FORMAT_NO_OUTPUT;
1902         diff_queued_diff.nr = 0;
1903         diff_queued_diff.queue = NULL;
1904         diff_flush(&opts);
1905         return ret;
1906 }
1907
1908 static int tree_has_path(struct repository *r, struct tree *tree,
1909                          const char *path)
1910 {
1911         struct object_id hashy;
1912         unsigned short mode_o;
1913
1914         return !get_tree_entry(r,
1915                                &tree->object.oid, path,
1916                                &hashy, &mode_o);
1917 }
1918
1919 /*
1920  * Return a new string that replaces the beginning portion (which matches
1921  * entry->dir), with entry->new_dir.  In perl-speak:
1922  *   new_path_name = (old_path =~ s/entry->dir/entry->new_dir/);
1923  * NOTE:
1924  *   Caller must ensure that old_path starts with entry->dir + '/'.
1925  */
1926 static char *apply_dir_rename(struct dir_rename_entry *entry,
1927                               const char *old_path)
1928 {
1929         struct strbuf new_path = STRBUF_INIT;
1930         int oldlen, newlen;
1931
1932         if (entry->non_unique_new_dir)
1933                 return NULL;
1934
1935         oldlen = strlen(entry->dir);
1936         newlen = entry->new_dir.len + (strlen(old_path) - oldlen) + 1;
1937         strbuf_grow(&new_path, newlen);
1938         strbuf_addbuf(&new_path, &entry->new_dir);
1939         strbuf_addstr(&new_path, &old_path[oldlen]);
1940
1941         return strbuf_detach(&new_path, NULL);
1942 }
1943
1944 static void get_renamed_dir_portion(const char *old_path, const char *new_path,
1945                                     char **old_dir, char **new_dir)
1946 {
1947         char *end_of_old, *end_of_new;
1948         int old_len, new_len;
1949
1950         *old_dir = NULL;
1951         *new_dir = NULL;
1952
1953         /*
1954          * For
1955          *    "a/b/c/d/e/foo.c" -> "a/b/some/thing/else/e/foo.c"
1956          * the "e/foo.c" part is the same, we just want to know that
1957          *    "a/b/c/d" was renamed to "a/b/some/thing/else"
1958          * so, for this example, this function returns "a/b/c/d" in
1959          * *old_dir and "a/b/some/thing/else" in *new_dir.
1960          *
1961          * Also, if the basename of the file changed, we don't care.  We
1962          * want to know which portion of the directory, if any, changed.
1963          */
1964         end_of_old = strrchr(old_path, '/');
1965         end_of_new = strrchr(new_path, '/');
1966
1967         if (end_of_old == NULL || end_of_new == NULL)
1968                 return;
1969         while (*--end_of_new == *--end_of_old &&
1970                end_of_old != old_path &&
1971                end_of_new != new_path)
1972                 ; /* Do nothing; all in the while loop */
1973         /*
1974          * We've found the first non-matching character in the directory
1975          * paths.  That means the current directory we were comparing
1976          * represents the rename.  Move end_of_old and end_of_new back
1977          * to the full directory name.
1978          */
1979         if (*end_of_old == '/')
1980                 end_of_old++;
1981         if (*end_of_old != '/')
1982                 end_of_new++;
1983         end_of_old = strchr(end_of_old, '/');
1984         end_of_new = strchr(end_of_new, '/');
1985
1986         /*
1987          * It may have been the case that old_path and new_path were the same
1988          * directory all along.  Don't claim a rename if they're the same.
1989          */
1990         old_len = end_of_old - old_path;
1991         new_len = end_of_new - new_path;
1992
1993         if (old_len != new_len || strncmp(old_path, new_path, old_len)) {
1994                 *old_dir = xstrndup(old_path, old_len);
1995                 *new_dir = xstrndup(new_path, new_len);
1996         }
1997 }
1998
1999 static void remove_hashmap_entries(struct hashmap *dir_renames,
2000                                    struct string_list *items_to_remove)
2001 {
2002         int i;
2003         struct dir_rename_entry *entry;
2004
2005         for (i = 0; i < items_to_remove->nr; i++) {
2006                 entry = items_to_remove->items[i].util;
2007                 hashmap_remove(dir_renames, entry, NULL);
2008         }
2009         string_list_clear(items_to_remove, 0);
2010 }
2011
2012 /*
2013  * See if there is a directory rename for path, and if there are any file
2014  * level conflicts for the renamed location.  If there is a rename and
2015  * there are no conflicts, return the new name.  Otherwise, return NULL.
2016  */
2017 static char *handle_path_level_conflicts(struct merge_options *opt,
2018                                          const char *path,
2019                                          struct dir_rename_entry *entry,
2020                                          struct hashmap *collisions,
2021                                          struct tree *tree)
2022 {
2023         char *new_path = NULL;
2024         struct collision_entry *collision_ent;
2025         int clean = 1;
2026         struct strbuf collision_paths = STRBUF_INIT;
2027
2028         /*
2029          * entry has the mapping of old directory name to new directory name
2030          * that we want to apply to path.
2031          */
2032         new_path = apply_dir_rename(entry, path);
2033
2034         if (!new_path) {
2035                 /* This should only happen when entry->non_unique_new_dir set */
2036                 if (!entry->non_unique_new_dir)
2037                         BUG("entry->non_unqiue_dir not set and !new_path");
2038                 output(opt, 1, _("CONFLICT (directory rename split): "
2039                                "Unclear where to place %s because directory "
2040                                "%s was renamed to multiple other directories, "
2041                                "with no destination getting a majority of the "
2042                                "files."),
2043                        path, entry->dir);
2044                 clean = 0;
2045                 return NULL;
2046         }
2047
2048         /*
2049          * The caller needs to have ensured that it has pre-populated
2050          * collisions with all paths that map to new_path.  Do a quick check
2051          * to ensure that's the case.
2052          */
2053         collision_ent = collision_find_entry(collisions, new_path);
2054         if (collision_ent == NULL)
2055                 BUG("collision_ent is NULL");
2056
2057         /*
2058          * Check for one-sided add/add/.../add conflicts, i.e.
2059          * where implicit renames from the other side doing
2060          * directory rename(s) can affect this side of history
2061          * to put multiple paths into the same location.  Warn
2062          * and bail on directory renames for such paths.
2063          */
2064         if (collision_ent->reported_already) {
2065                 clean = 0;
2066         } else if (tree_has_path(opt->repo, tree, new_path)) {
2067                 collision_ent->reported_already = 1;
2068                 strbuf_add_separated_string_list(&collision_paths, ", ",
2069                                                  &collision_ent->source_files);
2070                 output(opt, 1, _("CONFLICT (implicit dir rename): Existing "
2071                                "file/dir at %s in the way of implicit "
2072                                "directory rename(s) putting the following "
2073                                "path(s) there: %s."),
2074                        new_path, collision_paths.buf);
2075                 clean = 0;
2076         } else if (collision_ent->source_files.nr > 1) {
2077                 collision_ent->reported_already = 1;
2078                 strbuf_add_separated_string_list(&collision_paths, ", ",
2079                                                  &collision_ent->source_files);
2080                 output(opt, 1, _("CONFLICT (implicit dir rename): Cannot map "
2081                                "more than one path to %s; implicit directory "
2082                                "renames tried to put these paths there: %s"),
2083                        new_path, collision_paths.buf);
2084                 clean = 0;
2085         }
2086
2087         /* Free memory we no longer need */
2088         strbuf_release(&collision_paths);
2089         if (!clean && new_path) {
2090                 free(new_path);
2091                 return NULL;
2092         }
2093
2094         return new_path;
2095 }
2096
2097 /*
2098  * There are a couple things we want to do at the directory level:
2099  *   1. Check for both sides renaming to the same thing, in order to avoid
2100  *      implicit renaming of files that should be left in place.  (See
2101  *      testcase 6b in t6043 for details.)
2102  *   2. Prune directory renames if there are still files left in the
2103  *      the original directory.  These represent a partial directory rename,
2104  *      i.e. a rename where only some of the files within the directory
2105  *      were renamed elsewhere.  (Technically, this could be done earlier
2106  *      in get_directory_renames(), except that would prevent us from
2107  *      doing the previous check and thus failing testcase 6b.)
2108  *   3. Check for rename/rename(1to2) conflicts (at the directory level).
2109  *      In the future, we could potentially record this info as well and
2110  *      omit reporting rename/rename(1to2) conflicts for each path within
2111  *      the affected directories, thus cleaning up the merge output.
2112  *   NOTE: We do NOT check for rename/rename(2to1) conflicts at the
2113  *         directory level, because merging directories is fine.  If it
2114  *         causes conflicts for files within those merged directories, then
2115  *         that should be detected at the individual path level.
2116  */
2117 static void handle_directory_level_conflicts(struct merge_options *opt,
2118                                              struct hashmap *dir_re_head,
2119                                              struct tree *head,
2120                                              struct hashmap *dir_re_merge,
2121                                              struct tree *merge)
2122 {
2123         struct hashmap_iter iter;
2124         struct dir_rename_entry *head_ent;
2125         struct dir_rename_entry *merge_ent;
2126
2127         struct string_list remove_from_head = STRING_LIST_INIT_NODUP;
2128         struct string_list remove_from_merge = STRING_LIST_INIT_NODUP;
2129
2130         hashmap_iter_init(dir_re_head, &iter);
2131         while ((head_ent = hashmap_iter_next(&iter))) {
2132                 merge_ent = dir_rename_find_entry(dir_re_merge, head_ent->dir);
2133                 if (merge_ent &&
2134                     !head_ent->non_unique_new_dir &&
2135                     !merge_ent->non_unique_new_dir &&
2136                     !strbuf_cmp(&head_ent->new_dir, &merge_ent->new_dir)) {
2137                         /* 1. Renamed identically; remove it from both sides */
2138                         string_list_append(&remove_from_head,
2139                                            head_ent->dir)->util = head_ent;
2140                         strbuf_release(&head_ent->new_dir);
2141                         string_list_append(&remove_from_merge,
2142                                            merge_ent->dir)->util = merge_ent;
2143                         strbuf_release(&merge_ent->new_dir);
2144                 } else if (tree_has_path(opt->repo, head, head_ent->dir)) {
2145                         /* 2. This wasn't a directory rename after all */
2146                         string_list_append(&remove_from_head,
2147                                            head_ent->dir)->util = head_ent;
2148                         strbuf_release(&head_ent->new_dir);
2149                 }
2150         }
2151
2152         remove_hashmap_entries(dir_re_head, &remove_from_head);
2153         remove_hashmap_entries(dir_re_merge, &remove_from_merge);
2154
2155         hashmap_iter_init(dir_re_merge, &iter);
2156         while ((merge_ent = hashmap_iter_next(&iter))) {
2157                 head_ent = dir_rename_find_entry(dir_re_head, merge_ent->dir);
2158                 if (tree_has_path(opt->repo, merge, merge_ent->dir)) {
2159                         /* 2. This wasn't a directory rename after all */
2160                         string_list_append(&remove_from_merge,
2161                                            merge_ent->dir)->util = merge_ent;
2162                 } else if (head_ent &&
2163                            !head_ent->non_unique_new_dir &&
2164                            !merge_ent->non_unique_new_dir) {
2165                         /* 3. rename/rename(1to2) */
2166                         /*
2167                          * We can assume it's not rename/rename(1to1) because
2168                          * that was case (1), already checked above.  So we
2169                          * know that head_ent->new_dir and merge_ent->new_dir
2170                          * are different strings.
2171                          */
2172                         output(opt, 1, _("CONFLICT (rename/rename): "
2173                                        "Rename directory %s->%s in %s. "
2174                                        "Rename directory %s->%s in %s"),
2175                                head_ent->dir, head_ent->new_dir.buf, opt->branch1,
2176                                head_ent->dir, merge_ent->new_dir.buf, opt->branch2);
2177                         string_list_append(&remove_from_head,
2178                                            head_ent->dir)->util = head_ent;
2179                         strbuf_release(&head_ent->new_dir);
2180                         string_list_append(&remove_from_merge,
2181                                            merge_ent->dir)->util = merge_ent;
2182                         strbuf_release(&merge_ent->new_dir);
2183                 }
2184         }
2185
2186         remove_hashmap_entries(dir_re_head, &remove_from_head);
2187         remove_hashmap_entries(dir_re_merge, &remove_from_merge);
2188 }
2189
2190 static struct hashmap *get_directory_renames(struct diff_queue_struct *pairs)
2191 {
2192         struct hashmap *dir_renames;
2193         struct hashmap_iter iter;
2194         struct dir_rename_entry *entry;
2195         int i;
2196
2197         /*
2198          * Typically, we think of a directory rename as all files from a
2199          * certain directory being moved to a target directory.  However,
2200          * what if someone first moved two files from the original
2201          * directory in one commit, and then renamed the directory
2202          * somewhere else in a later commit?  At merge time, we just know
2203          * that files from the original directory went to two different
2204          * places, and that the bulk of them ended up in the same place.
2205          * We want each directory rename to represent where the bulk of the
2206          * files from that directory end up; this function exists to find
2207          * where the bulk of the files went.
2208          *
2209          * The first loop below simply iterates through the list of file
2210          * renames, finding out how often each directory rename pair
2211          * possibility occurs.
2212          */
2213         dir_renames = xmalloc(sizeof(*dir_renames));
2214         dir_rename_init(dir_renames);
2215         for (i = 0; i < pairs->nr; ++i) {
2216                 struct string_list_item *item;
2217                 int *count;
2218                 struct diff_filepair *pair = pairs->queue[i];
2219                 char *old_dir, *new_dir;
2220
2221                 /* File not part of directory rename if it wasn't renamed */
2222                 if (pair->status != 'R')
2223                         continue;
2224
2225                 get_renamed_dir_portion(pair->one->path, pair->two->path,
2226                                         &old_dir,        &new_dir);
2227                 if (!old_dir)
2228                         /* Directory didn't change at all; ignore this one. */
2229                         continue;
2230
2231                 entry = dir_rename_find_entry(dir_renames, old_dir);
2232                 if (!entry) {
2233                         entry = xmalloc(sizeof(*entry));
2234                         dir_rename_entry_init(entry, old_dir);
2235                         hashmap_put(dir_renames, entry);
2236                 } else {
2237                         free(old_dir);
2238                 }
2239                 item = string_list_lookup(&entry->possible_new_dirs, new_dir);
2240                 if (!item) {
2241                         item = string_list_insert(&entry->possible_new_dirs,
2242                                                   new_dir);
2243                         item->util = xcalloc(1, sizeof(int));
2244                 } else {
2245                         free(new_dir);
2246                 }
2247                 count = item->util;
2248                 *count += 1;
2249         }
2250
2251         /*
2252          * For each directory with files moved out of it, we find out which
2253          * target directory received the most files so we can declare it to
2254          * be the "winning" target location for the directory rename.  This
2255          * winner gets recorded in new_dir.  If there is no winner
2256          * (multiple target directories received the same number of files),
2257          * we set non_unique_new_dir.  Once we've determined the winner (or
2258          * that there is no winner), we no longer need possible_new_dirs.
2259          */
2260         hashmap_iter_init(dir_renames, &iter);
2261         while ((entry = hashmap_iter_next(&iter))) {
2262                 int max = 0;
2263                 int bad_max = 0;
2264                 char *best = NULL;
2265
2266                 for (i = 0; i < entry->possible_new_dirs.nr; i++) {
2267                         int *count = entry->possible_new_dirs.items[i].util;
2268
2269                         if (*count == max)
2270                                 bad_max = max;
2271                         else if (*count > max) {
2272                                 max = *count;
2273                                 best = entry->possible_new_dirs.items[i].string;
2274                         }
2275                 }
2276                 if (bad_max == max)
2277                         entry->non_unique_new_dir = 1;
2278                 else {
2279                         assert(entry->new_dir.len == 0);
2280                         strbuf_addstr(&entry->new_dir, best);
2281                 }
2282                 /*
2283                  * The relevant directory sub-portion of the original full
2284                  * filepaths were xstrndup'ed before inserting into
2285                  * possible_new_dirs, and instead of manually iterating the
2286                  * list and free'ing each, just lie and tell
2287                  * possible_new_dirs that it did the strdup'ing so that it
2288                  * will free them for us.
2289                  */
2290                 entry->possible_new_dirs.strdup_strings = 1;
2291                 string_list_clear(&entry->possible_new_dirs, 1);
2292         }
2293
2294         return dir_renames;
2295 }
2296
2297 static struct dir_rename_entry *check_dir_renamed(const char *path,
2298                                                   struct hashmap *dir_renames)
2299 {
2300         char *temp = xstrdup(path);
2301         char *end;
2302         struct dir_rename_entry *entry = NULL;
2303
2304         while ((end = strrchr(temp, '/'))) {
2305                 *end = '\0';
2306                 entry = dir_rename_find_entry(dir_renames, temp);
2307                 if (entry)
2308                         break;
2309         }
2310         free(temp);
2311         return entry;
2312 }
2313
2314 static void compute_collisions(struct hashmap *collisions,
2315                                struct hashmap *dir_renames,
2316                                struct diff_queue_struct *pairs)
2317 {
2318         int i;
2319
2320         /*
2321          * Multiple files can be mapped to the same path due to directory
2322          * renames done by the other side of history.  Since that other
2323          * side of history could have merged multiple directories into one,
2324          * if our side of history added the same file basename to each of
2325          * those directories, then all N of them would get implicitly
2326          * renamed by the directory rename detection into the same path,
2327          * and we'd get an add/add/.../add conflict, and all those adds
2328          * from *this* side of history.  This is not representable in the
2329          * index, and users aren't going to easily be able to make sense of
2330          * it.  So we need to provide a good warning about what's
2331          * happening, and fall back to no-directory-rename detection
2332          * behavior for those paths.
2333          *
2334          * See testcases 9e and all of section 5 from t6043 for examples.
2335          */
2336         collision_init(collisions);
2337
2338         for (i = 0; i < pairs->nr; ++i) {
2339                 struct dir_rename_entry *dir_rename_ent;
2340                 struct collision_entry *collision_ent;
2341                 char *new_path;
2342                 struct diff_filepair *pair = pairs->queue[i];
2343
2344                 if (pair->status != 'A' && pair->status != 'R')
2345                         continue;
2346                 dir_rename_ent = check_dir_renamed(pair->two->path,
2347                                                    dir_renames);
2348                 if (!dir_rename_ent)
2349                         continue;
2350
2351                 new_path = apply_dir_rename(dir_rename_ent, pair->two->path);
2352                 if (!new_path)
2353                         /*
2354                          * dir_rename_ent->non_unique_new_path is true, which
2355                          * means there is no directory rename for us to use,
2356                          * which means it won't cause us any additional
2357                          * collisions.
2358                          */
2359                         continue;
2360                 collision_ent = collision_find_entry(collisions, new_path);
2361                 if (!collision_ent) {
2362                         collision_ent = xcalloc(1,
2363                                                 sizeof(struct collision_entry));
2364                         hashmap_entry_init(collision_ent, strhash(new_path));
2365                         hashmap_put(collisions, collision_ent);
2366                         collision_ent->target_file = new_path;
2367                 } else {
2368                         free(new_path);
2369                 }
2370                 string_list_insert(&collision_ent->source_files,
2371                                    pair->two->path);
2372         }
2373 }
2374
2375 static char *check_for_directory_rename(struct merge_options *opt,
2376                                         const char *path,
2377                                         struct tree *tree,
2378                                         struct hashmap *dir_renames,
2379                                         struct hashmap *dir_rename_exclusions,
2380                                         struct hashmap *collisions,
2381                                         int *clean_merge)
2382 {
2383         char *new_path = NULL;
2384         struct dir_rename_entry *entry = check_dir_renamed(path, dir_renames);
2385         struct dir_rename_entry *oentry = NULL;
2386
2387         if (!entry)
2388                 return new_path;
2389
2390         /*
2391          * This next part is a little weird.  We do not want to do an
2392          * implicit rename into a directory we renamed on our side, because
2393          * that will result in a spurious rename/rename(1to2) conflict.  An
2394          * example:
2395          *   Base commit: dumbdir/afile, otherdir/bfile
2396          *   Side 1:      smrtdir/afile, otherdir/bfile
2397          *   Side 2:      dumbdir/afile, dumbdir/bfile
2398          * Here, while working on Side 1, we could notice that otherdir was
2399          * renamed/merged to dumbdir, and change the diff_filepair for
2400          * otherdir/bfile into a rename into dumbdir/bfile.  However, Side
2401          * 2 will notice the rename from dumbdir to smrtdir, and do the
2402          * transitive rename to move it from dumbdir/bfile to
2403          * smrtdir/bfile.  That gives us bfile in dumbdir vs being in
2404          * smrtdir, a rename/rename(1to2) conflict.  We really just want
2405          * the file to end up in smrtdir.  And the way to achieve that is
2406          * to not let Side1 do the rename to dumbdir, since we know that is
2407          * the source of one of our directory renames.
2408          *
2409          * That's why oentry and dir_rename_exclusions is here.
2410          *
2411          * As it turns out, this also prevents N-way transient rename
2412          * confusion; See testcases 9c and 9d of t6043.
2413          */
2414         oentry = dir_rename_find_entry(dir_rename_exclusions, entry->new_dir.buf);
2415         if (oentry) {
2416                 output(opt, 1, _("WARNING: Avoiding applying %s -> %s rename "
2417                                "to %s, because %s itself was renamed."),
2418                        entry->dir, entry->new_dir.buf, path, entry->new_dir.buf);
2419         } else {
2420                 new_path = handle_path_level_conflicts(opt, path, entry,
2421                                                        collisions, tree);
2422                 *clean_merge &= (new_path != NULL);
2423         }
2424
2425         return new_path;
2426 }
2427
2428 static void apply_directory_rename_modifications(struct merge_options *opt,
2429                                                  struct diff_filepair *pair,
2430                                                  char *new_path,
2431                                                  struct rename *re,
2432                                                  struct tree *tree,
2433                                                  struct tree *o_tree,
2434                                                  struct tree *a_tree,
2435                                                  struct tree *b_tree,
2436                                                  struct string_list *entries)
2437 {
2438         struct string_list_item *item;
2439         int stage = (tree == a_tree ? 2 : 3);
2440         int update_wd;
2441
2442         /*
2443          * In all cases where we can do directory rename detection,
2444          * unpack_trees() will have read pair->two->path into the
2445          * index and the working copy.  We need to remove it so that
2446          * we can instead place it at new_path.  It is guaranteed to
2447          * not be untracked (unpack_trees() would have errored out
2448          * saying the file would have been overwritten), but it might
2449          * be dirty, though.
2450          */
2451         update_wd = !was_dirty(opt, pair->two->path);
2452         if (!update_wd)
2453                 output(opt, 1, _("Refusing to lose dirty file at %s"),
2454                        pair->two->path);
2455         remove_file(opt, 1, pair->two->path, !update_wd);
2456
2457         /* Find or create a new re->dst_entry */
2458         item = string_list_lookup(entries, new_path);
2459         if (item) {
2460                 /*
2461                  * Since we're renaming on this side of history, and it's
2462                  * due to a directory rename on the other side of history
2463                  * (which we only allow when the directory in question no
2464                  * longer exists on the other side of history), the
2465                  * original entry for re->dst_entry is no longer
2466                  * necessary...
2467                  */
2468                 re->dst_entry->processed = 1;
2469
2470                 /*
2471                  * ...because we'll be using this new one.
2472                  */
2473                 re->dst_entry = item->util;
2474         } else {
2475                 /*
2476                  * re->dst_entry is for the before-dir-rename path, and we
2477                  * need it to hold information for the after-dir-rename
2478                  * path.  Before creating a new entry, we need to mark the
2479                  * old one as unnecessary (...unless it is shared by
2480                  * src_entry, i.e. this didn't use to be a rename, in which
2481                  * case we can just allow the normal processing to happen
2482                  * for it).
2483                  */
2484                 if (pair->status == 'R')
2485                         re->dst_entry->processed = 1;
2486
2487                 re->dst_entry = insert_stage_data(opt->repo, new_path,
2488                                                   o_tree, a_tree, b_tree,
2489                                                   entries);
2490                 item = string_list_insert(entries, new_path);
2491                 item->util = re->dst_entry;
2492         }
2493
2494         /*
2495          * Update the stage_data with the information about the path we are
2496          * moving into place.  That slot will be empty and available for us
2497          * to write to because of the collision checks in
2498          * handle_path_level_conflicts().  In other words,
2499          * re->dst_entry->stages[stage].oid will be the null_oid, so it's
2500          * open for us to write to.
2501          *
2502          * It may be tempting to actually update the index at this point as
2503          * well, using update_stages_for_stage_data(), but as per the big
2504          * "NOTE" in update_stages(), doing so will modify the current
2505          * in-memory index which will break calls to would_lose_untracked()
2506          * that we need to make.  Instead, we need to just make sure that
2507          * the various handle_rename_*() functions update the index
2508          * explicitly rather than relying on unpack_trees() to have done it.
2509          */
2510         get_tree_entry(opt->repo,
2511                        &tree->object.oid,
2512                        pair->two->path,
2513                        &re->dst_entry->stages[stage].oid,
2514                        &re->dst_entry->stages[stage].mode);
2515
2516         /*
2517          * Record the original change status (or 'type' of change).  If it
2518          * was originally an add ('A'), this lets us differentiate later
2519          * between a RENAME_DELETE conflict and RENAME_VIA_DIR (they
2520          * otherwise look the same).  If it was originally a rename ('R'),
2521          * this lets us remember and report accurately about the transitive
2522          * renaming that occurred via the directory rename detection.  Also,
2523          * record the original destination name.
2524          */
2525         re->dir_rename_original_type = pair->status;
2526         re->dir_rename_original_dest = pair->two->path;
2527
2528         /*
2529          * We don't actually look at pair->status again, but it seems
2530          * pedagogically correct to adjust it.
2531          */
2532         pair->status = 'R';
2533
2534         /*
2535          * Finally, record the new location.
2536          */
2537         pair->two->path = new_path;
2538 }
2539
2540 /*
2541  * Get information of all renames which occurred in 'pairs', making use of
2542  * any implicit directory renames inferred from the other side of history.
2543  * We need the three trees in the merge ('o_tree', 'a_tree' and 'b_tree')
2544  * to be able to associate the correct cache entries with the rename
2545  * information; tree is always equal to either a_tree or b_tree.
2546  */
2547 static struct string_list *get_renames(struct merge_options *opt,
2548                                        const char *branch,
2549                                        struct diff_queue_struct *pairs,
2550                                        struct hashmap *dir_renames,
2551                                        struct hashmap *dir_rename_exclusions,
2552                                        struct tree *tree,
2553                                        struct tree *o_tree,
2554                                        struct tree *a_tree,
2555                                        struct tree *b_tree,
2556                                        struct string_list *entries,
2557                                        int *clean_merge)
2558 {
2559         int i;
2560         struct hashmap collisions;
2561         struct hashmap_iter iter;
2562         struct collision_entry *e;
2563         struct string_list *renames;
2564
2565         compute_collisions(&collisions, dir_renames, pairs);
2566         renames = xcalloc(1, sizeof(struct string_list));
2567
2568         for (i = 0; i < pairs->nr; ++i) {
2569                 struct string_list_item *item;
2570                 struct rename *re;
2571                 struct diff_filepair *pair = pairs->queue[i];
2572                 char *new_path; /* non-NULL only with directory renames */
2573
2574                 if (pair->status != 'A' && pair->status != 'R') {
2575                         diff_free_filepair(pair);
2576                         continue;
2577                 }
2578                 new_path = check_for_directory_rename(opt, pair->two->path, tree,
2579                                                       dir_renames,
2580                                                       dir_rename_exclusions,
2581                                                       &collisions,
2582                                                       clean_merge);
2583                 if (pair->status != 'R' && !new_path) {
2584                         diff_free_filepair(pair);
2585                         continue;
2586                 }
2587
2588                 re = xmalloc(sizeof(*re));
2589                 re->processed = 0;
2590                 re->pair = pair;
2591                 re->branch = branch;
2592                 re->dir_rename_original_type = '\0';
2593                 re->dir_rename_original_dest = NULL;
2594                 item = string_list_lookup(entries, re->pair->one->path);
2595                 if (!item)
2596                         re->src_entry = insert_stage_data(opt->repo,
2597                                         re->pair->one->path,
2598                                         o_tree, a_tree, b_tree, entries);
2599                 else
2600                         re->src_entry = item->util;
2601
2602                 item = string_list_lookup(entries, re->pair->two->path);
2603                 if (!item)
2604                         re->dst_entry = insert_stage_data(opt->repo,
2605                                         re->pair->two->path,
2606                                         o_tree, a_tree, b_tree, entries);
2607                 else
2608                         re->dst_entry = item->util;
2609                 item = string_list_insert(renames, pair->one->path);
2610                 item->util = re;
2611                 if (new_path)
2612                         apply_directory_rename_modifications(opt, pair, new_path,
2613                                                              re, tree, o_tree,
2614                                                              a_tree, b_tree,
2615                                                              entries);
2616         }
2617
2618         hashmap_iter_init(&collisions, &iter);
2619         while ((e = hashmap_iter_next(&iter))) {
2620                 free(e->target_file);
2621                 string_list_clear(&e->source_files, 0);
2622         }
2623         hashmap_free(&collisions, 1);
2624         return renames;
2625 }
2626
2627 static int process_renames(struct merge_options *opt,
2628                            struct string_list *a_renames,
2629                            struct string_list *b_renames)
2630 {
2631         int clean_merge = 1, i, j;
2632         struct string_list a_by_dst = STRING_LIST_INIT_NODUP;
2633         struct string_list b_by_dst = STRING_LIST_INIT_NODUP;
2634         const struct rename *sre;
2635
2636         for (i = 0; i < a_renames->nr; i++) {
2637                 sre = a_renames->items[i].util;
2638                 string_list_insert(&a_by_dst, sre->pair->two->path)->util
2639                         = (void *)sre;
2640         }
2641         for (i = 0; i < b_renames->nr; i++) {
2642                 sre = b_renames->items[i].util;
2643                 string_list_insert(&b_by_dst, sre->pair->two->path)->util
2644                         = (void *)sre;
2645         }
2646
2647         for (i = 0, j = 0; i < a_renames->nr || j < b_renames->nr;) {
2648                 struct string_list *renames1, *renames2Dst;
2649                 struct rename *ren1 = NULL, *ren2 = NULL;
2650                 const char *ren1_src, *ren1_dst;
2651                 struct string_list_item *lookup;
2652
2653                 if (i >= a_renames->nr) {
2654                         ren2 = b_renames->items[j++].util;
2655                 } else if (j >= b_renames->nr) {
2656                         ren1 = a_renames->items[i++].util;
2657                 } else {
2658                         int compare = strcmp(a_renames->items[i].string,
2659                                              b_renames->items[j].string);
2660                         if (compare <= 0)
2661                                 ren1 = a_renames->items[i++].util;
2662                         if (compare >= 0)
2663                                 ren2 = b_renames->items[j++].util;
2664                 }
2665
2666                 /* TODO: refactor, so that 1/2 are not needed */
2667                 if (ren1) {
2668                         renames1 = a_renames;
2669                         renames2Dst = &b_by_dst;
2670                 } else {
2671                         renames1 = b_renames;
2672                         renames2Dst = &a_by_dst;
2673                         SWAP(ren2, ren1);
2674                 }
2675
2676                 if (ren1->processed)
2677                         continue;
2678                 ren1->processed = 1;
2679                 ren1->dst_entry->processed = 1;
2680                 /* BUG: We should only mark src_entry as processed if we
2681                  * are not dealing with a rename + add-source case.
2682                  */
2683                 ren1->src_entry->processed = 1;
2684
2685                 ren1_src = ren1->pair->one->path;
2686                 ren1_dst = ren1->pair->two->path;
2687
2688                 if (ren2) {
2689                         /* One file renamed on both sides */
2690                         const char *ren2_src = ren2->pair->one->path;
2691                         const char *ren2_dst = ren2->pair->two->path;
2692                         enum rename_type rename_type;
2693                         if (strcmp(ren1_src, ren2_src) != 0)
2694                                 BUG("ren1_src != ren2_src");
2695                         ren2->dst_entry->processed = 1;
2696                         ren2->processed = 1;
2697                         if (strcmp(ren1_dst, ren2_dst) != 0) {
2698                                 rename_type = RENAME_ONE_FILE_TO_TWO;
2699                                 clean_merge = 0;
2700                         } else {
2701                                 rename_type = RENAME_ONE_FILE_TO_ONE;
2702                                 /* BUG: We should only remove ren1_src in
2703                                  * the base stage (think of rename +
2704                                  * add-source cases).
2705                                  */
2706                                 remove_file(opt, 1, ren1_src, 1);
2707                                 update_entry(ren1->dst_entry,
2708                                              ren1->pair->one,
2709                                              ren1->pair->two,
2710                                              ren2->pair->two);
2711                         }
2712                         setup_rename_conflict_info(rename_type, opt, ren1, ren2);
2713                 } else if ((lookup = string_list_lookup(renames2Dst, ren1_dst))) {
2714                         /* Two different files renamed to the same thing */
2715                         char *ren2_dst;
2716                         ren2 = lookup->util;
2717                         ren2_dst = ren2->pair->two->path;
2718                         if (strcmp(ren1_dst, ren2_dst) != 0)
2719                                 BUG("ren1_dst != ren2_dst");
2720
2721                         clean_merge = 0;
2722                         ren2->processed = 1;
2723                         /*
2724                          * BUG: We should only mark src_entry as processed
2725                          * if we are not dealing with a rename + add-source
2726                          * case.
2727                          */
2728                         ren2->src_entry->processed = 1;
2729
2730                         setup_rename_conflict_info(RENAME_TWO_FILES_TO_ONE,
2731                                                    opt, ren1, ren2);
2732                 } else {
2733                         /* Renamed in 1, maybe changed in 2 */
2734                         /* we only use sha1 and mode of these */
2735                         struct diff_filespec src_other, dst_other;
2736                         int try_merge;
2737
2738                         /*
2739                          * unpack_trees loads entries from common-commit
2740                          * into stage 1, from head-commit into stage 2, and
2741                          * from merge-commit into stage 3.  We keep track
2742                          * of which side corresponds to the rename.
2743                          */
2744                         int renamed_stage = a_renames == renames1 ? 2 : 3;
2745                         int other_stage =   a_renames == renames1 ? 3 : 2;
2746
2747                         /* BUG: We should only remove ren1_src in the base
2748                          * stage and in other_stage (think of rename +
2749                          * add-source case).
2750                          */
2751                         remove_file(opt, 1, ren1_src,
2752                                     renamed_stage == 2 || !was_tracked(opt, ren1_src));
2753
2754                         oidcpy(&src_other.oid,
2755                                &ren1->src_entry->stages[other_stage].oid);
2756                         src_other.mode = ren1->src_entry->stages[other_stage].mode;
2757                         oidcpy(&dst_other.oid,
2758                                &ren1->dst_entry->stages[other_stage].oid);
2759                         dst_other.mode = ren1->dst_entry->stages[other_stage].mode;
2760                         try_merge = 0;
2761
2762                         if (oid_eq(&src_other.oid, &null_oid) &&
2763                             ren1->dir_rename_original_type == 'A') {
2764                                 setup_rename_conflict_info(RENAME_VIA_DIR,
2765                                                            opt, ren1, NULL);
2766                         } else if (oid_eq(&src_other.oid, &null_oid)) {
2767                                 setup_rename_conflict_info(RENAME_DELETE,
2768                                                            opt, ren1, NULL);
2769                         } else if ((dst_other.mode == ren1->pair->two->mode) &&
2770                                    oid_eq(&dst_other.oid, &ren1->pair->two->oid)) {
2771                                 /*
2772                                  * Added file on the other side identical to
2773                                  * the file being renamed: clean merge.
2774                                  * Also, there is no need to overwrite the
2775                                  * file already in the working copy, so call
2776                                  * update_file_flags() instead of
2777                                  * update_file().
2778                                  */
2779                                 if (update_file_flags(opt,
2780                                                       ren1->pair->two,
2781                                                       ren1_dst,
2782                                                       1, /* update_cache */
2783                                                       0  /* update_wd    */))
2784                                         clean_merge = -1;
2785                         } else if (!oid_eq(&dst_other.oid, &null_oid)) {
2786                                 /*
2787                                  * Probably not a clean merge, but it's
2788                                  * premature to set clean_merge to 0 here,
2789                                  * because if the rename merges cleanly and
2790                                  * the merge exactly matches the newly added
2791                                  * file, then the merge will be clean.
2792                                  */
2793                                 setup_rename_conflict_info(RENAME_ADD,
2794                                                            opt, ren1, NULL);
2795                         } else
2796                                 try_merge = 1;
2797
2798                         if (clean_merge < 0)
2799                                 goto cleanup_and_return;
2800                         if (try_merge) {
2801                                 struct diff_filespec *o, *a, *b;
2802                                 src_other.path = (char *)ren1_src;
2803
2804                                 o = ren1->pair->one;
2805                                 if (a_renames == renames1) {
2806                                         a = ren1->pair->two;
2807                                         b = &src_other;
2808                                 } else {
2809                                         b = ren1->pair->two;
2810                                         a = &src_other;
2811                                 }
2812                                 update_entry(ren1->dst_entry, o, a, b);
2813                                 setup_rename_conflict_info(RENAME_NORMAL,
2814                                                            opt, ren1, NULL);
2815                         }
2816                 }
2817         }
2818 cleanup_and_return:
2819         string_list_clear(&a_by_dst, 0);
2820         string_list_clear(&b_by_dst, 0);
2821
2822         return clean_merge;
2823 }
2824
2825 struct rename_info {
2826         struct string_list *head_renames;
2827         struct string_list *merge_renames;
2828 };
2829
2830 static void initial_cleanup_rename(struct diff_queue_struct *pairs,
2831                                    struct hashmap *dir_renames)
2832 {
2833         struct hashmap_iter iter;
2834         struct dir_rename_entry *e;
2835
2836         hashmap_iter_init(dir_renames, &iter);
2837         while ((e = hashmap_iter_next(&iter))) {
2838                 free(e->dir);
2839                 strbuf_release(&e->new_dir);
2840                 /* possible_new_dirs already cleared in get_directory_renames */
2841         }
2842         hashmap_free(dir_renames, 1);
2843         free(dir_renames);
2844
2845         free(pairs->queue);
2846         free(pairs);
2847 }
2848
2849 static int detect_and_process_renames(struct merge_options *opt,
2850                                       struct tree *common,
2851                                       struct tree *head,
2852                                       struct tree *merge,
2853                                       struct string_list *entries,
2854                                       struct rename_info *ri)
2855 {
2856         struct diff_queue_struct *head_pairs, *merge_pairs;
2857         struct hashmap *dir_re_head, *dir_re_merge;
2858         int clean = 1;
2859
2860         ri->head_renames = NULL;
2861         ri->merge_renames = NULL;
2862
2863         if (!merge_detect_rename(opt))
2864                 return 1;
2865
2866         head_pairs = get_diffpairs(opt, common, head);
2867         merge_pairs = get_diffpairs(opt, common, merge);
2868
2869         if ((opt->detect_directory_renames == MERGE_DIRECTORY_RENAMES_TRUE) ||
2870             (opt->detect_directory_renames == MERGE_DIRECTORY_RENAMES_CONFLICT &&
2871              !opt->call_depth)) {
2872                 dir_re_head = get_directory_renames(head_pairs);
2873                 dir_re_merge = get_directory_renames(merge_pairs);
2874
2875                 handle_directory_level_conflicts(opt,
2876                                                  dir_re_head, head,
2877                                                  dir_re_merge, merge);
2878         } else {
2879                 dir_re_head  = xmalloc(sizeof(*dir_re_head));
2880                 dir_re_merge = xmalloc(sizeof(*dir_re_merge));
2881                 dir_rename_init(dir_re_head);
2882                 dir_rename_init(dir_re_merge);
2883         }
2884
2885         ri->head_renames  = get_renames(opt, opt->branch1, head_pairs,
2886                                         dir_re_merge, dir_re_head, head,
2887                                         common, head, merge, entries,
2888                                         &clean);
2889         if (clean < 0)
2890                 goto cleanup;
2891         ri->merge_renames = get_renames(opt, opt->branch2, merge_pairs,
2892                                         dir_re_head, dir_re_merge, merge,
2893                                         common, head, merge, entries,
2894                                         &clean);
2895         if (clean < 0)
2896                 goto cleanup;
2897         clean &= process_renames(opt, ri->head_renames, ri->merge_renames);
2898
2899 cleanup:
2900         /*
2901          * Some cleanup is deferred until cleanup_renames() because the
2902          * data structures are still needed and referenced in
2903          * process_entry().  But there are a few things we can free now.
2904          */
2905         initial_cleanup_rename(head_pairs, dir_re_head);
2906         initial_cleanup_rename(merge_pairs, dir_re_merge);
2907
2908         return clean;
2909 }
2910
2911 static void final_cleanup_rename(struct string_list *rename)
2912 {
2913         const struct rename *re;
2914         int i;
2915
2916         if (rename == NULL)
2917                 return;
2918
2919         for (i = 0; i < rename->nr; i++) {
2920                 re = rename->items[i].util;
2921                 diff_free_filepair(re->pair);
2922         }
2923         string_list_clear(rename, 1);
2924         free(rename);
2925 }
2926
2927 static void final_cleanup_renames(struct rename_info *re_info)
2928 {
2929         final_cleanup_rename(re_info->head_renames);
2930         final_cleanup_rename(re_info->merge_renames);
2931 }
2932
2933 static int read_oid_strbuf(struct merge_options *opt,
2934                            const struct object_id *oid,
2935                            struct strbuf *dst)
2936 {
2937         void *buf;
2938         enum object_type type;
2939         unsigned long size;
2940         buf = read_object_file(oid, &type, &size);
2941         if (!buf)
2942                 return err(opt, _("cannot read object %s"), oid_to_hex(oid));
2943         if (type != OBJ_BLOB) {
2944                 free(buf);
2945                 return err(opt, _("object %s is not a blob"), oid_to_hex(oid));
2946         }
2947         strbuf_attach(dst, buf, size, size + 1);
2948         return 0;
2949 }
2950
2951 static int blob_unchanged(struct merge_options *opt,
2952                           const struct diff_filespec *o,
2953                           const struct diff_filespec *a,
2954                           int renormalize, const char *path)
2955 {
2956         struct strbuf obuf = STRBUF_INIT;
2957         struct strbuf abuf = STRBUF_INIT;
2958         int ret = 0; /* assume changed for safety */
2959         const struct index_state *idx = opt->repo->index;
2960
2961         if (a->mode != o->mode)
2962                 return 0;
2963         if (oid_eq(&o->oid, &a->oid))
2964                 return 1;
2965         if (!renormalize)
2966                 return 0;
2967
2968         if (read_oid_strbuf(opt, &o->oid, &obuf) ||
2969             read_oid_strbuf(opt, &a->oid, &abuf))
2970                 goto error_return;
2971         /*
2972          * Note: binary | is used so that both renormalizations are
2973          * performed.  Comparison can be skipped if both files are
2974          * unchanged since their sha1s have already been compared.
2975          */
2976         if (renormalize_buffer(idx, path, obuf.buf, obuf.len, &obuf) |
2977             renormalize_buffer(idx, path, abuf.buf, abuf.len, &abuf))
2978                 ret = (obuf.len == abuf.len && !memcmp(obuf.buf, abuf.buf, obuf.len));
2979
2980 error_return:
2981         strbuf_release(&obuf);
2982         strbuf_release(&abuf);
2983         return ret;
2984 }
2985
2986 static int handle_modify_delete(struct merge_options *opt,
2987                                 const char *path,
2988                                 const struct diff_filespec *o,
2989                                 const struct diff_filespec *a,
2990                                 const struct diff_filespec *b)
2991 {
2992         const char *modify_branch, *delete_branch;
2993         const struct diff_filespec *changed;
2994
2995         if (is_valid(a)) {
2996                 modify_branch = opt->branch1;
2997                 delete_branch = opt->branch2;
2998                 changed = a;
2999         } else {
3000                 modify_branch = opt->branch2;
3001                 delete_branch = opt->branch1;
3002                 changed = b;
3003         }
3004
3005         return handle_change_delete(opt,
3006                                     path, NULL,
3007                                     o, changed,
3008                                     modify_branch, delete_branch,
3009                                     _("modify"), _("modified"));
3010 }
3011
3012 static int handle_content_merge(struct merge_file_info *mfi,
3013                                 struct merge_options *opt,
3014                                 const char *path,
3015                                 int is_dirty,
3016                                 const struct diff_filespec *o,
3017                                 const struct diff_filespec *a,
3018                                 const struct diff_filespec *b,
3019                                 struct rename_conflict_info *ci)
3020 {
3021         const char *reason = _("content");
3022         unsigned df_conflict_remains = 0;
3023
3024         if (!is_valid(o))
3025                 reason = _("add/add");
3026
3027         assert(o->path && a->path && b->path);
3028         if (ci && dir_in_way(opt->repo->index, path, !opt->call_depth,
3029                              S_ISGITLINK(ci->ren1->pair->two->mode)))
3030                 df_conflict_remains = 1;
3031
3032         if (merge_mode_and_contents(opt, o, a, b, path,
3033                                     opt->branch1, opt->branch2,
3034                                     opt->call_depth * 2, mfi))
3035                 return -1;
3036
3037         /*
3038          * We can skip updating the working tree file iff:
3039          *   a) The merge is clean
3040          *   b) The merge matches what was in HEAD (content, mode, pathname)
3041          *   c) The target path is usable (i.e. not involved in D/F conflict)
3042          */
3043         if (mfi->clean && was_tracked_and_matches(opt, path, &mfi->blob) &&
3044             !df_conflict_remains) {
3045                 int pos;
3046                 struct cache_entry *ce;
3047
3048                 output(opt, 3, _("Skipped %s (merged same as existing)"), path);
3049                 if (add_cacheinfo(opt, &mfi->blob, path,
3050                                   0, (!opt->call_depth && !is_dirty), 0))
3051                         return -1;
3052                 /*
3053                  * However, add_cacheinfo() will delete the old cache entry
3054                  * and add a new one.  We need to copy over any skip_worktree
3055                  * flag to avoid making the file appear as if it were
3056                  * deleted by the user.
3057                  */
3058                 pos = index_name_pos(&opt->orig_index, path, strlen(path));
3059                 ce = opt->orig_index.cache[pos];
3060                 if (ce_skip_worktree(ce)) {
3061                         pos = index_name_pos(opt->repo->index, path, strlen(path));
3062                         ce = opt->repo->index->cache[pos];
3063                         ce->ce_flags |= CE_SKIP_WORKTREE;
3064                 }
3065                 return mfi->clean;
3066         }
3067
3068         if (!mfi->clean) {
3069                 if (S_ISGITLINK(mfi->blob.mode))
3070                         reason = _("submodule");
3071                 output(opt, 1, _("CONFLICT (%s): Merge conflict in %s"),
3072                                 reason, path);
3073                 if (ci && !df_conflict_remains)
3074                         if (update_stages(opt, path, o, a, b))
3075                                 return -1;
3076         }
3077
3078         if (df_conflict_remains || is_dirty) {
3079                 char *new_path;
3080                 if (opt->call_depth) {
3081                         remove_file_from_index(opt->repo->index, path);
3082                 } else {
3083                         if (!mfi->clean) {
3084                                 if (update_stages(opt, path, o, a, b))
3085                                         return -1;
3086                         } else {
3087                                 int file_from_stage2 = was_tracked(opt, path);
3088
3089                                 if (update_stages(opt, path, NULL,
3090                                                   file_from_stage2 ? &mfi->blob : NULL,
3091                                                   file_from_stage2 ? NULL : &mfi->blob))
3092                                         return -1;
3093                         }
3094
3095                 }
3096                 new_path = unique_path(opt, path, ci->ren1->branch);
3097                 if (is_dirty) {
3098                         output(opt, 1, _("Refusing to lose dirty file at %s"),
3099                                path);
3100                 }
3101                 output(opt, 1, _("Adding as %s instead"), new_path);
3102                 if (update_file(opt, 0, &mfi->blob, new_path)) {
3103                         free(new_path);
3104                         return -1;
3105                 }
3106                 free(new_path);
3107                 mfi->clean = 0;
3108         } else if (update_file(opt, mfi->clean, &mfi->blob, path))
3109                 return -1;
3110         return !is_dirty && mfi->clean;
3111 }
3112
3113 static int handle_rename_normal(struct merge_options *opt,
3114                                 const char *path,
3115                                 const struct diff_filespec *o,
3116                                 const struct diff_filespec *a,
3117                                 const struct diff_filespec *b,
3118                                 struct rename_conflict_info *ci)
3119 {
3120         struct rename *ren = ci->ren1;
3121         struct merge_file_info mfi;
3122         int clean;
3123         int side = (ren->branch == opt->branch1 ? 2 : 3);
3124
3125         /* Merge the content and write it out */
3126         clean = handle_content_merge(&mfi, opt, path, was_dirty(opt, path),
3127                                      o, a, b, ci);
3128
3129         if (clean &&
3130             opt->detect_directory_renames == MERGE_DIRECTORY_RENAMES_CONFLICT &&
3131             ren->dir_rename_original_dest) {
3132                 if (update_stages(opt, path,
3133                                   NULL,
3134                                   side == 2 ? &mfi.blob : NULL,
3135                                   side == 2 ? NULL : &mfi.blob))
3136                         return -1;
3137                 clean = 0; /* not clean, but conflicted */
3138         }
3139         return clean;
3140 }
3141
3142 static void dir_rename_warning(const char *msg,
3143                                int is_add,
3144                                int clean,
3145                                struct merge_options *opt,
3146                                struct rename *ren)
3147 {
3148         const char *other_branch;
3149         other_branch = (ren->branch == opt->branch1 ?
3150                         opt->branch2 : opt->branch1);
3151         if (is_add) {
3152                 output(opt, clean ? 2 : 1, msg,
3153                        ren->pair->one->path, ren->branch,
3154                        other_branch, ren->pair->two->path);
3155                 return;
3156         }
3157         output(opt, clean ? 2 : 1, msg,
3158                ren->pair->one->path, ren->dir_rename_original_dest, ren->branch,
3159                other_branch, ren->pair->two->path);
3160 }
3161 static int warn_about_dir_renamed_entries(struct merge_options *opt,
3162                                           struct rename *ren)
3163 {
3164         const char *msg;
3165         int clean = 1, is_add;
3166
3167         if (!ren)
3168                 return clean;
3169
3170         /* Return early if ren was not affected/created by a directory rename */
3171         if (!ren->dir_rename_original_dest)
3172                 return clean;
3173
3174         /* Sanity checks */
3175         assert(opt->detect_directory_renames > MERGE_DIRECTORY_RENAMES_NONE);
3176         assert(ren->dir_rename_original_type == 'A' ||
3177                ren->dir_rename_original_type == 'R');
3178
3179         /* Check whether to treat directory renames as a conflict */
3180         clean = (opt->detect_directory_renames == MERGE_DIRECTORY_RENAMES_TRUE);
3181
3182         is_add = (ren->dir_rename_original_type == 'A');
3183         if (ren->dir_rename_original_type == 'A' && clean) {
3184                 msg = _("Path updated: %s added in %s inside a "
3185                         "directory that was renamed in %s; moving it to %s.");
3186         } else if (ren->dir_rename_original_type == 'A' && !clean) {
3187                 msg = _("CONFLICT (file location): %s added in %s "
3188                         "inside a directory that was renamed in %s, "
3189                         "suggesting it should perhaps be moved to %s.");
3190         } else if (ren->dir_rename_original_type == 'R' && clean) {
3191                 msg = _("Path updated: %s renamed to %s in %s, inside a "
3192                         "directory that was renamed in %s; moving it to %s.");
3193         } else if (ren->dir_rename_original_type == 'R' && !clean) {
3194                 msg = _("CONFLICT (file location): %s renamed to %s in %s, "
3195                         "inside a directory that was renamed in %s, "
3196                         "suggesting it should perhaps be moved to %s.");
3197         } else {
3198                 BUG("Impossible dir_rename_original_type/clean combination");
3199         }
3200         dir_rename_warning(msg, is_add, clean, opt, ren);
3201
3202         return clean;
3203 }
3204
3205 /* Per entry merge function */
3206 static int process_entry(struct merge_options *opt,
3207                          const char *path, struct stage_data *entry)
3208 {
3209         int clean_merge = 1;
3210         int normalize = opt->renormalize;
3211
3212         struct diff_filespec *o = &entry->stages[1];
3213         struct diff_filespec *a = &entry->stages[2];
3214         struct diff_filespec *b = &entry->stages[3];
3215         int o_valid = is_valid(o);
3216         int a_valid = is_valid(a);
3217         int b_valid = is_valid(b);
3218         o->path = a->path = b->path = (char*)path;
3219
3220         entry->processed = 1;
3221         if (entry->rename_conflict_info) {
3222                 struct rename_conflict_info *ci = entry->rename_conflict_info;
3223                 struct diff_filespec *temp;
3224                 int path_clean;
3225
3226                 path_clean = warn_about_dir_renamed_entries(opt, ci->ren1);
3227                 path_clean &= warn_about_dir_renamed_entries(opt, ci->ren2);
3228
3229                 /*
3230                  * For cases with a single rename, {o,a,b}->path have all been
3231                  * set to the rename target path; we need to set two of these
3232                  * back to the rename source.
3233                  * For rename/rename conflicts, we'll manually fix paths below.
3234                  */
3235                 temp = (opt->branch1 == ci->ren1->branch) ? b : a;
3236                 o->path = temp->path = ci->ren1->pair->one->path;
3237                 if (ci->ren2) {
3238                         assert(opt->branch1 == ci->ren1->branch);
3239                 }
3240
3241                 switch (ci->rename_type) {
3242                 case RENAME_NORMAL:
3243                 case RENAME_ONE_FILE_TO_ONE:
3244                         clean_merge = handle_rename_normal(opt, path, o, a, b,
3245                                                            ci);
3246                         break;
3247                 case RENAME_VIA_DIR:
3248                         clean_merge = handle_rename_via_dir(opt, ci);
3249                         break;
3250                 case RENAME_ADD:
3251                         /*
3252                          * Probably unclean merge, but if the renamed file
3253                          * merges cleanly and the result can then be
3254                          * two-way merged cleanly with the added file, I
3255                          * guess it's a clean merge?
3256                          */
3257                         clean_merge = handle_rename_add(opt, ci);
3258                         break;
3259                 case RENAME_DELETE:
3260                         clean_merge = 0;
3261                         if (handle_rename_delete(opt, ci))
3262                                 clean_merge = -1;
3263                         break;
3264                 case RENAME_ONE_FILE_TO_TWO:
3265                         /*
3266                          * Manually fix up paths; note:
3267                          * ren[12]->pair->one->path are equal.
3268                          */
3269                         o->path = ci->ren1->pair->one->path;
3270                         a->path = ci->ren1->pair->two->path;
3271                         b->path = ci->ren2->pair->two->path;
3272
3273                         clean_merge = 0;
3274                         if (handle_rename_rename_1to2(opt, ci))
3275                                 clean_merge = -1;
3276                         break;
3277                 case RENAME_TWO_FILES_TO_ONE:
3278                         /*
3279                          * Manually fix up paths; note,
3280                          * ren[12]->pair->two->path are actually equal.
3281                          */
3282                         o->path = NULL;
3283                         a->path = ci->ren1->pair->two->path;
3284                         b->path = ci->ren2->pair->two->path;
3285
3286                         /*
3287                          * Probably unclean merge, but if the two renamed
3288                          * files merge cleanly and the two resulting files
3289                          * can then be two-way merged cleanly, I guess it's
3290                          * a clean merge?
3291                          */
3292                         clean_merge = handle_rename_rename_2to1(opt, ci);
3293                         break;
3294                 default:
3295                         entry->processed = 0;
3296                         break;
3297                 }
3298                 if (path_clean < clean_merge)
3299                         clean_merge = path_clean;
3300         } else if (o_valid && (!a_valid || !b_valid)) {
3301                 /* Case A: Deleted in one */
3302                 if ((!a_valid && !b_valid) ||
3303                     (!b_valid && blob_unchanged(opt, o, a, normalize, path)) ||
3304                     (!a_valid && blob_unchanged(opt, o, b, normalize, path))) {
3305                         /* Deleted in both or deleted in one and
3306                          * unchanged in the other */
3307                         if (a_valid)
3308                                 output(opt, 2, _("Removing %s"), path);
3309                         /* do not touch working file if it did not exist */
3310                         remove_file(opt, 1, path, !a_valid);
3311                 } else {
3312                         /* Modify/delete; deleted side may have put a directory in the way */
3313                         clean_merge = 0;
3314                         if (handle_modify_delete(opt, path, o, a, b))
3315                                 clean_merge = -1;
3316                 }
3317         } else if ((!o_valid && a_valid && !b_valid) ||
3318                    (!o_valid && !a_valid && b_valid)) {
3319                 /* Case B: Added in one. */
3320                 /* [nothing|directory] -> ([nothing|directory], file) */
3321
3322                 const char *add_branch;
3323                 const char *other_branch;
3324                 const char *conf;
3325                 const struct diff_filespec *contents;
3326
3327                 if (a_valid) {
3328                         add_branch = opt->branch1;
3329                         other_branch = opt->branch2;
3330                         contents = a;
3331                         conf = _("file/directory");
3332                 } else {
3333                         add_branch = opt->branch2;
3334                         other_branch = opt->branch1;
3335                         contents = b;
3336                         conf = _("directory/file");
3337                 }
3338                 if (dir_in_way(opt->repo->index, path,
3339                                !opt->call_depth && !S_ISGITLINK(a->mode),
3340                                0)) {
3341                         char *new_path = unique_path(opt, path, add_branch);
3342                         clean_merge = 0;
3343                         output(opt, 1, _("CONFLICT (%s): There is a directory with name %s in %s. "
3344                                "Adding %s as %s"),
3345                                conf, path, other_branch, path, new_path);
3346                         if (update_file(opt, 0, contents, new_path))
3347                                 clean_merge = -1;
3348                         else if (opt->call_depth)
3349                                 remove_file_from_index(opt->repo->index, path);
3350                         free(new_path);
3351                 } else {
3352                         output(opt, 2, _("Adding %s"), path);
3353                         /* do not overwrite file if already present */
3354                         if (update_file_flags(opt, contents, path, 1, !a_valid))
3355                                 clean_merge = -1;
3356                 }
3357         } else if (a_valid && b_valid) {
3358                 if (!o_valid) {
3359                         /* Case C: Added in both (check for same permissions) */
3360                         output(opt, 1,
3361                                _("CONFLICT (add/add): Merge conflict in %s"),
3362                                path);
3363                         clean_merge = handle_file_collision(opt,
3364                                                             path, NULL, NULL,
3365                                                             opt->branch1,
3366                                                             opt->branch2,
3367                                                             a, b);
3368                 } else {
3369                         /* case D: Modified in both, but differently. */
3370                         struct merge_file_info mfi;
3371                         int is_dirty = 0; /* unpack_trees would have bailed if dirty */
3372                         clean_merge = handle_content_merge(&mfi, opt, path,
3373                                                            is_dirty,
3374                                                            o, a, b, NULL);
3375                 }
3376         } else if (!o_valid && !a_valid && !b_valid) {
3377                 /*
3378                  * this entry was deleted altogether. a_mode == 0 means
3379                  * we had that path and want to actively remove it.
3380                  */
3381                 remove_file(opt, 1, path, !a->mode);
3382         } else
3383                 BUG("fatal merge failure, shouldn't happen.");
3384
3385         return clean_merge;
3386 }
3387
3388 static int merge_trees_internal(struct merge_options *opt,
3389                                 struct tree *head,
3390                                 struct tree *merge,
3391                                 struct tree *merge_base,
3392                                 struct tree **result)
3393 {
3394         struct index_state *istate = opt->repo->index;
3395         int code, clean;
3396
3397         if (opt->subtree_shift) {
3398                 merge = shift_tree_object(opt->repo, head, merge,
3399                                           opt->subtree_shift);
3400                 merge_base = shift_tree_object(opt->repo, head, merge_base,
3401                                                opt->subtree_shift);
3402         }
3403
3404         if (oid_eq(&merge_base->object.oid, &merge->object.oid)) {
3405                 output(opt, 0, _("Already up to date!"));
3406                 *result = head;
3407                 return 1;
3408         }
3409
3410         code = unpack_trees_start(opt, merge_base, head, merge);
3411
3412         if (code != 0) {
3413                 if (show(opt, 4) || opt->call_depth)
3414                         err(opt, _("merging of trees %s and %s failed"),
3415                             oid_to_hex(&head->object.oid),
3416                             oid_to_hex(&merge->object.oid));
3417                 unpack_trees_finish(opt);
3418                 return -1;
3419         }
3420
3421         if (unmerged_index(istate)) {
3422                 struct string_list *entries;
3423                 struct rename_info re_info;
3424                 int i;
3425                 /*
3426                  * Only need the hashmap while processing entries, so
3427                  * initialize it here and free it when we are done running
3428                  * through the entries. Keeping it in the merge_options as
3429                  * opposed to decaring a local hashmap is for convenience
3430                  * so that we don't have to pass it to around.
3431                  */
3432                 hashmap_init(&opt->current_file_dir_set, path_hashmap_cmp,
3433                              NULL, 512);
3434                 get_files_dirs(opt, head);
3435                 get_files_dirs(opt, merge);
3436
3437                 entries = get_unmerged(opt->repo->index);
3438                 clean = detect_and_process_renames(opt, merge_base, head, merge,
3439                                                    entries, &re_info);
3440                 record_df_conflict_files(opt, entries);
3441                 if (clean < 0)
3442                         goto cleanup;
3443                 for (i = entries->nr-1; 0 <= i; i--) {
3444                         const char *path = entries->items[i].string;
3445                         struct stage_data *e = entries->items[i].util;
3446                         if (!e->processed) {
3447                                 int ret = process_entry(opt, path, e);
3448                                 if (!ret)
3449                                         clean = 0;
3450                                 else if (ret < 0) {
3451                                         clean = ret;
3452                                         goto cleanup;
3453                                 }
3454                         }
3455                 }
3456                 for (i = 0; i < entries->nr; i++) {
3457                         struct stage_data *e = entries->items[i].util;
3458                         if (!e->processed)
3459                                 BUG("unprocessed path??? %s",
3460                                     entries->items[i].string);
3461                 }
3462
3463         cleanup:
3464                 final_cleanup_renames(&re_info);
3465
3466                 string_list_clear(entries, 1);
3467                 free(entries);
3468
3469                 hashmap_free(&opt->current_file_dir_set, 1);
3470
3471                 if (clean < 0) {
3472                         unpack_trees_finish(opt);
3473                         return clean;
3474                 }
3475         }
3476         else
3477                 clean = 1;
3478
3479         unpack_trees_finish(opt);
3480
3481         if (opt->call_depth &&
3482             !(*result = write_in_core_index_as_tree(opt->repo)))
3483                 return -1;
3484
3485         return clean;
3486 }
3487
3488 static struct commit_list *reverse_commit_list(struct commit_list *list)
3489 {
3490         struct commit_list *next = NULL, *current, *backup;
3491         for (current = list; current; current = backup) {
3492                 backup = current->next;
3493                 current->next = next;
3494                 next = current;
3495         }
3496         return next;
3497 }
3498
3499 /*
3500  * Merge the commits h1 and h2, return the resulting virtual
3501  * commit object and a flag indicating the cleanness of the merge.
3502  */
3503 static int merge_recursive_internal(struct merge_options *opt,
3504                                     struct commit *h1,
3505                                     struct commit *h2,
3506                                     struct commit_list *merge_bases,
3507                                     struct commit **result)
3508 {
3509         struct commit_list *iter;
3510         struct commit *merged_merge_bases;
3511         struct tree *result_tree;
3512         int clean;
3513         const char *ancestor_name;
3514         struct strbuf merge_base_abbrev = STRBUF_INIT;
3515
3516         if (show(opt, 4)) {
3517                 output(opt, 4, _("Merging:"));
3518                 output_commit_title(opt, h1);
3519                 output_commit_title(opt, h2);
3520         }
3521
3522         if (!merge_bases) {
3523                 merge_bases = get_merge_bases(h1, h2);
3524                 merge_bases = reverse_commit_list(merge_bases);
3525         }
3526
3527         if (show(opt, 5)) {
3528                 unsigned cnt = commit_list_count(merge_bases);
3529
3530                 output(opt, 5, Q_("found %u common ancestor:",
3531                                 "found %u common ancestors:", cnt), cnt);
3532                 for (iter = merge_bases; iter; iter = iter->next)
3533                         output_commit_title(opt, iter->item);
3534         }
3535
3536         merged_merge_bases = pop_commit(&merge_bases);
3537         if (merged_merge_bases == NULL) {
3538                 /* if there is no common ancestor, use an empty tree */
3539                 struct tree *tree;
3540
3541                 tree = lookup_tree(opt->repo, opt->repo->hash_algo->empty_tree);
3542                 merged_merge_bases = make_virtual_commit(opt->repo, tree,
3543                                                          "ancestor");
3544                 ancestor_name = "empty tree";
3545         } else if (merge_bases) {
3546                 ancestor_name = "merged common ancestors";
3547         } else {
3548                 strbuf_add_unique_abbrev(&merge_base_abbrev,
3549                                          &merged_merge_bases->object.oid,
3550                                          DEFAULT_ABBREV);
3551                 ancestor_name = merge_base_abbrev.buf;
3552         }
3553
3554         for (iter = merge_bases; iter; iter = iter->next) {
3555                 const char *saved_b1, *saved_b2;
3556                 opt->call_depth++;
3557                 /*
3558                  * When the merge fails, the result contains files
3559                  * with conflict markers. The cleanness flag is
3560                  * ignored (unless indicating an error), it was never
3561                  * actually used, as result of merge_trees has always
3562                  * overwritten it: the committed "conflicts" were
3563                  * already resolved.
3564                  */
3565                 discard_index(opt->repo->index);
3566                 saved_b1 = opt->branch1;
3567                 saved_b2 = opt->branch2;
3568                 opt->branch1 = "Temporary merge branch 1";
3569                 opt->branch2 = "Temporary merge branch 2";
3570                 if (merge_recursive_internal(opt, merged_merge_bases, iter->item,
3571                                              NULL, &merged_merge_bases) < 0)
3572                         return -1;
3573                 opt->branch1 = saved_b1;
3574                 opt->branch2 = saved_b2;
3575                 opt->call_depth--;
3576
3577                 if (!merged_merge_bases)
3578                         return err(opt, _("merge returned no commit"));
3579         }
3580
3581         discard_index(opt->repo->index);
3582         if (!opt->call_depth)
3583                 repo_read_index(opt->repo);
3584
3585         opt->ancestor = ancestor_name;
3586         clean = merge_trees_internal(opt,
3587                                      repo_get_commit_tree(opt->repo, h1),
3588                                      repo_get_commit_tree(opt->repo, h2),
3589                                      repo_get_commit_tree(opt->repo,
3590                                                           merged_merge_bases),
3591                                      &result_tree);
3592         strbuf_release(&merge_base_abbrev);
3593         if (clean < 0) {
3594                 flush_output(opt);
3595                 return clean;
3596         }
3597
3598         if (opt->call_depth) {
3599                 *result = make_virtual_commit(opt->repo, result_tree,
3600                                               "merged tree");
3601                 commit_list_insert(h1, &(*result)->parents);
3602                 commit_list_insert(h2, &(*result)->parents->next);
3603         }
3604         flush_output(opt);
3605         if (!opt->call_depth && opt->buffer_output < 2)
3606                 strbuf_release(&opt->obuf);
3607         return clean;
3608 }
3609
3610 static int merge_start(struct merge_options *opt, struct tree *head)
3611 {
3612         struct strbuf sb = STRBUF_INIT;
3613
3614         if (repo_index_has_changes(opt->repo, head, &sb)) {
3615                 err(opt, _("Your local changes to the following files would be overwritten by merge:\n  %s"),
3616                     sb.buf);
3617                 strbuf_release(&sb);
3618                 return -1;
3619         }
3620
3621         return 0;
3622 }
3623
3624 static void merge_finalize(struct merge_options *opt)
3625 {
3626         if (show(opt, 2))
3627                 diff_warn_rename_limit("merge.renamelimit",
3628                                        opt->needed_rename_limit, 0);
3629 }
3630
3631 int merge_trees(struct merge_options *opt,
3632                 struct tree *head,
3633                 struct tree *merge,
3634                 struct tree *merge_base)
3635 {
3636         int clean;
3637         struct tree *ignored;
3638
3639         assert(opt->ancestor != NULL);
3640
3641         if (merge_start(opt, head))
3642                 return -1;
3643         clean = merge_trees_internal(opt, head, merge, merge_base, &ignored);
3644         merge_finalize(opt);
3645
3646         return clean;
3647 }
3648
3649 int merge_recursive(struct merge_options *opt,
3650                     struct commit *h1,
3651                     struct commit *h2,
3652                     struct commit_list *merge_bases,
3653                     struct commit **result)
3654 {
3655         int clean;
3656
3657         assert(opt->ancestor == NULL);
3658
3659         if (merge_start(opt, repo_get_commit_tree(opt->repo, h1)))
3660                 return -1;
3661         clean = merge_recursive_internal(opt, h1, h2, merge_bases, result);
3662         merge_finalize(opt);
3663
3664         return clean;
3665 }
3666
3667 static struct commit *get_ref(struct repository *repo,
3668                               const struct object_id *oid,
3669                               const char *name)
3670 {
3671         struct object *object;
3672
3673         object = deref_tag(repo, parse_object(repo, oid),
3674                            name, strlen(name));
3675         if (!object)
3676                 return NULL;
3677         if (object->type == OBJ_TREE)
3678                 return make_virtual_commit(repo, (struct tree*)object, name);
3679         if (object->type != OBJ_COMMIT)
3680                 return NULL;
3681         if (parse_commit((struct commit *)object))
3682                 return NULL;
3683         return (struct commit *)object;
3684 }
3685
3686 int merge_recursive_generic(struct merge_options *opt,
3687                             const struct object_id *head,
3688                             const struct object_id *merge,
3689                             int num_merge_bases,
3690                             const struct object_id **merge_bases,
3691                             struct commit **result)
3692 {
3693         int clean;
3694         struct lock_file lock = LOCK_INIT;
3695         struct commit *head_commit = get_ref(opt->repo, head, opt->branch1);
3696         struct commit *next_commit = get_ref(opt->repo, merge, opt->branch2);
3697         struct commit_list *ca = NULL;
3698
3699         if (merge_bases) {
3700                 int i;
3701                 for (i = 0; i < num_merge_bases; ++i) {
3702                         struct commit *base;
3703                         if (!(base = get_ref(opt->repo, merge_bases[i],
3704                                              oid_to_hex(merge_bases[i]))))
3705                                 return err(opt, _("Could not parse object '%s'"),
3706                                            oid_to_hex(merge_bases[i]));
3707                         commit_list_insert(base, &ca);
3708                 }
3709         }
3710
3711         repo_hold_locked_index(opt->repo, &lock, LOCK_DIE_ON_ERROR);
3712         clean = merge_recursive(opt, head_commit, next_commit, ca,
3713                                 result);
3714         if (clean < 0) {
3715                 rollback_lock_file(&lock);
3716                 return clean;
3717         }
3718
3719         if (write_locked_index(opt->repo->index, &lock,
3720                                COMMIT_LOCK | SKIP_IF_UNCHANGED))
3721                 return err(opt, _("Unable to write index."));
3722
3723         return clean ? 0 : 1;
3724 }
3725
3726 static void merge_recursive_config(struct merge_options *opt)
3727 {
3728         char *value = NULL;
3729         git_config_get_int("merge.verbosity", &opt->verbosity);
3730         git_config_get_int("diff.renamelimit", &opt->diff_rename_limit);
3731         git_config_get_int("merge.renamelimit", &opt->merge_rename_limit);
3732         if (!git_config_get_string("diff.renames", &value)) {
3733                 opt->diff_detect_rename = git_config_rename("diff.renames", value);
3734                 free(value);
3735         }
3736         if (!git_config_get_string("merge.renames", &value)) {
3737                 opt->merge_detect_rename = git_config_rename("merge.renames", value);
3738                 free(value);
3739         }
3740         if (!git_config_get_string("merge.directoryrenames", &value)) {
3741                 int boolval = git_parse_maybe_bool(value);
3742                 if (0 <= boolval) {
3743                         opt->detect_directory_renames = boolval ?
3744                                 MERGE_DIRECTORY_RENAMES_TRUE :
3745                                 MERGE_DIRECTORY_RENAMES_NONE;
3746                 } else if (!strcasecmp(value, "conflict")) {
3747                         opt->detect_directory_renames =
3748                                 MERGE_DIRECTORY_RENAMES_CONFLICT;
3749                 } /* avoid erroring on values from future versions of git */
3750                 free(value);
3751         }
3752         git_config(git_xmerge_config, NULL);
3753 }
3754
3755 void init_merge_options(struct merge_options *opt,
3756                         struct repository *repo)
3757 {
3758         const char *merge_verbosity;
3759         memset(opt, 0, sizeof(struct merge_options));
3760         opt->repo = repo;
3761         opt->verbosity = 2;
3762         opt->buffer_output = 1;
3763         opt->diff_rename_limit = -1;
3764         opt->merge_rename_limit = -1;
3765         opt->renormalize = 0;
3766         opt->diff_detect_rename = -1;
3767         opt->merge_detect_rename = -1;
3768         opt->detect_directory_renames = MERGE_DIRECTORY_RENAMES_CONFLICT;
3769         merge_recursive_config(opt);
3770         merge_verbosity = getenv("GIT_MERGE_VERBOSITY");
3771         if (merge_verbosity)
3772                 opt->verbosity = strtol(merge_verbosity, NULL, 10);
3773         if (opt->verbosity >= 5)
3774                 opt->buffer_output = 0;
3775         strbuf_init(&opt->obuf, 0);
3776         string_list_init(&opt->df_conflict_file_set, 1);
3777 }
3778
3779 int parse_merge_opt(struct merge_options *opt, const char *s)
3780 {
3781         const char *arg;
3782
3783         if (!s || !*s)
3784                 return -1;
3785         if (!strcmp(s, "ours"))
3786                 opt->recursive_variant = MERGE_RECURSIVE_OURS;
3787         else if (!strcmp(s, "theirs"))
3788                 opt->recursive_variant = MERGE_RECURSIVE_THEIRS;
3789         else if (!strcmp(s, "subtree"))
3790                 opt->subtree_shift = "";
3791         else if (skip_prefix(s, "subtree=", &arg))
3792                 opt->subtree_shift = arg;
3793         else if (!strcmp(s, "patience"))
3794                 opt->xdl_opts = DIFF_WITH_ALG(opt, PATIENCE_DIFF);
3795         else if (!strcmp(s, "histogram"))
3796                 opt->xdl_opts = DIFF_WITH_ALG(opt, HISTOGRAM_DIFF);
3797         else if (skip_prefix(s, "diff-algorithm=", &arg)) {
3798                 long value = parse_algorithm_value(arg);
3799                 if (value < 0)
3800                         return -1;
3801                 /* clear out previous settings */
3802                 DIFF_XDL_CLR(opt, NEED_MINIMAL);
3803                 opt->xdl_opts &= ~XDF_DIFF_ALGORITHM_MASK;
3804                 opt->xdl_opts |= value;
3805         }
3806         else if (!strcmp(s, "ignore-space-change"))
3807                 DIFF_XDL_SET(opt, IGNORE_WHITESPACE_CHANGE);
3808         else if (!strcmp(s, "ignore-all-space"))
3809                 DIFF_XDL_SET(opt, IGNORE_WHITESPACE);
3810         else if (!strcmp(s, "ignore-space-at-eol"))
3811                 DIFF_XDL_SET(opt, IGNORE_WHITESPACE_AT_EOL);
3812         else if (!strcmp(s, "ignore-cr-at-eol"))
3813                 DIFF_XDL_SET(opt, IGNORE_CR_AT_EOL);
3814         else if (!strcmp(s, "renormalize"))
3815                 opt->renormalize = 1;
3816         else if (!strcmp(s, "no-renormalize"))
3817                 opt->renormalize = 0;
3818         else if (!strcmp(s, "no-renames"))
3819                 opt->merge_detect_rename = 0;
3820         else if (!strcmp(s, "find-renames")) {
3821                 opt->merge_detect_rename = 1;
3822                 opt->rename_score = 0;
3823         }
3824         else if (skip_prefix(s, "find-renames=", &arg) ||
3825                  skip_prefix(s, "rename-threshold=", &arg)) {
3826                 if ((opt->rename_score = parse_rename_score(&arg)) == -1 || *arg != 0)
3827                         return -1;
3828                 opt->merge_detect_rename = 1;
3829         }
3830         /*
3831          * Please update $__git_merge_strategy_options in
3832          * git-completion.bash when you add new options
3833          */
3834         else
3835                 return -1;
3836         return 0;
3837 }