merge-recursive: when comparing files, don't include trees
[git] / merge-recursive.c
1 /*
2  * Recursive Merge algorithm stolen from git-merge-recursive.py by
3  * Fredrik Kuivinen.
4  * The thieves were Alex Riesen and Johannes Schindelin, in June/July 2006
5  */
6 #include "cache.h"
7 #include "config.h"
8 #include "advice.h"
9 #include "lockfile.h"
10 #include "cache-tree.h"
11 #include "commit.h"
12 #include "blob.h"
13 #include "builtin.h"
14 #include "tree-walk.h"
15 #include "diff.h"
16 #include "diffcore.h"
17 #include "tag.h"
18 #include "unpack-trees.h"
19 #include "string-list.h"
20 #include "xdiff-interface.h"
21 #include "ll-merge.h"
22 #include "attr.h"
23 #include "merge-recursive.h"
24 #include "dir.h"
25 #include "submodule.h"
26
27 struct path_hashmap_entry {
28         struct hashmap_entry e;
29         char path[FLEX_ARRAY];
30 };
31
32 static int path_hashmap_cmp(const void *cmp_data,
33                             const void *entry,
34                             const void *entry_or_key,
35                             const void *keydata)
36 {
37         const struct path_hashmap_entry *a = entry;
38         const struct path_hashmap_entry *b = entry_or_key;
39         const char *key = keydata;
40
41         if (ignore_case)
42                 return strcasecmp(a->path, key ? key : b->path);
43         else
44                 return strcmp(a->path, key ? key : b->path);
45 }
46
47 static unsigned int path_hash(const char *path)
48 {
49         return ignore_case ? strihash(path) : strhash(path);
50 }
51
52 static struct dir_rename_entry *dir_rename_find_entry(struct hashmap *hashmap,
53                                                       char *dir)
54 {
55         struct dir_rename_entry key;
56
57         if (dir == NULL)
58                 return NULL;
59         hashmap_entry_init(&key, strhash(dir));
60         key.dir = dir;
61         return hashmap_get(hashmap, &key, NULL);
62 }
63
64 static int dir_rename_cmp(const void *unused_cmp_data,
65                           const void *entry,
66                           const void *entry_or_key,
67                           const void *unused_keydata)
68 {
69         const struct dir_rename_entry *e1 = entry;
70         const struct dir_rename_entry *e2 = entry_or_key;
71
72         return strcmp(e1->dir, e2->dir);
73 }
74
75 static void dir_rename_init(struct hashmap *map)
76 {
77         hashmap_init(map, dir_rename_cmp, NULL, 0);
78 }
79
80 static void dir_rename_entry_init(struct dir_rename_entry *entry,
81                                   char *directory)
82 {
83         hashmap_entry_init(entry, strhash(directory));
84         entry->dir = directory;
85         entry->non_unique_new_dir = 0;
86         strbuf_init(&entry->new_dir, 0);
87         string_list_init(&entry->possible_new_dirs, 0);
88 }
89
90 static struct collision_entry *collision_find_entry(struct hashmap *hashmap,
91                                                     char *target_file)
92 {
93         struct collision_entry key;
94
95         hashmap_entry_init(&key, strhash(target_file));
96         key.target_file = target_file;
97         return hashmap_get(hashmap, &key, NULL);
98 }
99
100 static int collision_cmp(void *unused_cmp_data,
101                          const struct collision_entry *e1,
102                          const struct collision_entry *e2,
103                          const void *unused_keydata)
104 {
105         return strcmp(e1->target_file, e2->target_file);
106 }
107
108 static void collision_init(struct hashmap *map)
109 {
110         hashmap_init(map, (hashmap_cmp_fn) collision_cmp, NULL, 0);
111 }
112
113 static void flush_output(struct merge_options *o)
114 {
115         if (o->buffer_output < 2 && o->obuf.len) {
116                 fputs(o->obuf.buf, stdout);
117                 strbuf_reset(&o->obuf);
118         }
119 }
120
121 static int err(struct merge_options *o, const char *err, ...)
122 {
123         va_list params;
124
125         if (o->buffer_output < 2)
126                 flush_output(o);
127         else {
128                 strbuf_complete(&o->obuf, '\n');
129                 strbuf_addstr(&o->obuf, "error: ");
130         }
131         va_start(params, err);
132         strbuf_vaddf(&o->obuf, err, params);
133         va_end(params);
134         if (o->buffer_output > 1)
135                 strbuf_addch(&o->obuf, '\n');
136         else {
137                 error("%s", o->obuf.buf);
138                 strbuf_reset(&o->obuf);
139         }
140
141         return -1;
142 }
143
144 static struct tree *shift_tree_object(struct tree *one, struct tree *two,
145                                       const char *subtree_shift)
146 {
147         struct object_id shifted;
148
149         if (!*subtree_shift) {
150                 shift_tree(&one->object.oid, &two->object.oid, &shifted, 0);
151         } else {
152                 shift_tree_by(&one->object.oid, &two->object.oid, &shifted,
153                               subtree_shift);
154         }
155         if (!oidcmp(&two->object.oid, &shifted))
156                 return two;
157         return lookup_tree(&shifted);
158 }
159
160 static struct commit *make_virtual_commit(struct tree *tree, const char *comment)
161 {
162         struct commit *commit = alloc_commit_node();
163
164         set_merge_remote_desc(commit, comment, (struct object *)commit);
165         commit->tree = tree;
166         commit->object.parsed = 1;
167         return commit;
168 }
169
170 /*
171  * Since we use get_tree_entry(), which does not put the read object into
172  * the object pool, we cannot rely on a == b.
173  */
174 static int oid_eq(const struct object_id *a, const struct object_id *b)
175 {
176         if (!a && !b)
177                 return 2;
178         return a && b && oidcmp(a, b) == 0;
179 }
180
181 enum rename_type {
182         RENAME_NORMAL = 0,
183         RENAME_DELETE,
184         RENAME_ONE_FILE_TO_ONE,
185         RENAME_ONE_FILE_TO_TWO,
186         RENAME_TWO_FILES_TO_ONE
187 };
188
189 struct rename_conflict_info {
190         enum rename_type rename_type;
191         struct diff_filepair *pair1;
192         struct diff_filepair *pair2;
193         const char *branch1;
194         const char *branch2;
195         struct stage_data *dst_entry1;
196         struct stage_data *dst_entry2;
197         struct diff_filespec ren1_other;
198         struct diff_filespec ren2_other;
199 };
200
201 /*
202  * Since we want to write the index eventually, we cannot reuse the index
203  * for these (temporary) data.
204  */
205 struct stage_data {
206         struct {
207                 unsigned mode;
208                 struct object_id oid;
209         } stages[4];
210         struct rename_conflict_info *rename_conflict_info;
211         unsigned processed:1;
212 };
213
214 static inline void setup_rename_conflict_info(enum rename_type rename_type,
215                                               struct diff_filepair *pair1,
216                                               struct diff_filepair *pair2,
217                                               const char *branch1,
218                                               const char *branch2,
219                                               struct stage_data *dst_entry1,
220                                               struct stage_data *dst_entry2,
221                                               struct merge_options *o,
222                                               struct stage_data *src_entry1,
223                                               struct stage_data *src_entry2)
224 {
225         struct rename_conflict_info *ci = xcalloc(1, sizeof(struct rename_conflict_info));
226         ci->rename_type = rename_type;
227         ci->pair1 = pair1;
228         ci->branch1 = branch1;
229         ci->branch2 = branch2;
230
231         ci->dst_entry1 = dst_entry1;
232         dst_entry1->rename_conflict_info = ci;
233         dst_entry1->processed = 0;
234
235         assert(!pair2 == !dst_entry2);
236         if (dst_entry2) {
237                 ci->dst_entry2 = dst_entry2;
238                 ci->pair2 = pair2;
239                 dst_entry2->rename_conflict_info = ci;
240         }
241
242         if (rename_type == RENAME_TWO_FILES_TO_ONE) {
243                 /*
244                  * For each rename, there could have been
245                  * modifications on the side of history where that
246                  * file was not renamed.
247                  */
248                 int ostage1 = o->branch1 == branch1 ? 3 : 2;
249                 int ostage2 = ostage1 ^ 1;
250
251                 ci->ren1_other.path = pair1->one->path;
252                 oidcpy(&ci->ren1_other.oid, &src_entry1->stages[ostage1].oid);
253                 ci->ren1_other.mode = src_entry1->stages[ostage1].mode;
254
255                 ci->ren2_other.path = pair2->one->path;
256                 oidcpy(&ci->ren2_other.oid, &src_entry2->stages[ostage2].oid);
257                 ci->ren2_other.mode = src_entry2->stages[ostage2].mode;
258         }
259 }
260
261 static int show(struct merge_options *o, int v)
262 {
263         return (!o->call_depth && o->verbosity >= v) || o->verbosity >= 5;
264 }
265
266 __attribute__((format (printf, 3, 4)))
267 static void output(struct merge_options *o, int v, const char *fmt, ...)
268 {
269         va_list ap;
270
271         if (!show(o, v))
272                 return;
273
274         strbuf_addchars(&o->obuf, ' ', o->call_depth * 2);
275
276         va_start(ap, fmt);
277         strbuf_vaddf(&o->obuf, fmt, ap);
278         va_end(ap);
279
280         strbuf_addch(&o->obuf, '\n');
281         if (!o->buffer_output)
282                 flush_output(o);
283 }
284
285 static void output_commit_title(struct merge_options *o, struct commit *commit)
286 {
287         strbuf_addchars(&o->obuf, ' ', o->call_depth * 2);
288         if (commit->util)
289                 strbuf_addf(&o->obuf, "virtual %s\n",
290                         merge_remote_util(commit)->name);
291         else {
292                 strbuf_add_unique_abbrev(&o->obuf, &commit->object.oid,
293                                          DEFAULT_ABBREV);
294                 strbuf_addch(&o->obuf, ' ');
295                 if (parse_commit(commit) != 0)
296                         strbuf_addstr(&o->obuf, _("(bad commit)\n"));
297                 else {
298                         const char *title;
299                         const char *msg = get_commit_buffer(commit, NULL);
300                         int len = find_commit_subject(msg, &title);
301                         if (len)
302                                 strbuf_addf(&o->obuf, "%.*s\n", len, title);
303                         unuse_commit_buffer(commit, msg);
304                 }
305         }
306         flush_output(o);
307 }
308
309 static int add_cacheinfo(struct merge_options *o,
310                 unsigned int mode, const struct object_id *oid,
311                 const char *path, int stage, int refresh, int options)
312 {
313         struct cache_entry *ce;
314         int ret;
315
316         ce = make_cache_entry(mode, oid ? oid->hash : null_sha1, path, stage, 0);
317         if (!ce)
318                 return err(o, _("addinfo_cache failed for path '%s'"), path);
319
320         ret = add_cache_entry(ce, options);
321         if (refresh) {
322                 struct cache_entry *nce;
323
324                 nce = refresh_cache_entry(ce, CE_MATCH_REFRESH | CE_MATCH_IGNORE_MISSING);
325                 if (!nce)
326                         return err(o, _("addinfo_cache failed for path '%s'"), path);
327                 if (nce != ce)
328                         ret = add_cache_entry(nce, options);
329         }
330         return ret;
331 }
332
333 static void init_tree_desc_from_tree(struct tree_desc *desc, struct tree *tree)
334 {
335         parse_tree(tree);
336         init_tree_desc(desc, tree->buffer, tree->size);
337 }
338
339 static int git_merge_trees(int index_only,
340                            struct tree *common,
341                            struct tree *head,
342                            struct tree *merge)
343 {
344         int rc;
345         struct tree_desc t[3];
346         struct unpack_trees_options opts;
347
348         memset(&opts, 0, sizeof(opts));
349         if (index_only)
350                 opts.index_only = 1;
351         else
352                 opts.update = 1;
353         opts.merge = 1;
354         opts.head_idx = 2;
355         opts.fn = threeway_merge;
356         opts.src_index = &the_index;
357         opts.dst_index = &the_index;
358         setup_unpack_trees_porcelain(&opts, "merge");
359
360         init_tree_desc_from_tree(t+0, common);
361         init_tree_desc_from_tree(t+1, head);
362         init_tree_desc_from_tree(t+2, merge);
363
364         rc = unpack_trees(3, t, &opts);
365         cache_tree_free(&active_cache_tree);
366         return rc;
367 }
368
369 struct tree *write_tree_from_memory(struct merge_options *o)
370 {
371         struct tree *result = NULL;
372
373         if (unmerged_cache()) {
374                 int i;
375                 fprintf(stderr, "BUG: There are unmerged index entries:\n");
376                 for (i = 0; i < active_nr; i++) {
377                         const struct cache_entry *ce = active_cache[i];
378                         if (ce_stage(ce))
379                                 fprintf(stderr, "BUG: %d %.*s\n", ce_stage(ce),
380                                         (int)ce_namelen(ce), ce->name);
381                 }
382                 die("BUG: unmerged index entries in merge-recursive.c");
383         }
384
385         if (!active_cache_tree)
386                 active_cache_tree = cache_tree();
387
388         if (!cache_tree_fully_valid(active_cache_tree) &&
389             cache_tree_update(&the_index, 0) < 0) {
390                 err(o, _("error building trees"));
391                 return NULL;
392         }
393
394         result = lookup_tree(&active_cache_tree->oid);
395
396         return result;
397 }
398
399 static int save_files_dirs(const struct object_id *oid,
400                 struct strbuf *base, const char *path,
401                 unsigned int mode, int stage, void *context)
402 {
403         struct path_hashmap_entry *entry;
404         int baselen = base->len;
405         struct merge_options *o = context;
406
407         strbuf_addstr(base, path);
408
409         FLEX_ALLOC_MEM(entry, path, base->buf, base->len);
410         hashmap_entry_init(entry, path_hash(entry->path));
411         hashmap_add(&o->current_file_dir_set, entry);
412
413         strbuf_setlen(base, baselen);
414         return (S_ISDIR(mode) ? READ_TREE_RECURSIVE : 0);
415 }
416
417 static void get_files_dirs(struct merge_options *o, struct tree *tree)
418 {
419         struct pathspec match_all;
420         memset(&match_all, 0, sizeof(match_all));
421         read_tree_recursive(tree, "", 0, 0, &match_all, save_files_dirs, o);
422 }
423
424 static int get_tree_entry_if_blob(const struct object_id *tree,
425                                   const char *path,
426                                   struct object_id *hashy,
427                                   unsigned int *mode_o)
428 {
429         int ret;
430
431         ret = get_tree_entry(tree, path, hashy, mode_o);
432         if (S_ISDIR(*mode_o)) {
433                 oidcpy(hashy, &null_oid);
434                 *mode_o = 0;
435         }
436         return ret;
437 }
438
439 /*
440  * Returns an index_entry instance which doesn't have to correspond to
441  * a real cache entry in Git's index.
442  */
443 static struct stage_data *insert_stage_data(const char *path,
444                 struct tree *o, struct tree *a, struct tree *b,
445                 struct string_list *entries)
446 {
447         struct string_list_item *item;
448         struct stage_data *e = xcalloc(1, sizeof(struct stage_data));
449         get_tree_entry_if_blob(&o->object.oid, path,
450                                &e->stages[1].oid, &e->stages[1].mode);
451         get_tree_entry_if_blob(&a->object.oid, path,
452                                &e->stages[2].oid, &e->stages[2].mode);
453         get_tree_entry_if_blob(&b->object.oid, path,
454                                &e->stages[3].oid, &e->stages[3].mode);
455         item = string_list_insert(entries, path);
456         item->util = e;
457         return e;
458 }
459
460 /*
461  * Create a dictionary mapping file names to stage_data objects. The
462  * dictionary contains one entry for every path with a non-zero stage entry.
463  */
464 static struct string_list *get_unmerged(void)
465 {
466         struct string_list *unmerged = xcalloc(1, sizeof(struct string_list));
467         int i;
468
469         unmerged->strdup_strings = 1;
470
471         for (i = 0; i < active_nr; i++) {
472                 struct string_list_item *item;
473                 struct stage_data *e;
474                 const struct cache_entry *ce = active_cache[i];
475                 if (!ce_stage(ce))
476                         continue;
477
478                 item = string_list_lookup(unmerged, ce->name);
479                 if (!item) {
480                         item = string_list_insert(unmerged, ce->name);
481                         item->util = xcalloc(1, sizeof(struct stage_data));
482                 }
483                 e = item->util;
484                 e->stages[ce_stage(ce)].mode = ce->ce_mode;
485                 oidcpy(&e->stages[ce_stage(ce)].oid, &ce->oid);
486         }
487
488         return unmerged;
489 }
490
491 static int string_list_df_name_compare(const char *one, const char *two)
492 {
493         int onelen = strlen(one);
494         int twolen = strlen(two);
495         /*
496          * Here we only care that entries for D/F conflicts are
497          * adjacent, in particular with the file of the D/F conflict
498          * appearing before files below the corresponding directory.
499          * The order of the rest of the list is irrelevant for us.
500          *
501          * To achieve this, we sort with df_name_compare and provide
502          * the mode S_IFDIR so that D/F conflicts will sort correctly.
503          * We use the mode S_IFDIR for everything else for simplicity,
504          * since in other cases any changes in their order due to
505          * sorting cause no problems for us.
506          */
507         int cmp = df_name_compare(one, onelen, S_IFDIR,
508                                   two, twolen, S_IFDIR);
509         /*
510          * Now that 'foo' and 'foo/bar' compare equal, we have to make sure
511          * that 'foo' comes before 'foo/bar'.
512          */
513         if (cmp)
514                 return cmp;
515         return onelen - twolen;
516 }
517
518 static void record_df_conflict_files(struct merge_options *o,
519                                      struct string_list *entries)
520 {
521         /* If there is a D/F conflict and the file for such a conflict
522          * currently exist in the working tree, we want to allow it to be
523          * removed to make room for the corresponding directory if needed.
524          * The files underneath the directories of such D/F conflicts will
525          * be processed before the corresponding file involved in the D/F
526          * conflict.  If the D/F directory ends up being removed by the
527          * merge, then we won't have to touch the D/F file.  If the D/F
528          * directory needs to be written to the working copy, then the D/F
529          * file will simply be removed (in make_room_for_path()) to make
530          * room for the necessary paths.  Note that if both the directory
531          * and the file need to be present, then the D/F file will be
532          * reinstated with a new unique name at the time it is processed.
533          */
534         struct string_list df_sorted_entries = STRING_LIST_INIT_NODUP;
535         const char *last_file = NULL;
536         int last_len = 0;
537         int i;
538
539         /*
540          * If we're merging merge-bases, we don't want to bother with
541          * any working directory changes.
542          */
543         if (o->call_depth)
544                 return;
545
546         /* Ensure D/F conflicts are adjacent in the entries list. */
547         for (i = 0; i < entries->nr; i++) {
548                 struct string_list_item *next = &entries->items[i];
549                 string_list_append(&df_sorted_entries, next->string)->util =
550                                    next->util;
551         }
552         df_sorted_entries.cmp = string_list_df_name_compare;
553         string_list_sort(&df_sorted_entries);
554
555         string_list_clear(&o->df_conflict_file_set, 1);
556         for (i = 0; i < df_sorted_entries.nr; i++) {
557                 const char *path = df_sorted_entries.items[i].string;
558                 int len = strlen(path);
559                 struct stage_data *e = df_sorted_entries.items[i].util;
560
561                 /*
562                  * Check if last_file & path correspond to a D/F conflict;
563                  * i.e. whether path is last_file+'/'+<something>.
564                  * If so, record that it's okay to remove last_file to make
565                  * room for path and friends if needed.
566                  */
567                 if (last_file &&
568                     len > last_len &&
569                     memcmp(path, last_file, last_len) == 0 &&
570                     path[last_len] == '/') {
571                         string_list_insert(&o->df_conflict_file_set, last_file);
572                 }
573
574                 /*
575                  * Determine whether path could exist as a file in the
576                  * working directory as a possible D/F conflict.  This
577                  * will only occur when it exists in stage 2 as a
578                  * file.
579                  */
580                 if (S_ISREG(e->stages[2].mode) || S_ISLNK(e->stages[2].mode)) {
581                         last_file = path;
582                         last_len = len;
583                 } else {
584                         last_file = NULL;
585                 }
586         }
587         string_list_clear(&df_sorted_entries, 0);
588 }
589
590 struct rename {
591         struct diff_filepair *pair;
592         /*
593          * Purpose of src_entry and dst_entry:
594          *
595          * If 'before' is renamed to 'after' then src_entry will contain
596          * the versions of 'before' from the merge_base, HEAD, and MERGE in
597          * stages 1, 2, and 3; dst_entry will contain the respective
598          * versions of 'after' in corresponding locations.  Thus, we have a
599          * total of six modes and oids, though some will be null.  (Stage 0
600          * is ignored; we're interested in handling conflicts.)
601          *
602          * Since we don't turn on break-rewrites by default, neither
603          * src_entry nor dst_entry can have all three of their stages have
604          * non-null oids, meaning at most four of the six will be non-null.
605          * Also, since this is a rename, both src_entry and dst_entry will
606          * have at least one non-null oid, meaning at least two will be
607          * non-null.  Of the six oids, a typical rename will have three be
608          * non-null.  Only two implies a rename/delete, and four implies a
609          * rename/add.
610          */
611         struct stage_data *src_entry;
612         struct stage_data *dst_entry;
613         unsigned processed:1;
614 };
615
616 static int update_stages(struct merge_options *opt, const char *path,
617                          const struct diff_filespec *o,
618                          const struct diff_filespec *a,
619                          const struct diff_filespec *b)
620 {
621
622         /*
623          * NOTE: It is usually a bad idea to call update_stages on a path
624          * before calling update_file on that same path, since it can
625          * sometimes lead to spurious "refusing to lose untracked file..."
626          * messages from update_file (via make_room_for path via
627          * would_lose_untracked).  Instead, reverse the order of the calls
628          * (executing update_file first and then update_stages).
629          */
630         int clear = 1;
631         int options = ADD_CACHE_OK_TO_ADD | ADD_CACHE_SKIP_DFCHECK;
632         if (clear)
633                 if (remove_file_from_cache(path))
634                         return -1;
635         if (o)
636                 if (add_cacheinfo(opt, o->mode, &o->oid, path, 1, 0, options))
637                         return -1;
638         if (a)
639                 if (add_cacheinfo(opt, a->mode, &a->oid, path, 2, 0, options))
640                         return -1;
641         if (b)
642                 if (add_cacheinfo(opt, b->mode, &b->oid, path, 3, 0, options))
643                         return -1;
644         return 0;
645 }
646
647 static void update_entry(struct stage_data *entry,
648                          struct diff_filespec *o,
649                          struct diff_filespec *a,
650                          struct diff_filespec *b)
651 {
652         entry->processed = 0;
653         entry->stages[1].mode = o->mode;
654         entry->stages[2].mode = a->mode;
655         entry->stages[3].mode = b->mode;
656         oidcpy(&entry->stages[1].oid, &o->oid);
657         oidcpy(&entry->stages[2].oid, &a->oid);
658         oidcpy(&entry->stages[3].oid, &b->oid);
659 }
660
661 static int remove_file(struct merge_options *o, int clean,
662                        const char *path, int no_wd)
663 {
664         int update_cache = o->call_depth || clean;
665         int update_working_directory = !o->call_depth && !no_wd;
666
667         if (update_cache) {
668                 if (remove_file_from_cache(path))
669                         return -1;
670         }
671         if (update_working_directory) {
672                 if (ignore_case) {
673                         struct cache_entry *ce;
674                         ce = cache_file_exists(path, strlen(path), ignore_case);
675                         if (ce && ce_stage(ce) == 0 && strcmp(path, ce->name))
676                                 return 0;
677                 }
678                 if (remove_path(path))
679                         return -1;
680         }
681         return 0;
682 }
683
684 /* add a string to a strbuf, but converting "/" to "_" */
685 static void add_flattened_path(struct strbuf *out, const char *s)
686 {
687         size_t i = out->len;
688         strbuf_addstr(out, s);
689         for (; i < out->len; i++)
690                 if (out->buf[i] == '/')
691                         out->buf[i] = '_';
692 }
693
694 static char *unique_path(struct merge_options *o, const char *path, const char *branch)
695 {
696         struct path_hashmap_entry *entry;
697         struct strbuf newpath = STRBUF_INIT;
698         int suffix = 0;
699         size_t base_len;
700
701         strbuf_addf(&newpath, "%s~", path);
702         add_flattened_path(&newpath, branch);
703
704         base_len = newpath.len;
705         while (hashmap_get_from_hash(&o->current_file_dir_set,
706                                      path_hash(newpath.buf), newpath.buf) ||
707                (!o->call_depth && file_exists(newpath.buf))) {
708                 strbuf_setlen(&newpath, base_len);
709                 strbuf_addf(&newpath, "_%d", suffix++);
710         }
711
712         FLEX_ALLOC_MEM(entry, path, newpath.buf, newpath.len);
713         hashmap_entry_init(entry, path_hash(entry->path));
714         hashmap_add(&o->current_file_dir_set, entry);
715         return strbuf_detach(&newpath, NULL);
716 }
717
718 /**
719  * Check whether a directory in the index is in the way of an incoming
720  * file.  Return 1 if so.  If check_working_copy is non-zero, also
721  * check the working directory.  If empty_ok is non-zero, also return
722  * 0 in the case where the working-tree dir exists but is empty.
723  */
724 static int dir_in_way(const char *path, int check_working_copy, int empty_ok)
725 {
726         int pos;
727         struct strbuf dirpath = STRBUF_INIT;
728         struct stat st;
729
730         strbuf_addstr(&dirpath, path);
731         strbuf_addch(&dirpath, '/');
732
733         pos = cache_name_pos(dirpath.buf, dirpath.len);
734
735         if (pos < 0)
736                 pos = -1 - pos;
737         if (pos < active_nr &&
738             !strncmp(dirpath.buf, active_cache[pos]->name, dirpath.len)) {
739                 strbuf_release(&dirpath);
740                 return 1;
741         }
742
743         strbuf_release(&dirpath);
744         return check_working_copy && !lstat(path, &st) && S_ISDIR(st.st_mode) &&
745                 !(empty_ok && is_empty_dir(path));
746 }
747
748 static int was_tracked(const char *path)
749 {
750         int pos = cache_name_pos(path, strlen(path));
751
752         if (0 <= pos)
753                 /* we have been tracking this path */
754                 return 1;
755
756         /*
757          * Look for an unmerged entry for the path,
758          * specifically stage #2, which would indicate
759          * that "our" side before the merge started
760          * had the path tracked (and resulted in a conflict).
761          */
762         for (pos = -1 - pos;
763              pos < active_nr && !strcmp(path, active_cache[pos]->name);
764              pos++)
765                 if (ce_stage(active_cache[pos]) == 2)
766                         return 1;
767         return 0;
768 }
769
770 static int would_lose_untracked(const char *path)
771 {
772         return !was_tracked(path) && file_exists(path);
773 }
774
775 static int make_room_for_path(struct merge_options *o, const char *path)
776 {
777         int status, i;
778         const char *msg = _("failed to create path '%s'%s");
779
780         /* Unlink any D/F conflict files that are in the way */
781         for (i = 0; i < o->df_conflict_file_set.nr; i++) {
782                 const char *df_path = o->df_conflict_file_set.items[i].string;
783                 size_t pathlen = strlen(path);
784                 size_t df_pathlen = strlen(df_path);
785                 if (df_pathlen < pathlen &&
786                     path[df_pathlen] == '/' &&
787                     strncmp(path, df_path, df_pathlen) == 0) {
788                         output(o, 3,
789                                _("Removing %s to make room for subdirectory\n"),
790                                df_path);
791                         unlink(df_path);
792                         unsorted_string_list_delete_item(&o->df_conflict_file_set,
793                                                          i, 0);
794                         break;
795                 }
796         }
797
798         /* Make sure leading directories are created */
799         status = safe_create_leading_directories_const(path);
800         if (status) {
801                 if (status == SCLD_EXISTS)
802                         /* something else exists */
803                         return err(o, msg, path, _(": perhaps a D/F conflict?"));
804                 return err(o, msg, path, "");
805         }
806
807         /*
808          * Do not unlink a file in the work tree if we are not
809          * tracking it.
810          */
811         if (would_lose_untracked(path))
812                 return err(o, _("refusing to lose untracked file at '%s'"),
813                              path);
814
815         /* Successful unlink is good.. */
816         if (!unlink(path))
817                 return 0;
818         /* .. and so is no existing file */
819         if (errno == ENOENT)
820                 return 0;
821         /* .. but not some other error (who really cares what?) */
822         return err(o, msg, path, _(": perhaps a D/F conflict?"));
823 }
824
825 static int update_file_flags(struct merge_options *o,
826                              const struct object_id *oid,
827                              unsigned mode,
828                              const char *path,
829                              int update_cache,
830                              int update_wd)
831 {
832         int ret = 0;
833
834         if (o->call_depth)
835                 update_wd = 0;
836
837         if (update_wd) {
838                 enum object_type type;
839                 void *buf;
840                 unsigned long size;
841
842                 if (S_ISGITLINK(mode)) {
843                         /*
844                          * We may later decide to recursively descend into
845                          * the submodule directory and update its index
846                          * and/or work tree, but we do not do that now.
847                          */
848                         update_wd = 0;
849                         goto update_index;
850                 }
851
852                 buf = read_object_file(oid, &type, &size);
853                 if (!buf)
854                         return err(o, _("cannot read object %s '%s'"), oid_to_hex(oid), path);
855                 if (type != OBJ_BLOB) {
856                         ret = err(o, _("blob expected for %s '%s'"), oid_to_hex(oid), path);
857                         goto free_buf;
858                 }
859                 if (S_ISREG(mode)) {
860                         struct strbuf strbuf = STRBUF_INIT;
861                         if (convert_to_working_tree(path, buf, size, &strbuf)) {
862                                 free(buf);
863                                 size = strbuf.len;
864                                 buf = strbuf_detach(&strbuf, NULL);
865                         }
866                 }
867
868                 if (make_room_for_path(o, path) < 0) {
869                         update_wd = 0;
870                         goto free_buf;
871                 }
872                 if (S_ISREG(mode) || (!has_symlinks && S_ISLNK(mode))) {
873                         int fd;
874                         if (mode & 0100)
875                                 mode = 0777;
876                         else
877                                 mode = 0666;
878                         fd = open(path, O_WRONLY | O_TRUNC | O_CREAT, mode);
879                         if (fd < 0) {
880                                 ret = err(o, _("failed to open '%s': %s"),
881                                           path, strerror(errno));
882                                 goto free_buf;
883                         }
884                         write_in_full(fd, buf, size);
885                         close(fd);
886                 } else if (S_ISLNK(mode)) {
887                         char *lnk = xmemdupz(buf, size);
888                         safe_create_leading_directories_const(path);
889                         unlink(path);
890                         if (symlink(lnk, path))
891                                 ret = err(o, _("failed to symlink '%s': %s"),
892                                         path, strerror(errno));
893                         free(lnk);
894                 } else
895                         ret = err(o,
896                                   _("do not know what to do with %06o %s '%s'"),
897                                   mode, oid_to_hex(oid), path);
898  free_buf:
899                 free(buf);
900         }
901  update_index:
902         if (!ret && update_cache)
903                 add_cacheinfo(o, mode, oid, path, 0, update_wd, ADD_CACHE_OK_TO_ADD);
904         return ret;
905 }
906
907 static int update_file(struct merge_options *o,
908                        int clean,
909                        const struct object_id *oid,
910                        unsigned mode,
911                        const char *path)
912 {
913         return update_file_flags(o, oid, mode, path, o->call_depth || clean, !o->call_depth);
914 }
915
916 /* Low level file merging, update and removal */
917
918 struct merge_file_info {
919         struct object_id oid;
920         unsigned mode;
921         unsigned clean:1,
922                  merge:1;
923 };
924
925 static int merge_3way(struct merge_options *o,
926                       mmbuffer_t *result_buf,
927                       const struct diff_filespec *one,
928                       const struct diff_filespec *a,
929                       const struct diff_filespec *b,
930                       const char *branch1,
931                       const char *branch2)
932 {
933         mmfile_t orig, src1, src2;
934         struct ll_merge_options ll_opts = {0};
935         char *base_name, *name1, *name2;
936         int merge_status;
937
938         ll_opts.renormalize = o->renormalize;
939         ll_opts.xdl_opts = o->xdl_opts;
940
941         if (o->call_depth) {
942                 ll_opts.virtual_ancestor = 1;
943                 ll_opts.variant = 0;
944         } else {
945                 switch (o->recursive_variant) {
946                 case MERGE_RECURSIVE_OURS:
947                         ll_opts.variant = XDL_MERGE_FAVOR_OURS;
948                         break;
949                 case MERGE_RECURSIVE_THEIRS:
950                         ll_opts.variant = XDL_MERGE_FAVOR_THEIRS;
951                         break;
952                 default:
953                         ll_opts.variant = 0;
954                         break;
955                 }
956         }
957
958         if (strcmp(a->path, b->path) ||
959             (o->ancestor != NULL && strcmp(a->path, one->path) != 0)) {
960                 base_name = o->ancestor == NULL ? NULL :
961                         mkpathdup("%s:%s", o->ancestor, one->path);
962                 name1 = mkpathdup("%s:%s", branch1, a->path);
963                 name2 = mkpathdup("%s:%s", branch2, b->path);
964         } else {
965                 base_name = o->ancestor == NULL ? NULL :
966                         mkpathdup("%s", o->ancestor);
967                 name1 = mkpathdup("%s", branch1);
968                 name2 = mkpathdup("%s", branch2);
969         }
970
971         read_mmblob(&orig, &one->oid);
972         read_mmblob(&src1, &a->oid);
973         read_mmblob(&src2, &b->oid);
974
975         merge_status = ll_merge(result_buf, a->path, &orig, base_name,
976                                 &src1, name1, &src2, name2, &ll_opts);
977
978         free(base_name);
979         free(name1);
980         free(name2);
981         free(orig.ptr);
982         free(src1.ptr);
983         free(src2.ptr);
984         return merge_status;
985 }
986
987 static int merge_file_1(struct merge_options *o,
988                                            const struct diff_filespec *one,
989                                            const struct diff_filespec *a,
990                                            const struct diff_filespec *b,
991                                            const char *branch1,
992                                            const char *branch2,
993                                            struct merge_file_info *result)
994 {
995         result->merge = 0;
996         result->clean = 1;
997
998         if ((S_IFMT & a->mode) != (S_IFMT & b->mode)) {
999                 result->clean = 0;
1000                 if (S_ISREG(a->mode)) {
1001                         result->mode = a->mode;
1002                         oidcpy(&result->oid, &a->oid);
1003                 } else {
1004                         result->mode = b->mode;
1005                         oidcpy(&result->oid, &b->oid);
1006                 }
1007         } else {
1008                 if (!oid_eq(&a->oid, &one->oid) && !oid_eq(&b->oid, &one->oid))
1009                         result->merge = 1;
1010
1011                 /*
1012                  * Merge modes
1013                  */
1014                 if (a->mode == b->mode || a->mode == one->mode)
1015                         result->mode = b->mode;
1016                 else {
1017                         result->mode = a->mode;
1018                         if (b->mode != one->mode) {
1019                                 result->clean = 0;
1020                                 result->merge = 1;
1021                         }
1022                 }
1023
1024                 if (oid_eq(&a->oid, &b->oid) || oid_eq(&a->oid, &one->oid))
1025                         oidcpy(&result->oid, &b->oid);
1026                 else if (oid_eq(&b->oid, &one->oid))
1027                         oidcpy(&result->oid, &a->oid);
1028                 else if (S_ISREG(a->mode)) {
1029                         mmbuffer_t result_buf;
1030                         int ret = 0, merge_status;
1031
1032                         merge_status = merge_3way(o, &result_buf, one, a, b,
1033                                                   branch1, branch2);
1034
1035                         if ((merge_status < 0) || !result_buf.ptr)
1036                                 ret = err(o, _("Failed to execute internal merge"));
1037
1038                         if (!ret &&
1039                             write_object_file(result_buf.ptr, result_buf.size,
1040                                               blob_type, &result->oid))
1041                                 ret = err(o, _("Unable to add %s to database"),
1042                                           a->path);
1043
1044                         free(result_buf.ptr);
1045                         if (ret)
1046                                 return ret;
1047                         result->clean = (merge_status == 0);
1048                 } else if (S_ISGITLINK(a->mode)) {
1049                         result->clean = merge_submodule(&result->oid,
1050                                                        one->path,
1051                                                        &one->oid,
1052                                                        &a->oid,
1053                                                        &b->oid,
1054                                                        !o->call_depth);
1055                 } else if (S_ISLNK(a->mode)) {
1056                         switch (o->recursive_variant) {
1057                         case MERGE_RECURSIVE_NORMAL:
1058                                 oidcpy(&result->oid, &a->oid);
1059                                 if (!oid_eq(&a->oid, &b->oid))
1060                                         result->clean = 0;
1061                                 break;
1062                         case MERGE_RECURSIVE_OURS:
1063                                 oidcpy(&result->oid, &a->oid);
1064                                 break;
1065                         case MERGE_RECURSIVE_THEIRS:
1066                                 oidcpy(&result->oid, &b->oid);
1067                                 break;
1068                         }
1069                 } else
1070                         die("BUG: unsupported object type in the tree");
1071         }
1072
1073         return 0;
1074 }
1075
1076 static int merge_file_special_markers(struct merge_options *o,
1077                            const struct diff_filespec *one,
1078                            const struct diff_filespec *a,
1079                            const struct diff_filespec *b,
1080                            const char *branch1,
1081                            const char *filename1,
1082                            const char *branch2,
1083                            const char *filename2,
1084                            struct merge_file_info *mfi)
1085 {
1086         char *side1 = NULL;
1087         char *side2 = NULL;
1088         int ret;
1089
1090         if (filename1)
1091                 side1 = xstrfmt("%s:%s", branch1, filename1);
1092         if (filename2)
1093                 side2 = xstrfmt("%s:%s", branch2, filename2);
1094
1095         ret = merge_file_1(o, one, a, b,
1096                            side1 ? side1 : branch1,
1097                            side2 ? side2 : branch2, mfi);
1098         free(side1);
1099         free(side2);
1100         return ret;
1101 }
1102
1103 static int merge_file_one(struct merge_options *o,
1104                                          const char *path,
1105                                          const struct object_id *o_oid, int o_mode,
1106                                          const struct object_id *a_oid, int a_mode,
1107                                          const struct object_id *b_oid, int b_mode,
1108                                          const char *branch1,
1109                                          const char *branch2,
1110                                          struct merge_file_info *mfi)
1111 {
1112         struct diff_filespec one, a, b;
1113
1114         one.path = a.path = b.path = (char *)path;
1115         oidcpy(&one.oid, o_oid);
1116         one.mode = o_mode;
1117         oidcpy(&a.oid, a_oid);
1118         a.mode = a_mode;
1119         oidcpy(&b.oid, b_oid);
1120         b.mode = b_mode;
1121         return merge_file_1(o, &one, &a, &b, branch1, branch2, mfi);
1122 }
1123
1124 static int handle_change_delete(struct merge_options *o,
1125                                  const char *path, const char *old_path,
1126                                  const struct object_id *o_oid, int o_mode,
1127                                  const struct object_id *changed_oid,
1128                                  int changed_mode,
1129                                  const char *change_branch,
1130                                  const char *delete_branch,
1131                                  const char *change, const char *change_past)
1132 {
1133         char *alt_path = NULL;
1134         const char *update_path = path;
1135         int ret = 0;
1136
1137         if (dir_in_way(path, !o->call_depth, 0)) {
1138                 update_path = alt_path = unique_path(o, path, change_branch);
1139         }
1140
1141         if (o->call_depth) {
1142                 /*
1143                  * We cannot arbitrarily accept either a_sha or b_sha as
1144                  * correct; since there is no true "middle point" between
1145                  * them, simply reuse the base version for virtual merge base.
1146                  */
1147                 ret = remove_file_from_cache(path);
1148                 if (!ret)
1149                         ret = update_file(o, 0, o_oid, o_mode, update_path);
1150         } else {
1151                 if (!alt_path) {
1152                         if (!old_path) {
1153                                 output(o, 1, _("CONFLICT (%s/delete): %s deleted in %s "
1154                                        "and %s in %s. Version %s of %s left in tree."),
1155                                        change, path, delete_branch, change_past,
1156                                        change_branch, change_branch, path);
1157                         } else {
1158                                 output(o, 1, _("CONFLICT (%s/delete): %s deleted in %s "
1159                                        "and %s to %s in %s. Version %s of %s left in tree."),
1160                                        change, old_path, delete_branch, change_past, path,
1161                                        change_branch, change_branch, path);
1162                         }
1163                 } else {
1164                         if (!old_path) {
1165                                 output(o, 1, _("CONFLICT (%s/delete): %s deleted in %s "
1166                                        "and %s in %s. Version %s of %s left in tree at %s."),
1167                                        change, path, delete_branch, change_past,
1168                                        change_branch, change_branch, path, alt_path);
1169                         } else {
1170                                 output(o, 1, _("CONFLICT (%s/delete): %s deleted in %s "
1171                                        "and %s to %s in %s. Version %s of %s left in tree at %s."),
1172                                        change, old_path, delete_branch, change_past, path,
1173                                        change_branch, change_branch, path, alt_path);
1174                         }
1175                 }
1176                 /*
1177                  * No need to call update_file() on path when change_branch ==
1178                  * o->branch1 && !alt_path, since that would needlessly touch
1179                  * path.  We could call update_file_flags() with update_cache=0
1180                  * and update_wd=0, but that's a no-op.
1181                  */
1182                 if (change_branch != o->branch1 || alt_path)
1183                         ret = update_file(o, 0, changed_oid, changed_mode, update_path);
1184         }
1185         free(alt_path);
1186
1187         return ret;
1188 }
1189
1190 static int conflict_rename_delete(struct merge_options *o,
1191                                    struct diff_filepair *pair,
1192                                    const char *rename_branch,
1193                                    const char *delete_branch)
1194 {
1195         const struct diff_filespec *orig = pair->one;
1196         const struct diff_filespec *dest = pair->two;
1197
1198         if (handle_change_delete(o,
1199                                  o->call_depth ? orig->path : dest->path,
1200                                  o->call_depth ? NULL : orig->path,
1201                                  &orig->oid, orig->mode,
1202                                  &dest->oid, dest->mode,
1203                                  rename_branch, delete_branch,
1204                                  _("rename"), _("renamed")))
1205                 return -1;
1206
1207         if (o->call_depth)
1208                 return remove_file_from_cache(dest->path);
1209         else
1210                 return update_stages(o, dest->path, NULL,
1211                                      rename_branch == o->branch1 ? dest : NULL,
1212                                      rename_branch == o->branch1 ? NULL : dest);
1213 }
1214
1215 static struct diff_filespec *filespec_from_entry(struct diff_filespec *target,
1216                                                  struct stage_data *entry,
1217                                                  int stage)
1218 {
1219         struct object_id *oid = &entry->stages[stage].oid;
1220         unsigned mode = entry->stages[stage].mode;
1221         if (mode == 0 || is_null_oid(oid))
1222                 return NULL;
1223         oidcpy(&target->oid, oid);
1224         target->mode = mode;
1225         return target;
1226 }
1227
1228 static int handle_file(struct merge_options *o,
1229                         struct diff_filespec *rename,
1230                         int stage,
1231                         struct rename_conflict_info *ci)
1232 {
1233         char *dst_name = rename->path;
1234         struct stage_data *dst_entry;
1235         const char *cur_branch, *other_branch;
1236         struct diff_filespec other;
1237         struct diff_filespec *add;
1238         int ret;
1239
1240         if (stage == 2) {
1241                 dst_entry = ci->dst_entry1;
1242                 cur_branch = ci->branch1;
1243                 other_branch = ci->branch2;
1244         } else {
1245                 dst_entry = ci->dst_entry2;
1246                 cur_branch = ci->branch2;
1247                 other_branch = ci->branch1;
1248         }
1249
1250         add = filespec_from_entry(&other, dst_entry, stage ^ 1);
1251         if (add) {
1252                 char *add_name = unique_path(o, rename->path, other_branch);
1253                 if (update_file(o, 0, &add->oid, add->mode, add_name))
1254                         return -1;
1255
1256                 remove_file(o, 0, rename->path, 0);
1257                 dst_name = unique_path(o, rename->path, cur_branch);
1258         } else {
1259                 if (dir_in_way(rename->path, !o->call_depth, 0)) {
1260                         dst_name = unique_path(o, rename->path, cur_branch);
1261                         output(o, 1, _("%s is a directory in %s adding as %s instead"),
1262                                rename->path, other_branch, dst_name);
1263                 }
1264         }
1265         if ((ret = update_file(o, 0, &rename->oid, rename->mode, dst_name)))
1266                 ; /* fall through, do allow dst_name to be released */
1267         else if (stage == 2)
1268                 ret = update_stages(o, rename->path, NULL, rename, add);
1269         else
1270                 ret = update_stages(o, rename->path, NULL, add, rename);
1271
1272         if (dst_name != rename->path)
1273                 free(dst_name);
1274
1275         return ret;
1276 }
1277
1278 static int conflict_rename_rename_1to2(struct merge_options *o,
1279                                         struct rename_conflict_info *ci)
1280 {
1281         /* One file was renamed in both branches, but to different names. */
1282         struct diff_filespec *one = ci->pair1->one;
1283         struct diff_filespec *a = ci->pair1->two;
1284         struct diff_filespec *b = ci->pair2->two;
1285
1286         output(o, 1, _("CONFLICT (rename/rename): "
1287                "Rename \"%s\"->\"%s\" in branch \"%s\" "
1288                "rename \"%s\"->\"%s\" in \"%s\"%s"),
1289                one->path, a->path, ci->branch1,
1290                one->path, b->path, ci->branch2,
1291                o->call_depth ? _(" (left unresolved)") : "");
1292         if (o->call_depth) {
1293                 struct merge_file_info mfi;
1294                 struct diff_filespec other;
1295                 struct diff_filespec *add;
1296                 if (merge_file_one(o, one->path,
1297                                  &one->oid, one->mode,
1298                                  &a->oid, a->mode,
1299                                  &b->oid, b->mode,
1300                                  ci->branch1, ci->branch2, &mfi))
1301                         return -1;
1302
1303                 /*
1304                  * FIXME: For rename/add-source conflicts (if we could detect
1305                  * such), this is wrong.  We should instead find a unique
1306                  * pathname and then either rename the add-source file to that
1307                  * unique path, or use that unique path instead of src here.
1308                  */
1309                 if (update_file(o, 0, &mfi.oid, mfi.mode, one->path))
1310                         return -1;
1311
1312                 /*
1313                  * Above, we put the merged content at the merge-base's
1314                  * path.  Now we usually need to delete both a->path and
1315                  * b->path.  However, the rename on each side of the merge
1316                  * could also be involved in a rename/add conflict.  In
1317                  * such cases, we should keep the added file around,
1318                  * resolving the conflict at that path in its favor.
1319                  */
1320                 add = filespec_from_entry(&other, ci->dst_entry1, 2 ^ 1);
1321                 if (add) {
1322                         if (update_file(o, 0, &add->oid, add->mode, a->path))
1323                                 return -1;
1324                 }
1325                 else
1326                         remove_file_from_cache(a->path);
1327                 add = filespec_from_entry(&other, ci->dst_entry2, 3 ^ 1);
1328                 if (add) {
1329                         if (update_file(o, 0, &add->oid, add->mode, b->path))
1330                                 return -1;
1331                 }
1332                 else
1333                         remove_file_from_cache(b->path);
1334         } else if (handle_file(o, a, 2, ci) || handle_file(o, b, 3, ci))
1335                 return -1;
1336
1337         return 0;
1338 }
1339
1340 static int conflict_rename_rename_2to1(struct merge_options *o,
1341                                         struct rename_conflict_info *ci)
1342 {
1343         /* Two files, a & b, were renamed to the same thing, c. */
1344         struct diff_filespec *a = ci->pair1->one;
1345         struct diff_filespec *b = ci->pair2->one;
1346         struct diff_filespec *c1 = ci->pair1->two;
1347         struct diff_filespec *c2 = ci->pair2->two;
1348         char *path = c1->path; /* == c2->path */
1349         struct merge_file_info mfi_c1;
1350         struct merge_file_info mfi_c2;
1351         int ret;
1352
1353         output(o, 1, _("CONFLICT (rename/rename): "
1354                "Rename %s->%s in %s. "
1355                "Rename %s->%s in %s"),
1356                a->path, c1->path, ci->branch1,
1357                b->path, c2->path, ci->branch2);
1358
1359         remove_file(o, 1, a->path, o->call_depth || would_lose_untracked(a->path));
1360         remove_file(o, 1, b->path, o->call_depth || would_lose_untracked(b->path));
1361
1362         if (merge_file_special_markers(o, a, c1, &ci->ren1_other,
1363                                        o->branch1, c1->path,
1364                                        o->branch2, ci->ren1_other.path, &mfi_c1) ||
1365             merge_file_special_markers(o, b, &ci->ren2_other, c2,
1366                                        o->branch1, ci->ren2_other.path,
1367                                        o->branch2, c2->path, &mfi_c2))
1368                 return -1;
1369
1370         if (o->call_depth) {
1371                 /*
1372                  * If mfi_c1.clean && mfi_c2.clean, then it might make
1373                  * sense to do a two-way merge of those results.  But, I
1374                  * think in all cases, it makes sense to have the virtual
1375                  * merge base just undo the renames; they can be detected
1376                  * again later for the non-recursive merge.
1377                  */
1378                 remove_file(o, 0, path, 0);
1379                 ret = update_file(o, 0, &mfi_c1.oid, mfi_c1.mode, a->path);
1380                 if (!ret)
1381                         ret = update_file(o, 0, &mfi_c2.oid, mfi_c2.mode,
1382                                           b->path);
1383         } else {
1384                 char *new_path1 = unique_path(o, path, ci->branch1);
1385                 char *new_path2 = unique_path(o, path, ci->branch2);
1386                 output(o, 1, _("Renaming %s to %s and %s to %s instead"),
1387                        a->path, new_path1, b->path, new_path2);
1388                 remove_file(o, 0, path, 0);
1389                 ret = update_file(o, 0, &mfi_c1.oid, mfi_c1.mode, new_path1);
1390                 if (!ret)
1391                         ret = update_file(o, 0, &mfi_c2.oid, mfi_c2.mode,
1392                                           new_path2);
1393                 free(new_path2);
1394                 free(new_path1);
1395         }
1396
1397         return ret;
1398 }
1399
1400 /*
1401  * Get the diff_filepairs changed between o_tree and tree.
1402  */
1403 static struct diff_queue_struct *get_diffpairs(struct merge_options *o,
1404                                                struct tree *o_tree,
1405                                                struct tree *tree)
1406 {
1407         struct diff_queue_struct *ret;
1408         struct diff_options opts;
1409
1410         diff_setup(&opts);
1411         opts.flags.recursive = 1;
1412         opts.flags.rename_empty = 0;
1413         opts.detect_rename = DIFF_DETECT_RENAME;
1414         opts.rename_limit = o->merge_rename_limit >= 0 ? o->merge_rename_limit :
1415                             o->diff_rename_limit >= 0 ? o->diff_rename_limit :
1416                             1000;
1417         opts.rename_score = o->rename_score;
1418         opts.show_rename_progress = o->show_rename_progress;
1419         opts.output_format = DIFF_FORMAT_NO_OUTPUT;
1420         diff_setup_done(&opts);
1421         diff_tree_oid(&o_tree->object.oid, &tree->object.oid, "", &opts);
1422         diffcore_std(&opts);
1423         if (opts.needed_rename_limit > o->needed_rename_limit)
1424                 o->needed_rename_limit = opts.needed_rename_limit;
1425
1426         ret = xmalloc(sizeof(*ret));
1427         *ret = diff_queued_diff;
1428
1429         opts.output_format = DIFF_FORMAT_NO_OUTPUT;
1430         diff_queued_diff.nr = 0;
1431         diff_queued_diff.queue = NULL;
1432         diff_flush(&opts);
1433         return ret;
1434 }
1435
1436 static int tree_has_path(struct tree *tree, const char *path)
1437 {
1438         struct object_id hashy;
1439         unsigned int mode_o;
1440
1441         return !get_tree_entry(&tree->object.oid, path,
1442                                &hashy, &mode_o);
1443 }
1444
1445 /*
1446  * Return a new string that replaces the beginning portion (which matches
1447  * entry->dir), with entry->new_dir.  In perl-speak:
1448  *   new_path_name = (old_path =~ s/entry->dir/entry->new_dir/);
1449  * NOTE:
1450  *   Caller must ensure that old_path starts with entry->dir + '/'.
1451  */
1452 static char *apply_dir_rename(struct dir_rename_entry *entry,
1453                               const char *old_path)
1454 {
1455         struct strbuf new_path = STRBUF_INIT;
1456         int oldlen, newlen;
1457
1458         if (entry->non_unique_new_dir)
1459                 return NULL;
1460
1461         oldlen = strlen(entry->dir);
1462         newlen = entry->new_dir.len + (strlen(old_path) - oldlen) + 1;
1463         strbuf_grow(&new_path, newlen);
1464         strbuf_addbuf(&new_path, &entry->new_dir);
1465         strbuf_addstr(&new_path, &old_path[oldlen]);
1466
1467         return strbuf_detach(&new_path, NULL);
1468 }
1469
1470 static void get_renamed_dir_portion(const char *old_path, const char *new_path,
1471                                     char **old_dir, char **new_dir)
1472 {
1473         char *end_of_old, *end_of_new;
1474         int old_len, new_len;
1475
1476         *old_dir = NULL;
1477         *new_dir = NULL;
1478
1479         /*
1480          * For
1481          *    "a/b/c/d/e/foo.c" -> "a/b/some/thing/else/e/foo.c"
1482          * the "e/foo.c" part is the same, we just want to know that
1483          *    "a/b/c/d" was renamed to "a/b/some/thing/else"
1484          * so, for this example, this function returns "a/b/c/d" in
1485          * *old_dir and "a/b/some/thing/else" in *new_dir.
1486          *
1487          * Also, if the basename of the file changed, we don't care.  We
1488          * want to know which portion of the directory, if any, changed.
1489          */
1490         end_of_old = strrchr(old_path, '/');
1491         end_of_new = strrchr(new_path, '/');
1492
1493         if (end_of_old == NULL || end_of_new == NULL)
1494                 return;
1495         while (*--end_of_new == *--end_of_old &&
1496                end_of_old != old_path &&
1497                end_of_new != new_path)
1498                 ; /* Do nothing; all in the while loop */
1499         /*
1500          * We've found the first non-matching character in the directory
1501          * paths.  That means the current directory we were comparing
1502          * represents the rename.  Move end_of_old and end_of_new back
1503          * to the full directory name.
1504          */
1505         if (*end_of_old == '/')
1506                 end_of_old++;
1507         if (*end_of_old != '/')
1508                 end_of_new++;
1509         end_of_old = strchr(end_of_old, '/');
1510         end_of_new = strchr(end_of_new, '/');
1511
1512         /*
1513          * It may have been the case that old_path and new_path were the same
1514          * directory all along.  Don't claim a rename if they're the same.
1515          */
1516         old_len = end_of_old - old_path;
1517         new_len = end_of_new - new_path;
1518
1519         if (old_len != new_len || strncmp(old_path, new_path, old_len)) {
1520                 *old_dir = xstrndup(old_path, old_len);
1521                 *new_dir = xstrndup(new_path, new_len);
1522         }
1523 }
1524
1525 static void remove_hashmap_entries(struct hashmap *dir_renames,
1526                                    struct string_list *items_to_remove)
1527 {
1528         int i;
1529         struct dir_rename_entry *entry;
1530
1531         for (i = 0; i < items_to_remove->nr; i++) {
1532                 entry = items_to_remove->items[i].util;
1533                 hashmap_remove(dir_renames, entry, NULL);
1534         }
1535         string_list_clear(items_to_remove, 0);
1536 }
1537
1538 /*
1539  * See if there is a directory rename for path, and if there are any file
1540  * level conflicts for the renamed location.  If there is a rename and
1541  * there are no conflicts, return the new name.  Otherwise, return NULL.
1542  */
1543 static char *handle_path_level_conflicts(struct merge_options *o,
1544                                          const char *path,
1545                                          struct dir_rename_entry *entry,
1546                                          struct hashmap *collisions,
1547                                          struct tree *tree)
1548 {
1549         char *new_path = NULL;
1550         struct collision_entry *collision_ent;
1551         int clean = 1;
1552         struct strbuf collision_paths = STRBUF_INIT;
1553
1554         /*
1555          * entry has the mapping of old directory name to new directory name
1556          * that we want to apply to path.
1557          */
1558         new_path = apply_dir_rename(entry, path);
1559
1560         if (!new_path) {
1561                 /* This should only happen when entry->non_unique_new_dir set */
1562                 if (!entry->non_unique_new_dir)
1563                         BUG("entry->non_unqiue_dir not set and !new_path");
1564                 output(o, 1, _("CONFLICT (directory rename split): "
1565                                "Unclear where to place %s because directory "
1566                                "%s was renamed to multiple other directories, "
1567                                "with no destination getting a majority of the "
1568                                "files."),
1569                        path, entry->dir);
1570                 clean = 0;
1571                 return NULL;
1572         }
1573
1574         /*
1575          * The caller needs to have ensured that it has pre-populated
1576          * collisions with all paths that map to new_path.  Do a quick check
1577          * to ensure that's the case.
1578          */
1579         collision_ent = collision_find_entry(collisions, new_path);
1580         if (collision_ent == NULL)
1581                 BUG("collision_ent is NULL");
1582
1583         /*
1584          * Check for one-sided add/add/.../add conflicts, i.e.
1585          * where implicit renames from the other side doing
1586          * directory rename(s) can affect this side of history
1587          * to put multiple paths into the same location.  Warn
1588          * and bail on directory renames for such paths.
1589          */
1590         if (collision_ent->reported_already) {
1591                 clean = 0;
1592         } else if (tree_has_path(tree, new_path)) {
1593                 collision_ent->reported_already = 1;
1594                 strbuf_add_separated_string_list(&collision_paths, ", ",
1595                                                  &collision_ent->source_files);
1596                 output(o, 1, _("CONFLICT (implicit dir rename): Existing "
1597                                "file/dir at %s in the way of implicit "
1598                                "directory rename(s) putting the following "
1599                                "path(s) there: %s."),
1600                        new_path, collision_paths.buf);
1601                 clean = 0;
1602         } else if (collision_ent->source_files.nr > 1) {
1603                 collision_ent->reported_already = 1;
1604                 strbuf_add_separated_string_list(&collision_paths, ", ",
1605                                                  &collision_ent->source_files);
1606                 output(o, 1, _("CONFLICT (implicit dir rename): Cannot map "
1607                                "more than one path to %s; implicit directory "
1608                                "renames tried to put these paths there: %s"),
1609                        new_path, collision_paths.buf);
1610                 clean = 0;
1611         }
1612
1613         /* Free memory we no longer need */
1614         strbuf_release(&collision_paths);
1615         if (!clean && new_path) {
1616                 free(new_path);
1617                 return NULL;
1618         }
1619
1620         return new_path;
1621 }
1622
1623 /*
1624  * There are a couple things we want to do at the directory level:
1625  *   1. Check for both sides renaming to the same thing, in order to avoid
1626  *      implicit renaming of files that should be left in place.  (See
1627  *      testcase 6b in t6043 for details.)
1628  *   2. Prune directory renames if there are still files left in the
1629  *      the original directory.  These represent a partial directory rename,
1630  *      i.e. a rename where only some of the files within the directory
1631  *      were renamed elsewhere.  (Technically, this could be done earlier
1632  *      in get_directory_renames(), except that would prevent us from
1633  *      doing the previous check and thus failing testcase 6b.)
1634  *   3. Check for rename/rename(1to2) conflicts (at the directory level).
1635  *      In the future, we could potentially record this info as well and
1636  *      omit reporting rename/rename(1to2) conflicts for each path within
1637  *      the affected directories, thus cleaning up the merge output.
1638  *   NOTE: We do NOT check for rename/rename(2to1) conflicts at the
1639  *         directory level, because merging directories is fine.  If it
1640  *         causes conflicts for files within those merged directories, then
1641  *         that should be detected at the individual path level.
1642  */
1643 static void handle_directory_level_conflicts(struct merge_options *o,
1644                                              struct hashmap *dir_re_head,
1645                                              struct tree *head,
1646                                              struct hashmap *dir_re_merge,
1647                                              struct tree *merge)
1648 {
1649         struct hashmap_iter iter;
1650         struct dir_rename_entry *head_ent;
1651         struct dir_rename_entry *merge_ent;
1652
1653         struct string_list remove_from_head = STRING_LIST_INIT_NODUP;
1654         struct string_list remove_from_merge = STRING_LIST_INIT_NODUP;
1655
1656         hashmap_iter_init(dir_re_head, &iter);
1657         while ((head_ent = hashmap_iter_next(&iter))) {
1658                 merge_ent = dir_rename_find_entry(dir_re_merge, head_ent->dir);
1659                 if (merge_ent &&
1660                     !head_ent->non_unique_new_dir &&
1661                     !merge_ent->non_unique_new_dir &&
1662                     !strbuf_cmp(&head_ent->new_dir, &merge_ent->new_dir)) {
1663                         /* 1. Renamed identically; remove it from both sides */
1664                         string_list_append(&remove_from_head,
1665                                            head_ent->dir)->util = head_ent;
1666                         strbuf_release(&head_ent->new_dir);
1667                         string_list_append(&remove_from_merge,
1668                                            merge_ent->dir)->util = merge_ent;
1669                         strbuf_release(&merge_ent->new_dir);
1670                 } else if (tree_has_path(head, head_ent->dir)) {
1671                         /* 2. This wasn't a directory rename after all */
1672                         string_list_append(&remove_from_head,
1673                                            head_ent->dir)->util = head_ent;
1674                         strbuf_release(&head_ent->new_dir);
1675                 }
1676         }
1677
1678         remove_hashmap_entries(dir_re_head, &remove_from_head);
1679         remove_hashmap_entries(dir_re_merge, &remove_from_merge);
1680
1681         hashmap_iter_init(dir_re_merge, &iter);
1682         while ((merge_ent = hashmap_iter_next(&iter))) {
1683                 head_ent = dir_rename_find_entry(dir_re_head, merge_ent->dir);
1684                 if (tree_has_path(merge, merge_ent->dir)) {
1685                         /* 2. This wasn't a directory rename after all */
1686                         string_list_append(&remove_from_merge,
1687                                            merge_ent->dir)->util = merge_ent;
1688                 } else if (head_ent &&
1689                            !head_ent->non_unique_new_dir &&
1690                            !merge_ent->non_unique_new_dir) {
1691                         /* 3. rename/rename(1to2) */
1692                         /*
1693                          * We can assume it's not rename/rename(1to1) because
1694                          * that was case (1), already checked above.  So we
1695                          * know that head_ent->new_dir and merge_ent->new_dir
1696                          * are different strings.
1697                          */
1698                         output(o, 1, _("CONFLICT (rename/rename): "
1699                                        "Rename directory %s->%s in %s. "
1700                                        "Rename directory %s->%s in %s"),
1701                                head_ent->dir, head_ent->new_dir.buf, o->branch1,
1702                                head_ent->dir, merge_ent->new_dir.buf, o->branch2);
1703                         string_list_append(&remove_from_head,
1704                                            head_ent->dir)->util = head_ent;
1705                         strbuf_release(&head_ent->new_dir);
1706                         string_list_append(&remove_from_merge,
1707                                            merge_ent->dir)->util = merge_ent;
1708                         strbuf_release(&merge_ent->new_dir);
1709                 }
1710         }
1711
1712         remove_hashmap_entries(dir_re_head, &remove_from_head);
1713         remove_hashmap_entries(dir_re_merge, &remove_from_merge);
1714 }
1715
1716 static struct hashmap *get_directory_renames(struct diff_queue_struct *pairs,
1717                                              struct tree *tree)
1718 {
1719         struct hashmap *dir_renames;
1720         struct hashmap_iter iter;
1721         struct dir_rename_entry *entry;
1722         int i;
1723
1724         /*
1725          * Typically, we think of a directory rename as all files from a
1726          * certain directory being moved to a target directory.  However,
1727          * what if someone first moved two files from the original
1728          * directory in one commit, and then renamed the directory
1729          * somewhere else in a later commit?  At merge time, we just know
1730          * that files from the original directory went to two different
1731          * places, and that the bulk of them ended up in the same place.
1732          * We want each directory rename to represent where the bulk of the
1733          * files from that directory end up; this function exists to find
1734          * where the bulk of the files went.
1735          *
1736          * The first loop below simply iterates through the list of file
1737          * renames, finding out how often each directory rename pair
1738          * possibility occurs.
1739          */
1740         dir_renames = xmalloc(sizeof(*dir_renames));
1741         dir_rename_init(dir_renames);
1742         for (i = 0; i < pairs->nr; ++i) {
1743                 struct string_list_item *item;
1744                 int *count;
1745                 struct diff_filepair *pair = pairs->queue[i];
1746                 char *old_dir, *new_dir;
1747
1748                 /* File not part of directory rename if it wasn't renamed */
1749                 if (pair->status != 'R')
1750                         continue;
1751
1752                 get_renamed_dir_portion(pair->one->path, pair->two->path,
1753                                         &old_dir,        &new_dir);
1754                 if (!old_dir)
1755                         /* Directory didn't change at all; ignore this one. */
1756                         continue;
1757
1758                 entry = dir_rename_find_entry(dir_renames, old_dir);
1759                 if (!entry) {
1760                         entry = xmalloc(sizeof(*entry));
1761                         dir_rename_entry_init(entry, old_dir);
1762                         hashmap_put(dir_renames, entry);
1763                 } else {
1764                         free(old_dir);
1765                 }
1766                 item = string_list_lookup(&entry->possible_new_dirs, new_dir);
1767                 if (!item) {
1768                         item = string_list_insert(&entry->possible_new_dirs,
1769                                                   new_dir);
1770                         item->util = xcalloc(1, sizeof(int));
1771                 } else {
1772                         free(new_dir);
1773                 }
1774                 count = item->util;
1775                 *count += 1;
1776         }
1777
1778         /*
1779          * For each directory with files moved out of it, we find out which
1780          * target directory received the most files so we can declare it to
1781          * be the "winning" target location for the directory rename.  This
1782          * winner gets recorded in new_dir.  If there is no winner
1783          * (multiple target directories received the same number of files),
1784          * we set non_unique_new_dir.  Once we've determined the winner (or
1785          * that there is no winner), we no longer need possible_new_dirs.
1786          */
1787         hashmap_iter_init(dir_renames, &iter);
1788         while ((entry = hashmap_iter_next(&iter))) {
1789                 int max = 0;
1790                 int bad_max = 0;
1791                 char *best = NULL;
1792
1793                 for (i = 0; i < entry->possible_new_dirs.nr; i++) {
1794                         int *count = entry->possible_new_dirs.items[i].util;
1795
1796                         if (*count == max)
1797                                 bad_max = max;
1798                         else if (*count > max) {
1799                                 max = *count;
1800                                 best = entry->possible_new_dirs.items[i].string;
1801                         }
1802                 }
1803                 if (bad_max == max)
1804                         entry->non_unique_new_dir = 1;
1805                 else {
1806                         assert(entry->new_dir.len == 0);
1807                         strbuf_addstr(&entry->new_dir, best);
1808                 }
1809                 /*
1810                  * The relevant directory sub-portion of the original full
1811                  * filepaths were xstrndup'ed before inserting into
1812                  * possible_new_dirs, and instead of manually iterating the
1813                  * list and free'ing each, just lie and tell
1814                  * possible_new_dirs that it did the strdup'ing so that it
1815                  * will free them for us.
1816                  */
1817                 entry->possible_new_dirs.strdup_strings = 1;
1818                 string_list_clear(&entry->possible_new_dirs, 1);
1819         }
1820
1821         return dir_renames;
1822 }
1823
1824 static struct dir_rename_entry *check_dir_renamed(const char *path,
1825                                                   struct hashmap *dir_renames)
1826 {
1827         char temp[PATH_MAX];
1828         char *end;
1829         struct dir_rename_entry *entry;
1830
1831         strcpy(temp, path);
1832         while ((end = strrchr(temp, '/'))) {
1833                 *end = '\0';
1834                 entry = dir_rename_find_entry(dir_renames, temp);
1835                 if (entry)
1836                         return entry;
1837         }
1838         return NULL;
1839 }
1840
1841 static void compute_collisions(struct hashmap *collisions,
1842                                struct hashmap *dir_renames,
1843                                struct diff_queue_struct *pairs)
1844 {
1845         int i;
1846
1847         /*
1848          * Multiple files can be mapped to the same path due to directory
1849          * renames done by the other side of history.  Since that other
1850          * side of history could have merged multiple directories into one,
1851          * if our side of history added the same file basename to each of
1852          * those directories, then all N of them would get implicitly
1853          * renamed by the directory rename detection into the same path,
1854          * and we'd get an add/add/.../add conflict, and all those adds
1855          * from *this* side of history.  This is not representable in the
1856          * index, and users aren't going to easily be able to make sense of
1857          * it.  So we need to provide a good warning about what's
1858          * happening, and fall back to no-directory-rename detection
1859          * behavior for those paths.
1860          *
1861          * See testcases 9e and all of section 5 from t6043 for examples.
1862          */
1863         collision_init(collisions);
1864
1865         for (i = 0; i < pairs->nr; ++i) {
1866                 struct dir_rename_entry *dir_rename_ent;
1867                 struct collision_entry *collision_ent;
1868                 char *new_path;
1869                 struct diff_filepair *pair = pairs->queue[i];
1870
1871                 if (pair->status == 'D')
1872                         continue;
1873                 dir_rename_ent = check_dir_renamed(pair->two->path,
1874                                                    dir_renames);
1875                 if (!dir_rename_ent)
1876                         continue;
1877
1878                 new_path = apply_dir_rename(dir_rename_ent, pair->two->path);
1879                 if (!new_path)
1880                         /*
1881                          * dir_rename_ent->non_unique_new_path is true, which
1882                          * means there is no directory rename for us to use,
1883                          * which means it won't cause us any additional
1884                          * collisions.
1885                          */
1886                         continue;
1887                 collision_ent = collision_find_entry(collisions, new_path);
1888                 if (!collision_ent) {
1889                         collision_ent = xcalloc(1,
1890                                                 sizeof(struct collision_entry));
1891                         hashmap_entry_init(collision_ent, strhash(new_path));
1892                         hashmap_put(collisions, collision_ent);
1893                         collision_ent->target_file = new_path;
1894                 } else {
1895                         free(new_path);
1896                 }
1897                 string_list_insert(&collision_ent->source_files,
1898                                    pair->two->path);
1899         }
1900 }
1901
1902 static char *check_for_directory_rename(struct merge_options *o,
1903                                         const char *path,
1904                                         struct tree *tree,
1905                                         struct hashmap *dir_renames,
1906                                         struct hashmap *dir_rename_exclusions,
1907                                         struct hashmap *collisions,
1908                                         int *clean_merge)
1909 {
1910         char *new_path = NULL;
1911         struct dir_rename_entry *entry = check_dir_renamed(path, dir_renames);
1912         struct dir_rename_entry *oentry = NULL;
1913
1914         if (!entry)
1915                 return new_path;
1916
1917         /*
1918          * This next part is a little weird.  We do not want to do an
1919          * implicit rename into a directory we renamed on our side, because
1920          * that will result in a spurious rename/rename(1to2) conflict.  An
1921          * example:
1922          *   Base commit: dumbdir/afile, otherdir/bfile
1923          *   Side 1:      smrtdir/afile, otherdir/bfile
1924          *   Side 2:      dumbdir/afile, dumbdir/bfile
1925          * Here, while working on Side 1, we could notice that otherdir was
1926          * renamed/merged to dumbdir, and change the diff_filepair for
1927          * otherdir/bfile into a rename into dumbdir/bfile.  However, Side
1928          * 2 will notice the rename from dumbdir to smrtdir, and do the
1929          * transitive rename to move it from dumbdir/bfile to
1930          * smrtdir/bfile.  That gives us bfile in dumbdir vs being in
1931          * smrtdir, a rename/rename(1to2) conflict.  We really just want
1932          * the file to end up in smrtdir.  And the way to achieve that is
1933          * to not let Side1 do the rename to dumbdir, since we know that is
1934          * the source of one of our directory renames.
1935          *
1936          * That's why oentry and dir_rename_exclusions is here.
1937          *
1938          * As it turns out, this also prevents N-way transient rename
1939          * confusion; See testcases 9c and 9d of t6043.
1940          */
1941         oentry = dir_rename_find_entry(dir_rename_exclusions, entry->new_dir.buf);
1942         if (oentry) {
1943                 output(o, 1, _("WARNING: Avoiding applying %s -> %s rename "
1944                                "to %s, because %s itself was renamed."),
1945                        entry->dir, entry->new_dir.buf, path, entry->new_dir.buf);
1946         } else {
1947                 new_path = handle_path_level_conflicts(o, path, entry,
1948                                                        collisions, tree);
1949                 *clean_merge &= (new_path != NULL);
1950         }
1951
1952         return new_path;
1953 }
1954
1955 /*
1956  * Get information of all renames which occurred in 'pairs', making use of
1957  * any implicit directory renames inferred from the other side of history.
1958  * We need the three trees in the merge ('o_tree', 'a_tree' and 'b_tree')
1959  * to be able to associate the correct cache entries with the rename
1960  * information; tree is always equal to either a_tree or b_tree.
1961  */
1962 static struct string_list *get_renames(struct merge_options *o,
1963                                        struct diff_queue_struct *pairs,
1964                                        struct hashmap *dir_renames,
1965                                        struct hashmap *dir_rename_exclusions,
1966                                        struct tree *tree,
1967                                        struct tree *o_tree,
1968                                        struct tree *a_tree,
1969                                        struct tree *b_tree,
1970                                        struct string_list *entries,
1971                                        int *clean_merge)
1972 {
1973         int i;
1974         struct hashmap collisions;
1975         struct hashmap_iter iter;
1976         struct collision_entry *e;
1977         struct string_list *renames;
1978
1979         compute_collisions(&collisions, dir_renames, pairs);
1980         renames = xcalloc(1, sizeof(struct string_list));
1981
1982         for (i = 0; i < pairs->nr; ++i) {
1983                 struct string_list_item *item;
1984                 struct rename *re;
1985                 struct diff_filepair *pair = pairs->queue[i];
1986                 char *new_path; /* non-NULL only with directory renames */
1987
1988                 if (pair->status == 'D') {
1989                         diff_free_filepair(pair);
1990                         continue;
1991                 }
1992                 new_path = check_for_directory_rename(o, pair->two->path, tree,
1993                                                       dir_renames,
1994                                                       dir_rename_exclusions,
1995                                                       &collisions,
1996                                                       clean_merge);
1997                 if (pair->status != 'R' && !new_path) {
1998                         diff_free_filepair(pair);
1999                         continue;
2000                 }
2001
2002                 re = xmalloc(sizeof(*re));
2003                 re->processed = 0;
2004                 re->pair = pair;
2005                 item = string_list_lookup(entries, re->pair->one->path);
2006                 if (!item)
2007                         re->src_entry = insert_stage_data(re->pair->one->path,
2008                                         o_tree, a_tree, b_tree, entries);
2009                 else
2010                         re->src_entry = item->util;
2011
2012                 item = string_list_lookup(entries, re->pair->two->path);
2013                 if (!item)
2014                         re->dst_entry = insert_stage_data(re->pair->two->path,
2015                                         o_tree, a_tree, b_tree, entries);
2016                 else
2017                         re->dst_entry = item->util;
2018                 item = string_list_insert(renames, pair->one->path);
2019                 item->util = re;
2020         }
2021
2022         hashmap_iter_init(&collisions, &iter);
2023         while ((e = hashmap_iter_next(&iter))) {
2024                 free(e->target_file);
2025                 string_list_clear(&e->source_files, 0);
2026         }
2027         hashmap_free(&collisions, 1);
2028         return renames;
2029 }
2030
2031 static int process_renames(struct merge_options *o,
2032                            struct string_list *a_renames,
2033                            struct string_list *b_renames)
2034 {
2035         int clean_merge = 1, i, j;
2036         struct string_list a_by_dst = STRING_LIST_INIT_NODUP;
2037         struct string_list b_by_dst = STRING_LIST_INIT_NODUP;
2038         const struct rename *sre;
2039
2040         for (i = 0; i < a_renames->nr; i++) {
2041                 sre = a_renames->items[i].util;
2042                 string_list_insert(&a_by_dst, sre->pair->two->path)->util
2043                         = (void *)sre;
2044         }
2045         for (i = 0; i < b_renames->nr; i++) {
2046                 sre = b_renames->items[i].util;
2047                 string_list_insert(&b_by_dst, sre->pair->two->path)->util
2048                         = (void *)sre;
2049         }
2050
2051         for (i = 0, j = 0; i < a_renames->nr || j < b_renames->nr;) {
2052                 struct string_list *renames1, *renames2Dst;
2053                 struct rename *ren1 = NULL, *ren2 = NULL;
2054                 const char *branch1, *branch2;
2055                 const char *ren1_src, *ren1_dst;
2056                 struct string_list_item *lookup;
2057
2058                 if (i >= a_renames->nr) {
2059                         ren2 = b_renames->items[j++].util;
2060                 } else if (j >= b_renames->nr) {
2061                         ren1 = a_renames->items[i++].util;
2062                 } else {
2063                         int compare = strcmp(a_renames->items[i].string,
2064                                              b_renames->items[j].string);
2065                         if (compare <= 0)
2066                                 ren1 = a_renames->items[i++].util;
2067                         if (compare >= 0)
2068                                 ren2 = b_renames->items[j++].util;
2069                 }
2070
2071                 /* TODO: refactor, so that 1/2 are not needed */
2072                 if (ren1) {
2073                         renames1 = a_renames;
2074                         renames2Dst = &b_by_dst;
2075                         branch1 = o->branch1;
2076                         branch2 = o->branch2;
2077                 } else {
2078                         renames1 = b_renames;
2079                         renames2Dst = &a_by_dst;
2080                         branch1 = o->branch2;
2081                         branch2 = o->branch1;
2082                         SWAP(ren2, ren1);
2083                 }
2084
2085                 if (ren1->processed)
2086                         continue;
2087                 ren1->processed = 1;
2088                 ren1->dst_entry->processed = 1;
2089                 /* BUG: We should only mark src_entry as processed if we
2090                  * are not dealing with a rename + add-source case.
2091                  */
2092                 ren1->src_entry->processed = 1;
2093
2094                 ren1_src = ren1->pair->one->path;
2095                 ren1_dst = ren1->pair->two->path;
2096
2097                 if (ren2) {
2098                         /* One file renamed on both sides */
2099                         const char *ren2_src = ren2->pair->one->path;
2100                         const char *ren2_dst = ren2->pair->two->path;
2101                         enum rename_type rename_type;
2102                         if (strcmp(ren1_src, ren2_src) != 0)
2103                                 die("BUG: ren1_src != ren2_src");
2104                         ren2->dst_entry->processed = 1;
2105                         ren2->processed = 1;
2106                         if (strcmp(ren1_dst, ren2_dst) != 0) {
2107                                 rename_type = RENAME_ONE_FILE_TO_TWO;
2108                                 clean_merge = 0;
2109                         } else {
2110                                 rename_type = RENAME_ONE_FILE_TO_ONE;
2111                                 /* BUG: We should only remove ren1_src in
2112                                  * the base stage (think of rename +
2113                                  * add-source cases).
2114                                  */
2115                                 remove_file(o, 1, ren1_src, 1);
2116                                 update_entry(ren1->dst_entry,
2117                                              ren1->pair->one,
2118                                              ren1->pair->two,
2119                                              ren2->pair->two);
2120                         }
2121                         setup_rename_conflict_info(rename_type,
2122                                                    ren1->pair,
2123                                                    ren2->pair,
2124                                                    branch1,
2125                                                    branch2,
2126                                                    ren1->dst_entry,
2127                                                    ren2->dst_entry,
2128                                                    o,
2129                                                    NULL,
2130                                                    NULL);
2131                 } else if ((lookup = string_list_lookup(renames2Dst, ren1_dst))) {
2132                         /* Two different files renamed to the same thing */
2133                         char *ren2_dst;
2134                         ren2 = lookup->util;
2135                         ren2_dst = ren2->pair->two->path;
2136                         if (strcmp(ren1_dst, ren2_dst) != 0)
2137                                 die("BUG: ren1_dst != ren2_dst");
2138
2139                         clean_merge = 0;
2140                         ren2->processed = 1;
2141                         /*
2142                          * BUG: We should only mark src_entry as processed
2143                          * if we are not dealing with a rename + add-source
2144                          * case.
2145                          */
2146                         ren2->src_entry->processed = 1;
2147
2148                         setup_rename_conflict_info(RENAME_TWO_FILES_TO_ONE,
2149                                                    ren1->pair,
2150                                                    ren2->pair,
2151                                                    branch1,
2152                                                    branch2,
2153                                                    ren1->dst_entry,
2154                                                    ren2->dst_entry,
2155                                                    o,
2156                                                    ren1->src_entry,
2157                                                    ren2->src_entry);
2158
2159                 } else {
2160                         /* Renamed in 1, maybe changed in 2 */
2161                         /* we only use sha1 and mode of these */
2162                         struct diff_filespec src_other, dst_other;
2163                         int try_merge;
2164
2165                         /*
2166                          * unpack_trees loads entries from common-commit
2167                          * into stage 1, from head-commit into stage 2, and
2168                          * from merge-commit into stage 3.  We keep track
2169                          * of which side corresponds to the rename.
2170                          */
2171                         int renamed_stage = a_renames == renames1 ? 2 : 3;
2172                         int other_stage =   a_renames == renames1 ? 3 : 2;
2173
2174                         /* BUG: We should only remove ren1_src in the base
2175                          * stage and in other_stage (think of rename +
2176                          * add-source case).
2177                          */
2178                         remove_file(o, 1, ren1_src,
2179                                     renamed_stage == 2 || !was_tracked(ren1_src));
2180
2181                         oidcpy(&src_other.oid,
2182                                &ren1->src_entry->stages[other_stage].oid);
2183                         src_other.mode = ren1->src_entry->stages[other_stage].mode;
2184                         oidcpy(&dst_other.oid,
2185                                &ren1->dst_entry->stages[other_stage].oid);
2186                         dst_other.mode = ren1->dst_entry->stages[other_stage].mode;
2187                         try_merge = 0;
2188
2189                         if (oid_eq(&src_other.oid, &null_oid)) {
2190                                 setup_rename_conflict_info(RENAME_DELETE,
2191                                                            ren1->pair,
2192                                                            NULL,
2193                                                            branch1,
2194                                                            branch2,
2195                                                            ren1->dst_entry,
2196                                                            NULL,
2197                                                            o,
2198                                                            NULL,
2199                                                            NULL);
2200                         } else if ((dst_other.mode == ren1->pair->two->mode) &&
2201                                    oid_eq(&dst_other.oid, &ren1->pair->two->oid)) {
2202                                 /*
2203                                  * Added file on the other side identical to
2204                                  * the file being renamed: clean merge.
2205                                  * Also, there is no need to overwrite the
2206                                  * file already in the working copy, so call
2207                                  * update_file_flags() instead of
2208                                  * update_file().
2209                                  */
2210                                 if (update_file_flags(o,
2211                                                       &ren1->pair->two->oid,
2212                                                       ren1->pair->two->mode,
2213                                                       ren1_dst,
2214                                                       1, /* update_cache */
2215                                                       0  /* update_wd    */))
2216                                         clean_merge = -1;
2217                         } else if (!oid_eq(&dst_other.oid, &null_oid)) {
2218                                 clean_merge = 0;
2219                                 try_merge = 1;
2220                                 output(o, 1, _("CONFLICT (rename/add): Rename %s->%s in %s. "
2221                                        "%s added in %s"),
2222                                        ren1_src, ren1_dst, branch1,
2223                                        ren1_dst, branch2);
2224                                 if (o->call_depth) {
2225                                         struct merge_file_info mfi;
2226                                         if (merge_file_one(o, ren1_dst, &null_oid, 0,
2227                                                            &ren1->pair->two->oid,
2228                                                            ren1->pair->two->mode,
2229                                                            &dst_other.oid,
2230                                                            dst_other.mode,
2231                                                            branch1, branch2, &mfi)) {
2232                                                 clean_merge = -1;
2233                                                 goto cleanup_and_return;
2234                                         }
2235                                         output(o, 1, _("Adding merged %s"), ren1_dst);
2236                                         if (update_file(o, 0, &mfi.oid,
2237                                                         mfi.mode, ren1_dst))
2238                                                 clean_merge = -1;
2239                                         try_merge = 0;
2240                                 } else {
2241                                         char *new_path = unique_path(o, ren1_dst, branch2);
2242                                         output(o, 1, _("Adding as %s instead"), new_path);
2243                                         if (update_file(o, 0, &dst_other.oid,
2244                                                         dst_other.mode, new_path))
2245                                                 clean_merge = -1;
2246                                         free(new_path);
2247                                 }
2248                         } else
2249                                 try_merge = 1;
2250
2251                         if (clean_merge < 0)
2252                                 goto cleanup_and_return;
2253                         if (try_merge) {
2254                                 struct diff_filespec *one, *a, *b;
2255                                 src_other.path = (char *)ren1_src;
2256
2257                                 one = ren1->pair->one;
2258                                 if (a_renames == renames1) {
2259                                         a = ren1->pair->two;
2260                                         b = &src_other;
2261                                 } else {
2262                                         b = ren1->pair->two;
2263                                         a = &src_other;
2264                                 }
2265                                 update_entry(ren1->dst_entry, one, a, b);
2266                                 setup_rename_conflict_info(RENAME_NORMAL,
2267                                                            ren1->pair,
2268                                                            NULL,
2269                                                            branch1,
2270                                                            NULL,
2271                                                            ren1->dst_entry,
2272                                                            NULL,
2273                                                            o,
2274                                                            NULL,
2275                                                            NULL);
2276                         }
2277                 }
2278         }
2279 cleanup_and_return:
2280         string_list_clear(&a_by_dst, 0);
2281         string_list_clear(&b_by_dst, 0);
2282
2283         return clean_merge;
2284 }
2285
2286 struct rename_info {
2287         struct string_list *head_renames;
2288         struct string_list *merge_renames;
2289 };
2290
2291 static void initial_cleanup_rename(struct diff_queue_struct *pairs,
2292                                    struct hashmap *dir_renames)
2293 {
2294         struct hashmap_iter iter;
2295         struct dir_rename_entry *e;
2296
2297         hashmap_iter_init(dir_renames, &iter);
2298         while ((e = hashmap_iter_next(&iter))) {
2299                 free(e->dir);
2300                 strbuf_release(&e->new_dir);
2301                 /* possible_new_dirs already cleared in get_directory_renames */
2302         }
2303         hashmap_free(dir_renames, 1);
2304         free(dir_renames);
2305
2306         free(pairs->queue);
2307         free(pairs);
2308 }
2309
2310 static int handle_renames(struct merge_options *o,
2311                           struct tree *common,
2312                           struct tree *head,
2313                           struct tree *merge,
2314                           struct string_list *entries,
2315                           struct rename_info *ri)
2316 {
2317         struct diff_queue_struct *head_pairs, *merge_pairs;
2318         struct hashmap *dir_re_head, *dir_re_merge;
2319         int clean = 1;
2320
2321         ri->head_renames = NULL;
2322         ri->merge_renames = NULL;
2323
2324         if (!o->detect_rename)
2325                 return 1;
2326
2327         head_pairs = get_diffpairs(o, common, head);
2328         merge_pairs = get_diffpairs(o, common, merge);
2329
2330         dir_re_head = get_directory_renames(head_pairs, head);
2331         dir_re_merge = get_directory_renames(merge_pairs, merge);
2332
2333         handle_directory_level_conflicts(o,
2334                                          dir_re_head, head,
2335                                          dir_re_merge, merge);
2336
2337         ri->head_renames  = get_renames(o, head_pairs,
2338                                         dir_re_merge, dir_re_head, head,
2339                                         common, head, merge, entries,
2340                                         &clean);
2341         if (clean < 0)
2342                 goto cleanup;
2343         ri->merge_renames = get_renames(o, merge_pairs,
2344                                         dir_re_head, dir_re_merge, merge,
2345                                         common, head, merge, entries,
2346                                         &clean);
2347         if (clean < 0)
2348                 goto cleanup;
2349         clean &= process_renames(o, ri->head_renames, ri->merge_renames);
2350
2351 cleanup:
2352         /*
2353          * Some cleanup is deferred until cleanup_renames() because the
2354          * data structures are still needed and referenced in
2355          * process_entry().  But there are a few things we can free now.
2356          */
2357         initial_cleanup_rename(head_pairs, dir_re_head);
2358         initial_cleanup_rename(merge_pairs, dir_re_merge);
2359
2360         return clean;
2361 }
2362
2363 static void final_cleanup_rename(struct string_list *rename)
2364 {
2365         const struct rename *re;
2366         int i;
2367
2368         if (rename == NULL)
2369                 return;
2370
2371         for (i = 0; i < rename->nr; i++) {
2372                 re = rename->items[i].util;
2373                 diff_free_filepair(re->pair);
2374         }
2375         string_list_clear(rename, 1);
2376         free(rename);
2377 }
2378
2379 static void final_cleanup_renames(struct rename_info *re_info)
2380 {
2381         final_cleanup_rename(re_info->head_renames);
2382         final_cleanup_rename(re_info->merge_renames);
2383 }
2384
2385 static struct object_id *stage_oid(const struct object_id *oid, unsigned mode)
2386 {
2387         return (is_null_oid(oid) || mode == 0) ? NULL: (struct object_id *)oid;
2388 }
2389
2390 static int read_oid_strbuf(struct merge_options *o,
2391         const struct object_id *oid, struct strbuf *dst)
2392 {
2393         void *buf;
2394         enum object_type type;
2395         unsigned long size;
2396         buf = read_object_file(oid, &type, &size);
2397         if (!buf)
2398                 return err(o, _("cannot read object %s"), oid_to_hex(oid));
2399         if (type != OBJ_BLOB) {
2400                 free(buf);
2401                 return err(o, _("object %s is not a blob"), oid_to_hex(oid));
2402         }
2403         strbuf_attach(dst, buf, size, size + 1);
2404         return 0;
2405 }
2406
2407 static int blob_unchanged(struct merge_options *opt,
2408                           const struct object_id *o_oid,
2409                           unsigned o_mode,
2410                           const struct object_id *a_oid,
2411                           unsigned a_mode,
2412                           int renormalize, const char *path)
2413 {
2414         struct strbuf o = STRBUF_INIT;
2415         struct strbuf a = STRBUF_INIT;
2416         int ret = 0; /* assume changed for safety */
2417
2418         if (a_mode != o_mode)
2419                 return 0;
2420         if (oid_eq(o_oid, a_oid))
2421                 return 1;
2422         if (!renormalize)
2423                 return 0;
2424
2425         assert(o_oid && a_oid);
2426         if (read_oid_strbuf(opt, o_oid, &o) || read_oid_strbuf(opt, a_oid, &a))
2427                 goto error_return;
2428         /*
2429          * Note: binary | is used so that both renormalizations are
2430          * performed.  Comparison can be skipped if both files are
2431          * unchanged since their sha1s have already been compared.
2432          */
2433         if (renormalize_buffer(&the_index, path, o.buf, o.len, &o) |
2434             renormalize_buffer(&the_index, path, a.buf, a.len, &a))
2435                 ret = (o.len == a.len && !memcmp(o.buf, a.buf, o.len));
2436
2437 error_return:
2438         strbuf_release(&o);
2439         strbuf_release(&a);
2440         return ret;
2441 }
2442
2443 static int handle_modify_delete(struct merge_options *o,
2444                                  const char *path,
2445                                  struct object_id *o_oid, int o_mode,
2446                                  struct object_id *a_oid, int a_mode,
2447                                  struct object_id *b_oid, int b_mode)
2448 {
2449         const char *modify_branch, *delete_branch;
2450         struct object_id *changed_oid;
2451         int changed_mode;
2452
2453         if (a_oid) {
2454                 modify_branch = o->branch1;
2455                 delete_branch = o->branch2;
2456                 changed_oid = a_oid;
2457                 changed_mode = a_mode;
2458         } else {
2459                 modify_branch = o->branch2;
2460                 delete_branch = o->branch1;
2461                 changed_oid = b_oid;
2462                 changed_mode = b_mode;
2463         }
2464
2465         return handle_change_delete(o,
2466                                     path, NULL,
2467                                     o_oid, o_mode,
2468                                     changed_oid, changed_mode,
2469                                     modify_branch, delete_branch,
2470                                     _("modify"), _("modified"));
2471 }
2472
2473 static int merge_content(struct merge_options *o,
2474                          const char *path,
2475                          struct object_id *o_oid, int o_mode,
2476                          struct object_id *a_oid, int a_mode,
2477                          struct object_id *b_oid, int b_mode,
2478                          struct rename_conflict_info *rename_conflict_info)
2479 {
2480         const char *reason = _("content");
2481         const char *path1 = NULL, *path2 = NULL;
2482         struct merge_file_info mfi;
2483         struct diff_filespec one, a, b;
2484         unsigned df_conflict_remains = 0;
2485
2486         if (!o_oid) {
2487                 reason = _("add/add");
2488                 o_oid = (struct object_id *)&null_oid;
2489         }
2490         one.path = a.path = b.path = (char *)path;
2491         oidcpy(&one.oid, o_oid);
2492         one.mode = o_mode;
2493         oidcpy(&a.oid, a_oid);
2494         a.mode = a_mode;
2495         oidcpy(&b.oid, b_oid);
2496         b.mode = b_mode;
2497
2498         if (rename_conflict_info) {
2499                 struct diff_filepair *pair1 = rename_conflict_info->pair1;
2500
2501                 path1 = (o->branch1 == rename_conflict_info->branch1) ?
2502                         pair1->two->path : pair1->one->path;
2503                 /* If rename_conflict_info->pair2 != NULL, we are in
2504                  * RENAME_ONE_FILE_TO_ONE case.  Otherwise, we have a
2505                  * normal rename.
2506                  */
2507                 path2 = (rename_conflict_info->pair2 ||
2508                          o->branch2 == rename_conflict_info->branch1) ?
2509                         pair1->two->path : pair1->one->path;
2510
2511                 if (dir_in_way(path, !o->call_depth,
2512                                S_ISGITLINK(pair1->two->mode)))
2513                         df_conflict_remains = 1;
2514         }
2515         if (merge_file_special_markers(o, &one, &a, &b,
2516                                        o->branch1, path1,
2517                                        o->branch2, path2, &mfi))
2518                 return -1;
2519
2520         if (mfi.clean && !df_conflict_remains &&
2521             oid_eq(&mfi.oid, a_oid) && mfi.mode == a_mode) {
2522                 int path_renamed_outside_HEAD;
2523                 output(o, 3, _("Skipped %s (merged same as existing)"), path);
2524                 /*
2525                  * The content merge resulted in the same file contents we
2526                  * already had.  We can return early if those file contents
2527                  * are recorded at the correct path (which may not be true
2528                  * if the merge involves a rename).
2529                  */
2530                 path_renamed_outside_HEAD = !path2 || !strcmp(path, path2);
2531                 if (!path_renamed_outside_HEAD) {
2532                         add_cacheinfo(o, mfi.mode, &mfi.oid, path,
2533                                       0, (!o->call_depth), 0);
2534                         return mfi.clean;
2535                 }
2536         } else
2537                 output(o, 2, _("Auto-merging %s"), path);
2538
2539         if (!mfi.clean) {
2540                 if (S_ISGITLINK(mfi.mode))
2541                         reason = _("submodule");
2542                 output(o, 1, _("CONFLICT (%s): Merge conflict in %s"),
2543                                 reason, path);
2544                 if (rename_conflict_info && !df_conflict_remains)
2545                         if (update_stages(o, path, &one, &a, &b))
2546                                 return -1;
2547         }
2548
2549         if (df_conflict_remains) {
2550                 char *new_path;
2551                 if (o->call_depth) {
2552                         remove_file_from_cache(path);
2553                 } else {
2554                         if (!mfi.clean) {
2555                                 if (update_stages(o, path, &one, &a, &b))
2556                                         return -1;
2557                         } else {
2558                                 int file_from_stage2 = was_tracked(path);
2559                                 struct diff_filespec merged;
2560                                 oidcpy(&merged.oid, &mfi.oid);
2561                                 merged.mode = mfi.mode;
2562
2563                                 if (update_stages(o, path, NULL,
2564                                                   file_from_stage2 ? &merged : NULL,
2565                                                   file_from_stage2 ? NULL : &merged))
2566                                         return -1;
2567                         }
2568
2569                 }
2570                 new_path = unique_path(o, path, rename_conflict_info->branch1);
2571                 output(o, 1, _("Adding as %s instead"), new_path);
2572                 if (update_file(o, 0, &mfi.oid, mfi.mode, new_path)) {
2573                         free(new_path);
2574                         return -1;
2575                 }
2576                 free(new_path);
2577                 mfi.clean = 0;
2578         } else if (update_file(o, mfi.clean, &mfi.oid, mfi.mode, path))
2579                 return -1;
2580         return mfi.clean;
2581 }
2582
2583 /* Per entry merge function */
2584 static int process_entry(struct merge_options *o,
2585                          const char *path, struct stage_data *entry)
2586 {
2587         int clean_merge = 1;
2588         int normalize = o->renormalize;
2589         unsigned o_mode = entry->stages[1].mode;
2590         unsigned a_mode = entry->stages[2].mode;
2591         unsigned b_mode = entry->stages[3].mode;
2592         struct object_id *o_oid = stage_oid(&entry->stages[1].oid, o_mode);
2593         struct object_id *a_oid = stage_oid(&entry->stages[2].oid, a_mode);
2594         struct object_id *b_oid = stage_oid(&entry->stages[3].oid, b_mode);
2595
2596         entry->processed = 1;
2597         if (entry->rename_conflict_info) {
2598                 struct rename_conflict_info *conflict_info = entry->rename_conflict_info;
2599                 switch (conflict_info->rename_type) {
2600                 case RENAME_NORMAL:
2601                 case RENAME_ONE_FILE_TO_ONE:
2602                         clean_merge = merge_content(o, path,
2603                                                     o_oid, o_mode, a_oid, a_mode, b_oid, b_mode,
2604                                                     conflict_info);
2605                         break;
2606                 case RENAME_DELETE:
2607                         clean_merge = 0;
2608                         if (conflict_rename_delete(o,
2609                                                    conflict_info->pair1,
2610                                                    conflict_info->branch1,
2611                                                    conflict_info->branch2))
2612                                 clean_merge = -1;
2613                         break;
2614                 case RENAME_ONE_FILE_TO_TWO:
2615                         clean_merge = 0;
2616                         if (conflict_rename_rename_1to2(o, conflict_info))
2617                                 clean_merge = -1;
2618                         break;
2619                 case RENAME_TWO_FILES_TO_ONE:
2620                         clean_merge = 0;
2621                         if (conflict_rename_rename_2to1(o, conflict_info))
2622                                 clean_merge = -1;
2623                         break;
2624                 default:
2625                         entry->processed = 0;
2626                         break;
2627                 }
2628         } else if (o_oid && (!a_oid || !b_oid)) {
2629                 /* Case A: Deleted in one */
2630                 if ((!a_oid && !b_oid) ||
2631                     (!b_oid && blob_unchanged(o, o_oid, o_mode, a_oid, a_mode, normalize, path)) ||
2632                     (!a_oid && blob_unchanged(o, o_oid, o_mode, b_oid, b_mode, normalize, path))) {
2633                         /* Deleted in both or deleted in one and
2634                          * unchanged in the other */
2635                         if (a_oid)
2636                                 output(o, 2, _("Removing %s"), path);
2637                         /* do not touch working file if it did not exist */
2638                         remove_file(o, 1, path, !a_oid);
2639                 } else {
2640                         /* Modify/delete; deleted side may have put a directory in the way */
2641                         clean_merge = 0;
2642                         if (handle_modify_delete(o, path, o_oid, o_mode,
2643                                                  a_oid, a_mode, b_oid, b_mode))
2644                                 clean_merge = -1;
2645                 }
2646         } else if ((!o_oid && a_oid && !b_oid) ||
2647                    (!o_oid && !a_oid && b_oid)) {
2648                 /* Case B: Added in one. */
2649                 /* [nothing|directory] -> ([nothing|directory], file) */
2650
2651                 const char *add_branch;
2652                 const char *other_branch;
2653                 unsigned mode;
2654                 const struct object_id *oid;
2655                 const char *conf;
2656
2657                 if (a_oid) {
2658                         add_branch = o->branch1;
2659                         other_branch = o->branch2;
2660                         mode = a_mode;
2661                         oid = a_oid;
2662                         conf = _("file/directory");
2663                 } else {
2664                         add_branch = o->branch2;
2665                         other_branch = o->branch1;
2666                         mode = b_mode;
2667                         oid = b_oid;
2668                         conf = _("directory/file");
2669                 }
2670                 if (dir_in_way(path,
2671                                !o->call_depth && !S_ISGITLINK(a_mode),
2672                                0)) {
2673                         char *new_path = unique_path(o, path, add_branch);
2674                         clean_merge = 0;
2675                         output(o, 1, _("CONFLICT (%s): There is a directory with name %s in %s. "
2676                                "Adding %s as %s"),
2677                                conf, path, other_branch, path, new_path);
2678                         if (update_file(o, 0, oid, mode, new_path))
2679                                 clean_merge = -1;
2680                         else if (o->call_depth)
2681                                 remove_file_from_cache(path);
2682                         free(new_path);
2683                 } else {
2684                         output(o, 2, _("Adding %s"), path);
2685                         /* do not overwrite file if already present */
2686                         if (update_file_flags(o, oid, mode, path, 1, !a_oid))
2687                                 clean_merge = -1;
2688                 }
2689         } else if (a_oid && b_oid) {
2690                 /* Case C: Added in both (check for same permissions) and */
2691                 /* case D: Modified in both, but differently. */
2692                 clean_merge = merge_content(o, path,
2693                                             o_oid, o_mode, a_oid, a_mode, b_oid, b_mode,
2694                                             NULL);
2695         } else if (!o_oid && !a_oid && !b_oid) {
2696                 /*
2697                  * this entry was deleted altogether. a_mode == 0 means
2698                  * we had that path and want to actively remove it.
2699                  */
2700                 remove_file(o, 1, path, !a_mode);
2701         } else
2702                 die("BUG: fatal merge failure, shouldn't happen.");
2703
2704         return clean_merge;
2705 }
2706
2707 int merge_trees(struct merge_options *o,
2708                 struct tree *head,
2709                 struct tree *merge,
2710                 struct tree *common,
2711                 struct tree **result)
2712 {
2713         int code, clean;
2714
2715         if (o->subtree_shift) {
2716                 merge = shift_tree_object(head, merge, o->subtree_shift);
2717                 common = shift_tree_object(head, common, o->subtree_shift);
2718         }
2719
2720         if (oid_eq(&common->object.oid, &merge->object.oid)) {
2721                 struct strbuf sb = STRBUF_INIT;
2722
2723                 if (!o->call_depth && index_has_changes(&sb)) {
2724                         err(o, _("Dirty index: cannot merge (dirty: %s)"),
2725                             sb.buf);
2726                         return 0;
2727                 }
2728                 output(o, 0, _("Already up to date!"));
2729                 *result = head;
2730                 return 1;
2731         }
2732
2733         code = git_merge_trees(o->call_depth, common, head, merge);
2734
2735         if (code != 0) {
2736                 if (show(o, 4) || o->call_depth)
2737                         err(o, _("merging of trees %s and %s failed"),
2738                             oid_to_hex(&head->object.oid),
2739                             oid_to_hex(&merge->object.oid));
2740                 return -1;
2741         }
2742
2743         if (unmerged_cache()) {
2744                 struct string_list *entries;
2745                 struct rename_info re_info;
2746                 int i;
2747                 /*
2748                  * Only need the hashmap while processing entries, so
2749                  * initialize it here and free it when we are done running
2750                  * through the entries. Keeping it in the merge_options as
2751                  * opposed to decaring a local hashmap is for convenience
2752                  * so that we don't have to pass it to around.
2753                  */
2754                 hashmap_init(&o->current_file_dir_set, path_hashmap_cmp, NULL, 512);
2755                 get_files_dirs(o, head);
2756                 get_files_dirs(o, merge);
2757
2758                 entries = get_unmerged();
2759                 clean = handle_renames(o, common, head, merge, entries,
2760                                        &re_info);
2761                 record_df_conflict_files(o, entries);
2762                 if (clean < 0)
2763                         goto cleanup;
2764                 for (i = entries->nr-1; 0 <= i; i--) {
2765                         const char *path = entries->items[i].string;
2766                         struct stage_data *e = entries->items[i].util;
2767                         if (!e->processed) {
2768                                 int ret = process_entry(o, path, e);
2769                                 if (!ret)
2770                                         clean = 0;
2771                                 else if (ret < 0) {
2772                                         clean = ret;
2773                                         goto cleanup;
2774                                 }
2775                         }
2776                 }
2777                 for (i = 0; i < entries->nr; i++) {
2778                         struct stage_data *e = entries->items[i].util;
2779                         if (!e->processed)
2780                                 die("BUG: unprocessed path??? %s",
2781                                     entries->items[i].string);
2782                 }
2783
2784 cleanup:
2785                 final_cleanup_renames(&re_info);
2786
2787                 string_list_clear(entries, 1);
2788                 free(entries);
2789
2790                 hashmap_free(&o->current_file_dir_set, 1);
2791
2792                 if (clean < 0)
2793                         return clean;
2794         }
2795         else
2796                 clean = 1;
2797
2798         if (o->call_depth && !(*result = write_tree_from_memory(o)))
2799                 return -1;
2800
2801         return clean;
2802 }
2803
2804 static struct commit_list *reverse_commit_list(struct commit_list *list)
2805 {
2806         struct commit_list *next = NULL, *current, *backup;
2807         for (current = list; current; current = backup) {
2808                 backup = current->next;
2809                 current->next = next;
2810                 next = current;
2811         }
2812         return next;
2813 }
2814
2815 /*
2816  * Merge the commits h1 and h2, return the resulting virtual
2817  * commit object and a flag indicating the cleanness of the merge.
2818  */
2819 int merge_recursive(struct merge_options *o,
2820                     struct commit *h1,
2821                     struct commit *h2,
2822                     struct commit_list *ca,
2823                     struct commit **result)
2824 {
2825         struct commit_list *iter;
2826         struct commit *merged_common_ancestors;
2827         struct tree *mrtree;
2828         int clean;
2829
2830         if (show(o, 4)) {
2831                 output(o, 4, _("Merging:"));
2832                 output_commit_title(o, h1);
2833                 output_commit_title(o, h2);
2834         }
2835
2836         if (!ca) {
2837                 ca = get_merge_bases(h1, h2);
2838                 ca = reverse_commit_list(ca);
2839         }
2840
2841         if (show(o, 5)) {
2842                 unsigned cnt = commit_list_count(ca);
2843
2844                 output(o, 5, Q_("found %u common ancestor:",
2845                                 "found %u common ancestors:", cnt), cnt);
2846                 for (iter = ca; iter; iter = iter->next)
2847                         output_commit_title(o, iter->item);
2848         }
2849
2850         merged_common_ancestors = pop_commit(&ca);
2851         if (merged_common_ancestors == NULL) {
2852                 /* if there is no common ancestor, use an empty tree */
2853                 struct tree *tree;
2854
2855                 tree = lookup_tree(the_hash_algo->empty_tree);
2856                 merged_common_ancestors = make_virtual_commit(tree, "ancestor");
2857         }
2858
2859         for (iter = ca; iter; iter = iter->next) {
2860                 const char *saved_b1, *saved_b2;
2861                 o->call_depth++;
2862                 /*
2863                  * When the merge fails, the result contains files
2864                  * with conflict markers. The cleanness flag is
2865                  * ignored (unless indicating an error), it was never
2866                  * actually used, as result of merge_trees has always
2867                  * overwritten it: the committed "conflicts" were
2868                  * already resolved.
2869                  */
2870                 discard_cache();
2871                 saved_b1 = o->branch1;
2872                 saved_b2 = o->branch2;
2873                 o->branch1 = "Temporary merge branch 1";
2874                 o->branch2 = "Temporary merge branch 2";
2875                 if (merge_recursive(o, merged_common_ancestors, iter->item,
2876                                     NULL, &merged_common_ancestors) < 0)
2877                         return -1;
2878                 o->branch1 = saved_b1;
2879                 o->branch2 = saved_b2;
2880                 o->call_depth--;
2881
2882                 if (!merged_common_ancestors)
2883                         return err(o, _("merge returned no commit"));
2884         }
2885
2886         discard_cache();
2887         if (!o->call_depth)
2888                 read_cache();
2889
2890         o->ancestor = "merged common ancestors";
2891         clean = merge_trees(o, h1->tree, h2->tree, merged_common_ancestors->tree,
2892                             &mrtree);
2893         if (clean < 0) {
2894                 flush_output(o);
2895                 return clean;
2896         }
2897
2898         if (o->call_depth) {
2899                 *result = make_virtual_commit(mrtree, "merged tree");
2900                 commit_list_insert(h1, &(*result)->parents);
2901                 commit_list_insert(h2, &(*result)->parents->next);
2902         }
2903         flush_output(o);
2904         if (!o->call_depth && o->buffer_output < 2)
2905                 strbuf_release(&o->obuf);
2906         if (show(o, 2))
2907                 diff_warn_rename_limit("merge.renamelimit",
2908                                        o->needed_rename_limit, 0);
2909         return clean;
2910 }
2911
2912 static struct commit *get_ref(const struct object_id *oid, const char *name)
2913 {
2914         struct object *object;
2915
2916         object = deref_tag(parse_object(oid), name, strlen(name));
2917         if (!object)
2918                 return NULL;
2919         if (object->type == OBJ_TREE)
2920                 return make_virtual_commit((struct tree*)object, name);
2921         if (object->type != OBJ_COMMIT)
2922                 return NULL;
2923         if (parse_commit((struct commit *)object))
2924                 return NULL;
2925         return (struct commit *)object;
2926 }
2927
2928 int merge_recursive_generic(struct merge_options *o,
2929                             const struct object_id *head,
2930                             const struct object_id *merge,
2931                             int num_base_list,
2932                             const struct object_id **base_list,
2933                             struct commit **result)
2934 {
2935         int clean;
2936         struct lock_file lock = LOCK_INIT;
2937         struct commit *head_commit = get_ref(head, o->branch1);
2938         struct commit *next_commit = get_ref(merge, o->branch2);
2939         struct commit_list *ca = NULL;
2940
2941         if (base_list) {
2942                 int i;
2943                 for (i = 0; i < num_base_list; ++i) {
2944                         struct commit *base;
2945                         if (!(base = get_ref(base_list[i], oid_to_hex(base_list[i]))))
2946                                 return err(o, _("Could not parse object '%s'"),
2947                                         oid_to_hex(base_list[i]));
2948                         commit_list_insert(base, &ca);
2949                 }
2950         }
2951
2952         hold_locked_index(&lock, LOCK_DIE_ON_ERROR);
2953         clean = merge_recursive(o, head_commit, next_commit, ca,
2954                         result);
2955         if (clean < 0) {
2956                 rollback_lock_file(&lock);
2957                 return clean;
2958         }
2959
2960         if (write_locked_index(&the_index, &lock,
2961                                COMMIT_LOCK | SKIP_IF_UNCHANGED))
2962                 return err(o, _("Unable to write index."));
2963
2964         return clean ? 0 : 1;
2965 }
2966
2967 static void merge_recursive_config(struct merge_options *o)
2968 {
2969         git_config_get_int("merge.verbosity", &o->verbosity);
2970         git_config_get_int("diff.renamelimit", &o->diff_rename_limit);
2971         git_config_get_int("merge.renamelimit", &o->merge_rename_limit);
2972         git_config(git_xmerge_config, NULL);
2973 }
2974
2975 void init_merge_options(struct merge_options *o)
2976 {
2977         const char *merge_verbosity;
2978         memset(o, 0, sizeof(struct merge_options));
2979         o->verbosity = 2;
2980         o->buffer_output = 1;
2981         o->diff_rename_limit = -1;
2982         o->merge_rename_limit = -1;
2983         o->renormalize = 0;
2984         o->detect_rename = 1;
2985         merge_recursive_config(o);
2986         merge_verbosity = getenv("GIT_MERGE_VERBOSITY");
2987         if (merge_verbosity)
2988                 o->verbosity = strtol(merge_verbosity, NULL, 10);
2989         if (o->verbosity >= 5)
2990                 o->buffer_output = 0;
2991         strbuf_init(&o->obuf, 0);
2992         string_list_init(&o->df_conflict_file_set, 1);
2993 }
2994
2995 int parse_merge_opt(struct merge_options *o, const char *s)
2996 {
2997         const char *arg;
2998
2999         if (!s || !*s)
3000                 return -1;
3001         if (!strcmp(s, "ours"))
3002                 o->recursive_variant = MERGE_RECURSIVE_OURS;
3003         else if (!strcmp(s, "theirs"))
3004                 o->recursive_variant = MERGE_RECURSIVE_THEIRS;
3005         else if (!strcmp(s, "subtree"))
3006                 o->subtree_shift = "";
3007         else if (skip_prefix(s, "subtree=", &arg))
3008                 o->subtree_shift = arg;
3009         else if (!strcmp(s, "patience"))
3010                 o->xdl_opts = DIFF_WITH_ALG(o, PATIENCE_DIFF);
3011         else if (!strcmp(s, "histogram"))
3012                 o->xdl_opts = DIFF_WITH_ALG(o, HISTOGRAM_DIFF);
3013         else if (skip_prefix(s, "diff-algorithm=", &arg)) {
3014                 long value = parse_algorithm_value(arg);
3015                 if (value < 0)
3016                         return -1;
3017                 /* clear out previous settings */
3018                 DIFF_XDL_CLR(o, NEED_MINIMAL);
3019                 o->xdl_opts &= ~XDF_DIFF_ALGORITHM_MASK;
3020                 o->xdl_opts |= value;
3021         }
3022         else if (!strcmp(s, "ignore-space-change"))
3023                 DIFF_XDL_SET(o, IGNORE_WHITESPACE_CHANGE);
3024         else if (!strcmp(s, "ignore-all-space"))
3025                 DIFF_XDL_SET(o, IGNORE_WHITESPACE);
3026         else if (!strcmp(s, "ignore-space-at-eol"))
3027                 DIFF_XDL_SET(o, IGNORE_WHITESPACE_AT_EOL);
3028         else if (!strcmp(s, "ignore-cr-at-eol"))
3029                 DIFF_XDL_SET(o, IGNORE_CR_AT_EOL);
3030         else if (!strcmp(s, "renormalize"))
3031                 o->renormalize = 1;
3032         else if (!strcmp(s, "no-renormalize"))
3033                 o->renormalize = 0;
3034         else if (!strcmp(s, "no-renames"))
3035                 o->detect_rename = 0;
3036         else if (!strcmp(s, "find-renames")) {
3037                 o->detect_rename = 1;
3038                 o->rename_score = 0;
3039         }
3040         else if (skip_prefix(s, "find-renames=", &arg) ||
3041                  skip_prefix(s, "rename-threshold=", &arg)) {
3042                 if ((o->rename_score = parse_rename_score(&arg)) == -1 || *arg != 0)
3043                         return -1;
3044                 o->detect_rename = 1;
3045         }
3046         else
3047                 return -1;
3048         return 0;
3049 }