merge-recursive: use common name for ancestors/common/base_list
[git] / merge-recursive.c
1 /*
2  * Recursive Merge algorithm stolen from git-merge-recursive.py by
3  * Fredrik Kuivinen.
4  * The thieves were Alex Riesen and Johannes Schindelin, in June/July 2006
5  */
6 #include "cache.h"
7 #include "config.h"
8 #include "advice.h"
9 #include "lockfile.h"
10 #include "cache-tree.h"
11 #include "object-store.h"
12 #include "repository.h"
13 #include "commit.h"
14 #include "blob.h"
15 #include "builtin.h"
16 #include "tree-walk.h"
17 #include "diff.h"
18 #include "diffcore.h"
19 #include "tag.h"
20 #include "alloc.h"
21 #include "unpack-trees.h"
22 #include "string-list.h"
23 #include "xdiff-interface.h"
24 #include "ll-merge.h"
25 #include "attr.h"
26 #include "merge-recursive.h"
27 #include "dir.h"
28 #include "submodule.h"
29 #include "revision.h"
30 #include "commit-reach.h"
31
32 struct path_hashmap_entry {
33         struct hashmap_entry e;
34         char path[FLEX_ARRAY];
35 };
36
37 static int path_hashmap_cmp(const void *cmp_data,
38                             const void *entry,
39                             const void *entry_or_key,
40                             const void *keydata)
41 {
42         const struct path_hashmap_entry *a = entry;
43         const struct path_hashmap_entry *b = entry_or_key;
44         const char *key = keydata;
45
46         if (ignore_case)
47                 return strcasecmp(a->path, key ? key : b->path);
48         else
49                 return strcmp(a->path, key ? key : b->path);
50 }
51
52 static unsigned int path_hash(const char *path)
53 {
54         return ignore_case ? strihash(path) : strhash(path);
55 }
56
57 static struct dir_rename_entry *dir_rename_find_entry(struct hashmap *hashmap,
58                                                       char *dir)
59 {
60         struct dir_rename_entry key;
61
62         if (dir == NULL)
63                 return NULL;
64         hashmap_entry_init(&key, strhash(dir));
65         key.dir = dir;
66         return hashmap_get(hashmap, &key, NULL);
67 }
68
69 static int dir_rename_cmp(const void *unused_cmp_data,
70                           const void *entry,
71                           const void *entry_or_key,
72                           const void *unused_keydata)
73 {
74         const struct dir_rename_entry *e1 = entry;
75         const struct dir_rename_entry *e2 = entry_or_key;
76
77         return strcmp(e1->dir, e2->dir);
78 }
79
80 static void dir_rename_init(struct hashmap *map)
81 {
82         hashmap_init(map, dir_rename_cmp, NULL, 0);
83 }
84
85 static void dir_rename_entry_init(struct dir_rename_entry *entry,
86                                   char *directory)
87 {
88         hashmap_entry_init(entry, strhash(directory));
89         entry->dir = directory;
90         entry->non_unique_new_dir = 0;
91         strbuf_init(&entry->new_dir, 0);
92         string_list_init(&entry->possible_new_dirs, 0);
93 }
94
95 static struct collision_entry *collision_find_entry(struct hashmap *hashmap,
96                                                     char *target_file)
97 {
98         struct collision_entry key;
99
100         hashmap_entry_init(&key, strhash(target_file));
101         key.target_file = target_file;
102         return hashmap_get(hashmap, &key, NULL);
103 }
104
105 static int collision_cmp(void *unused_cmp_data,
106                          const struct collision_entry *e1,
107                          const struct collision_entry *e2,
108                          const void *unused_keydata)
109 {
110         return strcmp(e1->target_file, e2->target_file);
111 }
112
113 static void collision_init(struct hashmap *map)
114 {
115         hashmap_init(map, (hashmap_cmp_fn) collision_cmp, NULL, 0);
116 }
117
118 static void flush_output(struct merge_options *opt)
119 {
120         if (opt->buffer_output < 2 && opt->obuf.len) {
121                 fputs(opt->obuf.buf, stdout);
122                 strbuf_reset(&opt->obuf);
123         }
124 }
125
126 static int err(struct merge_options *opt, const char *err, ...)
127 {
128         va_list params;
129
130         if (opt->buffer_output < 2)
131                 flush_output(opt);
132         else {
133                 strbuf_complete(&opt->obuf, '\n');
134                 strbuf_addstr(&opt->obuf, "error: ");
135         }
136         va_start(params, err);
137         strbuf_vaddf(&opt->obuf, err, params);
138         va_end(params);
139         if (opt->buffer_output > 1)
140                 strbuf_addch(&opt->obuf, '\n');
141         else {
142                 error("%s", opt->obuf.buf);
143                 strbuf_reset(&opt->obuf);
144         }
145
146         return -1;
147 }
148
149 static struct tree *shift_tree_object(struct repository *repo,
150                                       struct tree *one, struct tree *two,
151                                       const char *subtree_shift)
152 {
153         struct object_id shifted;
154
155         if (!*subtree_shift) {
156                 shift_tree(repo, &one->object.oid, &two->object.oid, &shifted, 0);
157         } else {
158                 shift_tree_by(repo, &one->object.oid, &two->object.oid, &shifted,
159                               subtree_shift);
160         }
161         if (oideq(&two->object.oid, &shifted))
162                 return two;
163         return lookup_tree(repo, &shifted);
164 }
165
166 static inline void set_commit_tree(struct commit *c, struct tree *t)
167 {
168         c->maybe_tree = t;
169 }
170
171 static struct commit *make_virtual_commit(struct repository *repo,
172                                           struct tree *tree,
173                                           const char *comment)
174 {
175         struct commit *commit = alloc_commit_node(repo);
176
177         set_merge_remote_desc(commit, comment, (struct object *)commit);
178         set_commit_tree(commit, tree);
179         commit->object.parsed = 1;
180         return commit;
181 }
182
183 /*
184  * Since we use get_tree_entry(), which does not put the read object into
185  * the object pool, we cannot rely on a == b.
186  */
187 static int oid_eq(const struct object_id *a, const struct object_id *b)
188 {
189         if (!a && !b)
190                 return 2;
191         return a && b && oideq(a, b);
192 }
193
194 enum rename_type {
195         RENAME_NORMAL = 0,
196         RENAME_VIA_DIR,
197         RENAME_ADD,
198         RENAME_DELETE,
199         RENAME_ONE_FILE_TO_ONE,
200         RENAME_ONE_FILE_TO_TWO,
201         RENAME_TWO_FILES_TO_ONE
202 };
203
204 /*
205  * Since we want to write the index eventually, we cannot reuse the index
206  * for these (temporary) data.
207  */
208 struct stage_data {
209         struct diff_filespec stages[4]; /* mostly for oid & mode; maybe path */
210         struct rename_conflict_info *rename_conflict_info;
211         unsigned processed:1;
212 };
213
214 struct rename {
215         unsigned processed:1;
216         struct diff_filepair *pair;
217         const char *branch; /* branch that the rename occurred on */
218         /*
219          * If directory rename detection affected this rename, what was its
220          * original type ('A' or 'R') and it's original destination before
221          * the directory rename (otherwise, '\0' and NULL for these two vars).
222          */
223         char dir_rename_original_type;
224         char *dir_rename_original_dest;
225         /*
226          * Purpose of src_entry and dst_entry:
227          *
228          * If 'before' is renamed to 'after' then src_entry will contain
229          * the versions of 'before' from the merge_base, HEAD, and MERGE in
230          * stages 1, 2, and 3; dst_entry will contain the respective
231          * versions of 'after' in corresponding locations.  Thus, we have a
232          * total of six modes and oids, though some will be null.  (Stage 0
233          * is ignored; we're interested in handling conflicts.)
234          *
235          * Since we don't turn on break-rewrites by default, neither
236          * src_entry nor dst_entry can have all three of their stages have
237          * non-null oids, meaning at most four of the six will be non-null.
238          * Also, since this is a rename, both src_entry and dst_entry will
239          * have at least one non-null oid, meaning at least two will be
240          * non-null.  Of the six oids, a typical rename will have three be
241          * non-null.  Only two implies a rename/delete, and four implies a
242          * rename/add.
243          */
244         struct stage_data *src_entry;
245         struct stage_data *dst_entry;
246 };
247
248 struct rename_conflict_info {
249         enum rename_type rename_type;
250         struct rename *ren1;
251         struct rename *ren2;
252 };
253
254 static inline void setup_rename_conflict_info(enum rename_type rename_type,
255                                               struct merge_options *opt,
256                                               struct rename *ren1,
257                                               struct rename *ren2)
258 {
259         struct rename_conflict_info *ci;
260
261         /*
262          * When we have two renames involved, it's easiest to get the
263          * correct things into stage 2 and 3, and to make sure that the
264          * content merge puts HEAD before the other branch if we just
265          * ensure that branch1 == opt->branch1.  So, simply flip arguments
266          * around if we don't have that.
267          */
268         if (ren2 && ren1->branch != opt->branch1) {
269                 setup_rename_conflict_info(rename_type, opt, ren2, ren1);
270                 return;
271         }
272
273         ci = xcalloc(1, sizeof(struct rename_conflict_info));
274         ci->rename_type = rename_type;
275         ci->ren1 = ren1;
276         ci->ren2 = ren2;
277
278         ci->ren1->dst_entry->processed = 0;
279         ci->ren1->dst_entry->rename_conflict_info = ci;
280         if (ren2) {
281                 ci->ren2->dst_entry->rename_conflict_info = ci;
282         }
283 }
284
285 static int show(struct merge_options *opt, int v)
286 {
287         return (!opt->call_depth && opt->verbosity >= v) || opt->verbosity >= 5;
288 }
289
290 __attribute__((format (printf, 3, 4)))
291 static void output(struct merge_options *opt, int v, const char *fmt, ...)
292 {
293         va_list ap;
294
295         if (!show(opt, v))
296                 return;
297
298         strbuf_addchars(&opt->obuf, ' ', opt->call_depth * 2);
299
300         va_start(ap, fmt);
301         strbuf_vaddf(&opt->obuf, fmt, ap);
302         va_end(ap);
303
304         strbuf_addch(&opt->obuf, '\n');
305         if (!opt->buffer_output)
306                 flush_output(opt);
307 }
308
309 static void output_commit_title(struct merge_options *opt, struct commit *commit)
310 {
311         struct merge_remote_desc *desc;
312
313         strbuf_addchars(&opt->obuf, ' ', opt->call_depth * 2);
314         desc = merge_remote_util(commit);
315         if (desc)
316                 strbuf_addf(&opt->obuf, "virtual %s\n", desc->name);
317         else {
318                 strbuf_add_unique_abbrev(&opt->obuf, &commit->object.oid,
319                                          DEFAULT_ABBREV);
320                 strbuf_addch(&opt->obuf, ' ');
321                 if (parse_commit(commit) != 0)
322                         strbuf_addstr(&opt->obuf, _("(bad commit)\n"));
323                 else {
324                         const char *title;
325                         const char *msg = get_commit_buffer(commit, NULL);
326                         int len = find_commit_subject(msg, &title);
327                         if (len)
328                                 strbuf_addf(&opt->obuf, "%.*s\n", len, title);
329                         unuse_commit_buffer(commit, msg);
330                 }
331         }
332         flush_output(opt);
333 }
334
335 static int add_cacheinfo(struct merge_options *opt,
336                          const struct diff_filespec *blob,
337                          const char *path, int stage, int refresh, int options)
338 {
339         struct index_state *istate = opt->repo->index;
340         struct cache_entry *ce;
341         int ret;
342
343         ce = make_cache_entry(istate, blob->mode, &blob->oid, path, stage, 0);
344         if (!ce)
345                 return err(opt, _("add_cacheinfo failed for path '%s'; merge aborting."), path);
346
347         ret = add_index_entry(istate, ce, options);
348         if (refresh) {
349                 struct cache_entry *nce;
350
351                 nce = refresh_cache_entry(istate, ce,
352                                           CE_MATCH_REFRESH | CE_MATCH_IGNORE_MISSING);
353                 if (!nce)
354                         return err(opt, _("add_cacheinfo failed to refresh for path '%s'; merge aborting."), path);
355                 if (nce != ce)
356                         ret = add_index_entry(istate, nce, options);
357         }
358         return ret;
359 }
360
361 static void init_tree_desc_from_tree(struct tree_desc *desc, struct tree *tree)
362 {
363         parse_tree(tree);
364         init_tree_desc(desc, tree->buffer, tree->size);
365 }
366
367 static int unpack_trees_start(struct merge_options *opt,
368                               struct tree *common,
369                               struct tree *head,
370                               struct tree *merge)
371 {
372         int rc;
373         struct tree_desc t[3];
374         struct index_state tmp_index = { NULL };
375
376         memset(&opt->unpack_opts, 0, sizeof(opt->unpack_opts));
377         if (opt->call_depth)
378                 opt->unpack_opts.index_only = 1;
379         else
380                 opt->unpack_opts.update = 1;
381         opt->unpack_opts.merge = 1;
382         opt->unpack_opts.head_idx = 2;
383         opt->unpack_opts.fn = threeway_merge;
384         opt->unpack_opts.src_index = opt->repo->index;
385         opt->unpack_opts.dst_index = &tmp_index;
386         opt->unpack_opts.aggressive = !merge_detect_rename(opt);
387         setup_unpack_trees_porcelain(&opt->unpack_opts, "merge");
388
389         init_tree_desc_from_tree(t+0, common);
390         init_tree_desc_from_tree(t+1, head);
391         init_tree_desc_from_tree(t+2, merge);
392
393         rc = unpack_trees(3, t, &opt->unpack_opts);
394         cache_tree_free(&opt->repo->index->cache_tree);
395
396         /*
397          * Update opt->repo->index to match the new results, AFTER saving a copy
398          * in opt->orig_index.  Update src_index to point to the saved copy.
399          * (verify_uptodate() checks src_index, and the original index is
400          * the one that had the necessary modification timestamps.)
401          */
402         opt->orig_index = *opt->repo->index;
403         *opt->repo->index = tmp_index;
404         opt->unpack_opts.src_index = &opt->orig_index;
405
406         return rc;
407 }
408
409 static void unpack_trees_finish(struct merge_options *opt)
410 {
411         discard_index(&opt->orig_index);
412         clear_unpack_trees_porcelain(&opt->unpack_opts);
413 }
414
415 static int save_files_dirs(const struct object_id *oid,
416                            struct strbuf *base, const char *path,
417                            unsigned int mode, int stage, void *context)
418 {
419         struct path_hashmap_entry *entry;
420         int baselen = base->len;
421         struct merge_options *opt = context;
422
423         strbuf_addstr(base, path);
424
425         FLEX_ALLOC_MEM(entry, path, base->buf, base->len);
426         hashmap_entry_init(entry, path_hash(entry->path));
427         hashmap_add(&opt->current_file_dir_set, entry);
428
429         strbuf_setlen(base, baselen);
430         return (S_ISDIR(mode) ? READ_TREE_RECURSIVE : 0);
431 }
432
433 static void get_files_dirs(struct merge_options *opt, struct tree *tree)
434 {
435         struct pathspec match_all;
436         memset(&match_all, 0, sizeof(match_all));
437         read_tree_recursive(opt->repo, tree, "", 0, 0,
438                             &match_all, save_files_dirs, opt);
439 }
440
441 static int get_tree_entry_if_blob(struct repository *r,
442                                   const struct object_id *tree,
443                                   const char *path,
444                                   struct diff_filespec *dfs)
445 {
446         int ret;
447
448         ret = get_tree_entry(r, tree, path, &dfs->oid, &dfs->mode);
449         if (S_ISDIR(dfs->mode)) {
450                 oidcpy(&dfs->oid, &null_oid);
451                 dfs->mode = 0;
452         }
453         return ret;
454 }
455
456 /*
457  * Returns an index_entry instance which doesn't have to correspond to
458  * a real cache entry in Git's index.
459  */
460 static struct stage_data *insert_stage_data(struct repository *r,
461                 const char *path,
462                 struct tree *o, struct tree *a, struct tree *b,
463                 struct string_list *entries)
464 {
465         struct string_list_item *item;
466         struct stage_data *e = xcalloc(1, sizeof(struct stage_data));
467         get_tree_entry_if_blob(r, &o->object.oid, path, &e->stages[1]);
468         get_tree_entry_if_blob(r, &a->object.oid, path, &e->stages[2]);
469         get_tree_entry_if_blob(r, &b->object.oid, path, &e->stages[3]);
470         item = string_list_insert(entries, path);
471         item->util = e;
472         return e;
473 }
474
475 /*
476  * Create a dictionary mapping file names to stage_data objects. The
477  * dictionary contains one entry for every path with a non-zero stage entry.
478  */
479 static struct string_list *get_unmerged(struct index_state *istate)
480 {
481         struct string_list *unmerged = xcalloc(1, sizeof(struct string_list));
482         int i;
483
484         unmerged->strdup_strings = 1;
485
486         for (i = 0; i < istate->cache_nr; i++) {
487                 struct string_list_item *item;
488                 struct stage_data *e;
489                 const struct cache_entry *ce = istate->cache[i];
490                 if (!ce_stage(ce))
491                         continue;
492
493                 item = string_list_lookup(unmerged, ce->name);
494                 if (!item) {
495                         item = string_list_insert(unmerged, ce->name);
496                         item->util = xcalloc(1, sizeof(struct stage_data));
497                 }
498                 e = item->util;
499                 e->stages[ce_stage(ce)].mode = ce->ce_mode;
500                 oidcpy(&e->stages[ce_stage(ce)].oid, &ce->oid);
501         }
502
503         return unmerged;
504 }
505
506 static int string_list_df_name_compare(const char *one, const char *two)
507 {
508         int onelen = strlen(one);
509         int twolen = strlen(two);
510         /*
511          * Here we only care that entries for D/F conflicts are
512          * adjacent, in particular with the file of the D/F conflict
513          * appearing before files below the corresponding directory.
514          * The order of the rest of the list is irrelevant for us.
515          *
516          * To achieve this, we sort with df_name_compare and provide
517          * the mode S_IFDIR so that D/F conflicts will sort correctly.
518          * We use the mode S_IFDIR for everything else for simplicity,
519          * since in other cases any changes in their order due to
520          * sorting cause no problems for us.
521          */
522         int cmp = df_name_compare(one, onelen, S_IFDIR,
523                                   two, twolen, S_IFDIR);
524         /*
525          * Now that 'foo' and 'foo/bar' compare equal, we have to make sure
526          * that 'foo' comes before 'foo/bar'.
527          */
528         if (cmp)
529                 return cmp;
530         return onelen - twolen;
531 }
532
533 static void record_df_conflict_files(struct merge_options *opt,
534                                      struct string_list *entries)
535 {
536         /* If there is a D/F conflict and the file for such a conflict
537          * currently exists in the working tree, we want to allow it to be
538          * removed to make room for the corresponding directory if needed.
539          * The files underneath the directories of such D/F conflicts will
540          * be processed before the corresponding file involved in the D/F
541          * conflict.  If the D/F directory ends up being removed by the
542          * merge, then we won't have to touch the D/F file.  If the D/F
543          * directory needs to be written to the working copy, then the D/F
544          * file will simply be removed (in make_room_for_path()) to make
545          * room for the necessary paths.  Note that if both the directory
546          * and the file need to be present, then the D/F file will be
547          * reinstated with a new unique name at the time it is processed.
548          */
549         struct string_list df_sorted_entries = STRING_LIST_INIT_NODUP;
550         const char *last_file = NULL;
551         int last_len = 0;
552         int i;
553
554         /*
555          * If we're merging merge-bases, we don't want to bother with
556          * any working directory changes.
557          */
558         if (opt->call_depth)
559                 return;
560
561         /* Ensure D/F conflicts are adjacent in the entries list. */
562         for (i = 0; i < entries->nr; i++) {
563                 struct string_list_item *next = &entries->items[i];
564                 string_list_append(&df_sorted_entries, next->string)->util =
565                                    next->util;
566         }
567         df_sorted_entries.cmp = string_list_df_name_compare;
568         string_list_sort(&df_sorted_entries);
569
570         string_list_clear(&opt->df_conflict_file_set, 1);
571         for (i = 0; i < df_sorted_entries.nr; i++) {
572                 const char *path = df_sorted_entries.items[i].string;
573                 int len = strlen(path);
574                 struct stage_data *e = df_sorted_entries.items[i].util;
575
576                 /*
577                  * Check if last_file & path correspond to a D/F conflict;
578                  * i.e. whether path is last_file+'/'+<something>.
579                  * If so, record that it's okay to remove last_file to make
580                  * room for path and friends if needed.
581                  */
582                 if (last_file &&
583                     len > last_len &&
584                     memcmp(path, last_file, last_len) == 0 &&
585                     path[last_len] == '/') {
586                         string_list_insert(&opt->df_conflict_file_set, last_file);
587                 }
588
589                 /*
590                  * Determine whether path could exist as a file in the
591                  * working directory as a possible D/F conflict.  This
592                  * will only occur when it exists in stage 2 as a
593                  * file.
594                  */
595                 if (S_ISREG(e->stages[2].mode) || S_ISLNK(e->stages[2].mode)) {
596                         last_file = path;
597                         last_len = len;
598                 } else {
599                         last_file = NULL;
600                 }
601         }
602         string_list_clear(&df_sorted_entries, 0);
603 }
604
605 static int update_stages(struct merge_options *opt, const char *path,
606                          const struct diff_filespec *o,
607                          const struct diff_filespec *a,
608                          const struct diff_filespec *b)
609 {
610
611         /*
612          * NOTE: It is usually a bad idea to call update_stages on a path
613          * before calling update_file on that same path, since it can
614          * sometimes lead to spurious "refusing to lose untracked file..."
615          * messages from update_file (via make_room_for path via
616          * would_lose_untracked).  Instead, reverse the order of the calls
617          * (executing update_file first and then update_stages).
618          */
619         int clear = 1;
620         int options = ADD_CACHE_OK_TO_ADD | ADD_CACHE_SKIP_DFCHECK;
621         if (clear)
622                 if (remove_file_from_index(opt->repo->index, path))
623                         return -1;
624         if (o)
625                 if (add_cacheinfo(opt, o, path, 1, 0, options))
626                         return -1;
627         if (a)
628                 if (add_cacheinfo(opt, a, path, 2, 0, options))
629                         return -1;
630         if (b)
631                 if (add_cacheinfo(opt, b, path, 3, 0, options))
632                         return -1;
633         return 0;
634 }
635
636 static void update_entry(struct stage_data *entry,
637                          struct diff_filespec *o,
638                          struct diff_filespec *a,
639                          struct diff_filespec *b)
640 {
641         entry->processed = 0;
642         entry->stages[1].mode = o->mode;
643         entry->stages[2].mode = a->mode;
644         entry->stages[3].mode = b->mode;
645         oidcpy(&entry->stages[1].oid, &o->oid);
646         oidcpy(&entry->stages[2].oid, &a->oid);
647         oidcpy(&entry->stages[3].oid, &b->oid);
648 }
649
650 static int remove_file(struct merge_options *opt, int clean,
651                        const char *path, int no_wd)
652 {
653         int update_cache = opt->call_depth || clean;
654         int update_working_directory = !opt->call_depth && !no_wd;
655
656         if (update_cache) {
657                 if (remove_file_from_index(opt->repo->index, path))
658                         return -1;
659         }
660         if (update_working_directory) {
661                 if (ignore_case) {
662                         struct cache_entry *ce;
663                         ce = index_file_exists(opt->repo->index, path, strlen(path),
664                                                ignore_case);
665                         if (ce && ce_stage(ce) == 0 && strcmp(path, ce->name))
666                                 return 0;
667                 }
668                 if (remove_path(path))
669                         return -1;
670         }
671         return 0;
672 }
673
674 /* add a string to a strbuf, but converting "/" to "_" */
675 static void add_flattened_path(struct strbuf *out, const char *s)
676 {
677         size_t i = out->len;
678         strbuf_addstr(out, s);
679         for (; i < out->len; i++)
680                 if (out->buf[i] == '/')
681                         out->buf[i] = '_';
682 }
683
684 static char *unique_path(struct merge_options *opt,
685                          const char *path,
686                          const char *branch)
687 {
688         struct path_hashmap_entry *entry;
689         struct strbuf newpath = STRBUF_INIT;
690         int suffix = 0;
691         size_t base_len;
692
693         strbuf_addf(&newpath, "%s~", path);
694         add_flattened_path(&newpath, branch);
695
696         base_len = newpath.len;
697         while (hashmap_get_from_hash(&opt->current_file_dir_set,
698                                      path_hash(newpath.buf), newpath.buf) ||
699                (!opt->call_depth && file_exists(newpath.buf))) {
700                 strbuf_setlen(&newpath, base_len);
701                 strbuf_addf(&newpath, "_%d", suffix++);
702         }
703
704         FLEX_ALLOC_MEM(entry, path, newpath.buf, newpath.len);
705         hashmap_entry_init(entry, path_hash(entry->path));
706         hashmap_add(&opt->current_file_dir_set, entry);
707         return strbuf_detach(&newpath, NULL);
708 }
709
710 /**
711  * Check whether a directory in the index is in the way of an incoming
712  * file.  Return 1 if so.  If check_working_copy is non-zero, also
713  * check the working directory.  If empty_ok is non-zero, also return
714  * 0 in the case where the working-tree dir exists but is empty.
715  */
716 static int dir_in_way(struct index_state *istate, const char *path,
717                       int check_working_copy, int empty_ok)
718 {
719         int pos;
720         struct strbuf dirpath = STRBUF_INIT;
721         struct stat st;
722
723         strbuf_addstr(&dirpath, path);
724         strbuf_addch(&dirpath, '/');
725
726         pos = index_name_pos(istate, dirpath.buf, dirpath.len);
727
728         if (pos < 0)
729                 pos = -1 - pos;
730         if (pos < istate->cache_nr &&
731             !strncmp(dirpath.buf, istate->cache[pos]->name, dirpath.len)) {
732                 strbuf_release(&dirpath);
733                 return 1;
734         }
735
736         strbuf_release(&dirpath);
737         return check_working_copy && !lstat(path, &st) && S_ISDIR(st.st_mode) &&
738                 !(empty_ok && is_empty_dir(path));
739 }
740
741 /*
742  * Returns whether path was tracked in the index before the merge started,
743  * and its oid and mode match the specified values
744  */
745 static int was_tracked_and_matches(struct merge_options *opt, const char *path,
746                                    const struct diff_filespec *blob)
747 {
748         int pos = index_name_pos(&opt->orig_index, path, strlen(path));
749         struct cache_entry *ce;
750
751         if (0 > pos)
752                 /* we were not tracking this path before the merge */
753                 return 0;
754
755         /* See if the file we were tracking before matches */
756         ce = opt->orig_index.cache[pos];
757         return (oid_eq(&ce->oid, &blob->oid) && ce->ce_mode == blob->mode);
758 }
759
760 /*
761  * Returns whether path was tracked in the index before the merge started
762  */
763 static int was_tracked(struct merge_options *opt, const char *path)
764 {
765         int pos = index_name_pos(&opt->orig_index, path, strlen(path));
766
767         if (0 <= pos)
768                 /* we were tracking this path before the merge */
769                 return 1;
770
771         return 0;
772 }
773
774 static int would_lose_untracked(struct merge_options *opt, const char *path)
775 {
776         struct index_state *istate = opt->repo->index;
777
778         /*
779          * This may look like it can be simplified to:
780          *   return !was_tracked(opt, path) && file_exists(path)
781          * but it can't.  This function needs to know whether path was in
782          * the working tree due to EITHER having been tracked in the index
783          * before the merge OR having been put into the working copy and
784          * index by unpack_trees().  Due to that either-or requirement, we
785          * check the current index instead of the original one.
786          *
787          * Note that we do not need to worry about merge-recursive itself
788          * updating the index after unpack_trees() and before calling this
789          * function, because we strictly require all code paths in
790          * merge-recursive to update the working tree first and the index
791          * second.  Doing otherwise would break
792          * update_file()/would_lose_untracked(); see every comment in this
793          * file which mentions "update_stages".
794          */
795         int pos = index_name_pos(istate, path, strlen(path));
796
797         if (pos < 0)
798                 pos = -1 - pos;
799         while (pos < istate->cache_nr &&
800                !strcmp(path, istate->cache[pos]->name)) {
801                 /*
802                  * If stage #0, it is definitely tracked.
803                  * If it has stage #2 then it was tracked
804                  * before this merge started.  All other
805                  * cases the path was not tracked.
806                  */
807                 switch (ce_stage(istate->cache[pos])) {
808                 case 0:
809                 case 2:
810                         return 0;
811                 }
812                 pos++;
813         }
814         return file_exists(path);
815 }
816
817 static int was_dirty(struct merge_options *opt, const char *path)
818 {
819         struct cache_entry *ce;
820         int dirty = 1;
821
822         if (opt->call_depth || !was_tracked(opt, path))
823                 return !dirty;
824
825         ce = index_file_exists(opt->unpack_opts.src_index,
826                                path, strlen(path), ignore_case);
827         dirty = verify_uptodate(ce, &opt->unpack_opts) != 0;
828         return dirty;
829 }
830
831 static int make_room_for_path(struct merge_options *opt, const char *path)
832 {
833         int status, i;
834         const char *msg = _("failed to create path '%s'%s");
835
836         /* Unlink any D/F conflict files that are in the way */
837         for (i = 0; i < opt->df_conflict_file_set.nr; i++) {
838                 const char *df_path = opt->df_conflict_file_set.items[i].string;
839                 size_t pathlen = strlen(path);
840                 size_t df_pathlen = strlen(df_path);
841                 if (df_pathlen < pathlen &&
842                     path[df_pathlen] == '/' &&
843                     strncmp(path, df_path, df_pathlen) == 0) {
844                         output(opt, 3,
845                                _("Removing %s to make room for subdirectory\n"),
846                                df_path);
847                         unlink(df_path);
848                         unsorted_string_list_delete_item(&opt->df_conflict_file_set,
849                                                          i, 0);
850                         break;
851                 }
852         }
853
854         /* Make sure leading directories are created */
855         status = safe_create_leading_directories_const(path);
856         if (status) {
857                 if (status == SCLD_EXISTS)
858                         /* something else exists */
859                         return err(opt, msg, path, _(": perhaps a D/F conflict?"));
860                 return err(opt, msg, path, "");
861         }
862
863         /*
864          * Do not unlink a file in the work tree if we are not
865          * tracking it.
866          */
867         if (would_lose_untracked(opt, path))
868                 return err(opt, _("refusing to lose untracked file at '%s'"),
869                            path);
870
871         /* Successful unlink is good.. */
872         if (!unlink(path))
873                 return 0;
874         /* .. and so is no existing file */
875         if (errno == ENOENT)
876                 return 0;
877         /* .. but not some other error (who really cares what?) */
878         return err(opt, msg, path, _(": perhaps a D/F conflict?"));
879 }
880
881 static int update_file_flags(struct merge_options *opt,
882                              const struct diff_filespec *contents,
883                              const char *path,
884                              int update_cache,
885                              int update_wd)
886 {
887         int ret = 0;
888
889         if (opt->call_depth)
890                 update_wd = 0;
891
892         if (update_wd) {
893                 enum object_type type;
894                 void *buf;
895                 unsigned long size;
896
897                 if (S_ISGITLINK(contents->mode)) {
898                         /*
899                          * We may later decide to recursively descend into
900                          * the submodule directory and update its index
901                          * and/or work tree, but we do not do that now.
902                          */
903                         update_wd = 0;
904                         goto update_index;
905                 }
906
907                 buf = read_object_file(&contents->oid, &type, &size);
908                 if (!buf) {
909                         ret = err(opt, _("cannot read object %s '%s'"),
910                                   oid_to_hex(&contents->oid), path);
911                         goto free_buf;
912                 }
913                 if (type != OBJ_BLOB) {
914                         ret = err(opt, _("blob expected for %s '%s'"),
915                                   oid_to_hex(&contents->oid), path);
916                         goto free_buf;
917                 }
918                 if (S_ISREG(contents->mode)) {
919                         struct strbuf strbuf = STRBUF_INIT;
920                         if (convert_to_working_tree(opt->repo->index,
921                                                     path, buf, size, &strbuf)) {
922                                 free(buf);
923                                 size = strbuf.len;
924                                 buf = strbuf_detach(&strbuf, NULL);
925                         }
926                 }
927
928                 if (make_room_for_path(opt, path) < 0) {
929                         update_wd = 0;
930                         goto free_buf;
931                 }
932                 if (S_ISREG(contents->mode) ||
933                     (!has_symlinks && S_ISLNK(contents->mode))) {
934                         int fd;
935                         int mode = (contents->mode & 0100 ? 0777 : 0666);
936
937                         fd = open(path, O_WRONLY | O_TRUNC | O_CREAT, mode);
938                         if (fd < 0) {
939                                 ret = err(opt, _("failed to open '%s': %s"),
940                                           path, strerror(errno));
941                                 goto free_buf;
942                         }
943                         write_in_full(fd, buf, size);
944                         close(fd);
945                 } else if (S_ISLNK(contents->mode)) {
946                         char *lnk = xmemdupz(buf, size);
947                         safe_create_leading_directories_const(path);
948                         unlink(path);
949                         if (symlink(lnk, path))
950                                 ret = err(opt, _("failed to symlink '%s': %s"),
951                                           path, strerror(errno));
952                         free(lnk);
953                 } else
954                         ret = err(opt,
955                                   _("do not know what to do with %06o %s '%s'"),
956                                   contents->mode, oid_to_hex(&contents->oid), path);
957         free_buf:
958                 free(buf);
959         }
960 update_index:
961         if (!ret && update_cache)
962                 if (add_cacheinfo(opt, contents, path, 0, update_wd,
963                                   ADD_CACHE_OK_TO_ADD))
964                         return -1;
965         return ret;
966 }
967
968 static int update_file(struct merge_options *opt,
969                        int clean,
970                        const struct diff_filespec *contents,
971                        const char *path)
972 {
973         return update_file_flags(opt, contents, path,
974                                  opt->call_depth || clean, !opt->call_depth);
975 }
976
977 /* Low level file merging, update and removal */
978
979 struct merge_file_info {
980         struct diff_filespec blob; /* mostly use oid & mode; sometimes path */
981         unsigned clean:1,
982                  merge:1;
983 };
984
985 static int merge_3way(struct merge_options *opt,
986                       mmbuffer_t *result_buf,
987                       const struct diff_filespec *o,
988                       const struct diff_filespec *a,
989                       const struct diff_filespec *b,
990                       const char *branch1,
991                       const char *branch2,
992                       const int extra_marker_size)
993 {
994         mmfile_t orig, src1, src2;
995         struct ll_merge_options ll_opts = {0};
996         char *base, *name1, *name2;
997         int merge_status;
998
999         ll_opts.renormalize = opt->renormalize;
1000         ll_opts.extra_marker_size = extra_marker_size;
1001         ll_opts.xdl_opts = opt->xdl_opts;
1002
1003         if (opt->call_depth) {
1004                 ll_opts.virtual_ancestor = 1;
1005                 ll_opts.variant = 0;
1006         } else {
1007                 switch (opt->recursive_variant) {
1008                 case MERGE_RECURSIVE_OURS:
1009                         ll_opts.variant = XDL_MERGE_FAVOR_OURS;
1010                         break;
1011                 case MERGE_RECURSIVE_THEIRS:
1012                         ll_opts.variant = XDL_MERGE_FAVOR_THEIRS;
1013                         break;
1014                 default:
1015                         ll_opts.variant = 0;
1016                         break;
1017                 }
1018         }
1019
1020         assert(a->path && b->path && o->path && opt->ancestor);
1021         if (strcmp(a->path, b->path) || strcmp(a->path, o->path) != 0) {
1022                 base  = mkpathdup("%s:%s", opt->ancestor, o->path);
1023                 name1 = mkpathdup("%s:%s", branch1, a->path);
1024                 name2 = mkpathdup("%s:%s", branch2, b->path);
1025         } else {
1026                 base  = mkpathdup("%s", opt->ancestor);
1027                 name1 = mkpathdup("%s", branch1);
1028                 name2 = mkpathdup("%s", branch2);
1029         }
1030
1031         read_mmblob(&orig, &o->oid);
1032         read_mmblob(&src1, &a->oid);
1033         read_mmblob(&src2, &b->oid);
1034
1035         merge_status = ll_merge(result_buf, a->path, &orig, base,
1036                                 &src1, name1, &src2, name2,
1037                                 opt->repo->index, &ll_opts);
1038
1039         free(base);
1040         free(name1);
1041         free(name2);
1042         free(orig.ptr);
1043         free(src1.ptr);
1044         free(src2.ptr);
1045         return merge_status;
1046 }
1047
1048 static int find_first_merges(struct repository *repo,
1049                              struct object_array *result, const char *path,
1050                              struct commit *a, struct commit *b)
1051 {
1052         int i, j;
1053         struct object_array merges = OBJECT_ARRAY_INIT;
1054         struct commit *commit;
1055         int contains_another;
1056
1057         char merged_revision[GIT_MAX_HEXSZ + 2];
1058         const char *rev_args[] = { "rev-list", "--merges", "--ancestry-path",
1059                                    "--all", merged_revision, NULL };
1060         struct rev_info revs;
1061         struct setup_revision_opt rev_opts;
1062
1063         memset(result, 0, sizeof(struct object_array));
1064         memset(&rev_opts, 0, sizeof(rev_opts));
1065
1066         /* get all revisions that merge commit a */
1067         xsnprintf(merged_revision, sizeof(merged_revision), "^%s",
1068                   oid_to_hex(&a->object.oid));
1069         repo_init_revisions(repo, &revs, NULL);
1070         rev_opts.submodule = path;
1071         /* FIXME: can't handle linked worktrees in submodules yet */
1072         revs.single_worktree = path != NULL;
1073         setup_revisions(ARRAY_SIZE(rev_args)-1, rev_args, &revs, &rev_opts);
1074
1075         /* save all revisions from the above list that contain b */
1076         if (prepare_revision_walk(&revs))
1077                 die("revision walk setup failed");
1078         while ((commit = get_revision(&revs)) != NULL) {
1079                 struct object *o = &(commit->object);
1080                 if (in_merge_bases(b, commit))
1081                         add_object_array(o, NULL, &merges);
1082         }
1083         reset_revision_walk();
1084
1085         /* Now we've got all merges that contain a and b. Prune all
1086          * merges that contain another found merge and save them in
1087          * result.
1088          */
1089         for (i = 0; i < merges.nr; i++) {
1090                 struct commit *m1 = (struct commit *) merges.objects[i].item;
1091
1092                 contains_another = 0;
1093                 for (j = 0; j < merges.nr; j++) {
1094                         struct commit *m2 = (struct commit *) merges.objects[j].item;
1095                         if (i != j && in_merge_bases(m2, m1)) {
1096                                 contains_another = 1;
1097                                 break;
1098                         }
1099                 }
1100
1101                 if (!contains_another)
1102                         add_object_array(merges.objects[i].item, NULL, result);
1103         }
1104
1105         object_array_clear(&merges);
1106         return result->nr;
1107 }
1108
1109 static void print_commit(struct commit *commit)
1110 {
1111         struct strbuf sb = STRBUF_INIT;
1112         struct pretty_print_context ctx = {0};
1113         ctx.date_mode.type = DATE_NORMAL;
1114         format_commit_message(commit, " %h: %m %s", &sb, &ctx);
1115         fprintf(stderr, "%s\n", sb.buf);
1116         strbuf_release(&sb);
1117 }
1118
1119 static int is_valid(const struct diff_filespec *dfs)
1120 {
1121         return dfs->mode != 0 && !is_null_oid(&dfs->oid);
1122 }
1123
1124 static int merge_submodule(struct merge_options *opt,
1125                            struct object_id *result, const char *path,
1126                            const struct object_id *base, const struct object_id *a,
1127                            const struct object_id *b)
1128 {
1129         struct commit *commit_base, *commit_a, *commit_b;
1130         int parent_count;
1131         struct object_array merges;
1132
1133         int i;
1134         int search = !opt->call_depth;
1135
1136         /* store a in result in case we fail */
1137         oidcpy(result, a);
1138
1139         /* we can not handle deletion conflicts */
1140         if (is_null_oid(base))
1141                 return 0;
1142         if (is_null_oid(a))
1143                 return 0;
1144         if (is_null_oid(b))
1145                 return 0;
1146
1147         if (add_submodule_odb(path)) {
1148                 output(opt, 1, _("Failed to merge submodule %s (not checked out)"), path);
1149                 return 0;
1150         }
1151
1152         if (!(commit_base = lookup_commit_reference(opt->repo, base)) ||
1153             !(commit_a = lookup_commit_reference(opt->repo, a)) ||
1154             !(commit_b = lookup_commit_reference(opt->repo, b))) {
1155                 output(opt, 1, _("Failed to merge submodule %s (commits not present)"), path);
1156                 return 0;
1157         }
1158
1159         /* check whether both changes are forward */
1160         if (!in_merge_bases(commit_base, commit_a) ||
1161             !in_merge_bases(commit_base, commit_b)) {
1162                 output(opt, 1, _("Failed to merge submodule %s (commits don't follow merge-base)"), path);
1163                 return 0;
1164         }
1165
1166         /* Case #1: a is contained in b or vice versa */
1167         if (in_merge_bases(commit_a, commit_b)) {
1168                 oidcpy(result, b);
1169                 if (show(opt, 3)) {
1170                         output(opt, 3, _("Fast-forwarding submodule %s to the following commit:"), path);
1171                         output_commit_title(opt, commit_b);
1172                 } else if (show(opt, 2))
1173                         output(opt, 2, _("Fast-forwarding submodule %s"), path);
1174                 else
1175                         ; /* no output */
1176
1177                 return 1;
1178         }
1179         if (in_merge_bases(commit_b, commit_a)) {
1180                 oidcpy(result, a);
1181                 if (show(opt, 3)) {
1182                         output(opt, 3, _("Fast-forwarding submodule %s to the following commit:"), path);
1183                         output_commit_title(opt, commit_a);
1184                 } else if (show(opt, 2))
1185                         output(opt, 2, _("Fast-forwarding submodule %s"), path);
1186                 else
1187                         ; /* no output */
1188
1189                 return 1;
1190         }
1191
1192         /*
1193          * Case #2: There are one or more merges that contain a and b in
1194          * the submodule. If there is only one, then present it as a
1195          * suggestion to the user, but leave it marked unmerged so the
1196          * user needs to confirm the resolution.
1197          */
1198
1199         /* Skip the search if makes no sense to the calling context.  */
1200         if (!search)
1201                 return 0;
1202
1203         /* find commit which merges them */
1204         parent_count = find_first_merges(opt->repo, &merges, path,
1205                                          commit_a, commit_b);
1206         switch (parent_count) {
1207         case 0:
1208                 output(opt, 1, _("Failed to merge submodule %s (merge following commits not found)"), path);
1209                 break;
1210
1211         case 1:
1212                 output(opt, 1, _("Failed to merge submodule %s (not fast-forward)"), path);
1213                 output(opt, 2, _("Found a possible merge resolution for the submodule:\n"));
1214                 print_commit((struct commit *) merges.objects[0].item);
1215                 output(opt, 2, _(
1216                        "If this is correct simply add it to the index "
1217                        "for example\n"
1218                        "by using:\n\n"
1219                        "  git update-index --cacheinfo 160000 %s \"%s\"\n\n"
1220                        "which will accept this suggestion.\n"),
1221                        oid_to_hex(&merges.objects[0].item->oid), path);
1222                 break;
1223
1224         default:
1225                 output(opt, 1, _("Failed to merge submodule %s (multiple merges found)"), path);
1226                 for (i = 0; i < merges.nr; i++)
1227                         print_commit((struct commit *) merges.objects[i].item);
1228         }
1229
1230         object_array_clear(&merges);
1231         return 0;
1232 }
1233
1234 static int merge_mode_and_contents(struct merge_options *opt,
1235                                    const struct diff_filespec *o,
1236                                    const struct diff_filespec *a,
1237                                    const struct diff_filespec *b,
1238                                    const char *filename,
1239                                    const char *branch1,
1240                                    const char *branch2,
1241                                    const int extra_marker_size,
1242                                    struct merge_file_info *result)
1243 {
1244         if (opt->branch1 != branch1) {
1245                 /*
1246                  * It's weird getting a reverse merge with HEAD on the bottom
1247                  * side of the conflict markers and the other branch on the
1248                  * top.  Fix that.
1249                  */
1250                 return merge_mode_and_contents(opt, o, b, a,
1251                                                filename,
1252                                                branch2, branch1,
1253                                                extra_marker_size, result);
1254         }
1255
1256         result->merge = 0;
1257         result->clean = 1;
1258
1259         if ((S_IFMT & a->mode) != (S_IFMT & b->mode)) {
1260                 result->clean = 0;
1261                 if (S_ISREG(a->mode)) {
1262                         result->blob.mode = a->mode;
1263                         oidcpy(&result->blob.oid, &a->oid);
1264                 } else {
1265                         result->blob.mode = b->mode;
1266                         oidcpy(&result->blob.oid, &b->oid);
1267                 }
1268         } else {
1269                 if (!oid_eq(&a->oid, &o->oid) && !oid_eq(&b->oid, &o->oid))
1270                         result->merge = 1;
1271
1272                 /*
1273                  * Merge modes
1274                  */
1275                 if (a->mode == b->mode || a->mode == o->mode)
1276                         result->blob.mode = b->mode;
1277                 else {
1278                         result->blob.mode = a->mode;
1279                         if (b->mode != o->mode) {
1280                                 result->clean = 0;
1281                                 result->merge = 1;
1282                         }
1283                 }
1284
1285                 if (oid_eq(&a->oid, &b->oid) || oid_eq(&a->oid, &o->oid))
1286                         oidcpy(&result->blob.oid, &b->oid);
1287                 else if (oid_eq(&b->oid, &o->oid))
1288                         oidcpy(&result->blob.oid, &a->oid);
1289                 else if (S_ISREG(a->mode)) {
1290                         mmbuffer_t result_buf;
1291                         int ret = 0, merge_status;
1292
1293                         merge_status = merge_3way(opt, &result_buf, o, a, b,
1294                                                   branch1, branch2,
1295                                                   extra_marker_size);
1296
1297                         if ((merge_status < 0) || !result_buf.ptr)
1298                                 ret = err(opt, _("Failed to execute internal merge"));
1299
1300                         if (!ret &&
1301                             write_object_file(result_buf.ptr, result_buf.size,
1302                                               blob_type, &result->blob.oid))
1303                                 ret = err(opt, _("Unable to add %s to database"),
1304                                           a->path);
1305
1306                         free(result_buf.ptr);
1307                         if (ret)
1308                                 return ret;
1309                         result->clean = (merge_status == 0);
1310                 } else if (S_ISGITLINK(a->mode)) {
1311                         result->clean = merge_submodule(opt, &result->blob.oid,
1312                                                         o->path,
1313                                                         &o->oid,
1314                                                         &a->oid,
1315                                                         &b->oid);
1316                 } else if (S_ISLNK(a->mode)) {
1317                         switch (opt->recursive_variant) {
1318                         case MERGE_RECURSIVE_NORMAL:
1319                                 oidcpy(&result->blob.oid, &a->oid);
1320                                 if (!oid_eq(&a->oid, &b->oid))
1321                                         result->clean = 0;
1322                                 break;
1323                         case MERGE_RECURSIVE_OURS:
1324                                 oidcpy(&result->blob.oid, &a->oid);
1325                                 break;
1326                         case MERGE_RECURSIVE_THEIRS:
1327                                 oidcpy(&result->blob.oid, &b->oid);
1328                                 break;
1329                         }
1330                 } else
1331                         BUG("unsupported object type in the tree");
1332         }
1333
1334         if (result->merge)
1335                 output(opt, 2, _("Auto-merging %s"), filename);
1336
1337         return 0;
1338 }
1339
1340 static int handle_rename_via_dir(struct merge_options *opt,
1341                                  struct rename_conflict_info *ci)
1342 {
1343         /*
1344          * Handle file adds that need to be renamed due to directory rename
1345          * detection.  This differs from handle_rename_normal, because
1346          * there is no content merge to do; just move the file into the
1347          * desired final location.
1348          */
1349         const struct rename *ren = ci->ren1;
1350         const struct diff_filespec *dest = ren->pair->two;
1351         char *file_path = dest->path;
1352         int mark_conflicted = (opt->detect_directory_renames ==
1353                                MERGE_DIRECTORY_RENAMES_CONFLICT);
1354         assert(ren->dir_rename_original_dest);
1355
1356         if (!opt->call_depth && would_lose_untracked(opt, dest->path)) {
1357                 mark_conflicted = 1;
1358                 file_path = unique_path(opt, dest->path, ren->branch);
1359                 output(opt, 1, _("Error: Refusing to lose untracked file at %s; "
1360                                  "writing to %s instead."),
1361                        dest->path, file_path);
1362         }
1363
1364         if (mark_conflicted) {
1365                 /*
1366                  * Write the file in worktree at file_path.  In the index,
1367                  * only record the file at dest->path in the appropriate
1368                  * higher stage.
1369                  */
1370                 if (update_file(opt, 0, dest, file_path))
1371                         return -1;
1372                 if (file_path != dest->path)
1373                         free(file_path);
1374                 if (update_stages(opt, dest->path, NULL,
1375                                   ren->branch == opt->branch1 ? dest : NULL,
1376                                   ren->branch == opt->branch1 ? NULL : dest))
1377                         return -1;
1378                 return 0; /* not clean, but conflicted */
1379         } else {
1380                 /* Update dest->path both in index and in worktree */
1381                 if (update_file(opt, 1, dest, dest->path))
1382                         return -1;
1383                 return 1; /* clean */
1384         }
1385 }
1386
1387 static int handle_change_delete(struct merge_options *opt,
1388                                 const char *path, const char *old_path,
1389                                 const struct diff_filespec *o,
1390                                 const struct diff_filespec *changed,
1391                                 const char *change_branch,
1392                                 const char *delete_branch,
1393                                 const char *change, const char *change_past)
1394 {
1395         char *alt_path = NULL;
1396         const char *update_path = path;
1397         int ret = 0;
1398
1399         if (dir_in_way(opt->repo->index, path, !opt->call_depth, 0) ||
1400             (!opt->call_depth && would_lose_untracked(opt, path))) {
1401                 update_path = alt_path = unique_path(opt, path, change_branch);
1402         }
1403
1404         if (opt->call_depth) {
1405                 /*
1406                  * We cannot arbitrarily accept either a_sha or b_sha as
1407                  * correct; since there is no true "middle point" between
1408                  * them, simply reuse the base version for virtual merge base.
1409                  */
1410                 ret = remove_file_from_index(opt->repo->index, path);
1411                 if (!ret)
1412                         ret = update_file(opt, 0, o, update_path);
1413         } else {
1414                 /*
1415                  * Despite the four nearly duplicate messages and argument
1416                  * lists below and the ugliness of the nested if-statements,
1417                  * having complete messages makes the job easier for
1418                  * translators.
1419                  *
1420                  * The slight variance among the cases is due to the fact
1421                  * that:
1422                  *   1) directory/file conflicts (in effect if
1423                  *      !alt_path) could cause us to need to write the
1424                  *      file to a different path.
1425                  *   2) renames (in effect if !old_path) could mean that
1426                  *      there are two names for the path that the user
1427                  *      may know the file by.
1428                  */
1429                 if (!alt_path) {
1430                         if (!old_path) {
1431                                 output(opt, 1, _("CONFLICT (%s/delete): %s deleted in %s "
1432                                        "and %s in %s. Version %s of %s left in tree."),
1433                                        change, path, delete_branch, change_past,
1434                                        change_branch, change_branch, path);
1435                         } else {
1436                                 output(opt, 1, _("CONFLICT (%s/delete): %s deleted in %s "
1437                                        "and %s to %s in %s. Version %s of %s left in tree."),
1438                                        change, old_path, delete_branch, change_past, path,
1439                                        change_branch, change_branch, path);
1440                         }
1441                 } else {
1442                         if (!old_path) {
1443                                 output(opt, 1, _("CONFLICT (%s/delete): %s deleted in %s "
1444                                        "and %s in %s. Version %s of %s left in tree at %s."),
1445                                        change, path, delete_branch, change_past,
1446                                        change_branch, change_branch, path, alt_path);
1447                         } else {
1448                                 output(opt, 1, _("CONFLICT (%s/delete): %s deleted in %s "
1449                                        "and %s to %s in %s. Version %s of %s left in tree at %s."),
1450                                        change, old_path, delete_branch, change_past, path,
1451                                        change_branch, change_branch, path, alt_path);
1452                         }
1453                 }
1454                 /*
1455                  * No need to call update_file() on path when change_branch ==
1456                  * opt->branch1 && !alt_path, since that would needlessly touch
1457                  * path.  We could call update_file_flags() with update_cache=0
1458                  * and update_wd=0, but that's a no-op.
1459                  */
1460                 if (change_branch != opt->branch1 || alt_path)
1461                         ret = update_file(opt, 0, changed, update_path);
1462         }
1463         free(alt_path);
1464
1465         return ret;
1466 }
1467
1468 static int handle_rename_delete(struct merge_options *opt,
1469                                 struct rename_conflict_info *ci)
1470 {
1471         const struct rename *ren = ci->ren1;
1472         const struct diff_filespec *orig = ren->pair->one;
1473         const struct diff_filespec *dest = ren->pair->two;
1474         const char *rename_branch = ren->branch;
1475         const char *delete_branch = (opt->branch1 == ren->branch ?
1476                                      opt->branch2 : opt->branch1);
1477
1478         if (handle_change_delete(opt,
1479                                  opt->call_depth ? orig->path : dest->path,
1480                                  opt->call_depth ? NULL : orig->path,
1481                                  orig, dest,
1482                                  rename_branch, delete_branch,
1483                                  _("rename"), _("renamed")))
1484                 return -1;
1485
1486         if (opt->call_depth)
1487                 return remove_file_from_index(opt->repo->index, dest->path);
1488         else
1489                 return update_stages(opt, dest->path, NULL,
1490                                      rename_branch == opt->branch1 ? dest : NULL,
1491                                      rename_branch == opt->branch1 ? NULL : dest);
1492 }
1493
1494 static int handle_file_collision(struct merge_options *opt,
1495                                  const char *collide_path,
1496                                  const char *prev_path1,
1497                                  const char *prev_path2,
1498                                  const char *branch1, const char *branch2,
1499                                  struct diff_filespec *a,
1500                                  struct diff_filespec *b)
1501 {
1502         struct merge_file_info mfi;
1503         struct diff_filespec null;
1504         char *alt_path = NULL;
1505         const char *update_path = collide_path;
1506
1507         /*
1508          * It's easiest to get the correct things into stage 2 and 3, and
1509          * to make sure that the content merge puts HEAD before the other
1510          * branch if we just ensure that branch1 == opt->branch1.  So, simply
1511          * flip arguments around if we don't have that.
1512          */
1513         if (branch1 != opt->branch1) {
1514                 return handle_file_collision(opt, collide_path,
1515                                              prev_path2, prev_path1,
1516                                              branch2, branch1,
1517                                              b, a);
1518         }
1519
1520         /*
1521          * In the recursive case, we just opt to undo renames
1522          */
1523         if (opt->call_depth && (prev_path1 || prev_path2)) {
1524                 /* Put first file (a->oid, a->mode) in its original spot */
1525                 if (prev_path1) {
1526                         if (update_file(opt, 1, a, prev_path1))
1527                                 return -1;
1528                 } else {
1529                         if (update_file(opt, 1, a, collide_path))
1530                                 return -1;
1531                 }
1532
1533                 /* Put second file (b->oid, b->mode) in its original spot */
1534                 if (prev_path2) {
1535                         if (update_file(opt, 1, b, prev_path2))
1536                                 return -1;
1537                 } else {
1538                         if (update_file(opt, 1, b, collide_path))
1539                                 return -1;
1540                 }
1541
1542                 /* Don't leave something at collision path if unrenaming both */
1543                 if (prev_path1 && prev_path2)
1544                         remove_file(opt, 1, collide_path, 0);
1545
1546                 return 0;
1547         }
1548
1549         /* Remove rename sources if rename/add or rename/rename(2to1) */
1550         if (prev_path1)
1551                 remove_file(opt, 1, prev_path1,
1552                             opt->call_depth || would_lose_untracked(opt, prev_path1));
1553         if (prev_path2)
1554                 remove_file(opt, 1, prev_path2,
1555                             opt->call_depth || would_lose_untracked(opt, prev_path2));
1556
1557         /*
1558          * Remove the collision path, if it wouldn't cause dirty contents
1559          * or an untracked file to get lost.  We'll either overwrite with
1560          * merged contents, or just write out to differently named files.
1561          */
1562         if (was_dirty(opt, collide_path)) {
1563                 output(opt, 1, _("Refusing to lose dirty file at %s"),
1564                        collide_path);
1565                 update_path = alt_path = unique_path(opt, collide_path, "merged");
1566         } else if (would_lose_untracked(opt, collide_path)) {
1567                 /*
1568                  * Only way we get here is if both renames were from
1569                  * a directory rename AND user had an untracked file
1570                  * at the location where both files end up after the
1571                  * two directory renames.  See testcase 10d of t6043.
1572                  */
1573                 output(opt, 1, _("Refusing to lose untracked file at "
1574                                "%s, even though it's in the way."),
1575                        collide_path);
1576                 update_path = alt_path = unique_path(opt, collide_path, "merged");
1577         } else {
1578                 /*
1579                  * FIXME: It's possible that the two files are identical
1580                  * and that the current working copy happens to match, in
1581                  * which case we are unnecessarily touching the working
1582                  * tree file.  It's not a likely enough scenario that I
1583                  * want to code up the checks for it and a better fix is
1584                  * available if we restructure how unpack_trees() and
1585                  * merge-recursive interoperate anyway, so punting for
1586                  * now...
1587                  */
1588                 remove_file(opt, 0, collide_path, 0);
1589         }
1590
1591         /* Store things in diff_filespecs for functions that need it */
1592         null.path = (char *)collide_path;
1593         oidcpy(&null.oid, &null_oid);
1594         null.mode = 0;
1595
1596         if (merge_mode_and_contents(opt, &null, a, b, collide_path,
1597                                     branch1, branch2, opt->call_depth * 2, &mfi))
1598                 return -1;
1599         mfi.clean &= !alt_path;
1600         if (update_file(opt, mfi.clean, &mfi.blob, update_path))
1601                 return -1;
1602         if (!mfi.clean && !opt->call_depth &&
1603             update_stages(opt, collide_path, NULL, a, b))
1604                 return -1;
1605         free(alt_path);
1606         /*
1607          * FIXME: If both a & b both started with conflicts (only possible
1608          * if they came from a rename/rename(2to1)), but had IDENTICAL
1609          * contents including those conflicts, then in the next line we claim
1610          * it was clean.  If someone cares about this case, we should have the
1611          * caller notify us if we started with conflicts.
1612          */
1613         return mfi.clean;
1614 }
1615
1616 static int handle_rename_add(struct merge_options *opt,
1617                              struct rename_conflict_info *ci)
1618 {
1619         /* a was renamed to c, and a separate c was added. */
1620         struct diff_filespec *a = ci->ren1->pair->one;
1621         struct diff_filespec *c = ci->ren1->pair->two;
1622         char *path = c->path;
1623         char *prev_path_desc;
1624         struct merge_file_info mfi;
1625
1626         const char *rename_branch = ci->ren1->branch;
1627         const char *add_branch = (opt->branch1 == rename_branch ?
1628                                   opt->branch2 : opt->branch1);
1629         int other_stage = (ci->ren1->branch == opt->branch1 ? 3 : 2);
1630
1631         output(opt, 1, _("CONFLICT (rename/add): "
1632                "Rename %s->%s in %s.  Added %s in %s"),
1633                a->path, c->path, rename_branch,
1634                c->path, add_branch);
1635
1636         prev_path_desc = xstrfmt("version of %s from %s", path, a->path);
1637         ci->ren1->src_entry->stages[other_stage].path = a->path;
1638         if (merge_mode_and_contents(opt, a, c,
1639                                     &ci->ren1->src_entry->stages[other_stage],
1640                                     prev_path_desc,
1641                                     opt->branch1, opt->branch2,
1642                                     1 + opt->call_depth * 2, &mfi))
1643                 return -1;
1644         free(prev_path_desc);
1645
1646         ci->ren1->dst_entry->stages[other_stage].path = mfi.blob.path = c->path;
1647         return handle_file_collision(opt,
1648                                      c->path, a->path, NULL,
1649                                      rename_branch, add_branch,
1650                                      &mfi.blob,
1651                                      &ci->ren1->dst_entry->stages[other_stage]);
1652 }
1653
1654 static char *find_path_for_conflict(struct merge_options *opt,
1655                                     const char *path,
1656                                     const char *branch1,
1657                                     const char *branch2)
1658 {
1659         char *new_path = NULL;
1660         if (dir_in_way(opt->repo->index, path, !opt->call_depth, 0)) {
1661                 new_path = unique_path(opt, path, branch1);
1662                 output(opt, 1, _("%s is a directory in %s adding "
1663                                "as %s instead"),
1664                        path, branch2, new_path);
1665         } else if (would_lose_untracked(opt, path)) {
1666                 new_path = unique_path(opt, path, branch1);
1667                 output(opt, 1, _("Refusing to lose untracked file"
1668                                " at %s; adding as %s instead"),
1669                        path, new_path);
1670         }
1671
1672         return new_path;
1673 }
1674
1675 static int handle_rename_rename_1to2(struct merge_options *opt,
1676                                      struct rename_conflict_info *ci)
1677 {
1678         /* One file was renamed in both branches, but to different names. */
1679         struct merge_file_info mfi;
1680         struct diff_filespec *add;
1681         struct diff_filespec *o = ci->ren1->pair->one;
1682         struct diff_filespec *a = ci->ren1->pair->two;
1683         struct diff_filespec *b = ci->ren2->pair->two;
1684         char *path_desc;
1685
1686         output(opt, 1, _("CONFLICT (rename/rename): "
1687                "Rename \"%s\"->\"%s\" in branch \"%s\" "
1688                "rename \"%s\"->\"%s\" in \"%s\"%s"),
1689                o->path, a->path, ci->ren1->branch,
1690                o->path, b->path, ci->ren2->branch,
1691                opt->call_depth ? _(" (left unresolved)") : "");
1692
1693         path_desc = xstrfmt("%s and %s, both renamed from %s",
1694                             a->path, b->path, o->path);
1695         if (merge_mode_and_contents(opt, o, a, b, path_desc,
1696                                     ci->ren1->branch, ci->ren2->branch,
1697                                     opt->call_depth * 2, &mfi))
1698                 return -1;
1699         free(path_desc);
1700
1701         if (opt->call_depth) {
1702                 /*
1703                  * FIXME: For rename/add-source conflicts (if we could detect
1704                  * such), this is wrong.  We should instead find a unique
1705                  * pathname and then either rename the add-source file to that
1706                  * unique path, or use that unique path instead of src here.
1707                  */
1708                 if (update_file(opt, 0, &mfi.blob, o->path))
1709                         return -1;
1710
1711                 /*
1712                  * Above, we put the merged content at the merge-base's
1713                  * path.  Now we usually need to delete both a->path and
1714                  * b->path.  However, the rename on each side of the merge
1715                  * could also be involved in a rename/add conflict.  In
1716                  * such cases, we should keep the added file around,
1717                  * resolving the conflict at that path in its favor.
1718                  */
1719                 add = &ci->ren1->dst_entry->stages[2 ^ 1];
1720                 if (is_valid(add)) {
1721                         if (update_file(opt, 0, add, a->path))
1722                                 return -1;
1723                 }
1724                 else
1725                         remove_file_from_index(opt->repo->index, a->path);
1726                 add = &ci->ren2->dst_entry->stages[3 ^ 1];
1727                 if (is_valid(add)) {
1728                         if (update_file(opt, 0, add, b->path))
1729                                 return -1;
1730                 }
1731                 else
1732                         remove_file_from_index(opt->repo->index, b->path);
1733         } else {
1734                 /*
1735                  * For each destination path, we need to see if there is a
1736                  * rename/add collision.  If not, we can write the file out
1737                  * to the specified location.
1738                  */
1739                 add = &ci->ren1->dst_entry->stages[2 ^ 1];
1740                 if (is_valid(add)) {
1741                         add->path = mfi.blob.path = a->path;
1742                         if (handle_file_collision(opt, a->path,
1743                                                   NULL, NULL,
1744                                                   ci->ren1->branch,
1745                                                   ci->ren2->branch,
1746                                                   &mfi.blob, add) < 0)
1747                                 return -1;
1748                 } else {
1749                         char *new_path = find_path_for_conflict(opt, a->path,
1750                                                                 ci->ren1->branch,
1751                                                                 ci->ren2->branch);
1752                         if (update_file(opt, 0, &mfi.blob,
1753                                         new_path ? new_path : a->path))
1754                                 return -1;
1755                         free(new_path);
1756                         if (update_stages(opt, a->path, NULL, a, NULL))
1757                                 return -1;
1758                 }
1759
1760                 add = &ci->ren2->dst_entry->stages[3 ^ 1];
1761                 if (is_valid(add)) {
1762                         add->path = mfi.blob.path = b->path;
1763                         if (handle_file_collision(opt, b->path,
1764                                                   NULL, NULL,
1765                                                   ci->ren1->branch,
1766                                                   ci->ren2->branch,
1767                                                   add, &mfi.blob) < 0)
1768                                 return -1;
1769                 } else {
1770                         char *new_path = find_path_for_conflict(opt, b->path,
1771                                                                 ci->ren2->branch,
1772                                                                 ci->ren1->branch);
1773                         if (update_file(opt, 0, &mfi.blob,
1774                                         new_path ? new_path : b->path))
1775                                 return -1;
1776                         free(new_path);
1777                         if (update_stages(opt, b->path, NULL, NULL, b))
1778                                 return -1;
1779                 }
1780         }
1781
1782         return 0;
1783 }
1784
1785 static int handle_rename_rename_2to1(struct merge_options *opt,
1786                                      struct rename_conflict_info *ci)
1787 {
1788         /* Two files, a & b, were renamed to the same thing, c. */
1789         struct diff_filespec *a = ci->ren1->pair->one;
1790         struct diff_filespec *b = ci->ren2->pair->one;
1791         struct diff_filespec *c1 = ci->ren1->pair->two;
1792         struct diff_filespec *c2 = ci->ren2->pair->two;
1793         char *path = c1->path; /* == c2->path */
1794         char *path_side_1_desc;
1795         char *path_side_2_desc;
1796         struct merge_file_info mfi_c1;
1797         struct merge_file_info mfi_c2;
1798         int ostage1, ostage2;
1799
1800         output(opt, 1, _("CONFLICT (rename/rename): "
1801                "Rename %s->%s in %s. "
1802                "Rename %s->%s in %s"),
1803                a->path, c1->path, ci->ren1->branch,
1804                b->path, c2->path, ci->ren2->branch);
1805
1806         path_side_1_desc = xstrfmt("version of %s from %s", path, a->path);
1807         path_side_2_desc = xstrfmt("version of %s from %s", path, b->path);
1808         ostage1 = ci->ren1->branch == opt->branch1 ? 3 : 2;
1809         ostage2 = ostage1 ^ 1;
1810         ci->ren1->src_entry->stages[ostage1].path = a->path;
1811         ci->ren2->src_entry->stages[ostage2].path = b->path;
1812         if (merge_mode_and_contents(opt, a, c1,
1813                                     &ci->ren1->src_entry->stages[ostage1],
1814                                     path_side_1_desc,
1815                                     opt->branch1, opt->branch2,
1816                                     1 + opt->call_depth * 2, &mfi_c1) ||
1817             merge_mode_and_contents(opt, b,
1818                                     &ci->ren2->src_entry->stages[ostage2],
1819                                     c2, path_side_2_desc,
1820                                     opt->branch1, opt->branch2,
1821                                     1 + opt->call_depth * 2, &mfi_c2))
1822                 return -1;
1823         free(path_side_1_desc);
1824         free(path_side_2_desc);
1825         mfi_c1.blob.path = path;
1826         mfi_c2.blob.path = path;
1827
1828         return handle_file_collision(opt, path, a->path, b->path,
1829                                      ci->ren1->branch, ci->ren2->branch,
1830                                      &mfi_c1.blob, &mfi_c2.blob);
1831 }
1832
1833 /*
1834  * Get the diff_filepairs changed between o_tree and tree.
1835  */
1836 static struct diff_queue_struct *get_diffpairs(struct merge_options *opt,
1837                                                struct tree *o_tree,
1838                                                struct tree *tree)
1839 {
1840         struct diff_queue_struct *ret;
1841         struct diff_options opts;
1842
1843         repo_diff_setup(opt->repo, &opts);
1844         opts.flags.recursive = 1;
1845         opts.flags.rename_empty = 0;
1846         opts.detect_rename = merge_detect_rename(opt);
1847         /*
1848          * We do not have logic to handle the detection of copies.  In
1849          * fact, it may not even make sense to add such logic: would we
1850          * really want a change to a base file to be propagated through
1851          * multiple other files by a merge?
1852          */
1853         if (opts.detect_rename > DIFF_DETECT_RENAME)
1854                 opts.detect_rename = DIFF_DETECT_RENAME;
1855         opts.rename_limit = opt->merge_rename_limit >= 0 ? opt->merge_rename_limit :
1856                             opt->diff_rename_limit >= 0 ? opt->diff_rename_limit :
1857                             1000;
1858         opts.rename_score = opt->rename_score;
1859         opts.show_rename_progress = opt->show_rename_progress;
1860         opts.output_format = DIFF_FORMAT_NO_OUTPUT;
1861         diff_setup_done(&opts);
1862         diff_tree_oid(&o_tree->object.oid, &tree->object.oid, "", &opts);
1863         diffcore_std(&opts);
1864         if (opts.needed_rename_limit > opt->needed_rename_limit)
1865                 opt->needed_rename_limit = opts.needed_rename_limit;
1866
1867         ret = xmalloc(sizeof(*ret));
1868         *ret = diff_queued_diff;
1869
1870         opts.output_format = DIFF_FORMAT_NO_OUTPUT;
1871         diff_queued_diff.nr = 0;
1872         diff_queued_diff.queue = NULL;
1873         diff_flush(&opts);
1874         return ret;
1875 }
1876
1877 static int tree_has_path(struct repository *r, struct tree *tree,
1878                          const char *path)
1879 {
1880         struct object_id hashy;
1881         unsigned short mode_o;
1882
1883         return !get_tree_entry(r,
1884                                &tree->object.oid, path,
1885                                &hashy, &mode_o);
1886 }
1887
1888 /*
1889  * Return a new string that replaces the beginning portion (which matches
1890  * entry->dir), with entry->new_dir.  In perl-speak:
1891  *   new_path_name = (old_path =~ s/entry->dir/entry->new_dir/);
1892  * NOTE:
1893  *   Caller must ensure that old_path starts with entry->dir + '/'.
1894  */
1895 static char *apply_dir_rename(struct dir_rename_entry *entry,
1896                               const char *old_path)
1897 {
1898         struct strbuf new_path = STRBUF_INIT;
1899         int oldlen, newlen;
1900
1901         if (entry->non_unique_new_dir)
1902                 return NULL;
1903
1904         oldlen = strlen(entry->dir);
1905         newlen = entry->new_dir.len + (strlen(old_path) - oldlen) + 1;
1906         strbuf_grow(&new_path, newlen);
1907         strbuf_addbuf(&new_path, &entry->new_dir);
1908         strbuf_addstr(&new_path, &old_path[oldlen]);
1909
1910         return strbuf_detach(&new_path, NULL);
1911 }
1912
1913 static void get_renamed_dir_portion(const char *old_path, const char *new_path,
1914                                     char **old_dir, char **new_dir)
1915 {
1916         char *end_of_old, *end_of_new;
1917         int old_len, new_len;
1918
1919         *old_dir = NULL;
1920         *new_dir = NULL;
1921
1922         /*
1923          * For
1924          *    "a/b/c/d/e/foo.c" -> "a/b/some/thing/else/e/foo.c"
1925          * the "e/foo.c" part is the same, we just want to know that
1926          *    "a/b/c/d" was renamed to "a/b/some/thing/else"
1927          * so, for this example, this function returns "a/b/c/d" in
1928          * *old_dir and "a/b/some/thing/else" in *new_dir.
1929          *
1930          * Also, if the basename of the file changed, we don't care.  We
1931          * want to know which portion of the directory, if any, changed.
1932          */
1933         end_of_old = strrchr(old_path, '/');
1934         end_of_new = strrchr(new_path, '/');
1935
1936         if (end_of_old == NULL || end_of_new == NULL)
1937                 return;
1938         while (*--end_of_new == *--end_of_old &&
1939                end_of_old != old_path &&
1940                end_of_new != new_path)
1941                 ; /* Do nothing; all in the while loop */
1942         /*
1943          * We've found the first non-matching character in the directory
1944          * paths.  That means the current directory we were comparing
1945          * represents the rename.  Move end_of_old and end_of_new back
1946          * to the full directory name.
1947          */
1948         if (*end_of_old == '/')
1949                 end_of_old++;
1950         if (*end_of_old != '/')
1951                 end_of_new++;
1952         end_of_old = strchr(end_of_old, '/');
1953         end_of_new = strchr(end_of_new, '/');
1954
1955         /*
1956          * It may have been the case that old_path and new_path were the same
1957          * directory all along.  Don't claim a rename if they're the same.
1958          */
1959         old_len = end_of_old - old_path;
1960         new_len = end_of_new - new_path;
1961
1962         if (old_len != new_len || strncmp(old_path, new_path, old_len)) {
1963                 *old_dir = xstrndup(old_path, old_len);
1964                 *new_dir = xstrndup(new_path, new_len);
1965         }
1966 }
1967
1968 static void remove_hashmap_entries(struct hashmap *dir_renames,
1969                                    struct string_list *items_to_remove)
1970 {
1971         int i;
1972         struct dir_rename_entry *entry;
1973
1974         for (i = 0; i < items_to_remove->nr; i++) {
1975                 entry = items_to_remove->items[i].util;
1976                 hashmap_remove(dir_renames, entry, NULL);
1977         }
1978         string_list_clear(items_to_remove, 0);
1979 }
1980
1981 /*
1982  * See if there is a directory rename for path, and if there are any file
1983  * level conflicts for the renamed location.  If there is a rename and
1984  * there are no conflicts, return the new name.  Otherwise, return NULL.
1985  */
1986 static char *handle_path_level_conflicts(struct merge_options *opt,
1987                                          const char *path,
1988                                          struct dir_rename_entry *entry,
1989                                          struct hashmap *collisions,
1990                                          struct tree *tree)
1991 {
1992         char *new_path = NULL;
1993         struct collision_entry *collision_ent;
1994         int clean = 1;
1995         struct strbuf collision_paths = STRBUF_INIT;
1996
1997         /*
1998          * entry has the mapping of old directory name to new directory name
1999          * that we want to apply to path.
2000          */
2001         new_path = apply_dir_rename(entry, path);
2002
2003         if (!new_path) {
2004                 /* This should only happen when entry->non_unique_new_dir set */
2005                 if (!entry->non_unique_new_dir)
2006                         BUG("entry->non_unqiue_dir not set and !new_path");
2007                 output(opt, 1, _("CONFLICT (directory rename split): "
2008                                "Unclear where to place %s because directory "
2009                                "%s was renamed to multiple other directories, "
2010                                "with no destination getting a majority of the "
2011                                "files."),
2012                        path, entry->dir);
2013                 clean = 0;
2014                 return NULL;
2015         }
2016
2017         /*
2018          * The caller needs to have ensured that it has pre-populated
2019          * collisions with all paths that map to new_path.  Do a quick check
2020          * to ensure that's the case.
2021          */
2022         collision_ent = collision_find_entry(collisions, new_path);
2023         if (collision_ent == NULL)
2024                 BUG("collision_ent is NULL");
2025
2026         /*
2027          * Check for one-sided add/add/.../add conflicts, i.e.
2028          * where implicit renames from the other side doing
2029          * directory rename(s) can affect this side of history
2030          * to put multiple paths into the same location.  Warn
2031          * and bail on directory renames for such paths.
2032          */
2033         if (collision_ent->reported_already) {
2034                 clean = 0;
2035         } else if (tree_has_path(opt->repo, tree, new_path)) {
2036                 collision_ent->reported_already = 1;
2037                 strbuf_add_separated_string_list(&collision_paths, ", ",
2038                                                  &collision_ent->source_files);
2039                 output(opt, 1, _("CONFLICT (implicit dir rename): Existing "
2040                                "file/dir at %s in the way of implicit "
2041                                "directory rename(s) putting the following "
2042                                "path(s) there: %s."),
2043                        new_path, collision_paths.buf);
2044                 clean = 0;
2045         } else if (collision_ent->source_files.nr > 1) {
2046                 collision_ent->reported_already = 1;
2047                 strbuf_add_separated_string_list(&collision_paths, ", ",
2048                                                  &collision_ent->source_files);
2049                 output(opt, 1, _("CONFLICT (implicit dir rename): Cannot map "
2050                                "more than one path to %s; implicit directory "
2051                                "renames tried to put these paths there: %s"),
2052                        new_path, collision_paths.buf);
2053                 clean = 0;
2054         }
2055
2056         /* Free memory we no longer need */
2057         strbuf_release(&collision_paths);
2058         if (!clean && new_path) {
2059                 free(new_path);
2060                 return NULL;
2061         }
2062
2063         return new_path;
2064 }
2065
2066 /*
2067  * There are a couple things we want to do at the directory level:
2068  *   1. Check for both sides renaming to the same thing, in order to avoid
2069  *      implicit renaming of files that should be left in place.  (See
2070  *      testcase 6b in t6043 for details.)
2071  *   2. Prune directory renames if there are still files left in the
2072  *      the original directory.  These represent a partial directory rename,
2073  *      i.e. a rename where only some of the files within the directory
2074  *      were renamed elsewhere.  (Technically, this could be done earlier
2075  *      in get_directory_renames(), except that would prevent us from
2076  *      doing the previous check and thus failing testcase 6b.)
2077  *   3. Check for rename/rename(1to2) conflicts (at the directory level).
2078  *      In the future, we could potentially record this info as well and
2079  *      omit reporting rename/rename(1to2) conflicts for each path within
2080  *      the affected directories, thus cleaning up the merge output.
2081  *   NOTE: We do NOT check for rename/rename(2to1) conflicts at the
2082  *         directory level, because merging directories is fine.  If it
2083  *         causes conflicts for files within those merged directories, then
2084  *         that should be detected at the individual path level.
2085  */
2086 static void handle_directory_level_conflicts(struct merge_options *opt,
2087                                              struct hashmap *dir_re_head,
2088                                              struct tree *head,
2089                                              struct hashmap *dir_re_merge,
2090                                              struct tree *merge)
2091 {
2092         struct hashmap_iter iter;
2093         struct dir_rename_entry *head_ent;
2094         struct dir_rename_entry *merge_ent;
2095
2096         struct string_list remove_from_head = STRING_LIST_INIT_NODUP;
2097         struct string_list remove_from_merge = STRING_LIST_INIT_NODUP;
2098
2099         hashmap_iter_init(dir_re_head, &iter);
2100         while ((head_ent = hashmap_iter_next(&iter))) {
2101                 merge_ent = dir_rename_find_entry(dir_re_merge, head_ent->dir);
2102                 if (merge_ent &&
2103                     !head_ent->non_unique_new_dir &&
2104                     !merge_ent->non_unique_new_dir &&
2105                     !strbuf_cmp(&head_ent->new_dir, &merge_ent->new_dir)) {
2106                         /* 1. Renamed identically; remove it from both sides */
2107                         string_list_append(&remove_from_head,
2108                                            head_ent->dir)->util = head_ent;
2109                         strbuf_release(&head_ent->new_dir);
2110                         string_list_append(&remove_from_merge,
2111                                            merge_ent->dir)->util = merge_ent;
2112                         strbuf_release(&merge_ent->new_dir);
2113                 } else if (tree_has_path(opt->repo, head, head_ent->dir)) {
2114                         /* 2. This wasn't a directory rename after all */
2115                         string_list_append(&remove_from_head,
2116                                            head_ent->dir)->util = head_ent;
2117                         strbuf_release(&head_ent->new_dir);
2118                 }
2119         }
2120
2121         remove_hashmap_entries(dir_re_head, &remove_from_head);
2122         remove_hashmap_entries(dir_re_merge, &remove_from_merge);
2123
2124         hashmap_iter_init(dir_re_merge, &iter);
2125         while ((merge_ent = hashmap_iter_next(&iter))) {
2126                 head_ent = dir_rename_find_entry(dir_re_head, merge_ent->dir);
2127                 if (tree_has_path(opt->repo, merge, merge_ent->dir)) {
2128                         /* 2. This wasn't a directory rename after all */
2129                         string_list_append(&remove_from_merge,
2130                                            merge_ent->dir)->util = merge_ent;
2131                 } else if (head_ent &&
2132                            !head_ent->non_unique_new_dir &&
2133                            !merge_ent->non_unique_new_dir) {
2134                         /* 3. rename/rename(1to2) */
2135                         /*
2136                          * We can assume it's not rename/rename(1to1) because
2137                          * that was case (1), already checked above.  So we
2138                          * know that head_ent->new_dir and merge_ent->new_dir
2139                          * are different strings.
2140                          */
2141                         output(opt, 1, _("CONFLICT (rename/rename): "
2142                                        "Rename directory %s->%s in %s. "
2143                                        "Rename directory %s->%s in %s"),
2144                                head_ent->dir, head_ent->new_dir.buf, opt->branch1,
2145                                head_ent->dir, merge_ent->new_dir.buf, opt->branch2);
2146                         string_list_append(&remove_from_head,
2147                                            head_ent->dir)->util = head_ent;
2148                         strbuf_release(&head_ent->new_dir);
2149                         string_list_append(&remove_from_merge,
2150                                            merge_ent->dir)->util = merge_ent;
2151                         strbuf_release(&merge_ent->new_dir);
2152                 }
2153         }
2154
2155         remove_hashmap_entries(dir_re_head, &remove_from_head);
2156         remove_hashmap_entries(dir_re_merge, &remove_from_merge);
2157 }
2158
2159 static struct hashmap *get_directory_renames(struct diff_queue_struct *pairs)
2160 {
2161         struct hashmap *dir_renames;
2162         struct hashmap_iter iter;
2163         struct dir_rename_entry *entry;
2164         int i;
2165
2166         /*
2167          * Typically, we think of a directory rename as all files from a
2168          * certain directory being moved to a target directory.  However,
2169          * what if someone first moved two files from the original
2170          * directory in one commit, and then renamed the directory
2171          * somewhere else in a later commit?  At merge time, we just know
2172          * that files from the original directory went to two different
2173          * places, and that the bulk of them ended up in the same place.
2174          * We want each directory rename to represent where the bulk of the
2175          * files from that directory end up; this function exists to find
2176          * where the bulk of the files went.
2177          *
2178          * The first loop below simply iterates through the list of file
2179          * renames, finding out how often each directory rename pair
2180          * possibility occurs.
2181          */
2182         dir_renames = xmalloc(sizeof(*dir_renames));
2183         dir_rename_init(dir_renames);
2184         for (i = 0; i < pairs->nr; ++i) {
2185                 struct string_list_item *item;
2186                 int *count;
2187                 struct diff_filepair *pair = pairs->queue[i];
2188                 char *old_dir, *new_dir;
2189
2190                 /* File not part of directory rename if it wasn't renamed */
2191                 if (pair->status != 'R')
2192                         continue;
2193
2194                 get_renamed_dir_portion(pair->one->path, pair->two->path,
2195                                         &old_dir,        &new_dir);
2196                 if (!old_dir)
2197                         /* Directory didn't change at all; ignore this one. */
2198                         continue;
2199
2200                 entry = dir_rename_find_entry(dir_renames, old_dir);
2201                 if (!entry) {
2202                         entry = xmalloc(sizeof(*entry));
2203                         dir_rename_entry_init(entry, old_dir);
2204                         hashmap_put(dir_renames, entry);
2205                 } else {
2206                         free(old_dir);
2207                 }
2208                 item = string_list_lookup(&entry->possible_new_dirs, new_dir);
2209                 if (!item) {
2210                         item = string_list_insert(&entry->possible_new_dirs,
2211                                                   new_dir);
2212                         item->util = xcalloc(1, sizeof(int));
2213                 } else {
2214                         free(new_dir);
2215                 }
2216                 count = item->util;
2217                 *count += 1;
2218         }
2219
2220         /*
2221          * For each directory with files moved out of it, we find out which
2222          * target directory received the most files so we can declare it to
2223          * be the "winning" target location for the directory rename.  This
2224          * winner gets recorded in new_dir.  If there is no winner
2225          * (multiple target directories received the same number of files),
2226          * we set non_unique_new_dir.  Once we've determined the winner (or
2227          * that there is no winner), we no longer need possible_new_dirs.
2228          */
2229         hashmap_iter_init(dir_renames, &iter);
2230         while ((entry = hashmap_iter_next(&iter))) {
2231                 int max = 0;
2232                 int bad_max = 0;
2233                 char *best = NULL;
2234
2235                 for (i = 0; i < entry->possible_new_dirs.nr; i++) {
2236                         int *count = entry->possible_new_dirs.items[i].util;
2237
2238                         if (*count == max)
2239                                 bad_max = max;
2240                         else if (*count > max) {
2241                                 max = *count;
2242                                 best = entry->possible_new_dirs.items[i].string;
2243                         }
2244                 }
2245                 if (bad_max == max)
2246                         entry->non_unique_new_dir = 1;
2247                 else {
2248                         assert(entry->new_dir.len == 0);
2249                         strbuf_addstr(&entry->new_dir, best);
2250                 }
2251                 /*
2252                  * The relevant directory sub-portion of the original full
2253                  * filepaths were xstrndup'ed before inserting into
2254                  * possible_new_dirs, and instead of manually iterating the
2255                  * list and free'ing each, just lie and tell
2256                  * possible_new_dirs that it did the strdup'ing so that it
2257                  * will free them for us.
2258                  */
2259                 entry->possible_new_dirs.strdup_strings = 1;
2260                 string_list_clear(&entry->possible_new_dirs, 1);
2261         }
2262
2263         return dir_renames;
2264 }
2265
2266 static struct dir_rename_entry *check_dir_renamed(const char *path,
2267                                                   struct hashmap *dir_renames)
2268 {
2269         char *temp = xstrdup(path);
2270         char *end;
2271         struct dir_rename_entry *entry = NULL;
2272
2273         while ((end = strrchr(temp, '/'))) {
2274                 *end = '\0';
2275                 entry = dir_rename_find_entry(dir_renames, temp);
2276                 if (entry)
2277                         break;
2278         }
2279         free(temp);
2280         return entry;
2281 }
2282
2283 static void compute_collisions(struct hashmap *collisions,
2284                                struct hashmap *dir_renames,
2285                                struct diff_queue_struct *pairs)
2286 {
2287         int i;
2288
2289         /*
2290          * Multiple files can be mapped to the same path due to directory
2291          * renames done by the other side of history.  Since that other
2292          * side of history could have merged multiple directories into one,
2293          * if our side of history added the same file basename to each of
2294          * those directories, then all N of them would get implicitly
2295          * renamed by the directory rename detection into the same path,
2296          * and we'd get an add/add/.../add conflict, and all those adds
2297          * from *this* side of history.  This is not representable in the
2298          * index, and users aren't going to easily be able to make sense of
2299          * it.  So we need to provide a good warning about what's
2300          * happening, and fall back to no-directory-rename detection
2301          * behavior for those paths.
2302          *
2303          * See testcases 9e and all of section 5 from t6043 for examples.
2304          */
2305         collision_init(collisions);
2306
2307         for (i = 0; i < pairs->nr; ++i) {
2308                 struct dir_rename_entry *dir_rename_ent;
2309                 struct collision_entry *collision_ent;
2310                 char *new_path;
2311                 struct diff_filepair *pair = pairs->queue[i];
2312
2313                 if (pair->status != 'A' && pair->status != 'R')
2314                         continue;
2315                 dir_rename_ent = check_dir_renamed(pair->two->path,
2316                                                    dir_renames);
2317                 if (!dir_rename_ent)
2318                         continue;
2319
2320                 new_path = apply_dir_rename(dir_rename_ent, pair->two->path);
2321                 if (!new_path)
2322                         /*
2323                          * dir_rename_ent->non_unique_new_path is true, which
2324                          * means there is no directory rename for us to use,
2325                          * which means it won't cause us any additional
2326                          * collisions.
2327                          */
2328                         continue;
2329                 collision_ent = collision_find_entry(collisions, new_path);
2330                 if (!collision_ent) {
2331                         collision_ent = xcalloc(1,
2332                                                 sizeof(struct collision_entry));
2333                         hashmap_entry_init(collision_ent, strhash(new_path));
2334                         hashmap_put(collisions, collision_ent);
2335                         collision_ent->target_file = new_path;
2336                 } else {
2337                         free(new_path);
2338                 }
2339                 string_list_insert(&collision_ent->source_files,
2340                                    pair->two->path);
2341         }
2342 }
2343
2344 static char *check_for_directory_rename(struct merge_options *opt,
2345                                         const char *path,
2346                                         struct tree *tree,
2347                                         struct hashmap *dir_renames,
2348                                         struct hashmap *dir_rename_exclusions,
2349                                         struct hashmap *collisions,
2350                                         int *clean_merge)
2351 {
2352         char *new_path = NULL;
2353         struct dir_rename_entry *entry = check_dir_renamed(path, dir_renames);
2354         struct dir_rename_entry *oentry = NULL;
2355
2356         if (!entry)
2357                 return new_path;
2358
2359         /*
2360          * This next part is a little weird.  We do not want to do an
2361          * implicit rename into a directory we renamed on our side, because
2362          * that will result in a spurious rename/rename(1to2) conflict.  An
2363          * example:
2364          *   Base commit: dumbdir/afile, otherdir/bfile
2365          *   Side 1:      smrtdir/afile, otherdir/bfile
2366          *   Side 2:      dumbdir/afile, dumbdir/bfile
2367          * Here, while working on Side 1, we could notice that otherdir was
2368          * renamed/merged to dumbdir, and change the diff_filepair for
2369          * otherdir/bfile into a rename into dumbdir/bfile.  However, Side
2370          * 2 will notice the rename from dumbdir to smrtdir, and do the
2371          * transitive rename to move it from dumbdir/bfile to
2372          * smrtdir/bfile.  That gives us bfile in dumbdir vs being in
2373          * smrtdir, a rename/rename(1to2) conflict.  We really just want
2374          * the file to end up in smrtdir.  And the way to achieve that is
2375          * to not let Side1 do the rename to dumbdir, since we know that is
2376          * the source of one of our directory renames.
2377          *
2378          * That's why oentry and dir_rename_exclusions is here.
2379          *
2380          * As it turns out, this also prevents N-way transient rename
2381          * confusion; See testcases 9c and 9d of t6043.
2382          */
2383         oentry = dir_rename_find_entry(dir_rename_exclusions, entry->new_dir.buf);
2384         if (oentry) {
2385                 output(opt, 1, _("WARNING: Avoiding applying %s -> %s rename "
2386                                "to %s, because %s itself was renamed."),
2387                        entry->dir, entry->new_dir.buf, path, entry->new_dir.buf);
2388         } else {
2389                 new_path = handle_path_level_conflicts(opt, path, entry,
2390                                                        collisions, tree);
2391                 *clean_merge &= (new_path != NULL);
2392         }
2393
2394         return new_path;
2395 }
2396
2397 static void apply_directory_rename_modifications(struct merge_options *opt,
2398                                                  struct diff_filepair *pair,
2399                                                  char *new_path,
2400                                                  struct rename *re,
2401                                                  struct tree *tree,
2402                                                  struct tree *o_tree,
2403                                                  struct tree *a_tree,
2404                                                  struct tree *b_tree,
2405                                                  struct string_list *entries)
2406 {
2407         struct string_list_item *item;
2408         int stage = (tree == a_tree ? 2 : 3);
2409         int update_wd;
2410
2411         /*
2412          * In all cases where we can do directory rename detection,
2413          * unpack_trees() will have read pair->two->path into the
2414          * index and the working copy.  We need to remove it so that
2415          * we can instead place it at new_path.  It is guaranteed to
2416          * not be untracked (unpack_trees() would have errored out
2417          * saying the file would have been overwritten), but it might
2418          * be dirty, though.
2419          */
2420         update_wd = !was_dirty(opt, pair->two->path);
2421         if (!update_wd)
2422                 output(opt, 1, _("Refusing to lose dirty file at %s"),
2423                        pair->two->path);
2424         remove_file(opt, 1, pair->two->path, !update_wd);
2425
2426         /* Find or create a new re->dst_entry */
2427         item = string_list_lookup(entries, new_path);
2428         if (item) {
2429                 /*
2430                  * Since we're renaming on this side of history, and it's
2431                  * due to a directory rename on the other side of history
2432                  * (which we only allow when the directory in question no
2433                  * longer exists on the other side of history), the
2434                  * original entry for re->dst_entry is no longer
2435                  * necessary...
2436                  */
2437                 re->dst_entry->processed = 1;
2438
2439                 /*
2440                  * ...because we'll be using this new one.
2441                  */
2442                 re->dst_entry = item->util;
2443         } else {
2444                 /*
2445                  * re->dst_entry is for the before-dir-rename path, and we
2446                  * need it to hold information for the after-dir-rename
2447                  * path.  Before creating a new entry, we need to mark the
2448                  * old one as unnecessary (...unless it is shared by
2449                  * src_entry, i.e. this didn't use to be a rename, in which
2450                  * case we can just allow the normal processing to happen
2451                  * for it).
2452                  */
2453                 if (pair->status == 'R')
2454                         re->dst_entry->processed = 1;
2455
2456                 re->dst_entry = insert_stage_data(opt->repo, new_path,
2457                                                   o_tree, a_tree, b_tree,
2458                                                   entries);
2459                 item = string_list_insert(entries, new_path);
2460                 item->util = re->dst_entry;
2461         }
2462
2463         /*
2464          * Update the stage_data with the information about the path we are
2465          * moving into place.  That slot will be empty and available for us
2466          * to write to because of the collision checks in
2467          * handle_path_level_conflicts().  In other words,
2468          * re->dst_entry->stages[stage].oid will be the null_oid, so it's
2469          * open for us to write to.
2470          *
2471          * It may be tempting to actually update the index at this point as
2472          * well, using update_stages_for_stage_data(), but as per the big
2473          * "NOTE" in update_stages(), doing so will modify the current
2474          * in-memory index which will break calls to would_lose_untracked()
2475          * that we need to make.  Instead, we need to just make sure that
2476          * the various handle_rename_*() functions update the index
2477          * explicitly rather than relying on unpack_trees() to have done it.
2478          */
2479         get_tree_entry(opt->repo,
2480                        &tree->object.oid,
2481                        pair->two->path,
2482                        &re->dst_entry->stages[stage].oid,
2483                        &re->dst_entry->stages[stage].mode);
2484
2485         /*
2486          * Record the original change status (or 'type' of change).  If it
2487          * was originally an add ('A'), this lets us differentiate later
2488          * between a RENAME_DELETE conflict and RENAME_VIA_DIR (they
2489          * otherwise look the same).  If it was originally a rename ('R'),
2490          * this lets us remember and report accurately about the transitive
2491          * renaming that occurred via the directory rename detection.  Also,
2492          * record the original destination name.
2493          */
2494         re->dir_rename_original_type = pair->status;
2495         re->dir_rename_original_dest = pair->two->path;
2496
2497         /*
2498          * We don't actually look at pair->status again, but it seems
2499          * pedagogically correct to adjust it.
2500          */
2501         pair->status = 'R';
2502
2503         /*
2504          * Finally, record the new location.
2505          */
2506         pair->two->path = new_path;
2507 }
2508
2509 /*
2510  * Get information of all renames which occurred in 'pairs', making use of
2511  * any implicit directory renames inferred from the other side of history.
2512  * We need the three trees in the merge ('o_tree', 'a_tree' and 'b_tree')
2513  * to be able to associate the correct cache entries with the rename
2514  * information; tree is always equal to either a_tree or b_tree.
2515  */
2516 static struct string_list *get_renames(struct merge_options *opt,
2517                                        const char *branch,
2518                                        struct diff_queue_struct *pairs,
2519                                        struct hashmap *dir_renames,
2520                                        struct hashmap *dir_rename_exclusions,
2521                                        struct tree *tree,
2522                                        struct tree *o_tree,
2523                                        struct tree *a_tree,
2524                                        struct tree *b_tree,
2525                                        struct string_list *entries,
2526                                        int *clean_merge)
2527 {
2528         int i;
2529         struct hashmap collisions;
2530         struct hashmap_iter iter;
2531         struct collision_entry *e;
2532         struct string_list *renames;
2533
2534         compute_collisions(&collisions, dir_renames, pairs);
2535         renames = xcalloc(1, sizeof(struct string_list));
2536
2537         for (i = 0; i < pairs->nr; ++i) {
2538                 struct string_list_item *item;
2539                 struct rename *re;
2540                 struct diff_filepair *pair = pairs->queue[i];
2541                 char *new_path; /* non-NULL only with directory renames */
2542
2543                 if (pair->status != 'A' && pair->status != 'R') {
2544                         diff_free_filepair(pair);
2545                         continue;
2546                 }
2547                 new_path = check_for_directory_rename(opt, pair->two->path, tree,
2548                                                       dir_renames,
2549                                                       dir_rename_exclusions,
2550                                                       &collisions,
2551                                                       clean_merge);
2552                 if (pair->status != 'R' && !new_path) {
2553                         diff_free_filepair(pair);
2554                         continue;
2555                 }
2556
2557                 re = xmalloc(sizeof(*re));
2558                 re->processed = 0;
2559                 re->pair = pair;
2560                 re->branch = branch;
2561                 re->dir_rename_original_type = '\0';
2562                 re->dir_rename_original_dest = NULL;
2563                 item = string_list_lookup(entries, re->pair->one->path);
2564                 if (!item)
2565                         re->src_entry = insert_stage_data(opt->repo,
2566                                         re->pair->one->path,
2567                                         o_tree, a_tree, b_tree, entries);
2568                 else
2569                         re->src_entry = item->util;
2570
2571                 item = string_list_lookup(entries, re->pair->two->path);
2572                 if (!item)
2573                         re->dst_entry = insert_stage_data(opt->repo,
2574                                         re->pair->two->path,
2575                                         o_tree, a_tree, b_tree, entries);
2576                 else
2577                         re->dst_entry = item->util;
2578                 item = string_list_insert(renames, pair->one->path);
2579                 item->util = re;
2580                 if (new_path)
2581                         apply_directory_rename_modifications(opt, pair, new_path,
2582                                                              re, tree, o_tree,
2583                                                              a_tree, b_tree,
2584                                                              entries);
2585         }
2586
2587         hashmap_iter_init(&collisions, &iter);
2588         while ((e = hashmap_iter_next(&iter))) {
2589                 free(e->target_file);
2590                 string_list_clear(&e->source_files, 0);
2591         }
2592         hashmap_free(&collisions, 1);
2593         return renames;
2594 }
2595
2596 static int process_renames(struct merge_options *opt,
2597                            struct string_list *a_renames,
2598                            struct string_list *b_renames)
2599 {
2600         int clean_merge = 1, i, j;
2601         struct string_list a_by_dst = STRING_LIST_INIT_NODUP;
2602         struct string_list b_by_dst = STRING_LIST_INIT_NODUP;
2603         const struct rename *sre;
2604
2605         for (i = 0; i < a_renames->nr; i++) {
2606                 sre = a_renames->items[i].util;
2607                 string_list_insert(&a_by_dst, sre->pair->two->path)->util
2608                         = (void *)sre;
2609         }
2610         for (i = 0; i < b_renames->nr; i++) {
2611                 sre = b_renames->items[i].util;
2612                 string_list_insert(&b_by_dst, sre->pair->two->path)->util
2613                         = (void *)sre;
2614         }
2615
2616         for (i = 0, j = 0; i < a_renames->nr || j < b_renames->nr;) {
2617                 struct string_list *renames1, *renames2Dst;
2618                 struct rename *ren1 = NULL, *ren2 = NULL;
2619                 const char *ren1_src, *ren1_dst;
2620                 struct string_list_item *lookup;
2621
2622                 if (i >= a_renames->nr) {
2623                         ren2 = b_renames->items[j++].util;
2624                 } else if (j >= b_renames->nr) {
2625                         ren1 = a_renames->items[i++].util;
2626                 } else {
2627                         int compare = strcmp(a_renames->items[i].string,
2628                                              b_renames->items[j].string);
2629                         if (compare <= 0)
2630                                 ren1 = a_renames->items[i++].util;
2631                         if (compare >= 0)
2632                                 ren2 = b_renames->items[j++].util;
2633                 }
2634
2635                 /* TODO: refactor, so that 1/2 are not needed */
2636                 if (ren1) {
2637                         renames1 = a_renames;
2638                         renames2Dst = &b_by_dst;
2639                 } else {
2640                         renames1 = b_renames;
2641                         renames2Dst = &a_by_dst;
2642                         SWAP(ren2, ren1);
2643                 }
2644
2645                 if (ren1->processed)
2646                         continue;
2647                 ren1->processed = 1;
2648                 ren1->dst_entry->processed = 1;
2649                 /* BUG: We should only mark src_entry as processed if we
2650                  * are not dealing with a rename + add-source case.
2651                  */
2652                 ren1->src_entry->processed = 1;
2653
2654                 ren1_src = ren1->pair->one->path;
2655                 ren1_dst = ren1->pair->two->path;
2656
2657                 if (ren2) {
2658                         /* One file renamed on both sides */
2659                         const char *ren2_src = ren2->pair->one->path;
2660                         const char *ren2_dst = ren2->pair->two->path;
2661                         enum rename_type rename_type;
2662                         if (strcmp(ren1_src, ren2_src) != 0)
2663                                 BUG("ren1_src != ren2_src");
2664                         ren2->dst_entry->processed = 1;
2665                         ren2->processed = 1;
2666                         if (strcmp(ren1_dst, ren2_dst) != 0) {
2667                                 rename_type = RENAME_ONE_FILE_TO_TWO;
2668                                 clean_merge = 0;
2669                         } else {
2670                                 rename_type = RENAME_ONE_FILE_TO_ONE;
2671                                 /* BUG: We should only remove ren1_src in
2672                                  * the base stage (think of rename +
2673                                  * add-source cases).
2674                                  */
2675                                 remove_file(opt, 1, ren1_src, 1);
2676                                 update_entry(ren1->dst_entry,
2677                                              ren1->pair->one,
2678                                              ren1->pair->two,
2679                                              ren2->pair->two);
2680                         }
2681                         setup_rename_conflict_info(rename_type, opt, ren1, ren2);
2682                 } else if ((lookup = string_list_lookup(renames2Dst, ren1_dst))) {
2683                         /* Two different files renamed to the same thing */
2684                         char *ren2_dst;
2685                         ren2 = lookup->util;
2686                         ren2_dst = ren2->pair->two->path;
2687                         if (strcmp(ren1_dst, ren2_dst) != 0)
2688                                 BUG("ren1_dst != ren2_dst");
2689
2690                         clean_merge = 0;
2691                         ren2->processed = 1;
2692                         /*
2693                          * BUG: We should only mark src_entry as processed
2694                          * if we are not dealing with a rename + add-source
2695                          * case.
2696                          */
2697                         ren2->src_entry->processed = 1;
2698
2699                         setup_rename_conflict_info(RENAME_TWO_FILES_TO_ONE,
2700                                                    opt, ren1, ren2);
2701                 } else {
2702                         /* Renamed in 1, maybe changed in 2 */
2703                         /* we only use sha1 and mode of these */
2704                         struct diff_filespec src_other, dst_other;
2705                         int try_merge;
2706
2707                         /*
2708                          * unpack_trees loads entries from common-commit
2709                          * into stage 1, from head-commit into stage 2, and
2710                          * from merge-commit into stage 3.  We keep track
2711                          * of which side corresponds to the rename.
2712                          */
2713                         int renamed_stage = a_renames == renames1 ? 2 : 3;
2714                         int other_stage =   a_renames == renames1 ? 3 : 2;
2715
2716                         /* BUG: We should only remove ren1_src in the base
2717                          * stage and in other_stage (think of rename +
2718                          * add-source case).
2719                          */
2720                         remove_file(opt, 1, ren1_src,
2721                                     renamed_stage == 2 || !was_tracked(opt, ren1_src));
2722
2723                         oidcpy(&src_other.oid,
2724                                &ren1->src_entry->stages[other_stage].oid);
2725                         src_other.mode = ren1->src_entry->stages[other_stage].mode;
2726                         oidcpy(&dst_other.oid,
2727                                &ren1->dst_entry->stages[other_stage].oid);
2728                         dst_other.mode = ren1->dst_entry->stages[other_stage].mode;
2729                         try_merge = 0;
2730
2731                         if (oid_eq(&src_other.oid, &null_oid) &&
2732                             ren1->dir_rename_original_type == 'A') {
2733                                 setup_rename_conflict_info(RENAME_VIA_DIR,
2734                                                            opt, ren1, NULL);
2735                         } else if (oid_eq(&src_other.oid, &null_oid)) {
2736                                 setup_rename_conflict_info(RENAME_DELETE,
2737                                                            opt, ren1, NULL);
2738                         } else if ((dst_other.mode == ren1->pair->two->mode) &&
2739                                    oid_eq(&dst_other.oid, &ren1->pair->two->oid)) {
2740                                 /*
2741                                  * Added file on the other side identical to
2742                                  * the file being renamed: clean merge.
2743                                  * Also, there is no need to overwrite the
2744                                  * file already in the working copy, so call
2745                                  * update_file_flags() instead of
2746                                  * update_file().
2747                                  */
2748                                 if (update_file_flags(opt,
2749                                                       ren1->pair->two,
2750                                                       ren1_dst,
2751                                                       1, /* update_cache */
2752                                                       0  /* update_wd    */))
2753                                         clean_merge = -1;
2754                         } else if (!oid_eq(&dst_other.oid, &null_oid)) {
2755                                 /*
2756                                  * Probably not a clean merge, but it's
2757                                  * premature to set clean_merge to 0 here,
2758                                  * because if the rename merges cleanly and
2759                                  * the merge exactly matches the newly added
2760                                  * file, then the merge will be clean.
2761                                  */
2762                                 setup_rename_conflict_info(RENAME_ADD,
2763                                                            opt, ren1, NULL);
2764                         } else
2765                                 try_merge = 1;
2766
2767                         if (clean_merge < 0)
2768                                 goto cleanup_and_return;
2769                         if (try_merge) {
2770                                 struct diff_filespec *o, *a, *b;
2771                                 src_other.path = (char *)ren1_src;
2772
2773                                 o = ren1->pair->one;
2774                                 if (a_renames == renames1) {
2775                                         a = ren1->pair->two;
2776                                         b = &src_other;
2777                                 } else {
2778                                         b = ren1->pair->two;
2779                                         a = &src_other;
2780                                 }
2781                                 update_entry(ren1->dst_entry, o, a, b);
2782                                 setup_rename_conflict_info(RENAME_NORMAL,
2783                                                            opt, ren1, NULL);
2784                         }
2785                 }
2786         }
2787 cleanup_and_return:
2788         string_list_clear(&a_by_dst, 0);
2789         string_list_clear(&b_by_dst, 0);
2790
2791         return clean_merge;
2792 }
2793
2794 struct rename_info {
2795         struct string_list *head_renames;
2796         struct string_list *merge_renames;
2797 };
2798
2799 static void initial_cleanup_rename(struct diff_queue_struct *pairs,
2800                                    struct hashmap *dir_renames)
2801 {
2802         struct hashmap_iter iter;
2803         struct dir_rename_entry *e;
2804
2805         hashmap_iter_init(dir_renames, &iter);
2806         while ((e = hashmap_iter_next(&iter))) {
2807                 free(e->dir);
2808                 strbuf_release(&e->new_dir);
2809                 /* possible_new_dirs already cleared in get_directory_renames */
2810         }
2811         hashmap_free(dir_renames, 1);
2812         free(dir_renames);
2813
2814         free(pairs->queue);
2815         free(pairs);
2816 }
2817
2818 static int detect_and_process_renames(struct merge_options *opt,
2819                                       struct tree *common,
2820                                       struct tree *head,
2821                                       struct tree *merge,
2822                                       struct string_list *entries,
2823                                       struct rename_info *ri)
2824 {
2825         struct diff_queue_struct *head_pairs, *merge_pairs;
2826         struct hashmap *dir_re_head, *dir_re_merge;
2827         int clean = 1;
2828
2829         ri->head_renames = NULL;
2830         ri->merge_renames = NULL;
2831
2832         if (!merge_detect_rename(opt))
2833                 return 1;
2834
2835         head_pairs = get_diffpairs(opt, common, head);
2836         merge_pairs = get_diffpairs(opt, common, merge);
2837
2838         if ((opt->detect_directory_renames == MERGE_DIRECTORY_RENAMES_TRUE) ||
2839             (opt->detect_directory_renames == MERGE_DIRECTORY_RENAMES_CONFLICT &&
2840              !opt->call_depth)) {
2841                 dir_re_head = get_directory_renames(head_pairs);
2842                 dir_re_merge = get_directory_renames(merge_pairs);
2843
2844                 handle_directory_level_conflicts(opt,
2845                                                  dir_re_head, head,
2846                                                  dir_re_merge, merge);
2847         } else {
2848                 dir_re_head  = xmalloc(sizeof(*dir_re_head));
2849                 dir_re_merge = xmalloc(sizeof(*dir_re_merge));
2850                 dir_rename_init(dir_re_head);
2851                 dir_rename_init(dir_re_merge);
2852         }
2853
2854         ri->head_renames  = get_renames(opt, opt->branch1, head_pairs,
2855                                         dir_re_merge, dir_re_head, head,
2856                                         common, head, merge, entries,
2857                                         &clean);
2858         if (clean < 0)
2859                 goto cleanup;
2860         ri->merge_renames = get_renames(opt, opt->branch2, merge_pairs,
2861                                         dir_re_head, dir_re_merge, merge,
2862                                         common, head, merge, entries,
2863                                         &clean);
2864         if (clean < 0)
2865                 goto cleanup;
2866         clean &= process_renames(opt, ri->head_renames, ri->merge_renames);
2867
2868 cleanup:
2869         /*
2870          * Some cleanup is deferred until cleanup_renames() because the
2871          * data structures are still needed and referenced in
2872          * process_entry().  But there are a few things we can free now.
2873          */
2874         initial_cleanup_rename(head_pairs, dir_re_head);
2875         initial_cleanup_rename(merge_pairs, dir_re_merge);
2876
2877         return clean;
2878 }
2879
2880 static void final_cleanup_rename(struct string_list *rename)
2881 {
2882         const struct rename *re;
2883         int i;
2884
2885         if (rename == NULL)
2886                 return;
2887
2888         for (i = 0; i < rename->nr; i++) {
2889                 re = rename->items[i].util;
2890                 diff_free_filepair(re->pair);
2891         }
2892         string_list_clear(rename, 1);
2893         free(rename);
2894 }
2895
2896 static void final_cleanup_renames(struct rename_info *re_info)
2897 {
2898         final_cleanup_rename(re_info->head_renames);
2899         final_cleanup_rename(re_info->merge_renames);
2900 }
2901
2902 static int read_oid_strbuf(struct merge_options *opt,
2903                            const struct object_id *oid,
2904                            struct strbuf *dst)
2905 {
2906         void *buf;
2907         enum object_type type;
2908         unsigned long size;
2909         buf = read_object_file(oid, &type, &size);
2910         if (!buf)
2911                 return err(opt, _("cannot read object %s"), oid_to_hex(oid));
2912         if (type != OBJ_BLOB) {
2913                 free(buf);
2914                 return err(opt, _("object %s is not a blob"), oid_to_hex(oid));
2915         }
2916         strbuf_attach(dst, buf, size, size + 1);
2917         return 0;
2918 }
2919
2920 static int blob_unchanged(struct merge_options *opt,
2921                           const struct diff_filespec *o,
2922                           const struct diff_filespec *a,
2923                           int renormalize, const char *path)
2924 {
2925         struct strbuf obuf = STRBUF_INIT;
2926         struct strbuf abuf = STRBUF_INIT;
2927         int ret = 0; /* assume changed for safety */
2928         const struct index_state *idx = opt->repo->index;
2929
2930         if (a->mode != o->mode)
2931                 return 0;
2932         if (oid_eq(&o->oid, &a->oid))
2933                 return 1;
2934         if (!renormalize)
2935                 return 0;
2936
2937         if (read_oid_strbuf(opt, &o->oid, &obuf) ||
2938             read_oid_strbuf(opt, &a->oid, &abuf))
2939                 goto error_return;
2940         /*
2941          * Note: binary | is used so that both renormalizations are
2942          * performed.  Comparison can be skipped if both files are
2943          * unchanged since their sha1s have already been compared.
2944          */
2945         if (renormalize_buffer(idx, path, obuf.buf, obuf.len, &obuf) |
2946             renormalize_buffer(idx, path, abuf.buf, abuf.len, &abuf))
2947                 ret = (obuf.len == abuf.len && !memcmp(obuf.buf, abuf.buf, obuf.len));
2948
2949 error_return:
2950         strbuf_release(&obuf);
2951         strbuf_release(&abuf);
2952         return ret;
2953 }
2954
2955 static int handle_modify_delete(struct merge_options *opt,
2956                                 const char *path,
2957                                 const struct diff_filespec *o,
2958                                 const struct diff_filespec *a,
2959                                 const struct diff_filespec *b)
2960 {
2961         const char *modify_branch, *delete_branch;
2962         const struct diff_filespec *changed;
2963
2964         if (is_valid(a)) {
2965                 modify_branch = opt->branch1;
2966                 delete_branch = opt->branch2;
2967                 changed = a;
2968         } else {
2969                 modify_branch = opt->branch2;
2970                 delete_branch = opt->branch1;
2971                 changed = b;
2972         }
2973
2974         return handle_change_delete(opt,
2975                                     path, NULL,
2976                                     o, changed,
2977                                     modify_branch, delete_branch,
2978                                     _("modify"), _("modified"));
2979 }
2980
2981 static int handle_content_merge(struct merge_file_info *mfi,
2982                                 struct merge_options *opt,
2983                                 const char *path,
2984                                 int is_dirty,
2985                                 const struct diff_filespec *o,
2986                                 const struct diff_filespec *a,
2987                                 const struct diff_filespec *b,
2988                                 struct rename_conflict_info *ci)
2989 {
2990         const char *reason = _("content");
2991         unsigned df_conflict_remains = 0;
2992
2993         if (!is_valid(o))
2994                 reason = _("add/add");
2995
2996         assert(o->path && a->path && b->path);
2997         if (ci && dir_in_way(opt->repo->index, path, !opt->call_depth,
2998                              S_ISGITLINK(ci->ren1->pair->two->mode)))
2999                 df_conflict_remains = 1;
3000
3001         if (merge_mode_and_contents(opt, o, a, b, path,
3002                                     opt->branch1, opt->branch2,
3003                                     opt->call_depth * 2, mfi))
3004                 return -1;
3005
3006         /*
3007          * We can skip updating the working tree file iff:
3008          *   a) The merge is clean
3009          *   b) The merge matches what was in HEAD (content, mode, pathname)
3010          *   c) The target path is usable (i.e. not involved in D/F conflict)
3011          */
3012         if (mfi->clean && was_tracked_and_matches(opt, path, &mfi->blob) &&
3013             !df_conflict_remains) {
3014                 int pos;
3015                 struct cache_entry *ce;
3016
3017                 output(opt, 3, _("Skipped %s (merged same as existing)"), path);
3018                 if (add_cacheinfo(opt, &mfi->blob, path,
3019                                   0, (!opt->call_depth && !is_dirty), 0))
3020                         return -1;
3021                 /*
3022                  * However, add_cacheinfo() will delete the old cache entry
3023                  * and add a new one.  We need to copy over any skip_worktree
3024                  * flag to avoid making the file appear as if it were
3025                  * deleted by the user.
3026                  */
3027                 pos = index_name_pos(&opt->orig_index, path, strlen(path));
3028                 ce = opt->orig_index.cache[pos];
3029                 if (ce_skip_worktree(ce)) {
3030                         pos = index_name_pos(opt->repo->index, path, strlen(path));
3031                         ce = opt->repo->index->cache[pos];
3032                         ce->ce_flags |= CE_SKIP_WORKTREE;
3033                 }
3034                 return mfi->clean;
3035         }
3036
3037         if (!mfi->clean) {
3038                 if (S_ISGITLINK(mfi->blob.mode))
3039                         reason = _("submodule");
3040                 output(opt, 1, _("CONFLICT (%s): Merge conflict in %s"),
3041                                 reason, path);
3042                 if (ci && !df_conflict_remains)
3043                         if (update_stages(opt, path, o, a, b))
3044                                 return -1;
3045         }
3046
3047         if (df_conflict_remains || is_dirty) {
3048                 char *new_path;
3049                 if (opt->call_depth) {
3050                         remove_file_from_index(opt->repo->index, path);
3051                 } else {
3052                         if (!mfi->clean) {
3053                                 if (update_stages(opt, path, o, a, b))
3054                                         return -1;
3055                         } else {
3056                                 int file_from_stage2 = was_tracked(opt, path);
3057
3058                                 if (update_stages(opt, path, NULL,
3059                                                   file_from_stage2 ? &mfi->blob : NULL,
3060                                                   file_from_stage2 ? NULL : &mfi->blob))
3061                                         return -1;
3062                         }
3063
3064                 }
3065                 new_path = unique_path(opt, path, ci->ren1->branch);
3066                 if (is_dirty) {
3067                         output(opt, 1, _("Refusing to lose dirty file at %s"),
3068                                path);
3069                 }
3070                 output(opt, 1, _("Adding as %s instead"), new_path);
3071                 if (update_file(opt, 0, &mfi->blob, new_path)) {
3072                         free(new_path);
3073                         return -1;
3074                 }
3075                 free(new_path);
3076                 mfi->clean = 0;
3077         } else if (update_file(opt, mfi->clean, &mfi->blob, path))
3078                 return -1;
3079         return !is_dirty && mfi->clean;
3080 }
3081
3082 static int handle_rename_normal(struct merge_options *opt,
3083                                 const char *path,
3084                                 const struct diff_filespec *o,
3085                                 const struct diff_filespec *a,
3086                                 const struct diff_filespec *b,
3087                                 struct rename_conflict_info *ci)
3088 {
3089         struct rename *ren = ci->ren1;
3090         struct merge_file_info mfi;
3091         int clean;
3092         int side = (ren->branch == opt->branch1 ? 2 : 3);
3093
3094         /* Merge the content and write it out */
3095         clean = handle_content_merge(&mfi, opt, path, was_dirty(opt, path),
3096                                      o, a, b, ci);
3097
3098         if (clean &&
3099             opt->detect_directory_renames == MERGE_DIRECTORY_RENAMES_CONFLICT &&
3100             ren->dir_rename_original_dest) {
3101                 if (update_stages(opt, path,
3102                                   NULL,
3103                                   side == 2 ? &mfi.blob : NULL,
3104                                   side == 2 ? NULL : &mfi.blob))
3105                         return -1;
3106                 clean = 0; /* not clean, but conflicted */
3107         }
3108         return clean;
3109 }
3110
3111 static void dir_rename_warning(const char *msg,
3112                                int is_add,
3113                                int clean,
3114                                struct merge_options *opt,
3115                                struct rename *ren)
3116 {
3117         const char *other_branch;
3118         other_branch = (ren->branch == opt->branch1 ?
3119                         opt->branch2 : opt->branch1);
3120         if (is_add) {
3121                 output(opt, clean ? 2 : 1, msg,
3122                        ren->pair->one->path, ren->branch,
3123                        other_branch, ren->pair->two->path);
3124                 return;
3125         }
3126         output(opt, clean ? 2 : 1, msg,
3127                ren->pair->one->path, ren->dir_rename_original_dest, ren->branch,
3128                other_branch, ren->pair->two->path);
3129 }
3130 static int warn_about_dir_renamed_entries(struct merge_options *opt,
3131                                           struct rename *ren)
3132 {
3133         const char *msg;
3134         int clean = 1, is_add;
3135
3136         if (!ren)
3137                 return clean;
3138
3139         /* Return early if ren was not affected/created by a directory rename */
3140         if (!ren->dir_rename_original_dest)
3141                 return clean;
3142
3143         /* Sanity checks */
3144         assert(opt->detect_directory_renames > MERGE_DIRECTORY_RENAMES_NONE);
3145         assert(ren->dir_rename_original_type == 'A' ||
3146                ren->dir_rename_original_type == 'R');
3147
3148         /* Check whether to treat directory renames as a conflict */
3149         clean = (opt->detect_directory_renames == MERGE_DIRECTORY_RENAMES_TRUE);
3150
3151         is_add = (ren->dir_rename_original_type == 'A');
3152         if (ren->dir_rename_original_type == 'A' && clean) {
3153                 msg = _("Path updated: %s added in %s inside a "
3154                         "directory that was renamed in %s; moving it to %s.");
3155         } else if (ren->dir_rename_original_type == 'A' && !clean) {
3156                 msg = _("CONFLICT (file location): %s added in %s "
3157                         "inside a directory that was renamed in %s, "
3158                         "suggesting it should perhaps be moved to %s.");
3159         } else if (ren->dir_rename_original_type == 'R' && clean) {
3160                 msg = _("Path updated: %s renamed to %s in %s, inside a "
3161                         "directory that was renamed in %s; moving it to %s.");
3162         } else if (ren->dir_rename_original_type == 'R' && !clean) {
3163                 msg = _("CONFLICT (file location): %s renamed to %s in %s, "
3164                         "inside a directory that was renamed in %s, "
3165                         "suggesting it should perhaps be moved to %s.");
3166         } else {
3167                 BUG("Impossible dir_rename_original_type/clean combination");
3168         }
3169         dir_rename_warning(msg, is_add, clean, opt, ren);
3170
3171         return clean;
3172 }
3173
3174 /* Per entry merge function */
3175 static int process_entry(struct merge_options *opt,
3176                          const char *path, struct stage_data *entry)
3177 {
3178         int clean_merge = 1;
3179         int normalize = opt->renormalize;
3180
3181         struct diff_filespec *o = &entry->stages[1];
3182         struct diff_filespec *a = &entry->stages[2];
3183         struct diff_filespec *b = &entry->stages[3];
3184         int o_valid = is_valid(o);
3185         int a_valid = is_valid(a);
3186         int b_valid = is_valid(b);
3187         o->path = a->path = b->path = (char*)path;
3188
3189         entry->processed = 1;
3190         if (entry->rename_conflict_info) {
3191                 struct rename_conflict_info *ci = entry->rename_conflict_info;
3192                 struct diff_filespec *temp;
3193                 int path_clean;
3194
3195                 path_clean = warn_about_dir_renamed_entries(opt, ci->ren1);
3196                 path_clean &= warn_about_dir_renamed_entries(opt, ci->ren2);
3197
3198                 /*
3199                  * For cases with a single rename, {o,a,b}->path have all been
3200                  * set to the rename target path; we need to set two of these
3201                  * back to the rename source.
3202                  * For rename/rename conflicts, we'll manually fix paths below.
3203                  */
3204                 temp = (opt->branch1 == ci->ren1->branch) ? b : a;
3205                 o->path = temp->path = ci->ren1->pair->one->path;
3206                 if (ci->ren2) {
3207                         assert(opt->branch1 == ci->ren1->branch);
3208                 }
3209
3210                 switch (ci->rename_type) {
3211                 case RENAME_NORMAL:
3212                 case RENAME_ONE_FILE_TO_ONE:
3213                         clean_merge = handle_rename_normal(opt, path, o, a, b,
3214                                                            ci);
3215                         break;
3216                 case RENAME_VIA_DIR:
3217                         clean_merge = handle_rename_via_dir(opt, ci);
3218                         break;
3219                 case RENAME_ADD:
3220                         /*
3221                          * Probably unclean merge, but if the renamed file
3222                          * merges cleanly and the result can then be
3223                          * two-way merged cleanly with the added file, I
3224                          * guess it's a clean merge?
3225                          */
3226                         clean_merge = handle_rename_add(opt, ci);
3227                         break;
3228                 case RENAME_DELETE:
3229                         clean_merge = 0;
3230                         if (handle_rename_delete(opt, ci))
3231                                 clean_merge = -1;
3232                         break;
3233                 case RENAME_ONE_FILE_TO_TWO:
3234                         /*
3235                          * Manually fix up paths; note:
3236                          * ren[12]->pair->one->path are equal.
3237                          */
3238                         o->path = ci->ren1->pair->one->path;
3239                         a->path = ci->ren1->pair->two->path;
3240                         b->path = ci->ren2->pair->two->path;
3241
3242                         clean_merge = 0;
3243                         if (handle_rename_rename_1to2(opt, ci))
3244                                 clean_merge = -1;
3245                         break;
3246                 case RENAME_TWO_FILES_TO_ONE:
3247                         /*
3248                          * Manually fix up paths; note,
3249                          * ren[12]->pair->two->path are actually equal.
3250                          */
3251                         o->path = NULL;
3252                         a->path = ci->ren1->pair->two->path;
3253                         b->path = ci->ren2->pair->two->path;
3254
3255                         /*
3256                          * Probably unclean merge, but if the two renamed
3257                          * files merge cleanly and the two resulting files
3258                          * can then be two-way merged cleanly, I guess it's
3259                          * a clean merge?
3260                          */
3261                         clean_merge = handle_rename_rename_2to1(opt, ci);
3262                         break;
3263                 default:
3264                         entry->processed = 0;
3265                         break;
3266                 }
3267                 if (path_clean < clean_merge)
3268                         clean_merge = path_clean;
3269         } else if (o_valid && (!a_valid || !b_valid)) {
3270                 /* Case A: Deleted in one */
3271                 if ((!a_valid && !b_valid) ||
3272                     (!b_valid && blob_unchanged(opt, o, a, normalize, path)) ||
3273                     (!a_valid && blob_unchanged(opt, o, b, normalize, path))) {
3274                         /* Deleted in both or deleted in one and
3275                          * unchanged in the other */
3276                         if (a_valid)
3277                                 output(opt, 2, _("Removing %s"), path);
3278                         /* do not touch working file if it did not exist */
3279                         remove_file(opt, 1, path, !a_valid);
3280                 } else {
3281                         /* Modify/delete; deleted side may have put a directory in the way */
3282                         clean_merge = 0;
3283                         if (handle_modify_delete(opt, path, o, a, b))
3284                                 clean_merge = -1;
3285                 }
3286         } else if ((!o_valid && a_valid && !b_valid) ||
3287                    (!o_valid && !a_valid && b_valid)) {
3288                 /* Case B: Added in one. */
3289                 /* [nothing|directory] -> ([nothing|directory], file) */
3290
3291                 const char *add_branch;
3292                 const char *other_branch;
3293                 const char *conf;
3294                 const struct diff_filespec *contents;
3295
3296                 if (a_valid) {
3297                         add_branch = opt->branch1;
3298                         other_branch = opt->branch2;
3299                         contents = a;
3300                         conf = _("file/directory");
3301                 } else {
3302                         add_branch = opt->branch2;
3303                         other_branch = opt->branch1;
3304                         contents = b;
3305                         conf = _("directory/file");
3306                 }
3307                 if (dir_in_way(opt->repo->index, path,
3308                                !opt->call_depth && !S_ISGITLINK(a->mode),
3309                                0)) {
3310                         char *new_path = unique_path(opt, path, add_branch);
3311                         clean_merge = 0;
3312                         output(opt, 1, _("CONFLICT (%s): There is a directory with name %s in %s. "
3313                                "Adding %s as %s"),
3314                                conf, path, other_branch, path, new_path);
3315                         if (update_file(opt, 0, contents, new_path))
3316                                 clean_merge = -1;
3317                         else if (opt->call_depth)
3318                                 remove_file_from_index(opt->repo->index, path);
3319                         free(new_path);
3320                 } else {
3321                         output(opt, 2, _("Adding %s"), path);
3322                         /* do not overwrite file if already present */
3323                         if (update_file_flags(opt, contents, path, 1, !a_valid))
3324                                 clean_merge = -1;
3325                 }
3326         } else if (a_valid && b_valid) {
3327                 if (!o_valid) {
3328                         /* Case C: Added in both (check for same permissions) */
3329                         output(opt, 1,
3330                                _("CONFLICT (add/add): Merge conflict in %s"),
3331                                path);
3332                         clean_merge = handle_file_collision(opt,
3333                                                             path, NULL, NULL,
3334                                                             opt->branch1,
3335                                                             opt->branch2,
3336                                                             a, b);
3337                 } else {
3338                         /* case D: Modified in both, but differently. */
3339                         struct merge_file_info mfi;
3340                         int is_dirty = 0; /* unpack_trees would have bailed if dirty */
3341                         clean_merge = handle_content_merge(&mfi, opt, path,
3342                                                            is_dirty,
3343                                                            o, a, b, NULL);
3344                 }
3345         } else if (!o_valid && !a_valid && !b_valid) {
3346                 /*
3347                  * this entry was deleted altogether. a_mode == 0 means
3348                  * we had that path and want to actively remove it.
3349                  */
3350                 remove_file(opt, 1, path, !a->mode);
3351         } else
3352                 BUG("fatal merge failure, shouldn't happen.");
3353
3354         return clean_merge;
3355 }
3356
3357 static int merge_trees_internal(struct merge_options *opt,
3358                                 struct tree *head,
3359                                 struct tree *merge,
3360                                 struct tree *merge_base,
3361                                 struct tree **result)
3362 {
3363         struct index_state *istate = opt->repo->index;
3364         int code, clean;
3365
3366         if (opt->subtree_shift) {
3367                 merge = shift_tree_object(opt->repo, head, merge,
3368                                           opt->subtree_shift);
3369                 merge_base = shift_tree_object(opt->repo, head, merge_base,
3370                                                opt->subtree_shift);
3371         }
3372
3373         if (oid_eq(&merge_base->object.oid, &merge->object.oid)) {
3374                 output(opt, 0, _("Already up to date!"));
3375                 *result = head;
3376                 return 1;
3377         }
3378
3379         code = unpack_trees_start(opt, merge_base, head, merge);
3380
3381         if (code != 0) {
3382                 if (show(opt, 4) || opt->call_depth)
3383                         err(opt, _("merging of trees %s and %s failed"),
3384                             oid_to_hex(&head->object.oid),
3385                             oid_to_hex(&merge->object.oid));
3386                 unpack_trees_finish(opt);
3387                 return -1;
3388         }
3389
3390         if (unmerged_index(istate)) {
3391                 struct string_list *entries;
3392                 struct rename_info re_info;
3393                 int i;
3394                 /*
3395                  * Only need the hashmap while processing entries, so
3396                  * initialize it here and free it when we are done running
3397                  * through the entries. Keeping it in the merge_options as
3398                  * opposed to decaring a local hashmap is for convenience
3399                  * so that we don't have to pass it to around.
3400                  */
3401                 hashmap_init(&opt->current_file_dir_set, path_hashmap_cmp,
3402                              NULL, 512);
3403                 get_files_dirs(opt, head);
3404                 get_files_dirs(opt, merge);
3405
3406                 entries = get_unmerged(opt->repo->index);
3407                 clean = detect_and_process_renames(opt, merge_base, head, merge,
3408                                                    entries, &re_info);
3409                 record_df_conflict_files(opt, entries);
3410                 if (clean < 0)
3411                         goto cleanup;
3412                 for (i = entries->nr-1; 0 <= i; i--) {
3413                         const char *path = entries->items[i].string;
3414                         struct stage_data *e = entries->items[i].util;
3415                         if (!e->processed) {
3416                                 int ret = process_entry(opt, path, e);
3417                                 if (!ret)
3418                                         clean = 0;
3419                                 else if (ret < 0) {
3420                                         clean = ret;
3421                                         goto cleanup;
3422                                 }
3423                         }
3424                 }
3425                 for (i = 0; i < entries->nr; i++) {
3426                         struct stage_data *e = entries->items[i].util;
3427                         if (!e->processed)
3428                                 BUG("unprocessed path??? %s",
3429                                     entries->items[i].string);
3430                 }
3431
3432         cleanup:
3433                 final_cleanup_renames(&re_info);
3434
3435                 string_list_clear(entries, 1);
3436                 free(entries);
3437
3438                 hashmap_free(&opt->current_file_dir_set, 1);
3439
3440                 if (clean < 0) {
3441                         unpack_trees_finish(opt);
3442                         return clean;
3443                 }
3444         }
3445         else
3446                 clean = 1;
3447
3448         unpack_trees_finish(opt);
3449
3450         if (opt->call_depth &&
3451             !(*result = write_in_core_index_as_tree(opt->repo)))
3452                 return -1;
3453
3454         return clean;
3455 }
3456
3457 static struct commit_list *reverse_commit_list(struct commit_list *list)
3458 {
3459         struct commit_list *next = NULL, *current, *backup;
3460         for (current = list; current; current = backup) {
3461                 backup = current->next;
3462                 current->next = next;
3463                 next = current;
3464         }
3465         return next;
3466 }
3467
3468 /*
3469  * Merge the commits h1 and h2, return the resulting virtual
3470  * commit object and a flag indicating the cleanness of the merge.
3471  */
3472 static int merge_recursive_internal(struct merge_options *opt,
3473                                     struct commit *h1,
3474                                     struct commit *h2,
3475                                     struct commit_list *merge_bases,
3476                                     struct commit **result)
3477 {
3478         struct commit_list *iter;
3479         struct commit *merged_merge_bases;
3480         struct tree *mrtree;
3481         int clean;
3482         const char *ancestor_name;
3483         struct strbuf merge_base_abbrev = STRBUF_INIT;
3484
3485         if (show(opt, 4)) {
3486                 output(opt, 4, _("Merging:"));
3487                 output_commit_title(opt, h1);
3488                 output_commit_title(opt, h2);
3489         }
3490
3491         if (!merge_bases) {
3492                 merge_bases = get_merge_bases(h1, h2);
3493                 merge_bases = reverse_commit_list(merge_bases);
3494         }
3495
3496         if (show(opt, 5)) {
3497                 unsigned cnt = commit_list_count(merge_bases);
3498
3499                 output(opt, 5, Q_("found %u common ancestor:",
3500                                 "found %u common ancestors:", cnt), cnt);
3501                 for (iter = merge_bases; iter; iter = iter->next)
3502                         output_commit_title(opt, iter->item);
3503         }
3504
3505         merged_merge_bases = pop_commit(&merge_bases);
3506         if (merged_merge_bases == NULL) {
3507                 /* if there is no common ancestor, use an empty tree */
3508                 struct tree *tree;
3509
3510                 tree = lookup_tree(opt->repo, opt->repo->hash_algo->empty_tree);
3511                 merged_merge_bases = make_virtual_commit(opt->repo, tree,
3512                                                          "ancestor");
3513                 ancestor_name = "empty tree";
3514         } else if (merge_bases) {
3515                 ancestor_name = "merged common ancestors";
3516         } else {
3517                 strbuf_add_unique_abbrev(&merge_base_abbrev,
3518                                          &merged_merge_bases->object.oid,
3519                                          DEFAULT_ABBREV);
3520                 ancestor_name = merge_base_abbrev.buf;
3521         }
3522
3523         for (iter = merge_bases; iter; iter = iter->next) {
3524                 const char *saved_b1, *saved_b2;
3525                 opt->call_depth++;
3526                 /*
3527                  * When the merge fails, the result contains files
3528                  * with conflict markers. The cleanness flag is
3529                  * ignored (unless indicating an error), it was never
3530                  * actually used, as result of merge_trees has always
3531                  * overwritten it: the committed "conflicts" were
3532                  * already resolved.
3533                  */
3534                 discard_index(opt->repo->index);
3535                 saved_b1 = opt->branch1;
3536                 saved_b2 = opt->branch2;
3537                 opt->branch1 = "Temporary merge branch 1";
3538                 opt->branch2 = "Temporary merge branch 2";
3539                 if (merge_recursive_internal(opt, merged_merge_bases, iter->item,
3540                                              NULL, &merged_merge_bases) < 0)
3541                         return -1;
3542                 opt->branch1 = saved_b1;
3543                 opt->branch2 = saved_b2;
3544                 opt->call_depth--;
3545
3546                 if (!merged_merge_bases)
3547                         return err(opt, _("merge returned no commit"));
3548         }
3549
3550         discard_index(opt->repo->index);
3551         if (!opt->call_depth)
3552                 repo_read_index(opt->repo);
3553
3554         opt->ancestor = ancestor_name;
3555         clean = merge_trees_internal(opt,
3556                                      repo_get_commit_tree(opt->repo, h1),
3557                                      repo_get_commit_tree(opt->repo, h2),
3558                                      repo_get_commit_tree(opt->repo,
3559                                                           merged_merge_bases),
3560                                      &mrtree);
3561         strbuf_release(&merge_base_abbrev);
3562         if (clean < 0) {
3563                 flush_output(opt);
3564                 return clean;
3565         }
3566
3567         if (opt->call_depth) {
3568                 *result = make_virtual_commit(opt->repo, mrtree, "merged tree");
3569                 commit_list_insert(h1, &(*result)->parents);
3570                 commit_list_insert(h2, &(*result)->parents->next);
3571         }
3572         flush_output(opt);
3573         if (!opt->call_depth && opt->buffer_output < 2)
3574                 strbuf_release(&opt->obuf);
3575         return clean;
3576 }
3577
3578 static int merge_start(struct merge_options *opt, struct tree *head)
3579 {
3580         struct strbuf sb = STRBUF_INIT;
3581
3582         if (repo_index_has_changes(opt->repo, head, &sb)) {
3583                 err(opt, _("Your local changes to the following files would be overwritten by merge:\n  %s"),
3584                     sb.buf);
3585                 strbuf_release(&sb);
3586                 return -1;
3587         }
3588
3589         return 0;
3590 }
3591
3592 static void merge_finalize(struct merge_options *opt)
3593 {
3594         if (show(opt, 2))
3595                 diff_warn_rename_limit("merge.renamelimit",
3596                                        opt->needed_rename_limit, 0);
3597 }
3598
3599 int merge_trees(struct merge_options *opt,
3600                 struct tree *head,
3601                 struct tree *merge,
3602                 struct tree *merge_base)
3603 {
3604         int clean;
3605         struct tree *ignored;
3606
3607         assert(opt->ancestor != NULL);
3608
3609         if (merge_start(opt, head))
3610                 return -1;
3611         clean = merge_trees_internal(opt, head, merge, merge_base, &ignored);
3612         merge_finalize(opt);
3613
3614         return clean;
3615 }
3616
3617 int merge_recursive(struct merge_options *opt,
3618                     struct commit *h1,
3619                     struct commit *h2,
3620                     struct commit_list *merge_bases,
3621                     struct commit **result)
3622 {
3623         int clean;
3624
3625         assert(opt->ancestor == NULL);
3626
3627         if (merge_start(opt, repo_get_commit_tree(opt->repo, h1)))
3628                 return -1;
3629         clean = merge_recursive_internal(opt, h1, h2, merge_bases, result);
3630         merge_finalize(opt);
3631
3632         return clean;
3633 }
3634
3635 static struct commit *get_ref(struct repository *repo,
3636                               const struct object_id *oid,
3637                               const char *name)
3638 {
3639         struct object *object;
3640
3641         object = deref_tag(repo, parse_object(repo, oid),
3642                            name, strlen(name));
3643         if (!object)
3644                 return NULL;
3645         if (object->type == OBJ_TREE)
3646                 return make_virtual_commit(repo, (struct tree*)object, name);
3647         if (object->type != OBJ_COMMIT)
3648                 return NULL;
3649         if (parse_commit((struct commit *)object))
3650                 return NULL;
3651         return (struct commit *)object;
3652 }
3653
3654 int merge_recursive_generic(struct merge_options *opt,
3655                             const struct object_id *head,
3656                             const struct object_id *merge,
3657                             int num_merge_bases,
3658                             const struct object_id **merge_bases,
3659                             struct commit **result)
3660 {
3661         int clean;
3662         struct lock_file lock = LOCK_INIT;
3663         struct commit *head_commit = get_ref(opt->repo, head, opt->branch1);
3664         struct commit *next_commit = get_ref(opt->repo, merge, opt->branch2);
3665         struct commit_list *ca = NULL;
3666
3667         if (merge_bases) {
3668                 int i;
3669                 for (i = 0; i < num_merge_bases; ++i) {
3670                         struct commit *base;
3671                         if (!(base = get_ref(opt->repo, merge_bases[i],
3672                                              oid_to_hex(merge_bases[i]))))
3673                                 return err(opt, _("Could not parse object '%s'"),
3674                                            oid_to_hex(merge_bases[i]));
3675                         commit_list_insert(base, &ca);
3676                 }
3677         }
3678
3679         repo_hold_locked_index(opt->repo, &lock, LOCK_DIE_ON_ERROR);
3680         clean = merge_recursive(opt, head_commit, next_commit, ca,
3681                                 result);
3682         if (clean < 0) {
3683                 rollback_lock_file(&lock);
3684                 return clean;
3685         }
3686
3687         if (write_locked_index(opt->repo->index, &lock,
3688                                COMMIT_LOCK | SKIP_IF_UNCHANGED))
3689                 return err(opt, _("Unable to write index."));
3690
3691         return clean ? 0 : 1;
3692 }
3693
3694 static void merge_recursive_config(struct merge_options *opt)
3695 {
3696         char *value = NULL;
3697         git_config_get_int("merge.verbosity", &opt->verbosity);
3698         git_config_get_int("diff.renamelimit", &opt->diff_rename_limit);
3699         git_config_get_int("merge.renamelimit", &opt->merge_rename_limit);
3700         if (!git_config_get_string("diff.renames", &value)) {
3701                 opt->diff_detect_rename = git_config_rename("diff.renames", value);
3702                 free(value);
3703         }
3704         if (!git_config_get_string("merge.renames", &value)) {
3705                 opt->merge_detect_rename = git_config_rename("merge.renames", value);
3706                 free(value);
3707         }
3708         if (!git_config_get_string("merge.directoryrenames", &value)) {
3709                 int boolval = git_parse_maybe_bool(value);
3710                 if (0 <= boolval) {
3711                         opt->detect_directory_renames = boolval ?
3712                                 MERGE_DIRECTORY_RENAMES_TRUE :
3713                                 MERGE_DIRECTORY_RENAMES_NONE;
3714                 } else if (!strcasecmp(value, "conflict")) {
3715                         opt->detect_directory_renames =
3716                                 MERGE_DIRECTORY_RENAMES_CONFLICT;
3717                 } /* avoid erroring on values from future versions of git */
3718                 free(value);
3719         }
3720         git_config(git_xmerge_config, NULL);
3721 }
3722
3723 void init_merge_options(struct merge_options *opt,
3724                         struct repository *repo)
3725 {
3726         const char *merge_verbosity;
3727         memset(opt, 0, sizeof(struct merge_options));
3728         opt->repo = repo;
3729         opt->verbosity = 2;
3730         opt->buffer_output = 1;
3731         opt->diff_rename_limit = -1;
3732         opt->merge_rename_limit = -1;
3733         opt->renormalize = 0;
3734         opt->diff_detect_rename = -1;
3735         opt->merge_detect_rename = -1;
3736         opt->detect_directory_renames = MERGE_DIRECTORY_RENAMES_CONFLICT;
3737         merge_recursive_config(opt);
3738         merge_verbosity = getenv("GIT_MERGE_VERBOSITY");
3739         if (merge_verbosity)
3740                 opt->verbosity = strtol(merge_verbosity, NULL, 10);
3741         if (opt->verbosity >= 5)
3742                 opt->buffer_output = 0;
3743         strbuf_init(&opt->obuf, 0);
3744         string_list_init(&opt->df_conflict_file_set, 1);
3745 }
3746
3747 int parse_merge_opt(struct merge_options *opt, const char *s)
3748 {
3749         const char *arg;
3750
3751         if (!s || !*s)
3752                 return -1;
3753         if (!strcmp(s, "ours"))
3754                 opt->recursive_variant = MERGE_RECURSIVE_OURS;
3755         else if (!strcmp(s, "theirs"))
3756                 opt->recursive_variant = MERGE_RECURSIVE_THEIRS;
3757         else if (!strcmp(s, "subtree"))
3758                 opt->subtree_shift = "";
3759         else if (skip_prefix(s, "subtree=", &arg))
3760                 opt->subtree_shift = arg;
3761         else if (!strcmp(s, "patience"))
3762                 opt->xdl_opts = DIFF_WITH_ALG(opt, PATIENCE_DIFF);
3763         else if (!strcmp(s, "histogram"))
3764                 opt->xdl_opts = DIFF_WITH_ALG(opt, HISTOGRAM_DIFF);
3765         else if (skip_prefix(s, "diff-algorithm=", &arg)) {
3766                 long value = parse_algorithm_value(arg);
3767                 if (value < 0)
3768                         return -1;
3769                 /* clear out previous settings */
3770                 DIFF_XDL_CLR(opt, NEED_MINIMAL);
3771                 opt->xdl_opts &= ~XDF_DIFF_ALGORITHM_MASK;
3772                 opt->xdl_opts |= value;
3773         }
3774         else if (!strcmp(s, "ignore-space-change"))
3775                 DIFF_XDL_SET(opt, IGNORE_WHITESPACE_CHANGE);
3776         else if (!strcmp(s, "ignore-all-space"))
3777                 DIFF_XDL_SET(opt, IGNORE_WHITESPACE);
3778         else if (!strcmp(s, "ignore-space-at-eol"))
3779                 DIFF_XDL_SET(opt, IGNORE_WHITESPACE_AT_EOL);
3780         else if (!strcmp(s, "ignore-cr-at-eol"))
3781                 DIFF_XDL_SET(opt, IGNORE_CR_AT_EOL);
3782         else if (!strcmp(s, "renormalize"))
3783                 opt->renormalize = 1;
3784         else if (!strcmp(s, "no-renormalize"))
3785                 opt->renormalize = 0;
3786         else if (!strcmp(s, "no-renames"))
3787                 opt->merge_detect_rename = 0;
3788         else if (!strcmp(s, "find-renames")) {
3789                 opt->merge_detect_rename = 1;
3790                 opt->rename_score = 0;
3791         }
3792         else if (skip_prefix(s, "find-renames=", &arg) ||
3793                  skip_prefix(s, "rename-threshold=", &arg)) {
3794                 if ((opt->rename_score = parse_rename_score(&arg)) == -1 || *arg != 0)
3795                         return -1;
3796                 opt->merge_detect_rename = 1;
3797         }
3798         /*
3799          * Please update $__git_merge_strategy_options in
3800          * git-completion.bash when you add new options
3801          */
3802         else
3803                 return -1;
3804         return 0;
3805 }