merge-recursive: add get_directory_renames()
[git] / merge-recursive.c
1 /*
2  * Recursive Merge algorithm stolen from git-merge-recursive.py by
3  * Fredrik Kuivinen.
4  * The thieves were Alex Riesen and Johannes Schindelin, in June/July 2006
5  */
6 #include "cache.h"
7 #include "config.h"
8 #include "advice.h"
9 #include "lockfile.h"
10 #include "cache-tree.h"
11 #include "commit.h"
12 #include "blob.h"
13 #include "builtin.h"
14 #include "tree-walk.h"
15 #include "diff.h"
16 #include "diffcore.h"
17 #include "tag.h"
18 #include "unpack-trees.h"
19 #include "string-list.h"
20 #include "xdiff-interface.h"
21 #include "ll-merge.h"
22 #include "attr.h"
23 #include "merge-recursive.h"
24 #include "dir.h"
25 #include "submodule.h"
26
27 struct path_hashmap_entry {
28         struct hashmap_entry e;
29         char path[FLEX_ARRAY];
30 };
31
32 static int path_hashmap_cmp(const void *cmp_data,
33                             const void *entry,
34                             const void *entry_or_key,
35                             const void *keydata)
36 {
37         const struct path_hashmap_entry *a = entry;
38         const struct path_hashmap_entry *b = entry_or_key;
39         const char *key = keydata;
40
41         if (ignore_case)
42                 return strcasecmp(a->path, key ? key : b->path);
43         else
44                 return strcmp(a->path, key ? key : b->path);
45 }
46
47 static unsigned int path_hash(const char *path)
48 {
49         return ignore_case ? strihash(path) : strhash(path);
50 }
51
52 static struct dir_rename_entry *dir_rename_find_entry(struct hashmap *hashmap,
53                                                       char *dir)
54 {
55         struct dir_rename_entry key;
56
57         if (dir == NULL)
58                 return NULL;
59         hashmap_entry_init(&key, strhash(dir));
60         key.dir = dir;
61         return hashmap_get(hashmap, &key, NULL);
62 }
63
64 static int dir_rename_cmp(const void *unused_cmp_data,
65                           const void *entry,
66                           const void *entry_or_key,
67                           const void *unused_keydata)
68 {
69         const struct dir_rename_entry *e1 = entry;
70         const struct dir_rename_entry *e2 = entry_or_key;
71
72         return strcmp(e1->dir, e2->dir);
73 }
74
75 static void dir_rename_init(struct hashmap *map)
76 {
77         hashmap_init(map, dir_rename_cmp, NULL, 0);
78 }
79
80 static void dir_rename_entry_init(struct dir_rename_entry *entry,
81                                   char *directory)
82 {
83         hashmap_entry_init(entry, strhash(directory));
84         entry->dir = directory;
85         entry->non_unique_new_dir = 0;
86         strbuf_init(&entry->new_dir, 0);
87         string_list_init(&entry->possible_new_dirs, 0);
88 }
89
90 static void flush_output(struct merge_options *o)
91 {
92         if (o->buffer_output < 2 && o->obuf.len) {
93                 fputs(o->obuf.buf, stdout);
94                 strbuf_reset(&o->obuf);
95         }
96 }
97
98 static int err(struct merge_options *o, const char *err, ...)
99 {
100         va_list params;
101
102         if (o->buffer_output < 2)
103                 flush_output(o);
104         else {
105                 strbuf_complete(&o->obuf, '\n');
106                 strbuf_addstr(&o->obuf, "error: ");
107         }
108         va_start(params, err);
109         strbuf_vaddf(&o->obuf, err, params);
110         va_end(params);
111         if (o->buffer_output > 1)
112                 strbuf_addch(&o->obuf, '\n');
113         else {
114                 error("%s", o->obuf.buf);
115                 strbuf_reset(&o->obuf);
116         }
117
118         return -1;
119 }
120
121 static struct tree *shift_tree_object(struct tree *one, struct tree *two,
122                                       const char *subtree_shift)
123 {
124         struct object_id shifted;
125
126         if (!*subtree_shift) {
127                 shift_tree(&one->object.oid, &two->object.oid, &shifted, 0);
128         } else {
129                 shift_tree_by(&one->object.oid, &two->object.oid, &shifted,
130                               subtree_shift);
131         }
132         if (!oidcmp(&two->object.oid, &shifted))
133                 return two;
134         return lookup_tree(&shifted);
135 }
136
137 static struct commit *make_virtual_commit(struct tree *tree, const char *comment)
138 {
139         struct commit *commit = alloc_commit_node();
140
141         set_merge_remote_desc(commit, comment, (struct object *)commit);
142         commit->tree = tree;
143         commit->object.parsed = 1;
144         return commit;
145 }
146
147 /*
148  * Since we use get_tree_entry(), which does not put the read object into
149  * the object pool, we cannot rely on a == b.
150  */
151 static int oid_eq(const struct object_id *a, const struct object_id *b)
152 {
153         if (!a && !b)
154                 return 2;
155         return a && b && oidcmp(a, b) == 0;
156 }
157
158 enum rename_type {
159         RENAME_NORMAL = 0,
160         RENAME_DELETE,
161         RENAME_ONE_FILE_TO_ONE,
162         RENAME_ONE_FILE_TO_TWO,
163         RENAME_TWO_FILES_TO_ONE
164 };
165
166 struct rename_conflict_info {
167         enum rename_type rename_type;
168         struct diff_filepair *pair1;
169         struct diff_filepair *pair2;
170         const char *branch1;
171         const char *branch2;
172         struct stage_data *dst_entry1;
173         struct stage_data *dst_entry2;
174         struct diff_filespec ren1_other;
175         struct diff_filespec ren2_other;
176 };
177
178 /*
179  * Since we want to write the index eventually, we cannot reuse the index
180  * for these (temporary) data.
181  */
182 struct stage_data {
183         struct {
184                 unsigned mode;
185                 struct object_id oid;
186         } stages[4];
187         struct rename_conflict_info *rename_conflict_info;
188         unsigned processed:1;
189 };
190
191 static inline void setup_rename_conflict_info(enum rename_type rename_type,
192                                               struct diff_filepair *pair1,
193                                               struct diff_filepair *pair2,
194                                               const char *branch1,
195                                               const char *branch2,
196                                               struct stage_data *dst_entry1,
197                                               struct stage_data *dst_entry2,
198                                               struct merge_options *o,
199                                               struct stage_data *src_entry1,
200                                               struct stage_data *src_entry2)
201 {
202         struct rename_conflict_info *ci = xcalloc(1, sizeof(struct rename_conflict_info));
203         ci->rename_type = rename_type;
204         ci->pair1 = pair1;
205         ci->branch1 = branch1;
206         ci->branch2 = branch2;
207
208         ci->dst_entry1 = dst_entry1;
209         dst_entry1->rename_conflict_info = ci;
210         dst_entry1->processed = 0;
211
212         assert(!pair2 == !dst_entry2);
213         if (dst_entry2) {
214                 ci->dst_entry2 = dst_entry2;
215                 ci->pair2 = pair2;
216                 dst_entry2->rename_conflict_info = ci;
217         }
218
219         if (rename_type == RENAME_TWO_FILES_TO_ONE) {
220                 /*
221                  * For each rename, there could have been
222                  * modifications on the side of history where that
223                  * file was not renamed.
224                  */
225                 int ostage1 = o->branch1 == branch1 ? 3 : 2;
226                 int ostage2 = ostage1 ^ 1;
227
228                 ci->ren1_other.path = pair1->one->path;
229                 oidcpy(&ci->ren1_other.oid, &src_entry1->stages[ostage1].oid);
230                 ci->ren1_other.mode = src_entry1->stages[ostage1].mode;
231
232                 ci->ren2_other.path = pair2->one->path;
233                 oidcpy(&ci->ren2_other.oid, &src_entry2->stages[ostage2].oid);
234                 ci->ren2_other.mode = src_entry2->stages[ostage2].mode;
235         }
236 }
237
238 static int show(struct merge_options *o, int v)
239 {
240         return (!o->call_depth && o->verbosity >= v) || o->verbosity >= 5;
241 }
242
243 __attribute__((format (printf, 3, 4)))
244 static void output(struct merge_options *o, int v, const char *fmt, ...)
245 {
246         va_list ap;
247
248         if (!show(o, v))
249                 return;
250
251         strbuf_addchars(&o->obuf, ' ', o->call_depth * 2);
252
253         va_start(ap, fmt);
254         strbuf_vaddf(&o->obuf, fmt, ap);
255         va_end(ap);
256
257         strbuf_addch(&o->obuf, '\n');
258         if (!o->buffer_output)
259                 flush_output(o);
260 }
261
262 static void output_commit_title(struct merge_options *o, struct commit *commit)
263 {
264         strbuf_addchars(&o->obuf, ' ', o->call_depth * 2);
265         if (commit->util)
266                 strbuf_addf(&o->obuf, "virtual %s\n",
267                         merge_remote_util(commit)->name);
268         else {
269                 strbuf_add_unique_abbrev(&o->obuf, &commit->object.oid,
270                                          DEFAULT_ABBREV);
271                 strbuf_addch(&o->obuf, ' ');
272                 if (parse_commit(commit) != 0)
273                         strbuf_addstr(&o->obuf, _("(bad commit)\n"));
274                 else {
275                         const char *title;
276                         const char *msg = get_commit_buffer(commit, NULL);
277                         int len = find_commit_subject(msg, &title);
278                         if (len)
279                                 strbuf_addf(&o->obuf, "%.*s\n", len, title);
280                         unuse_commit_buffer(commit, msg);
281                 }
282         }
283         flush_output(o);
284 }
285
286 static int add_cacheinfo(struct merge_options *o,
287                 unsigned int mode, const struct object_id *oid,
288                 const char *path, int stage, int refresh, int options)
289 {
290         struct cache_entry *ce;
291         int ret;
292
293         ce = make_cache_entry(mode, oid ? oid->hash : null_sha1, path, stage, 0);
294         if (!ce)
295                 return err(o, _("addinfo_cache failed for path '%s'"), path);
296
297         ret = add_cache_entry(ce, options);
298         if (refresh) {
299                 struct cache_entry *nce;
300
301                 nce = refresh_cache_entry(ce, CE_MATCH_REFRESH | CE_MATCH_IGNORE_MISSING);
302                 if (!nce)
303                         return err(o, _("addinfo_cache failed for path '%s'"), path);
304                 if (nce != ce)
305                         ret = add_cache_entry(nce, options);
306         }
307         return ret;
308 }
309
310 static void init_tree_desc_from_tree(struct tree_desc *desc, struct tree *tree)
311 {
312         parse_tree(tree);
313         init_tree_desc(desc, tree->buffer, tree->size);
314 }
315
316 static int git_merge_trees(int index_only,
317                            struct tree *common,
318                            struct tree *head,
319                            struct tree *merge)
320 {
321         int rc;
322         struct tree_desc t[3];
323         struct unpack_trees_options opts;
324
325         memset(&opts, 0, sizeof(opts));
326         if (index_only)
327                 opts.index_only = 1;
328         else
329                 opts.update = 1;
330         opts.merge = 1;
331         opts.head_idx = 2;
332         opts.fn = threeway_merge;
333         opts.src_index = &the_index;
334         opts.dst_index = &the_index;
335         setup_unpack_trees_porcelain(&opts, "merge");
336
337         init_tree_desc_from_tree(t+0, common);
338         init_tree_desc_from_tree(t+1, head);
339         init_tree_desc_from_tree(t+2, merge);
340
341         rc = unpack_trees(3, t, &opts);
342         cache_tree_free(&active_cache_tree);
343         return rc;
344 }
345
346 struct tree *write_tree_from_memory(struct merge_options *o)
347 {
348         struct tree *result = NULL;
349
350         if (unmerged_cache()) {
351                 int i;
352                 fprintf(stderr, "BUG: There are unmerged index entries:\n");
353                 for (i = 0; i < active_nr; i++) {
354                         const struct cache_entry *ce = active_cache[i];
355                         if (ce_stage(ce))
356                                 fprintf(stderr, "BUG: %d %.*s\n", ce_stage(ce),
357                                         (int)ce_namelen(ce), ce->name);
358                 }
359                 die("BUG: unmerged index entries in merge-recursive.c");
360         }
361
362         if (!active_cache_tree)
363                 active_cache_tree = cache_tree();
364
365         if (!cache_tree_fully_valid(active_cache_tree) &&
366             cache_tree_update(&the_index, 0) < 0) {
367                 err(o, _("error building trees"));
368                 return NULL;
369         }
370
371         result = lookup_tree(&active_cache_tree->oid);
372
373         return result;
374 }
375
376 static int save_files_dirs(const struct object_id *oid,
377                 struct strbuf *base, const char *path,
378                 unsigned int mode, int stage, void *context)
379 {
380         struct path_hashmap_entry *entry;
381         int baselen = base->len;
382         struct merge_options *o = context;
383
384         strbuf_addstr(base, path);
385
386         FLEX_ALLOC_MEM(entry, path, base->buf, base->len);
387         hashmap_entry_init(entry, path_hash(entry->path));
388         hashmap_add(&o->current_file_dir_set, entry);
389
390         strbuf_setlen(base, baselen);
391         return (S_ISDIR(mode) ? READ_TREE_RECURSIVE : 0);
392 }
393
394 static void get_files_dirs(struct merge_options *o, struct tree *tree)
395 {
396         struct pathspec match_all;
397         memset(&match_all, 0, sizeof(match_all));
398         read_tree_recursive(tree, "", 0, 0, &match_all, save_files_dirs, o);
399 }
400
401 /*
402  * Returns an index_entry instance which doesn't have to correspond to
403  * a real cache entry in Git's index.
404  */
405 static struct stage_data *insert_stage_data(const char *path,
406                 struct tree *o, struct tree *a, struct tree *b,
407                 struct string_list *entries)
408 {
409         struct string_list_item *item;
410         struct stage_data *e = xcalloc(1, sizeof(struct stage_data));
411         get_tree_entry(&o->object.oid, path,
412                         &e->stages[1].oid, &e->stages[1].mode);
413         get_tree_entry(&a->object.oid, path,
414                         &e->stages[2].oid, &e->stages[2].mode);
415         get_tree_entry(&b->object.oid, path,
416                         &e->stages[3].oid, &e->stages[3].mode);
417         item = string_list_insert(entries, path);
418         item->util = e;
419         return e;
420 }
421
422 /*
423  * Create a dictionary mapping file names to stage_data objects. The
424  * dictionary contains one entry for every path with a non-zero stage entry.
425  */
426 static struct string_list *get_unmerged(void)
427 {
428         struct string_list *unmerged = xcalloc(1, sizeof(struct string_list));
429         int i;
430
431         unmerged->strdup_strings = 1;
432
433         for (i = 0; i < active_nr; i++) {
434                 struct string_list_item *item;
435                 struct stage_data *e;
436                 const struct cache_entry *ce = active_cache[i];
437                 if (!ce_stage(ce))
438                         continue;
439
440                 item = string_list_lookup(unmerged, ce->name);
441                 if (!item) {
442                         item = string_list_insert(unmerged, ce->name);
443                         item->util = xcalloc(1, sizeof(struct stage_data));
444                 }
445                 e = item->util;
446                 e->stages[ce_stage(ce)].mode = ce->ce_mode;
447                 oidcpy(&e->stages[ce_stage(ce)].oid, &ce->oid);
448         }
449
450         return unmerged;
451 }
452
453 static int string_list_df_name_compare(const char *one, const char *two)
454 {
455         int onelen = strlen(one);
456         int twolen = strlen(two);
457         /*
458          * Here we only care that entries for D/F conflicts are
459          * adjacent, in particular with the file of the D/F conflict
460          * appearing before files below the corresponding directory.
461          * The order of the rest of the list is irrelevant for us.
462          *
463          * To achieve this, we sort with df_name_compare and provide
464          * the mode S_IFDIR so that D/F conflicts will sort correctly.
465          * We use the mode S_IFDIR for everything else for simplicity,
466          * since in other cases any changes in their order due to
467          * sorting cause no problems for us.
468          */
469         int cmp = df_name_compare(one, onelen, S_IFDIR,
470                                   two, twolen, S_IFDIR);
471         /*
472          * Now that 'foo' and 'foo/bar' compare equal, we have to make sure
473          * that 'foo' comes before 'foo/bar'.
474          */
475         if (cmp)
476                 return cmp;
477         return onelen - twolen;
478 }
479
480 static void record_df_conflict_files(struct merge_options *o,
481                                      struct string_list *entries)
482 {
483         /* If there is a D/F conflict and the file for such a conflict
484          * currently exist in the working tree, we want to allow it to be
485          * removed to make room for the corresponding directory if needed.
486          * The files underneath the directories of such D/F conflicts will
487          * be processed before the corresponding file involved in the D/F
488          * conflict.  If the D/F directory ends up being removed by the
489          * merge, then we won't have to touch the D/F file.  If the D/F
490          * directory needs to be written to the working copy, then the D/F
491          * file will simply be removed (in make_room_for_path()) to make
492          * room for the necessary paths.  Note that if both the directory
493          * and the file need to be present, then the D/F file will be
494          * reinstated with a new unique name at the time it is processed.
495          */
496         struct string_list df_sorted_entries = STRING_LIST_INIT_NODUP;
497         const char *last_file = NULL;
498         int last_len = 0;
499         int i;
500
501         /*
502          * If we're merging merge-bases, we don't want to bother with
503          * any working directory changes.
504          */
505         if (o->call_depth)
506                 return;
507
508         /* Ensure D/F conflicts are adjacent in the entries list. */
509         for (i = 0; i < entries->nr; i++) {
510                 struct string_list_item *next = &entries->items[i];
511                 string_list_append(&df_sorted_entries, next->string)->util =
512                                    next->util;
513         }
514         df_sorted_entries.cmp = string_list_df_name_compare;
515         string_list_sort(&df_sorted_entries);
516
517         string_list_clear(&o->df_conflict_file_set, 1);
518         for (i = 0; i < df_sorted_entries.nr; i++) {
519                 const char *path = df_sorted_entries.items[i].string;
520                 int len = strlen(path);
521                 struct stage_data *e = df_sorted_entries.items[i].util;
522
523                 /*
524                  * Check if last_file & path correspond to a D/F conflict;
525                  * i.e. whether path is last_file+'/'+<something>.
526                  * If so, record that it's okay to remove last_file to make
527                  * room for path and friends if needed.
528                  */
529                 if (last_file &&
530                     len > last_len &&
531                     memcmp(path, last_file, last_len) == 0 &&
532                     path[last_len] == '/') {
533                         string_list_insert(&o->df_conflict_file_set, last_file);
534                 }
535
536                 /*
537                  * Determine whether path could exist as a file in the
538                  * working directory as a possible D/F conflict.  This
539                  * will only occur when it exists in stage 2 as a
540                  * file.
541                  */
542                 if (S_ISREG(e->stages[2].mode) || S_ISLNK(e->stages[2].mode)) {
543                         last_file = path;
544                         last_len = len;
545                 } else {
546                         last_file = NULL;
547                 }
548         }
549         string_list_clear(&df_sorted_entries, 0);
550 }
551
552 struct rename {
553         struct diff_filepair *pair;
554         /*
555          * Purpose of src_entry and dst_entry:
556          *
557          * If 'before' is renamed to 'after' then src_entry will contain
558          * the versions of 'before' from the merge_base, HEAD, and MERGE in
559          * stages 1, 2, and 3; dst_entry will contain the respective
560          * versions of 'after' in corresponding locations.  Thus, we have a
561          * total of six modes and oids, though some will be null.  (Stage 0
562          * is ignored; we're interested in handling conflicts.)
563          *
564          * Since we don't turn on break-rewrites by default, neither
565          * src_entry nor dst_entry can have all three of their stages have
566          * non-null oids, meaning at most four of the six will be non-null.
567          * Also, since this is a rename, both src_entry and dst_entry will
568          * have at least one non-null oid, meaning at least two will be
569          * non-null.  Of the six oids, a typical rename will have three be
570          * non-null.  Only two implies a rename/delete, and four implies a
571          * rename/add.
572          */
573         struct stage_data *src_entry;
574         struct stage_data *dst_entry;
575         unsigned processed:1;
576 };
577
578 static int update_stages(struct merge_options *opt, const char *path,
579                          const struct diff_filespec *o,
580                          const struct diff_filespec *a,
581                          const struct diff_filespec *b)
582 {
583
584         /*
585          * NOTE: It is usually a bad idea to call update_stages on a path
586          * before calling update_file on that same path, since it can
587          * sometimes lead to spurious "refusing to lose untracked file..."
588          * messages from update_file (via make_room_for path via
589          * would_lose_untracked).  Instead, reverse the order of the calls
590          * (executing update_file first and then update_stages).
591          */
592         int clear = 1;
593         int options = ADD_CACHE_OK_TO_ADD | ADD_CACHE_SKIP_DFCHECK;
594         if (clear)
595                 if (remove_file_from_cache(path))
596                         return -1;
597         if (o)
598                 if (add_cacheinfo(opt, o->mode, &o->oid, path, 1, 0, options))
599                         return -1;
600         if (a)
601                 if (add_cacheinfo(opt, a->mode, &a->oid, path, 2, 0, options))
602                         return -1;
603         if (b)
604                 if (add_cacheinfo(opt, b->mode, &b->oid, path, 3, 0, options))
605                         return -1;
606         return 0;
607 }
608
609 static void update_entry(struct stage_data *entry,
610                          struct diff_filespec *o,
611                          struct diff_filespec *a,
612                          struct diff_filespec *b)
613 {
614         entry->processed = 0;
615         entry->stages[1].mode = o->mode;
616         entry->stages[2].mode = a->mode;
617         entry->stages[3].mode = b->mode;
618         oidcpy(&entry->stages[1].oid, &o->oid);
619         oidcpy(&entry->stages[2].oid, &a->oid);
620         oidcpy(&entry->stages[3].oid, &b->oid);
621 }
622
623 static int remove_file(struct merge_options *o, int clean,
624                        const char *path, int no_wd)
625 {
626         int update_cache = o->call_depth || clean;
627         int update_working_directory = !o->call_depth && !no_wd;
628
629         if (update_cache) {
630                 if (remove_file_from_cache(path))
631                         return -1;
632         }
633         if (update_working_directory) {
634                 if (ignore_case) {
635                         struct cache_entry *ce;
636                         ce = cache_file_exists(path, strlen(path), ignore_case);
637                         if (ce && ce_stage(ce) == 0 && strcmp(path, ce->name))
638                                 return 0;
639                 }
640                 if (remove_path(path))
641                         return -1;
642         }
643         return 0;
644 }
645
646 /* add a string to a strbuf, but converting "/" to "_" */
647 static void add_flattened_path(struct strbuf *out, const char *s)
648 {
649         size_t i = out->len;
650         strbuf_addstr(out, s);
651         for (; i < out->len; i++)
652                 if (out->buf[i] == '/')
653                         out->buf[i] = '_';
654 }
655
656 static char *unique_path(struct merge_options *o, const char *path, const char *branch)
657 {
658         struct path_hashmap_entry *entry;
659         struct strbuf newpath = STRBUF_INIT;
660         int suffix = 0;
661         size_t base_len;
662
663         strbuf_addf(&newpath, "%s~", path);
664         add_flattened_path(&newpath, branch);
665
666         base_len = newpath.len;
667         while (hashmap_get_from_hash(&o->current_file_dir_set,
668                                      path_hash(newpath.buf), newpath.buf) ||
669                (!o->call_depth && file_exists(newpath.buf))) {
670                 strbuf_setlen(&newpath, base_len);
671                 strbuf_addf(&newpath, "_%d", suffix++);
672         }
673
674         FLEX_ALLOC_MEM(entry, path, newpath.buf, newpath.len);
675         hashmap_entry_init(entry, path_hash(entry->path));
676         hashmap_add(&o->current_file_dir_set, entry);
677         return strbuf_detach(&newpath, NULL);
678 }
679
680 /**
681  * Check whether a directory in the index is in the way of an incoming
682  * file.  Return 1 if so.  If check_working_copy is non-zero, also
683  * check the working directory.  If empty_ok is non-zero, also return
684  * 0 in the case where the working-tree dir exists but is empty.
685  */
686 static int dir_in_way(const char *path, int check_working_copy, int empty_ok)
687 {
688         int pos;
689         struct strbuf dirpath = STRBUF_INIT;
690         struct stat st;
691
692         strbuf_addstr(&dirpath, path);
693         strbuf_addch(&dirpath, '/');
694
695         pos = cache_name_pos(dirpath.buf, dirpath.len);
696
697         if (pos < 0)
698                 pos = -1 - pos;
699         if (pos < active_nr &&
700             !strncmp(dirpath.buf, active_cache[pos]->name, dirpath.len)) {
701                 strbuf_release(&dirpath);
702                 return 1;
703         }
704
705         strbuf_release(&dirpath);
706         return check_working_copy && !lstat(path, &st) && S_ISDIR(st.st_mode) &&
707                 !(empty_ok && is_empty_dir(path));
708 }
709
710 static int was_tracked(const char *path)
711 {
712         int pos = cache_name_pos(path, strlen(path));
713
714         if (0 <= pos)
715                 /* we have been tracking this path */
716                 return 1;
717
718         /*
719          * Look for an unmerged entry for the path,
720          * specifically stage #2, which would indicate
721          * that "our" side before the merge started
722          * had the path tracked (and resulted in a conflict).
723          */
724         for (pos = -1 - pos;
725              pos < active_nr && !strcmp(path, active_cache[pos]->name);
726              pos++)
727                 if (ce_stage(active_cache[pos]) == 2)
728                         return 1;
729         return 0;
730 }
731
732 static int would_lose_untracked(const char *path)
733 {
734         return !was_tracked(path) && file_exists(path);
735 }
736
737 static int make_room_for_path(struct merge_options *o, const char *path)
738 {
739         int status, i;
740         const char *msg = _("failed to create path '%s'%s");
741
742         /* Unlink any D/F conflict files that are in the way */
743         for (i = 0; i < o->df_conflict_file_set.nr; i++) {
744                 const char *df_path = o->df_conflict_file_set.items[i].string;
745                 size_t pathlen = strlen(path);
746                 size_t df_pathlen = strlen(df_path);
747                 if (df_pathlen < pathlen &&
748                     path[df_pathlen] == '/' &&
749                     strncmp(path, df_path, df_pathlen) == 0) {
750                         output(o, 3,
751                                _("Removing %s to make room for subdirectory\n"),
752                                df_path);
753                         unlink(df_path);
754                         unsorted_string_list_delete_item(&o->df_conflict_file_set,
755                                                          i, 0);
756                         break;
757                 }
758         }
759
760         /* Make sure leading directories are created */
761         status = safe_create_leading_directories_const(path);
762         if (status) {
763                 if (status == SCLD_EXISTS)
764                         /* something else exists */
765                         return err(o, msg, path, _(": perhaps a D/F conflict?"));
766                 return err(o, msg, path, "");
767         }
768
769         /*
770          * Do not unlink a file in the work tree if we are not
771          * tracking it.
772          */
773         if (would_lose_untracked(path))
774                 return err(o, _("refusing to lose untracked file at '%s'"),
775                              path);
776
777         /* Successful unlink is good.. */
778         if (!unlink(path))
779                 return 0;
780         /* .. and so is no existing file */
781         if (errno == ENOENT)
782                 return 0;
783         /* .. but not some other error (who really cares what?) */
784         return err(o, msg, path, _(": perhaps a D/F conflict?"));
785 }
786
787 static int update_file_flags(struct merge_options *o,
788                              const struct object_id *oid,
789                              unsigned mode,
790                              const char *path,
791                              int update_cache,
792                              int update_wd)
793 {
794         int ret = 0;
795
796         if (o->call_depth)
797                 update_wd = 0;
798
799         if (update_wd) {
800                 enum object_type type;
801                 void *buf;
802                 unsigned long size;
803
804                 if (S_ISGITLINK(mode)) {
805                         /*
806                          * We may later decide to recursively descend into
807                          * the submodule directory and update its index
808                          * and/or work tree, but we do not do that now.
809                          */
810                         update_wd = 0;
811                         goto update_index;
812                 }
813
814                 buf = read_object_file(oid, &type, &size);
815                 if (!buf)
816                         return err(o, _("cannot read object %s '%s'"), oid_to_hex(oid), path);
817                 if (type != OBJ_BLOB) {
818                         ret = err(o, _("blob expected for %s '%s'"), oid_to_hex(oid), path);
819                         goto free_buf;
820                 }
821                 if (S_ISREG(mode)) {
822                         struct strbuf strbuf = STRBUF_INIT;
823                         if (convert_to_working_tree(path, buf, size, &strbuf)) {
824                                 free(buf);
825                                 size = strbuf.len;
826                                 buf = strbuf_detach(&strbuf, NULL);
827                         }
828                 }
829
830                 if (make_room_for_path(o, path) < 0) {
831                         update_wd = 0;
832                         goto free_buf;
833                 }
834                 if (S_ISREG(mode) || (!has_symlinks && S_ISLNK(mode))) {
835                         int fd;
836                         if (mode & 0100)
837                                 mode = 0777;
838                         else
839                                 mode = 0666;
840                         fd = open(path, O_WRONLY | O_TRUNC | O_CREAT, mode);
841                         if (fd < 0) {
842                                 ret = err(o, _("failed to open '%s': %s"),
843                                           path, strerror(errno));
844                                 goto free_buf;
845                         }
846                         write_in_full(fd, buf, size);
847                         close(fd);
848                 } else if (S_ISLNK(mode)) {
849                         char *lnk = xmemdupz(buf, size);
850                         safe_create_leading_directories_const(path);
851                         unlink(path);
852                         if (symlink(lnk, path))
853                                 ret = err(o, _("failed to symlink '%s': %s"),
854                                         path, strerror(errno));
855                         free(lnk);
856                 } else
857                         ret = err(o,
858                                   _("do not know what to do with %06o %s '%s'"),
859                                   mode, oid_to_hex(oid), path);
860  free_buf:
861                 free(buf);
862         }
863  update_index:
864         if (!ret && update_cache)
865                 add_cacheinfo(o, mode, oid, path, 0, update_wd, ADD_CACHE_OK_TO_ADD);
866         return ret;
867 }
868
869 static int update_file(struct merge_options *o,
870                        int clean,
871                        const struct object_id *oid,
872                        unsigned mode,
873                        const char *path)
874 {
875         return update_file_flags(o, oid, mode, path, o->call_depth || clean, !o->call_depth);
876 }
877
878 /* Low level file merging, update and removal */
879
880 struct merge_file_info {
881         struct object_id oid;
882         unsigned mode;
883         unsigned clean:1,
884                  merge:1;
885 };
886
887 static int merge_3way(struct merge_options *o,
888                       mmbuffer_t *result_buf,
889                       const struct diff_filespec *one,
890                       const struct diff_filespec *a,
891                       const struct diff_filespec *b,
892                       const char *branch1,
893                       const char *branch2)
894 {
895         mmfile_t orig, src1, src2;
896         struct ll_merge_options ll_opts = {0};
897         char *base_name, *name1, *name2;
898         int merge_status;
899
900         ll_opts.renormalize = o->renormalize;
901         ll_opts.xdl_opts = o->xdl_opts;
902
903         if (o->call_depth) {
904                 ll_opts.virtual_ancestor = 1;
905                 ll_opts.variant = 0;
906         } else {
907                 switch (o->recursive_variant) {
908                 case MERGE_RECURSIVE_OURS:
909                         ll_opts.variant = XDL_MERGE_FAVOR_OURS;
910                         break;
911                 case MERGE_RECURSIVE_THEIRS:
912                         ll_opts.variant = XDL_MERGE_FAVOR_THEIRS;
913                         break;
914                 default:
915                         ll_opts.variant = 0;
916                         break;
917                 }
918         }
919
920         if (strcmp(a->path, b->path) ||
921             (o->ancestor != NULL && strcmp(a->path, one->path) != 0)) {
922                 base_name = o->ancestor == NULL ? NULL :
923                         mkpathdup("%s:%s", o->ancestor, one->path);
924                 name1 = mkpathdup("%s:%s", branch1, a->path);
925                 name2 = mkpathdup("%s:%s", branch2, b->path);
926         } else {
927                 base_name = o->ancestor == NULL ? NULL :
928                         mkpathdup("%s", o->ancestor);
929                 name1 = mkpathdup("%s", branch1);
930                 name2 = mkpathdup("%s", branch2);
931         }
932
933         read_mmblob(&orig, &one->oid);
934         read_mmblob(&src1, &a->oid);
935         read_mmblob(&src2, &b->oid);
936
937         merge_status = ll_merge(result_buf, a->path, &orig, base_name,
938                                 &src1, name1, &src2, name2, &ll_opts);
939
940         free(base_name);
941         free(name1);
942         free(name2);
943         free(orig.ptr);
944         free(src1.ptr);
945         free(src2.ptr);
946         return merge_status;
947 }
948
949 static int merge_file_1(struct merge_options *o,
950                                            const struct diff_filespec *one,
951                                            const struct diff_filespec *a,
952                                            const struct diff_filespec *b,
953                                            const char *branch1,
954                                            const char *branch2,
955                                            struct merge_file_info *result)
956 {
957         result->merge = 0;
958         result->clean = 1;
959
960         if ((S_IFMT & a->mode) != (S_IFMT & b->mode)) {
961                 result->clean = 0;
962                 if (S_ISREG(a->mode)) {
963                         result->mode = a->mode;
964                         oidcpy(&result->oid, &a->oid);
965                 } else {
966                         result->mode = b->mode;
967                         oidcpy(&result->oid, &b->oid);
968                 }
969         } else {
970                 if (!oid_eq(&a->oid, &one->oid) && !oid_eq(&b->oid, &one->oid))
971                         result->merge = 1;
972
973                 /*
974                  * Merge modes
975                  */
976                 if (a->mode == b->mode || a->mode == one->mode)
977                         result->mode = b->mode;
978                 else {
979                         result->mode = a->mode;
980                         if (b->mode != one->mode) {
981                                 result->clean = 0;
982                                 result->merge = 1;
983                         }
984                 }
985
986                 if (oid_eq(&a->oid, &b->oid) || oid_eq(&a->oid, &one->oid))
987                         oidcpy(&result->oid, &b->oid);
988                 else if (oid_eq(&b->oid, &one->oid))
989                         oidcpy(&result->oid, &a->oid);
990                 else if (S_ISREG(a->mode)) {
991                         mmbuffer_t result_buf;
992                         int ret = 0, merge_status;
993
994                         merge_status = merge_3way(o, &result_buf, one, a, b,
995                                                   branch1, branch2);
996
997                         if ((merge_status < 0) || !result_buf.ptr)
998                                 ret = err(o, _("Failed to execute internal merge"));
999
1000                         if (!ret &&
1001                             write_object_file(result_buf.ptr, result_buf.size,
1002                                               blob_type, &result->oid))
1003                                 ret = err(o, _("Unable to add %s to database"),
1004                                           a->path);
1005
1006                         free(result_buf.ptr);
1007                         if (ret)
1008                                 return ret;
1009                         result->clean = (merge_status == 0);
1010                 } else if (S_ISGITLINK(a->mode)) {
1011                         result->clean = merge_submodule(&result->oid,
1012                                                        one->path,
1013                                                        &one->oid,
1014                                                        &a->oid,
1015                                                        &b->oid,
1016                                                        !o->call_depth);
1017                 } else if (S_ISLNK(a->mode)) {
1018                         switch (o->recursive_variant) {
1019                         case MERGE_RECURSIVE_NORMAL:
1020                                 oidcpy(&result->oid, &a->oid);
1021                                 if (!oid_eq(&a->oid, &b->oid))
1022                                         result->clean = 0;
1023                                 break;
1024                         case MERGE_RECURSIVE_OURS:
1025                                 oidcpy(&result->oid, &a->oid);
1026                                 break;
1027                         case MERGE_RECURSIVE_THEIRS:
1028                                 oidcpy(&result->oid, &b->oid);
1029                                 break;
1030                         }
1031                 } else
1032                         die("BUG: unsupported object type in the tree");
1033         }
1034
1035         return 0;
1036 }
1037
1038 static int merge_file_special_markers(struct merge_options *o,
1039                            const struct diff_filespec *one,
1040                            const struct diff_filespec *a,
1041                            const struct diff_filespec *b,
1042                            const char *branch1,
1043                            const char *filename1,
1044                            const char *branch2,
1045                            const char *filename2,
1046                            struct merge_file_info *mfi)
1047 {
1048         char *side1 = NULL;
1049         char *side2 = NULL;
1050         int ret;
1051
1052         if (filename1)
1053                 side1 = xstrfmt("%s:%s", branch1, filename1);
1054         if (filename2)
1055                 side2 = xstrfmt("%s:%s", branch2, filename2);
1056
1057         ret = merge_file_1(o, one, a, b,
1058                            side1 ? side1 : branch1,
1059                            side2 ? side2 : branch2, mfi);
1060         free(side1);
1061         free(side2);
1062         return ret;
1063 }
1064
1065 static int merge_file_one(struct merge_options *o,
1066                                          const char *path,
1067                                          const struct object_id *o_oid, int o_mode,
1068                                          const struct object_id *a_oid, int a_mode,
1069                                          const struct object_id *b_oid, int b_mode,
1070                                          const char *branch1,
1071                                          const char *branch2,
1072                                          struct merge_file_info *mfi)
1073 {
1074         struct diff_filespec one, a, b;
1075
1076         one.path = a.path = b.path = (char *)path;
1077         oidcpy(&one.oid, o_oid);
1078         one.mode = o_mode;
1079         oidcpy(&a.oid, a_oid);
1080         a.mode = a_mode;
1081         oidcpy(&b.oid, b_oid);
1082         b.mode = b_mode;
1083         return merge_file_1(o, &one, &a, &b, branch1, branch2, mfi);
1084 }
1085
1086 static int handle_change_delete(struct merge_options *o,
1087                                  const char *path, const char *old_path,
1088                                  const struct object_id *o_oid, int o_mode,
1089                                  const struct object_id *changed_oid,
1090                                  int changed_mode,
1091                                  const char *change_branch,
1092                                  const char *delete_branch,
1093                                  const char *change, const char *change_past)
1094 {
1095         char *alt_path = NULL;
1096         const char *update_path = path;
1097         int ret = 0;
1098
1099         if (dir_in_way(path, !o->call_depth, 0)) {
1100                 update_path = alt_path = unique_path(o, path, change_branch);
1101         }
1102
1103         if (o->call_depth) {
1104                 /*
1105                  * We cannot arbitrarily accept either a_sha or b_sha as
1106                  * correct; since there is no true "middle point" between
1107                  * them, simply reuse the base version for virtual merge base.
1108                  */
1109                 ret = remove_file_from_cache(path);
1110                 if (!ret)
1111                         ret = update_file(o, 0, o_oid, o_mode, update_path);
1112         } else {
1113                 if (!alt_path) {
1114                         if (!old_path) {
1115                                 output(o, 1, _("CONFLICT (%s/delete): %s deleted in %s "
1116                                        "and %s in %s. Version %s of %s left in tree."),
1117                                        change, path, delete_branch, change_past,
1118                                        change_branch, change_branch, path);
1119                         } else {
1120                                 output(o, 1, _("CONFLICT (%s/delete): %s deleted in %s "
1121                                        "and %s to %s in %s. Version %s of %s left in tree."),
1122                                        change, old_path, delete_branch, change_past, path,
1123                                        change_branch, change_branch, path);
1124                         }
1125                 } else {
1126                         if (!old_path) {
1127                                 output(o, 1, _("CONFLICT (%s/delete): %s deleted in %s "
1128                                        "and %s in %s. Version %s of %s left in tree at %s."),
1129                                        change, path, delete_branch, change_past,
1130                                        change_branch, change_branch, path, alt_path);
1131                         } else {
1132                                 output(o, 1, _("CONFLICT (%s/delete): %s deleted in %s "
1133                                        "and %s to %s in %s. Version %s of %s left in tree at %s."),
1134                                        change, old_path, delete_branch, change_past, path,
1135                                        change_branch, change_branch, path, alt_path);
1136                         }
1137                 }
1138                 /*
1139                  * No need to call update_file() on path when change_branch ==
1140                  * o->branch1 && !alt_path, since that would needlessly touch
1141                  * path.  We could call update_file_flags() with update_cache=0
1142                  * and update_wd=0, but that's a no-op.
1143                  */
1144                 if (change_branch != o->branch1 || alt_path)
1145                         ret = update_file(o, 0, changed_oid, changed_mode, update_path);
1146         }
1147         free(alt_path);
1148
1149         return ret;
1150 }
1151
1152 static int conflict_rename_delete(struct merge_options *o,
1153                                    struct diff_filepair *pair,
1154                                    const char *rename_branch,
1155                                    const char *delete_branch)
1156 {
1157         const struct diff_filespec *orig = pair->one;
1158         const struct diff_filespec *dest = pair->two;
1159
1160         if (handle_change_delete(o,
1161                                  o->call_depth ? orig->path : dest->path,
1162                                  o->call_depth ? NULL : orig->path,
1163                                  &orig->oid, orig->mode,
1164                                  &dest->oid, dest->mode,
1165                                  rename_branch, delete_branch,
1166                                  _("rename"), _("renamed")))
1167                 return -1;
1168
1169         if (o->call_depth)
1170                 return remove_file_from_cache(dest->path);
1171         else
1172                 return update_stages(o, dest->path, NULL,
1173                                      rename_branch == o->branch1 ? dest : NULL,
1174                                      rename_branch == o->branch1 ? NULL : dest);
1175 }
1176
1177 static struct diff_filespec *filespec_from_entry(struct diff_filespec *target,
1178                                                  struct stage_data *entry,
1179                                                  int stage)
1180 {
1181         struct object_id *oid = &entry->stages[stage].oid;
1182         unsigned mode = entry->stages[stage].mode;
1183         if (mode == 0 || is_null_oid(oid))
1184                 return NULL;
1185         oidcpy(&target->oid, oid);
1186         target->mode = mode;
1187         return target;
1188 }
1189
1190 static int handle_file(struct merge_options *o,
1191                         struct diff_filespec *rename,
1192                         int stage,
1193                         struct rename_conflict_info *ci)
1194 {
1195         char *dst_name = rename->path;
1196         struct stage_data *dst_entry;
1197         const char *cur_branch, *other_branch;
1198         struct diff_filespec other;
1199         struct diff_filespec *add;
1200         int ret;
1201
1202         if (stage == 2) {
1203                 dst_entry = ci->dst_entry1;
1204                 cur_branch = ci->branch1;
1205                 other_branch = ci->branch2;
1206         } else {
1207                 dst_entry = ci->dst_entry2;
1208                 cur_branch = ci->branch2;
1209                 other_branch = ci->branch1;
1210         }
1211
1212         add = filespec_from_entry(&other, dst_entry, stage ^ 1);
1213         if (add) {
1214                 char *add_name = unique_path(o, rename->path, other_branch);
1215                 if (update_file(o, 0, &add->oid, add->mode, add_name))
1216                         return -1;
1217
1218                 remove_file(o, 0, rename->path, 0);
1219                 dst_name = unique_path(o, rename->path, cur_branch);
1220         } else {
1221                 if (dir_in_way(rename->path, !o->call_depth, 0)) {
1222                         dst_name = unique_path(o, rename->path, cur_branch);
1223                         output(o, 1, _("%s is a directory in %s adding as %s instead"),
1224                                rename->path, other_branch, dst_name);
1225                 }
1226         }
1227         if ((ret = update_file(o, 0, &rename->oid, rename->mode, dst_name)))
1228                 ; /* fall through, do allow dst_name to be released */
1229         else if (stage == 2)
1230                 ret = update_stages(o, rename->path, NULL, rename, add);
1231         else
1232                 ret = update_stages(o, rename->path, NULL, add, rename);
1233
1234         if (dst_name != rename->path)
1235                 free(dst_name);
1236
1237         return ret;
1238 }
1239
1240 static int conflict_rename_rename_1to2(struct merge_options *o,
1241                                         struct rename_conflict_info *ci)
1242 {
1243         /* One file was renamed in both branches, but to different names. */
1244         struct diff_filespec *one = ci->pair1->one;
1245         struct diff_filespec *a = ci->pair1->two;
1246         struct diff_filespec *b = ci->pair2->two;
1247
1248         output(o, 1, _("CONFLICT (rename/rename): "
1249                "Rename \"%s\"->\"%s\" in branch \"%s\" "
1250                "rename \"%s\"->\"%s\" in \"%s\"%s"),
1251                one->path, a->path, ci->branch1,
1252                one->path, b->path, ci->branch2,
1253                o->call_depth ? _(" (left unresolved)") : "");
1254         if (o->call_depth) {
1255                 struct merge_file_info mfi;
1256                 struct diff_filespec other;
1257                 struct diff_filespec *add;
1258                 if (merge_file_one(o, one->path,
1259                                  &one->oid, one->mode,
1260                                  &a->oid, a->mode,
1261                                  &b->oid, b->mode,
1262                                  ci->branch1, ci->branch2, &mfi))
1263                         return -1;
1264
1265                 /*
1266                  * FIXME: For rename/add-source conflicts (if we could detect
1267                  * such), this is wrong.  We should instead find a unique
1268                  * pathname and then either rename the add-source file to that
1269                  * unique path, or use that unique path instead of src here.
1270                  */
1271                 if (update_file(o, 0, &mfi.oid, mfi.mode, one->path))
1272                         return -1;
1273
1274                 /*
1275                  * Above, we put the merged content at the merge-base's
1276                  * path.  Now we usually need to delete both a->path and
1277                  * b->path.  However, the rename on each side of the merge
1278                  * could also be involved in a rename/add conflict.  In
1279                  * such cases, we should keep the added file around,
1280                  * resolving the conflict at that path in its favor.
1281                  */
1282                 add = filespec_from_entry(&other, ci->dst_entry1, 2 ^ 1);
1283                 if (add) {
1284                         if (update_file(o, 0, &add->oid, add->mode, a->path))
1285                                 return -1;
1286                 }
1287                 else
1288                         remove_file_from_cache(a->path);
1289                 add = filespec_from_entry(&other, ci->dst_entry2, 3 ^ 1);
1290                 if (add) {
1291                         if (update_file(o, 0, &add->oid, add->mode, b->path))
1292                                 return -1;
1293                 }
1294                 else
1295                         remove_file_from_cache(b->path);
1296         } else if (handle_file(o, a, 2, ci) || handle_file(o, b, 3, ci))
1297                 return -1;
1298
1299         return 0;
1300 }
1301
1302 static int conflict_rename_rename_2to1(struct merge_options *o,
1303                                         struct rename_conflict_info *ci)
1304 {
1305         /* Two files, a & b, were renamed to the same thing, c. */
1306         struct diff_filespec *a = ci->pair1->one;
1307         struct diff_filespec *b = ci->pair2->one;
1308         struct diff_filespec *c1 = ci->pair1->two;
1309         struct diff_filespec *c2 = ci->pair2->two;
1310         char *path = c1->path; /* == c2->path */
1311         struct merge_file_info mfi_c1;
1312         struct merge_file_info mfi_c2;
1313         int ret;
1314
1315         output(o, 1, _("CONFLICT (rename/rename): "
1316                "Rename %s->%s in %s. "
1317                "Rename %s->%s in %s"),
1318                a->path, c1->path, ci->branch1,
1319                b->path, c2->path, ci->branch2);
1320
1321         remove_file(o, 1, a->path, o->call_depth || would_lose_untracked(a->path));
1322         remove_file(o, 1, b->path, o->call_depth || would_lose_untracked(b->path));
1323
1324         if (merge_file_special_markers(o, a, c1, &ci->ren1_other,
1325                                        o->branch1, c1->path,
1326                                        o->branch2, ci->ren1_other.path, &mfi_c1) ||
1327             merge_file_special_markers(o, b, &ci->ren2_other, c2,
1328                                        o->branch1, ci->ren2_other.path,
1329                                        o->branch2, c2->path, &mfi_c2))
1330                 return -1;
1331
1332         if (o->call_depth) {
1333                 /*
1334                  * If mfi_c1.clean && mfi_c2.clean, then it might make
1335                  * sense to do a two-way merge of those results.  But, I
1336                  * think in all cases, it makes sense to have the virtual
1337                  * merge base just undo the renames; they can be detected
1338                  * again later for the non-recursive merge.
1339                  */
1340                 remove_file(o, 0, path, 0);
1341                 ret = update_file(o, 0, &mfi_c1.oid, mfi_c1.mode, a->path);
1342                 if (!ret)
1343                         ret = update_file(o, 0, &mfi_c2.oid, mfi_c2.mode,
1344                                           b->path);
1345         } else {
1346                 char *new_path1 = unique_path(o, path, ci->branch1);
1347                 char *new_path2 = unique_path(o, path, ci->branch2);
1348                 output(o, 1, _("Renaming %s to %s and %s to %s instead"),
1349                        a->path, new_path1, b->path, new_path2);
1350                 remove_file(o, 0, path, 0);
1351                 ret = update_file(o, 0, &mfi_c1.oid, mfi_c1.mode, new_path1);
1352                 if (!ret)
1353                         ret = update_file(o, 0, &mfi_c2.oid, mfi_c2.mode,
1354                                           new_path2);
1355                 free(new_path2);
1356                 free(new_path1);
1357         }
1358
1359         return ret;
1360 }
1361
1362 /*
1363  * Get the diff_filepairs changed between o_tree and tree.
1364  */
1365 static struct diff_queue_struct *get_diffpairs(struct merge_options *o,
1366                                                struct tree *o_tree,
1367                                                struct tree *tree)
1368 {
1369         struct diff_queue_struct *ret;
1370         struct diff_options opts;
1371
1372         diff_setup(&opts);
1373         opts.flags.recursive = 1;
1374         opts.flags.rename_empty = 0;
1375         opts.detect_rename = DIFF_DETECT_RENAME;
1376         opts.rename_limit = o->merge_rename_limit >= 0 ? o->merge_rename_limit :
1377                             o->diff_rename_limit >= 0 ? o->diff_rename_limit :
1378                             1000;
1379         opts.rename_score = o->rename_score;
1380         opts.show_rename_progress = o->show_rename_progress;
1381         opts.output_format = DIFF_FORMAT_NO_OUTPUT;
1382         diff_setup_done(&opts);
1383         diff_tree_oid(&o_tree->object.oid, &tree->object.oid, "", &opts);
1384         diffcore_std(&opts);
1385         if (opts.needed_rename_limit > o->needed_rename_limit)
1386                 o->needed_rename_limit = opts.needed_rename_limit;
1387
1388         ret = xmalloc(sizeof(*ret));
1389         *ret = diff_queued_diff;
1390
1391         opts.output_format = DIFF_FORMAT_NO_OUTPUT;
1392         diff_queued_diff.nr = 0;
1393         diff_queued_diff.queue = NULL;
1394         diff_flush(&opts);
1395         return ret;
1396 }
1397
1398 static void get_renamed_dir_portion(const char *old_path, const char *new_path,
1399                                     char **old_dir, char **new_dir)
1400 {
1401         char *end_of_old, *end_of_new;
1402         int old_len, new_len;
1403
1404         *old_dir = NULL;
1405         *new_dir = NULL;
1406
1407         /*
1408          * For
1409          *    "a/b/c/d/e/foo.c" -> "a/b/some/thing/else/e/foo.c"
1410          * the "e/foo.c" part is the same, we just want to know that
1411          *    "a/b/c/d" was renamed to "a/b/some/thing/else"
1412          * so, for this example, this function returns "a/b/c/d" in
1413          * *old_dir and "a/b/some/thing/else" in *new_dir.
1414          *
1415          * Also, if the basename of the file changed, we don't care.  We
1416          * want to know which portion of the directory, if any, changed.
1417          */
1418         end_of_old = strrchr(old_path, '/');
1419         end_of_new = strrchr(new_path, '/');
1420
1421         if (end_of_old == NULL || end_of_new == NULL)
1422                 return;
1423         while (*--end_of_new == *--end_of_old &&
1424                end_of_old != old_path &&
1425                end_of_new != new_path)
1426                 ; /* Do nothing; all in the while loop */
1427         /*
1428          * We've found the first non-matching character in the directory
1429          * paths.  That means the current directory we were comparing
1430          * represents the rename.  Move end_of_old and end_of_new back
1431          * to the full directory name.
1432          */
1433         if (*end_of_old == '/')
1434                 end_of_old++;
1435         if (*end_of_old != '/')
1436                 end_of_new++;
1437         end_of_old = strchr(end_of_old, '/');
1438         end_of_new = strchr(end_of_new, '/');
1439
1440         /*
1441          * It may have been the case that old_path and new_path were the same
1442          * directory all along.  Don't claim a rename if they're the same.
1443          */
1444         old_len = end_of_old - old_path;
1445         new_len = end_of_new - new_path;
1446
1447         if (old_len != new_len || strncmp(old_path, new_path, old_len)) {
1448                 *old_dir = xstrndup(old_path, old_len);
1449                 *new_dir = xstrndup(new_path, new_len);
1450         }
1451 }
1452
1453 static struct hashmap *get_directory_renames(struct diff_queue_struct *pairs,
1454                                              struct tree *tree)
1455 {
1456         struct hashmap *dir_renames;
1457         struct hashmap_iter iter;
1458         struct dir_rename_entry *entry;
1459         int i;
1460
1461         /*
1462          * Typically, we think of a directory rename as all files from a
1463          * certain directory being moved to a target directory.  However,
1464          * what if someone first moved two files from the original
1465          * directory in one commit, and then renamed the directory
1466          * somewhere else in a later commit?  At merge time, we just know
1467          * that files from the original directory went to two different
1468          * places, and that the bulk of them ended up in the same place.
1469          * We want each directory rename to represent where the bulk of the
1470          * files from that directory end up; this function exists to find
1471          * where the bulk of the files went.
1472          *
1473          * The first loop below simply iterates through the list of file
1474          * renames, finding out how often each directory rename pair
1475          * possibility occurs.
1476          */
1477         dir_renames = xmalloc(sizeof(*dir_renames));
1478         dir_rename_init(dir_renames);
1479         for (i = 0; i < pairs->nr; ++i) {
1480                 struct string_list_item *item;
1481                 int *count;
1482                 struct diff_filepair *pair = pairs->queue[i];
1483                 char *old_dir, *new_dir;
1484
1485                 /* File not part of directory rename if it wasn't renamed */
1486                 if (pair->status != 'R')
1487                         continue;
1488
1489                 get_renamed_dir_portion(pair->one->path, pair->two->path,
1490                                         &old_dir,        &new_dir);
1491                 if (!old_dir)
1492                         /* Directory didn't change at all; ignore this one. */
1493                         continue;
1494
1495                 entry = dir_rename_find_entry(dir_renames, old_dir);
1496                 if (!entry) {
1497                         entry = xmalloc(sizeof(*entry));
1498                         dir_rename_entry_init(entry, old_dir);
1499                         hashmap_put(dir_renames, entry);
1500                 } else {
1501                         free(old_dir);
1502                 }
1503                 item = string_list_lookup(&entry->possible_new_dirs, new_dir);
1504                 if (!item) {
1505                         item = string_list_insert(&entry->possible_new_dirs,
1506                                                   new_dir);
1507                         item->util = xcalloc(1, sizeof(int));
1508                 } else {
1509                         free(new_dir);
1510                 }
1511                 count = item->util;
1512                 *count += 1;
1513         }
1514
1515         /*
1516          * For each directory with files moved out of it, we find out which
1517          * target directory received the most files so we can declare it to
1518          * be the "winning" target location for the directory rename.  This
1519          * winner gets recorded in new_dir.  If there is no winner
1520          * (multiple target directories received the same number of files),
1521          * we set non_unique_new_dir.  Once we've determined the winner (or
1522          * that there is no winner), we no longer need possible_new_dirs.
1523          */
1524         hashmap_iter_init(dir_renames, &iter);
1525         while ((entry = hashmap_iter_next(&iter))) {
1526                 int max = 0;
1527                 int bad_max = 0;
1528                 char *best = NULL;
1529
1530                 for (i = 0; i < entry->possible_new_dirs.nr; i++) {
1531                         int *count = entry->possible_new_dirs.items[i].util;
1532
1533                         if (*count == max)
1534                                 bad_max = max;
1535                         else if (*count > max) {
1536                                 max = *count;
1537                                 best = entry->possible_new_dirs.items[i].string;
1538                         }
1539                 }
1540                 if (bad_max == max)
1541                         entry->non_unique_new_dir = 1;
1542                 else {
1543                         assert(entry->new_dir.len == 0);
1544                         strbuf_addstr(&entry->new_dir, best);
1545                 }
1546                 /*
1547                  * The relevant directory sub-portion of the original full
1548                  * filepaths were xstrndup'ed before inserting into
1549                  * possible_new_dirs, and instead of manually iterating the
1550                  * list and free'ing each, just lie and tell
1551                  * possible_new_dirs that it did the strdup'ing so that it
1552                  * will free them for us.
1553                  */
1554                 entry->possible_new_dirs.strdup_strings = 1;
1555                 string_list_clear(&entry->possible_new_dirs, 1);
1556         }
1557
1558         return dir_renames;
1559 }
1560
1561 /*
1562  * Get information of all renames which occurred in 'pairs', making use of
1563  * any implicit directory renames inferred from the other side of history.
1564  * We need the three trees in the merge ('o_tree', 'a_tree' and 'b_tree')
1565  * to be able to associate the correct cache entries with the rename
1566  * information; tree is always equal to either a_tree or b_tree.
1567  */
1568 static struct string_list *get_renames(struct merge_options *o,
1569                                        struct diff_queue_struct *pairs,
1570                                        struct tree *tree,
1571                                        struct tree *o_tree,
1572                                        struct tree *a_tree,
1573                                        struct tree *b_tree,
1574                                        struct string_list *entries)
1575 {
1576         int i;
1577         struct string_list *renames;
1578
1579         renames = xcalloc(1, sizeof(struct string_list));
1580
1581         for (i = 0; i < pairs->nr; ++i) {
1582                 struct string_list_item *item;
1583                 struct rename *re;
1584                 struct diff_filepair *pair = pairs->queue[i];
1585
1586                 if (pair->status != 'R') {
1587                         diff_free_filepair(pair);
1588                         continue;
1589                 }
1590                 re = xmalloc(sizeof(*re));
1591                 re->processed = 0;
1592                 re->pair = pair;
1593                 item = string_list_lookup(entries, re->pair->one->path);
1594                 if (!item)
1595                         re->src_entry = insert_stage_data(re->pair->one->path,
1596                                         o_tree, a_tree, b_tree, entries);
1597                 else
1598                         re->src_entry = item->util;
1599
1600                 item = string_list_lookup(entries, re->pair->two->path);
1601                 if (!item)
1602                         re->dst_entry = insert_stage_data(re->pair->two->path,
1603                                         o_tree, a_tree, b_tree, entries);
1604                 else
1605                         re->dst_entry = item->util;
1606                 item = string_list_insert(renames, pair->one->path);
1607                 item->util = re;
1608         }
1609         return renames;
1610 }
1611
1612 static int process_renames(struct merge_options *o,
1613                            struct string_list *a_renames,
1614                            struct string_list *b_renames)
1615 {
1616         int clean_merge = 1, i, j;
1617         struct string_list a_by_dst = STRING_LIST_INIT_NODUP;
1618         struct string_list b_by_dst = STRING_LIST_INIT_NODUP;
1619         const struct rename *sre;
1620
1621         for (i = 0; i < a_renames->nr; i++) {
1622                 sre = a_renames->items[i].util;
1623                 string_list_insert(&a_by_dst, sre->pair->two->path)->util
1624                         = (void *)sre;
1625         }
1626         for (i = 0; i < b_renames->nr; i++) {
1627                 sre = b_renames->items[i].util;
1628                 string_list_insert(&b_by_dst, sre->pair->two->path)->util
1629                         = (void *)sre;
1630         }
1631
1632         for (i = 0, j = 0; i < a_renames->nr || j < b_renames->nr;) {
1633                 struct string_list *renames1, *renames2Dst;
1634                 struct rename *ren1 = NULL, *ren2 = NULL;
1635                 const char *branch1, *branch2;
1636                 const char *ren1_src, *ren1_dst;
1637                 struct string_list_item *lookup;
1638
1639                 if (i >= a_renames->nr) {
1640                         ren2 = b_renames->items[j++].util;
1641                 } else if (j >= b_renames->nr) {
1642                         ren1 = a_renames->items[i++].util;
1643                 } else {
1644                         int compare = strcmp(a_renames->items[i].string,
1645                                              b_renames->items[j].string);
1646                         if (compare <= 0)
1647                                 ren1 = a_renames->items[i++].util;
1648                         if (compare >= 0)
1649                                 ren2 = b_renames->items[j++].util;
1650                 }
1651
1652                 /* TODO: refactor, so that 1/2 are not needed */
1653                 if (ren1) {
1654                         renames1 = a_renames;
1655                         renames2Dst = &b_by_dst;
1656                         branch1 = o->branch1;
1657                         branch2 = o->branch2;
1658                 } else {
1659                         renames1 = b_renames;
1660                         renames2Dst = &a_by_dst;
1661                         branch1 = o->branch2;
1662                         branch2 = o->branch1;
1663                         SWAP(ren2, ren1);
1664                 }
1665
1666                 if (ren1->processed)
1667                         continue;
1668                 ren1->processed = 1;
1669                 ren1->dst_entry->processed = 1;
1670                 /* BUG: We should only mark src_entry as processed if we
1671                  * are not dealing with a rename + add-source case.
1672                  */
1673                 ren1->src_entry->processed = 1;
1674
1675                 ren1_src = ren1->pair->one->path;
1676                 ren1_dst = ren1->pair->two->path;
1677
1678                 if (ren2) {
1679                         /* One file renamed on both sides */
1680                         const char *ren2_src = ren2->pair->one->path;
1681                         const char *ren2_dst = ren2->pair->two->path;
1682                         enum rename_type rename_type;
1683                         if (strcmp(ren1_src, ren2_src) != 0)
1684                                 die("BUG: ren1_src != ren2_src");
1685                         ren2->dst_entry->processed = 1;
1686                         ren2->processed = 1;
1687                         if (strcmp(ren1_dst, ren2_dst) != 0) {
1688                                 rename_type = RENAME_ONE_FILE_TO_TWO;
1689                                 clean_merge = 0;
1690                         } else {
1691                                 rename_type = RENAME_ONE_FILE_TO_ONE;
1692                                 /* BUG: We should only remove ren1_src in
1693                                  * the base stage (think of rename +
1694                                  * add-source cases).
1695                                  */
1696                                 remove_file(o, 1, ren1_src, 1);
1697                                 update_entry(ren1->dst_entry,
1698                                              ren1->pair->one,
1699                                              ren1->pair->two,
1700                                              ren2->pair->two);
1701                         }
1702                         setup_rename_conflict_info(rename_type,
1703                                                    ren1->pair,
1704                                                    ren2->pair,
1705                                                    branch1,
1706                                                    branch2,
1707                                                    ren1->dst_entry,
1708                                                    ren2->dst_entry,
1709                                                    o,
1710                                                    NULL,
1711                                                    NULL);
1712                 } else if ((lookup = string_list_lookup(renames2Dst, ren1_dst))) {
1713                         /* Two different files renamed to the same thing */
1714                         char *ren2_dst;
1715                         ren2 = lookup->util;
1716                         ren2_dst = ren2->pair->two->path;
1717                         if (strcmp(ren1_dst, ren2_dst) != 0)
1718                                 die("BUG: ren1_dst != ren2_dst");
1719
1720                         clean_merge = 0;
1721                         ren2->processed = 1;
1722                         /*
1723                          * BUG: We should only mark src_entry as processed
1724                          * if we are not dealing with a rename + add-source
1725                          * case.
1726                          */
1727                         ren2->src_entry->processed = 1;
1728
1729                         setup_rename_conflict_info(RENAME_TWO_FILES_TO_ONE,
1730                                                    ren1->pair,
1731                                                    ren2->pair,
1732                                                    branch1,
1733                                                    branch2,
1734                                                    ren1->dst_entry,
1735                                                    ren2->dst_entry,
1736                                                    o,
1737                                                    ren1->src_entry,
1738                                                    ren2->src_entry);
1739
1740                 } else {
1741                         /* Renamed in 1, maybe changed in 2 */
1742                         /* we only use sha1 and mode of these */
1743                         struct diff_filespec src_other, dst_other;
1744                         int try_merge;
1745
1746                         /*
1747                          * unpack_trees loads entries from common-commit
1748                          * into stage 1, from head-commit into stage 2, and
1749                          * from merge-commit into stage 3.  We keep track
1750                          * of which side corresponds to the rename.
1751                          */
1752                         int renamed_stage = a_renames == renames1 ? 2 : 3;
1753                         int other_stage =   a_renames == renames1 ? 3 : 2;
1754
1755                         /* BUG: We should only remove ren1_src in the base
1756                          * stage and in other_stage (think of rename +
1757                          * add-source case).
1758                          */
1759                         remove_file(o, 1, ren1_src,
1760                                     renamed_stage == 2 || !was_tracked(ren1_src));
1761
1762                         oidcpy(&src_other.oid,
1763                                &ren1->src_entry->stages[other_stage].oid);
1764                         src_other.mode = ren1->src_entry->stages[other_stage].mode;
1765                         oidcpy(&dst_other.oid,
1766                                &ren1->dst_entry->stages[other_stage].oid);
1767                         dst_other.mode = ren1->dst_entry->stages[other_stage].mode;
1768                         try_merge = 0;
1769
1770                         if (oid_eq(&src_other.oid, &null_oid)) {
1771                                 setup_rename_conflict_info(RENAME_DELETE,
1772                                                            ren1->pair,
1773                                                            NULL,
1774                                                            branch1,
1775                                                            branch2,
1776                                                            ren1->dst_entry,
1777                                                            NULL,
1778                                                            o,
1779                                                            NULL,
1780                                                            NULL);
1781                         } else if ((dst_other.mode == ren1->pair->two->mode) &&
1782                                    oid_eq(&dst_other.oid, &ren1->pair->two->oid)) {
1783                                 /*
1784                                  * Added file on the other side identical to
1785                                  * the file being renamed: clean merge.
1786                                  * Also, there is no need to overwrite the
1787                                  * file already in the working copy, so call
1788                                  * update_file_flags() instead of
1789                                  * update_file().
1790                                  */
1791                                 if (update_file_flags(o,
1792                                                       &ren1->pair->two->oid,
1793                                                       ren1->pair->two->mode,
1794                                                       ren1_dst,
1795                                                       1, /* update_cache */
1796                                                       0  /* update_wd    */))
1797                                         clean_merge = -1;
1798                         } else if (!oid_eq(&dst_other.oid, &null_oid)) {
1799                                 clean_merge = 0;
1800                                 try_merge = 1;
1801                                 output(o, 1, _("CONFLICT (rename/add): Rename %s->%s in %s. "
1802                                        "%s added in %s"),
1803                                        ren1_src, ren1_dst, branch1,
1804                                        ren1_dst, branch2);
1805                                 if (o->call_depth) {
1806                                         struct merge_file_info mfi;
1807                                         if (merge_file_one(o, ren1_dst, &null_oid, 0,
1808                                                            &ren1->pair->two->oid,
1809                                                            ren1->pair->two->mode,
1810                                                            &dst_other.oid,
1811                                                            dst_other.mode,
1812                                                            branch1, branch2, &mfi)) {
1813                                                 clean_merge = -1;
1814                                                 goto cleanup_and_return;
1815                                         }
1816                                         output(o, 1, _("Adding merged %s"), ren1_dst);
1817                                         if (update_file(o, 0, &mfi.oid,
1818                                                         mfi.mode, ren1_dst))
1819                                                 clean_merge = -1;
1820                                         try_merge = 0;
1821                                 } else {
1822                                         char *new_path = unique_path(o, ren1_dst, branch2);
1823                                         output(o, 1, _("Adding as %s instead"), new_path);
1824                                         if (update_file(o, 0, &dst_other.oid,
1825                                                         dst_other.mode, new_path))
1826                                                 clean_merge = -1;
1827                                         free(new_path);
1828                                 }
1829                         } else
1830                                 try_merge = 1;
1831
1832                         if (clean_merge < 0)
1833                                 goto cleanup_and_return;
1834                         if (try_merge) {
1835                                 struct diff_filespec *one, *a, *b;
1836                                 src_other.path = (char *)ren1_src;
1837
1838                                 one = ren1->pair->one;
1839                                 if (a_renames == renames1) {
1840                                         a = ren1->pair->two;
1841                                         b = &src_other;
1842                                 } else {
1843                                         b = ren1->pair->two;
1844                                         a = &src_other;
1845                                 }
1846                                 update_entry(ren1->dst_entry, one, a, b);
1847                                 setup_rename_conflict_info(RENAME_NORMAL,
1848                                                            ren1->pair,
1849                                                            NULL,
1850                                                            branch1,
1851                                                            NULL,
1852                                                            ren1->dst_entry,
1853                                                            NULL,
1854                                                            o,
1855                                                            NULL,
1856                                                            NULL);
1857                         }
1858                 }
1859         }
1860 cleanup_and_return:
1861         string_list_clear(&a_by_dst, 0);
1862         string_list_clear(&b_by_dst, 0);
1863
1864         return clean_merge;
1865 }
1866
1867 struct rename_info {
1868         struct string_list *head_renames;
1869         struct string_list *merge_renames;
1870 };
1871
1872 static void initial_cleanup_rename(struct diff_queue_struct *pairs,
1873                                    struct hashmap *dir_renames)
1874 {
1875         struct hashmap_iter iter;
1876         struct dir_rename_entry *e;
1877
1878         hashmap_iter_init(dir_renames, &iter);
1879         while ((e = hashmap_iter_next(&iter))) {
1880                 free(e->dir);
1881                 strbuf_release(&e->new_dir);
1882                 /* possible_new_dirs already cleared in get_directory_renames */
1883         }
1884         hashmap_free(dir_renames, 1);
1885         free(dir_renames);
1886
1887         free(pairs->queue);
1888         free(pairs);
1889 }
1890
1891 static int handle_renames(struct merge_options *o,
1892                           struct tree *common,
1893                           struct tree *head,
1894                           struct tree *merge,
1895                           struct string_list *entries,
1896                           struct rename_info *ri)
1897 {
1898         struct diff_queue_struct *head_pairs, *merge_pairs;
1899         struct hashmap *dir_re_head, *dir_re_merge;
1900         int clean;
1901
1902         ri->head_renames = NULL;
1903         ri->merge_renames = NULL;
1904
1905         if (!o->detect_rename)
1906                 return 1;
1907
1908         head_pairs = get_diffpairs(o, common, head);
1909         merge_pairs = get_diffpairs(o, common, merge);
1910
1911         dir_re_head = get_directory_renames(head_pairs, head);
1912         dir_re_merge = get_directory_renames(merge_pairs, merge);
1913
1914         ri->head_renames  = get_renames(o, head_pairs, head,
1915                                          common, head, merge, entries);
1916         ri->merge_renames = get_renames(o, merge_pairs, merge,
1917                                          common, head, merge, entries);
1918         clean = process_renames(o, ri->head_renames, ri->merge_renames);
1919
1920         /*
1921          * Some cleanup is deferred until cleanup_renames() because the
1922          * data structures are still needed and referenced in
1923          * process_entry().  But there are a few things we can free now.
1924          */
1925         initial_cleanup_rename(head_pairs, dir_re_head);
1926         initial_cleanup_rename(merge_pairs, dir_re_merge);
1927
1928         return clean;
1929 }
1930
1931 static void final_cleanup_rename(struct string_list *rename)
1932 {
1933         const struct rename *re;
1934         int i;
1935
1936         if (rename == NULL)
1937                 return;
1938
1939         for (i = 0; i < rename->nr; i++) {
1940                 re = rename->items[i].util;
1941                 diff_free_filepair(re->pair);
1942         }
1943         string_list_clear(rename, 1);
1944         free(rename);
1945 }
1946
1947 static void final_cleanup_renames(struct rename_info *re_info)
1948 {
1949         final_cleanup_rename(re_info->head_renames);
1950         final_cleanup_rename(re_info->merge_renames);
1951 }
1952
1953 static struct object_id *stage_oid(const struct object_id *oid, unsigned mode)
1954 {
1955         return (is_null_oid(oid) || mode == 0) ? NULL: (struct object_id *)oid;
1956 }
1957
1958 static int read_oid_strbuf(struct merge_options *o,
1959         const struct object_id *oid, struct strbuf *dst)
1960 {
1961         void *buf;
1962         enum object_type type;
1963         unsigned long size;
1964         buf = read_object_file(oid, &type, &size);
1965         if (!buf)
1966                 return err(o, _("cannot read object %s"), oid_to_hex(oid));
1967         if (type != OBJ_BLOB) {
1968                 free(buf);
1969                 return err(o, _("object %s is not a blob"), oid_to_hex(oid));
1970         }
1971         strbuf_attach(dst, buf, size, size + 1);
1972         return 0;
1973 }
1974
1975 static int blob_unchanged(struct merge_options *opt,
1976                           const struct object_id *o_oid,
1977                           unsigned o_mode,
1978                           const struct object_id *a_oid,
1979                           unsigned a_mode,
1980                           int renormalize, const char *path)
1981 {
1982         struct strbuf o = STRBUF_INIT;
1983         struct strbuf a = STRBUF_INIT;
1984         int ret = 0; /* assume changed for safety */
1985
1986         if (a_mode != o_mode)
1987                 return 0;
1988         if (oid_eq(o_oid, a_oid))
1989                 return 1;
1990         if (!renormalize)
1991                 return 0;
1992
1993         assert(o_oid && a_oid);
1994         if (read_oid_strbuf(opt, o_oid, &o) || read_oid_strbuf(opt, a_oid, &a))
1995                 goto error_return;
1996         /*
1997          * Note: binary | is used so that both renormalizations are
1998          * performed.  Comparison can be skipped if both files are
1999          * unchanged since their sha1s have already been compared.
2000          */
2001         if (renormalize_buffer(&the_index, path, o.buf, o.len, &o) |
2002             renormalize_buffer(&the_index, path, a.buf, a.len, &a))
2003                 ret = (o.len == a.len && !memcmp(o.buf, a.buf, o.len));
2004
2005 error_return:
2006         strbuf_release(&o);
2007         strbuf_release(&a);
2008         return ret;
2009 }
2010
2011 static int handle_modify_delete(struct merge_options *o,
2012                                  const char *path,
2013                                  struct object_id *o_oid, int o_mode,
2014                                  struct object_id *a_oid, int a_mode,
2015                                  struct object_id *b_oid, int b_mode)
2016 {
2017         const char *modify_branch, *delete_branch;
2018         struct object_id *changed_oid;
2019         int changed_mode;
2020
2021         if (a_oid) {
2022                 modify_branch = o->branch1;
2023                 delete_branch = o->branch2;
2024                 changed_oid = a_oid;
2025                 changed_mode = a_mode;
2026         } else {
2027                 modify_branch = o->branch2;
2028                 delete_branch = o->branch1;
2029                 changed_oid = b_oid;
2030                 changed_mode = b_mode;
2031         }
2032
2033         return handle_change_delete(o,
2034                                     path, NULL,
2035                                     o_oid, o_mode,
2036                                     changed_oid, changed_mode,
2037                                     modify_branch, delete_branch,
2038                                     _("modify"), _("modified"));
2039 }
2040
2041 static int merge_content(struct merge_options *o,
2042                          const char *path,
2043                          struct object_id *o_oid, int o_mode,
2044                          struct object_id *a_oid, int a_mode,
2045                          struct object_id *b_oid, int b_mode,
2046                          struct rename_conflict_info *rename_conflict_info)
2047 {
2048         const char *reason = _("content");
2049         const char *path1 = NULL, *path2 = NULL;
2050         struct merge_file_info mfi;
2051         struct diff_filespec one, a, b;
2052         unsigned df_conflict_remains = 0;
2053
2054         if (!o_oid) {
2055                 reason = _("add/add");
2056                 o_oid = (struct object_id *)&null_oid;
2057         }
2058         one.path = a.path = b.path = (char *)path;
2059         oidcpy(&one.oid, o_oid);
2060         one.mode = o_mode;
2061         oidcpy(&a.oid, a_oid);
2062         a.mode = a_mode;
2063         oidcpy(&b.oid, b_oid);
2064         b.mode = b_mode;
2065
2066         if (rename_conflict_info) {
2067                 struct diff_filepair *pair1 = rename_conflict_info->pair1;
2068
2069                 path1 = (o->branch1 == rename_conflict_info->branch1) ?
2070                         pair1->two->path : pair1->one->path;
2071                 /* If rename_conflict_info->pair2 != NULL, we are in
2072                  * RENAME_ONE_FILE_TO_ONE case.  Otherwise, we have a
2073                  * normal rename.
2074                  */
2075                 path2 = (rename_conflict_info->pair2 ||
2076                          o->branch2 == rename_conflict_info->branch1) ?
2077                         pair1->two->path : pair1->one->path;
2078
2079                 if (dir_in_way(path, !o->call_depth,
2080                                S_ISGITLINK(pair1->two->mode)))
2081                         df_conflict_remains = 1;
2082         }
2083         if (merge_file_special_markers(o, &one, &a, &b,
2084                                        o->branch1, path1,
2085                                        o->branch2, path2, &mfi))
2086                 return -1;
2087
2088         if (mfi.clean && !df_conflict_remains &&
2089             oid_eq(&mfi.oid, a_oid) && mfi.mode == a_mode) {
2090                 int path_renamed_outside_HEAD;
2091                 output(o, 3, _("Skipped %s (merged same as existing)"), path);
2092                 /*
2093                  * The content merge resulted in the same file contents we
2094                  * already had.  We can return early if those file contents
2095                  * are recorded at the correct path (which may not be true
2096                  * if the merge involves a rename).
2097                  */
2098                 path_renamed_outside_HEAD = !path2 || !strcmp(path, path2);
2099                 if (!path_renamed_outside_HEAD) {
2100                         add_cacheinfo(o, mfi.mode, &mfi.oid, path,
2101                                       0, (!o->call_depth), 0);
2102                         return mfi.clean;
2103                 }
2104         } else
2105                 output(o, 2, _("Auto-merging %s"), path);
2106
2107         if (!mfi.clean) {
2108                 if (S_ISGITLINK(mfi.mode))
2109                         reason = _("submodule");
2110                 output(o, 1, _("CONFLICT (%s): Merge conflict in %s"),
2111                                 reason, path);
2112                 if (rename_conflict_info && !df_conflict_remains)
2113                         if (update_stages(o, path, &one, &a, &b))
2114                                 return -1;
2115         }
2116
2117         if (df_conflict_remains) {
2118                 char *new_path;
2119                 if (o->call_depth) {
2120                         remove_file_from_cache(path);
2121                 } else {
2122                         if (!mfi.clean) {
2123                                 if (update_stages(o, path, &one, &a, &b))
2124                                         return -1;
2125                         } else {
2126                                 int file_from_stage2 = was_tracked(path);
2127                                 struct diff_filespec merged;
2128                                 oidcpy(&merged.oid, &mfi.oid);
2129                                 merged.mode = mfi.mode;
2130
2131                                 if (update_stages(o, path, NULL,
2132                                                   file_from_stage2 ? &merged : NULL,
2133                                                   file_from_stage2 ? NULL : &merged))
2134                                         return -1;
2135                         }
2136
2137                 }
2138                 new_path = unique_path(o, path, rename_conflict_info->branch1);
2139                 output(o, 1, _("Adding as %s instead"), new_path);
2140                 if (update_file(o, 0, &mfi.oid, mfi.mode, new_path)) {
2141                         free(new_path);
2142                         return -1;
2143                 }
2144                 free(new_path);
2145                 mfi.clean = 0;
2146         } else if (update_file(o, mfi.clean, &mfi.oid, mfi.mode, path))
2147                 return -1;
2148         return mfi.clean;
2149 }
2150
2151 /* Per entry merge function */
2152 static int process_entry(struct merge_options *o,
2153                          const char *path, struct stage_data *entry)
2154 {
2155         int clean_merge = 1;
2156         int normalize = o->renormalize;
2157         unsigned o_mode = entry->stages[1].mode;
2158         unsigned a_mode = entry->stages[2].mode;
2159         unsigned b_mode = entry->stages[3].mode;
2160         struct object_id *o_oid = stage_oid(&entry->stages[1].oid, o_mode);
2161         struct object_id *a_oid = stage_oid(&entry->stages[2].oid, a_mode);
2162         struct object_id *b_oid = stage_oid(&entry->stages[3].oid, b_mode);
2163
2164         entry->processed = 1;
2165         if (entry->rename_conflict_info) {
2166                 struct rename_conflict_info *conflict_info = entry->rename_conflict_info;
2167                 switch (conflict_info->rename_type) {
2168                 case RENAME_NORMAL:
2169                 case RENAME_ONE_FILE_TO_ONE:
2170                         clean_merge = merge_content(o, path,
2171                                                     o_oid, o_mode, a_oid, a_mode, b_oid, b_mode,
2172                                                     conflict_info);
2173                         break;
2174                 case RENAME_DELETE:
2175                         clean_merge = 0;
2176                         if (conflict_rename_delete(o,
2177                                                    conflict_info->pair1,
2178                                                    conflict_info->branch1,
2179                                                    conflict_info->branch2))
2180                                 clean_merge = -1;
2181                         break;
2182                 case RENAME_ONE_FILE_TO_TWO:
2183                         clean_merge = 0;
2184                         if (conflict_rename_rename_1to2(o, conflict_info))
2185                                 clean_merge = -1;
2186                         break;
2187                 case RENAME_TWO_FILES_TO_ONE:
2188                         clean_merge = 0;
2189                         if (conflict_rename_rename_2to1(o, conflict_info))
2190                                 clean_merge = -1;
2191                         break;
2192                 default:
2193                         entry->processed = 0;
2194                         break;
2195                 }
2196         } else if (o_oid && (!a_oid || !b_oid)) {
2197                 /* Case A: Deleted in one */
2198                 if ((!a_oid && !b_oid) ||
2199                     (!b_oid && blob_unchanged(o, o_oid, o_mode, a_oid, a_mode, normalize, path)) ||
2200                     (!a_oid && blob_unchanged(o, o_oid, o_mode, b_oid, b_mode, normalize, path))) {
2201                         /* Deleted in both or deleted in one and
2202                          * unchanged in the other */
2203                         if (a_oid)
2204                                 output(o, 2, _("Removing %s"), path);
2205                         /* do not touch working file if it did not exist */
2206                         remove_file(o, 1, path, !a_oid);
2207                 } else {
2208                         /* Modify/delete; deleted side may have put a directory in the way */
2209                         clean_merge = 0;
2210                         if (handle_modify_delete(o, path, o_oid, o_mode,
2211                                                  a_oid, a_mode, b_oid, b_mode))
2212                                 clean_merge = -1;
2213                 }
2214         } else if ((!o_oid && a_oid && !b_oid) ||
2215                    (!o_oid && !a_oid && b_oid)) {
2216                 /* Case B: Added in one. */
2217                 /* [nothing|directory] -> ([nothing|directory], file) */
2218
2219                 const char *add_branch;
2220                 const char *other_branch;
2221                 unsigned mode;
2222                 const struct object_id *oid;
2223                 const char *conf;
2224
2225                 if (a_oid) {
2226                         add_branch = o->branch1;
2227                         other_branch = o->branch2;
2228                         mode = a_mode;
2229                         oid = a_oid;
2230                         conf = _("file/directory");
2231                 } else {
2232                         add_branch = o->branch2;
2233                         other_branch = o->branch1;
2234                         mode = b_mode;
2235                         oid = b_oid;
2236                         conf = _("directory/file");
2237                 }
2238                 if (dir_in_way(path,
2239                                !o->call_depth && !S_ISGITLINK(a_mode),
2240                                0)) {
2241                         char *new_path = unique_path(o, path, add_branch);
2242                         clean_merge = 0;
2243                         output(o, 1, _("CONFLICT (%s): There is a directory with name %s in %s. "
2244                                "Adding %s as %s"),
2245                                conf, path, other_branch, path, new_path);
2246                         if (update_file(o, 0, oid, mode, new_path))
2247                                 clean_merge = -1;
2248                         else if (o->call_depth)
2249                                 remove_file_from_cache(path);
2250                         free(new_path);
2251                 } else {
2252                         output(o, 2, _("Adding %s"), path);
2253                         /* do not overwrite file if already present */
2254                         if (update_file_flags(o, oid, mode, path, 1, !a_oid))
2255                                 clean_merge = -1;
2256                 }
2257         } else if (a_oid && b_oid) {
2258                 /* Case C: Added in both (check for same permissions) and */
2259                 /* case D: Modified in both, but differently. */
2260                 clean_merge = merge_content(o, path,
2261                                             o_oid, o_mode, a_oid, a_mode, b_oid, b_mode,
2262                                             NULL);
2263         } else if (!o_oid && !a_oid && !b_oid) {
2264                 /*
2265                  * this entry was deleted altogether. a_mode == 0 means
2266                  * we had that path and want to actively remove it.
2267                  */
2268                 remove_file(o, 1, path, !a_mode);
2269         } else
2270                 die("BUG: fatal merge failure, shouldn't happen.");
2271
2272         return clean_merge;
2273 }
2274
2275 int merge_trees(struct merge_options *o,
2276                 struct tree *head,
2277                 struct tree *merge,
2278                 struct tree *common,
2279                 struct tree **result)
2280 {
2281         int code, clean;
2282
2283         if (o->subtree_shift) {
2284                 merge = shift_tree_object(head, merge, o->subtree_shift);
2285                 common = shift_tree_object(head, common, o->subtree_shift);
2286         }
2287
2288         if (oid_eq(&common->object.oid, &merge->object.oid)) {
2289                 struct strbuf sb = STRBUF_INIT;
2290
2291                 if (!o->call_depth && index_has_changes(&sb)) {
2292                         err(o, _("Dirty index: cannot merge (dirty: %s)"),
2293                             sb.buf);
2294                         return 0;
2295                 }
2296                 output(o, 0, _("Already up to date!"));
2297                 *result = head;
2298                 return 1;
2299         }
2300
2301         code = git_merge_trees(o->call_depth, common, head, merge);
2302
2303         if (code != 0) {
2304                 if (show(o, 4) || o->call_depth)
2305                         err(o, _("merging of trees %s and %s failed"),
2306                             oid_to_hex(&head->object.oid),
2307                             oid_to_hex(&merge->object.oid));
2308                 return -1;
2309         }
2310
2311         if (unmerged_cache()) {
2312                 struct string_list *entries;
2313                 struct rename_info re_info;
2314                 int i;
2315                 /*
2316                  * Only need the hashmap while processing entries, so
2317                  * initialize it here and free it when we are done running
2318                  * through the entries. Keeping it in the merge_options as
2319                  * opposed to decaring a local hashmap is for convenience
2320                  * so that we don't have to pass it to around.
2321                  */
2322                 hashmap_init(&o->current_file_dir_set, path_hashmap_cmp, NULL, 512);
2323                 get_files_dirs(o, head);
2324                 get_files_dirs(o, merge);
2325
2326                 entries = get_unmerged();
2327                 clean = handle_renames(o, common, head, merge, entries,
2328                                        &re_info);
2329                 record_df_conflict_files(o, entries);
2330                 if (clean < 0)
2331                         goto cleanup;
2332                 for (i = entries->nr-1; 0 <= i; i--) {
2333                         const char *path = entries->items[i].string;
2334                         struct stage_data *e = entries->items[i].util;
2335                         if (!e->processed) {
2336                                 int ret = process_entry(o, path, e);
2337                                 if (!ret)
2338                                         clean = 0;
2339                                 else if (ret < 0) {
2340                                         clean = ret;
2341                                         goto cleanup;
2342                                 }
2343                         }
2344                 }
2345                 for (i = 0; i < entries->nr; i++) {
2346                         struct stage_data *e = entries->items[i].util;
2347                         if (!e->processed)
2348                                 die("BUG: unprocessed path??? %s",
2349                                     entries->items[i].string);
2350                 }
2351
2352 cleanup:
2353                 final_cleanup_renames(&re_info);
2354
2355                 string_list_clear(entries, 1);
2356                 free(entries);
2357
2358                 hashmap_free(&o->current_file_dir_set, 1);
2359
2360                 if (clean < 0)
2361                         return clean;
2362         }
2363         else
2364                 clean = 1;
2365
2366         if (o->call_depth && !(*result = write_tree_from_memory(o)))
2367                 return -1;
2368
2369         return clean;
2370 }
2371
2372 static struct commit_list *reverse_commit_list(struct commit_list *list)
2373 {
2374         struct commit_list *next = NULL, *current, *backup;
2375         for (current = list; current; current = backup) {
2376                 backup = current->next;
2377                 current->next = next;
2378                 next = current;
2379         }
2380         return next;
2381 }
2382
2383 /*
2384  * Merge the commits h1 and h2, return the resulting virtual
2385  * commit object and a flag indicating the cleanness of the merge.
2386  */
2387 int merge_recursive(struct merge_options *o,
2388                     struct commit *h1,
2389                     struct commit *h2,
2390                     struct commit_list *ca,
2391                     struct commit **result)
2392 {
2393         struct commit_list *iter;
2394         struct commit *merged_common_ancestors;
2395         struct tree *mrtree;
2396         int clean;
2397
2398         if (show(o, 4)) {
2399                 output(o, 4, _("Merging:"));
2400                 output_commit_title(o, h1);
2401                 output_commit_title(o, h2);
2402         }
2403
2404         if (!ca) {
2405                 ca = get_merge_bases(h1, h2);
2406                 ca = reverse_commit_list(ca);
2407         }
2408
2409         if (show(o, 5)) {
2410                 unsigned cnt = commit_list_count(ca);
2411
2412                 output(o, 5, Q_("found %u common ancestor:",
2413                                 "found %u common ancestors:", cnt), cnt);
2414                 for (iter = ca; iter; iter = iter->next)
2415                         output_commit_title(o, iter->item);
2416         }
2417
2418         merged_common_ancestors = pop_commit(&ca);
2419         if (merged_common_ancestors == NULL) {
2420                 /* if there is no common ancestor, use an empty tree */
2421                 struct tree *tree;
2422
2423                 tree = lookup_tree(the_hash_algo->empty_tree);
2424                 merged_common_ancestors = make_virtual_commit(tree, "ancestor");
2425         }
2426
2427         for (iter = ca; iter; iter = iter->next) {
2428                 const char *saved_b1, *saved_b2;
2429                 o->call_depth++;
2430                 /*
2431                  * When the merge fails, the result contains files
2432                  * with conflict markers. The cleanness flag is
2433                  * ignored (unless indicating an error), it was never
2434                  * actually used, as result of merge_trees has always
2435                  * overwritten it: the committed "conflicts" were
2436                  * already resolved.
2437                  */
2438                 discard_cache();
2439                 saved_b1 = o->branch1;
2440                 saved_b2 = o->branch2;
2441                 o->branch1 = "Temporary merge branch 1";
2442                 o->branch2 = "Temporary merge branch 2";
2443                 if (merge_recursive(o, merged_common_ancestors, iter->item,
2444                                     NULL, &merged_common_ancestors) < 0)
2445                         return -1;
2446                 o->branch1 = saved_b1;
2447                 o->branch2 = saved_b2;
2448                 o->call_depth--;
2449
2450                 if (!merged_common_ancestors)
2451                         return err(o, _("merge returned no commit"));
2452         }
2453
2454         discard_cache();
2455         if (!o->call_depth)
2456                 read_cache();
2457
2458         o->ancestor = "merged common ancestors";
2459         clean = merge_trees(o, h1->tree, h2->tree, merged_common_ancestors->tree,
2460                             &mrtree);
2461         if (clean < 0) {
2462                 flush_output(o);
2463                 return clean;
2464         }
2465
2466         if (o->call_depth) {
2467                 *result = make_virtual_commit(mrtree, "merged tree");
2468                 commit_list_insert(h1, &(*result)->parents);
2469                 commit_list_insert(h2, &(*result)->parents->next);
2470         }
2471         flush_output(o);
2472         if (!o->call_depth && o->buffer_output < 2)
2473                 strbuf_release(&o->obuf);
2474         if (show(o, 2))
2475                 diff_warn_rename_limit("merge.renamelimit",
2476                                        o->needed_rename_limit, 0);
2477         return clean;
2478 }
2479
2480 static struct commit *get_ref(const struct object_id *oid, const char *name)
2481 {
2482         struct object *object;
2483
2484         object = deref_tag(parse_object(oid), name, strlen(name));
2485         if (!object)
2486                 return NULL;
2487         if (object->type == OBJ_TREE)
2488                 return make_virtual_commit((struct tree*)object, name);
2489         if (object->type != OBJ_COMMIT)
2490                 return NULL;
2491         if (parse_commit((struct commit *)object))
2492                 return NULL;
2493         return (struct commit *)object;
2494 }
2495
2496 int merge_recursive_generic(struct merge_options *o,
2497                             const struct object_id *head,
2498                             const struct object_id *merge,
2499                             int num_base_list,
2500                             const struct object_id **base_list,
2501                             struct commit **result)
2502 {
2503         int clean;
2504         struct lock_file lock = LOCK_INIT;
2505         struct commit *head_commit = get_ref(head, o->branch1);
2506         struct commit *next_commit = get_ref(merge, o->branch2);
2507         struct commit_list *ca = NULL;
2508
2509         if (base_list) {
2510                 int i;
2511                 for (i = 0; i < num_base_list; ++i) {
2512                         struct commit *base;
2513                         if (!(base = get_ref(base_list[i], oid_to_hex(base_list[i]))))
2514                                 return err(o, _("Could not parse object '%s'"),
2515                                         oid_to_hex(base_list[i]));
2516                         commit_list_insert(base, &ca);
2517                 }
2518         }
2519
2520         hold_locked_index(&lock, LOCK_DIE_ON_ERROR);
2521         clean = merge_recursive(o, head_commit, next_commit, ca,
2522                         result);
2523         if (clean < 0) {
2524                 rollback_lock_file(&lock);
2525                 return clean;
2526         }
2527
2528         if (write_locked_index(&the_index, &lock,
2529                                COMMIT_LOCK | SKIP_IF_UNCHANGED))
2530                 return err(o, _("Unable to write index."));
2531
2532         return clean ? 0 : 1;
2533 }
2534
2535 static void merge_recursive_config(struct merge_options *o)
2536 {
2537         git_config_get_int("merge.verbosity", &o->verbosity);
2538         git_config_get_int("diff.renamelimit", &o->diff_rename_limit);
2539         git_config_get_int("merge.renamelimit", &o->merge_rename_limit);
2540         git_config(git_xmerge_config, NULL);
2541 }
2542
2543 void init_merge_options(struct merge_options *o)
2544 {
2545         const char *merge_verbosity;
2546         memset(o, 0, sizeof(struct merge_options));
2547         o->verbosity = 2;
2548         o->buffer_output = 1;
2549         o->diff_rename_limit = -1;
2550         o->merge_rename_limit = -1;
2551         o->renormalize = 0;
2552         o->detect_rename = 1;
2553         merge_recursive_config(o);
2554         merge_verbosity = getenv("GIT_MERGE_VERBOSITY");
2555         if (merge_verbosity)
2556                 o->verbosity = strtol(merge_verbosity, NULL, 10);
2557         if (o->verbosity >= 5)
2558                 o->buffer_output = 0;
2559         strbuf_init(&o->obuf, 0);
2560         string_list_init(&o->df_conflict_file_set, 1);
2561 }
2562
2563 int parse_merge_opt(struct merge_options *o, const char *s)
2564 {
2565         const char *arg;
2566
2567         if (!s || !*s)
2568                 return -1;
2569         if (!strcmp(s, "ours"))
2570                 o->recursive_variant = MERGE_RECURSIVE_OURS;
2571         else if (!strcmp(s, "theirs"))
2572                 o->recursive_variant = MERGE_RECURSIVE_THEIRS;
2573         else if (!strcmp(s, "subtree"))
2574                 o->subtree_shift = "";
2575         else if (skip_prefix(s, "subtree=", &arg))
2576                 o->subtree_shift = arg;
2577         else if (!strcmp(s, "patience"))
2578                 o->xdl_opts = DIFF_WITH_ALG(o, PATIENCE_DIFF);
2579         else if (!strcmp(s, "histogram"))
2580                 o->xdl_opts = DIFF_WITH_ALG(o, HISTOGRAM_DIFF);
2581         else if (skip_prefix(s, "diff-algorithm=", &arg)) {
2582                 long value = parse_algorithm_value(arg);
2583                 if (value < 0)
2584                         return -1;
2585                 /* clear out previous settings */
2586                 DIFF_XDL_CLR(o, NEED_MINIMAL);
2587                 o->xdl_opts &= ~XDF_DIFF_ALGORITHM_MASK;
2588                 o->xdl_opts |= value;
2589         }
2590         else if (!strcmp(s, "ignore-space-change"))
2591                 DIFF_XDL_SET(o, IGNORE_WHITESPACE_CHANGE);
2592         else if (!strcmp(s, "ignore-all-space"))
2593                 DIFF_XDL_SET(o, IGNORE_WHITESPACE);
2594         else if (!strcmp(s, "ignore-space-at-eol"))
2595                 DIFF_XDL_SET(o, IGNORE_WHITESPACE_AT_EOL);
2596         else if (!strcmp(s, "ignore-cr-at-eol"))
2597                 DIFF_XDL_SET(o, IGNORE_CR_AT_EOL);
2598         else if (!strcmp(s, "renormalize"))
2599                 o->renormalize = 1;
2600         else if (!strcmp(s, "no-renormalize"))
2601                 o->renormalize = 0;
2602         else if (!strcmp(s, "no-renames"))
2603                 o->detect_rename = 0;
2604         else if (!strcmp(s, "find-renames")) {
2605                 o->detect_rename = 1;
2606                 o->rename_score = 0;
2607         }
2608         else if (skip_prefix(s, "find-renames=", &arg) ||
2609                  skip_prefix(s, "rename-threshold=", &arg)) {
2610                 if ((o->rename_score = parse_rename_score(&arg)) == -1 || *arg != 0)
2611                         return -1;
2612                 o->detect_rename = 1;
2613         }
2614         else
2615                 return -1;
2616         return 0;
2617 }