2 * Copyright (C) 2005 Junio C Hamano
12 static const char *diff_opts = "-pu";
13 static unsigned char null_sha1[20] = { 0, };
15 static const char *external_diff(void)
17 static const char *external_diff_cmd = NULL;
18 static int done_preparing = 0;
21 return external_diff_cmd;
24 * Default values above are meant to match the
25 * Linux kernel development style. Examples of
26 * alternative styles you can specify via environment
31 if (gitenv("GIT_EXTERNAL_DIFF"))
32 external_diff_cmd = gitenv("GIT_EXTERNAL_DIFF");
34 /* In case external diff fails... */
35 diff_opts = gitenv("GIT_DIFF_OPTS") ? : diff_opts;
38 return external_diff_cmd;
41 /* Help to copy the thing properly quoted for the shell safety.
42 * any single quote is replaced with '\'', and the caller is
43 * expected to enclose the result within a single quote pair.
46 * original sq_expand result
47 * name ==> name ==> 'name'
48 * a b ==> a b ==> 'a b'
49 * a'b ==> a'\''b ==> 'a'\''b'
51 static char *sq_expand(const char *src)
53 static char *buf = NULL;
58 /* count bytes needed to store the quoted string. */
59 for (cnt = 1, cp = src; *cp; cnt++, cp++)
65 while ((c = *src++)) {
69 bp = strcpy(bp, "'\\''");
77 static struct diff_tempfile {
85 unsigned char blob_sha1[20];
86 unsigned short mode; /* file mode */
87 unsigned sha1_valid : 1; /* if true, use blob_sha1 and trust mode;
88 * if false, use the name and read from
91 unsigned file_valid : 1; /* if false the file does not exist */
94 static void builtin_diff(const char *name_a,
96 struct diff_tempfile *temp)
98 int i, next_at, cmd_size;
99 const char *diff_cmd = "diff -L'%s%s' -L'%s%s'";
100 const char *diff_arg = "'%s' '%s'||:"; /* "||:" is to return 0 */
101 const char *input_name_sq[2];
102 const char *path0[2];
103 const char *path1[2];
104 const char *name_sq[2];
107 name_sq[0] = sq_expand(name_a);
108 name_sq[1] = sq_expand(name_b);
110 /* diff_cmd and diff_arg have 6 %s in total which makes
111 * the sum of these strings 12 bytes larger than required.
112 * we use 2 spaces around diff-opts, and we need to count
113 * terminating NUL, so we subtract 9 here.
115 cmd_size = (strlen(diff_cmd) + strlen(diff_opts) +
116 strlen(diff_arg) - 9);
117 for (i = 0; i < 2; i++) {
118 input_name_sq[i] = sq_expand(temp[i].name);
119 if (!strcmp(temp[i].name, "/dev/null")) {
120 path0[i] = "/dev/null";
123 path0[i] = i ? "b/" : "a/";
124 path1[i] = name_sq[i];
126 cmd_size += (strlen(path0[i]) + strlen(path1[i]) +
127 strlen(input_name_sq[i]));
130 cmd = xmalloc(cmd_size);
133 next_at += snprintf(cmd+next_at, cmd_size-next_at,
135 path0[0], path1[0], path0[1], path1[1]);
136 next_at += snprintf(cmd+next_at, cmd_size-next_at,
138 next_at += snprintf(cmd+next_at, cmd_size-next_at,
139 diff_arg, input_name_sq[0], input_name_sq[1]);
141 printf("diff --git a/%s b/%s\n", name_a, name_b);
143 printf("new file mode %s\n", temp[1].mode);
144 else if (!path1[1][0])
145 printf("deleted file mode %s\n", temp[0].mode);
147 if (strcmp(temp[0].mode, temp[1].mode)) {
148 printf("old mode %s\n", temp[0].mode);
149 printf("new mode %s\n", temp[1].mode);
151 if (strcmp(name_a, name_b)) {
152 printf("rename old %s\n", name_a);
153 printf("rename new %s\n", name_b);
155 if (strncmp(temp[0].mode, temp[1].mode, 3))
156 /* we do not run diff between different kind
162 execlp("/bin/sh","sh", "-c", cmd, NULL);
166 * Given a name and sha1 pair, if the dircache tells us the file in
167 * the work tree has that object contents, return true, so that
168 * prepare_temp_file() does not have to inflate and extract.
170 static int work_tree_matches(const char *name, const unsigned char *sha1)
172 struct cache_entry *ce;
176 /* We do not read the cache ourselves here, because the
177 * benchmark with my previous version that always reads cache
178 * shows that it makes things worse for diff-tree comparing
179 * two linux-2.6 kernel trees in an already checked out work
180 * tree. This is because most diff-tree comparisons deal with
181 * only a small number of files, while reading the cache is
182 * expensive for a large project, and its cost outweighs the
183 * savings we get by not inflating the object to a temporary
184 * file. Practically, this code only helps when we are used
185 * by diff-cache --cached, which does read the cache before
192 pos = cache_name_pos(name, len);
195 ce = active_cache[pos];
196 if ((lstat(name, &st) < 0) ||
197 !S_ISREG(st.st_mode) ||
198 ce_match_stat(ce, &st) ||
199 memcmp(sha1, ce->sha1, 20))
204 static void prep_temp_blob(struct diff_tempfile *temp,
212 strcpy(temp->tmp_path, ".diff_XXXXXX");
213 fd = mkstemp(temp->tmp_path);
215 die("unable to create temp-file");
216 if (write(fd, blob, size) != size)
217 die("unable to write temp-file");
219 temp->name = temp->tmp_path;
220 strcpy(temp->hex, sha1_to_hex(sha1));
222 sprintf(temp->mode, "%06o", mode);
225 static void prepare_temp_file(const char *name,
226 struct diff_tempfile *temp,
227 struct diff_spec *one)
229 if (!one->file_valid) {
231 /* A '-' entry produces this for file-2, and
232 * a '+' entry produces this for file-1.
234 temp->name = "/dev/null";
235 strcpy(temp->hex, ".");
236 strcpy(temp->mode, ".");
240 if (!one->sha1_valid ||
241 work_tree_matches(name, one->blob_sha1)) {
244 if (lstat(temp->name, &st) < 0) {
246 goto not_a_valid_file;
247 die("stat(%s): %s", temp->name, strerror(errno));
249 if (S_ISLNK(st.st_mode)) {
251 char *buf, buf_[1024];
252 buf = ((sizeof(buf_) < st.st_size) ?
253 xmalloc(st.st_size) : buf_);
254 ret = readlink(name, buf, st.st_size);
256 die("readlink(%s)", name);
257 prep_temp_blob(temp, buf, st.st_size,
259 one->blob_sha1 : null_sha1),
261 one->mode : S_IFLNK));
264 if (!one->sha1_valid)
265 strcpy(temp->hex, sha1_to_hex(null_sha1));
267 strcpy(temp->hex, sha1_to_hex(one->blob_sha1));
268 sprintf(temp->mode, "%06o",
269 S_IFREG |ce_permissions(st.st_mode));
278 blob = read_sha1_file(one->blob_sha1, type, &size);
279 if (!blob || strcmp(type, "blob"))
280 die("unable to read blob object for %s (%s)",
281 name, sha1_to_hex(one->blob_sha1));
282 prep_temp_blob(temp, blob, size, one->blob_sha1, one->mode);
287 static void remove_tempfile(void)
291 for (i = 0; i < 2; i++)
292 if (diff_temp[i].name == diff_temp[i].tmp_path) {
293 unlink(diff_temp[i].name);
294 diff_temp[i].name = NULL;
298 static void remove_tempfile_on_signal(int signo)
303 static int detect_rename;
304 static int reverse_diff;
305 static const char **pathspec;
307 static int diff_rename_minimum_score;
309 static int matches_pathspec(const char *name)
317 namelen = strlen(name);
318 for (i = 0; i < speccnt; i++) {
319 int speclen = strlen(pathspec[i]);
320 if (! strncmp(pathspec[i], name, speclen) &&
321 speclen <= namelen &&
322 (name[speclen] == 0 || name[speclen] == '/'))
328 /* An external diff command takes:
330 * diff-cmd name infile1 infile1-sha1 infile1-mode \
331 * infile2 infile2-sha1 infile2-mode [ rename-to ]
334 static void run_external_diff(const char *name,
336 struct diff_spec *one,
337 struct diff_spec *two)
339 struct diff_tempfile *temp = diff_temp;
342 static int atexit_asked = 0;
345 struct diff_spec *tmp_spec;
346 tmp_spec = one; one = two; two = tmp_spec;
349 tmp = name; name = other; other = tmp;
353 if (!matches_pathspec(name) && (!other || !matches_pathspec(other)))
357 prepare_temp_file(name, &temp[0], one);
358 prepare_temp_file(other ? : name, &temp[1], two);
359 if (! atexit_asked &&
360 (temp[0].name == temp[0].tmp_path ||
361 temp[1].name == temp[1].tmp_path)) {
363 atexit(remove_tempfile);
365 signal(SIGINT, remove_tempfile_on_signal);
371 die("unable to fork");
373 const char *pgm = external_diff();
376 const char *exec_arg[9];
377 const char **arg = &exec_arg[0];
380 *arg++ = temp[0].name;
381 *arg++ = temp[0].hex;
382 *arg++ = temp[0].mode;
383 *arg++ = temp[1].name;
384 *arg++ = temp[1].hex;
385 *arg++ = temp[1].mode;
389 execvp(pgm, (char *const*) exec_arg);
392 execlp(pgm, pgm, name, NULL);
395 * otherwise we use the built-in one.
398 builtin_diff(name, other ? : name, temp);
400 printf("* Unmerged path %s\n", name);
403 if (waitpid(pid, &status, 0) < 0 ||
404 !WIFEXITED(status) || WEXITSTATUS(status)) {
405 /* Earlier we did not check the exit status because
406 * diff exits non-zero if files are different, and
407 * we are not interested in knowing that. It was a
408 * mistake which made it harder to quit a diff-*
409 * session that uses the git-apply-patch-script as
410 * the GIT_EXTERNAL_DIFF. A custom GIT_EXTERNAL_DIFF
411 * should also exit non-zero only when it wants to
412 * abort the entire diff-* session.
415 fprintf(stderr, "external diff died, stopping at %s.\n", name);
422 * We do not detect circular renames. Just hold created and deleted
423 * entries and later attempt to match them up. If they do not match,
424 * then spit them out as deletes or creates as original.
427 static struct diff_spec_hold {
428 struct diff_spec_hold *next;
433 #define SHOULD_FREE 2
434 #define SHOULD_MUNMAP 4
437 } *createdfile, *deletedfile;
439 static void hold_diff(const char *name,
440 struct diff_spec *one,
441 struct diff_spec *two)
443 struct diff_spec_hold **list, *elem;
445 if (one->file_valid && two->file_valid)
446 die("internal error");
448 if (!detect_rename) {
449 run_external_diff(name, NULL, one, two);
452 elem = xmalloc(sizeof(*elem) + strlen(name));
453 strcpy(elem->path, name);
457 if (one->file_valid) {
469 static int populate_data(struct diff_spec_hold *s)
475 if (s->it.sha1_valid) {
476 s->data = read_sha1_file(s->it.blob_sha1, type, &s->size);
477 s->flags |= SHOULD_FREE;
482 fd = open(s->path, O_RDONLY);
485 if (fstat(fd, &st)) {
489 s->size = st.st_size;
490 s->data = mmap(NULL, s->size, PROT_READ, MAP_PRIVATE, fd, 0);
495 s->flags |= SHOULD_MUNMAP;
500 static void free_data(struct diff_spec_hold *s)
502 if (s->flags & SHOULD_FREE)
504 else if (s->flags & SHOULD_MUNMAP)
505 munmap(s->data, s->size);
506 s->flags &= ~(SHOULD_FREE|SHOULD_MUNMAP);
510 static void flush_remaining_diff(struct diff_spec_hold *elem,
513 static struct diff_spec null_file_spec;
515 null_file_spec.file_valid = 0;
516 for ( ; elem ; elem = elem->next) {
518 if (elem->flags & MATCHED)
521 run_external_diff(elem->path, NULL,
522 &null_file_spec, &elem->it);
524 run_external_diff(elem->path, NULL,
525 &elem->it, &null_file_spec);
529 static int is_exact_match(struct diff_spec_hold *src,
530 struct diff_spec_hold *dst)
532 if (src->it.sha1_valid && dst->it.sha1_valid &&
533 !memcmp(src->it.blob_sha1, dst->it.blob_sha1, 20))
535 if (populate_data(src) || populate_data(dst))
536 /* this is an error but will be caught downstream */
538 if (src->size == dst->size &&
539 !memcmp(src->data, dst->data, src->size))
544 #define MINIMUM_SCORE 5000
545 int estimate_similarity(struct diff_spec_hold *src, struct diff_spec_hold *dst)
547 /* src points at a deleted file and dst points at a created
548 * file. They may be quite similar, in which case we want to
549 * say src is renamed to dst.
551 * Compare them and return how similar they are, representing
552 * the score as an integer between 0 and 10000. 10000 is
553 * reserved for the case where they match exactly.
556 unsigned long delta_size;
558 delta_size = ((src->size < dst->size) ?
559 (dst->size - src->size) : (src->size - dst->size));
561 /* We would not consider rename followed by more than
562 * 20% edits; that is, delta_size must be smaller than
563 * (src->size + dst->size)/2 * 0.2, which means...
565 if ((src->size + dst->size) < delta_size * 10)
568 delta = diff_delta(src->data, src->size,
569 dst->data, dst->size,
573 /* This "delta" is really xdiff with adler32 and all the
574 * overheads but it is a quick and dirty approximation.
576 * Now we will give some score to it. Let's say 20% edit gets
577 * 5000 points and 0% edit gets 9000 points. That is, every
578 * 1/20000 edit gets 1 point penalty. The amount of penalty is:
580 * (delta_size * 2 / (src->size + dst->size)) * 20000
583 return 9000 - (40000 * delta_size / (src->size+dst->size));
587 struct diff_spec_hold *src;
588 struct diff_spec_hold *dst;
592 static int score_compare(const void *a_, const void *b_)
594 const struct diff_score *a = a_, *b = b_;
595 return b->score - a->score;
598 static void flush_rename_pair(struct diff_spec_hold *src,
599 struct diff_spec_hold *dst)
601 src->flags |= MATCHED;
602 dst->flags |= MATCHED;
605 run_external_diff(src->path, dst->path,
609 static void free_held_diff(struct diff_spec_hold *list)
611 struct diff_spec_hold *h;
612 for (h = list; list; list = h) {
619 void diff_flush(void)
621 int num_create, num_delete, c, d;
622 struct diff_spec_hold *elem, *src, *dst;
623 struct diff_score *mx;
625 /* We really want to cull the candidates list early
626 * with cheap tests in order to avoid doing deltas.
628 * With the current callers, we should not have already
629 * matched entries at this point, but it is nonetheless
630 * checked for sanity.
632 for (dst = createdfile; dst; dst = dst->next) {
633 if (dst->flags & MATCHED)
635 for (src = deletedfile; src; src = src->next) {
636 if (src->flags & MATCHED)
638 if (! is_exact_match(src, dst))
640 flush_rename_pair(src, dst);
645 /* Count surviving candidates */
646 for (num_create = 0, elem = createdfile; elem; elem = elem->next)
647 if (!(elem->flags & MATCHED))
650 for (num_delete = 0, elem = deletedfile; elem; elem = elem->next)
651 if (!(elem->flags & MATCHED))
654 if (num_create == 0 || num_delete == 0)
657 mx = xmalloc(sizeof(*mx) * num_create * num_delete);
658 for (c = 0, dst = createdfile; dst; dst = dst->next) {
659 int base = c * num_delete;
660 if (dst->flags & MATCHED)
662 for (d = 0, src = deletedfile; src; src = src->next) {
663 struct diff_score *m = &mx[base+d];
664 if (src->flags & MATCHED)
668 m->score = estimate_similarity(src, dst);
673 qsort(mx, num_create*num_delete, sizeof(*mx), score_compare);
676 for (c = 0; c < num_create * num_delete; c++) {
679 if ((src->flags & MATCHED) || (dst->flags & MATCHED))
682 "**score ** %d %s %s\n",
683 mx[c].score, src->path, dst->path);
687 for (c = 0; c < num_create * num_delete; c++) {
690 if ((src->flags & MATCHED) || (dst->flags & MATCHED))
692 if (mx[c].score < diff_rename_minimum_score)
694 flush_rename_pair(src, dst);
699 flush_remaining_diff(createdfile, 1);
700 flush_remaining_diff(deletedfile, 0);
701 free_held_diff(createdfile);
702 free_held_diff(deletedfile);
703 createdfile = deletedfile = NULL;
706 void diff_setup(int detect_rename_, int minimum_score_, int reverse_diff_,
707 const char **pathspec_, int speccnt_)
709 free_held_diff(createdfile);
710 free_held_diff(deletedfile);
711 createdfile = deletedfile = NULL;
713 detect_rename = detect_rename_;
714 reverse_diff = reverse_diff_;
715 pathspec = pathspec_;
717 diff_rename_minimum_score = minimum_score_ ? : MINIMUM_SCORE;
720 void diff_addremove(int addremove, unsigned mode,
721 const unsigned char *sha1,
722 const char *base, const char *path)
724 char concatpath[PATH_MAX];
725 struct diff_spec spec[2], *one, *two;
727 memcpy(spec[0].blob_sha1, sha1, 20);
729 spec[0].sha1_valid = !!memcmp(sha1, null_sha1, 20);
730 spec[0].file_valid = 1;
731 spec[1].file_valid = 0;
733 if (addremove == '+') {
734 one = spec + 1; two = spec;
736 one = spec; two = one + 1;
740 strcpy(concatpath, base);
741 strcat(concatpath, path);
743 hold_diff(path ? concatpath : base, one, two);
746 void diff_change(unsigned old_mode, unsigned new_mode,
747 const unsigned char *old_sha1,
748 const unsigned char *new_sha1,
749 const char *base, const char *path) {
750 char concatpath[PATH_MAX];
751 struct diff_spec spec[2];
754 strcpy(concatpath, base);
755 strcat(concatpath, path);
758 memcpy(spec[0].blob_sha1, old_sha1, 20);
759 spec[0].mode = old_mode;
760 memcpy(spec[1].blob_sha1, new_sha1, 20);
761 spec[1].mode = new_mode;
762 spec[0].sha1_valid = !!memcmp(old_sha1, null_sha1, 20);
763 spec[1].sha1_valid = !!memcmp(new_sha1, null_sha1, 20);
764 spec[1].file_valid = spec[0].file_valid = 1;
766 /* We do not look at changed files as candidate for
767 * rename detection ever.
769 run_external_diff(path ? concatpath : base, NULL, &spec[0], &spec[1]);
772 void diff_unmerge(const char *path)
774 run_external_diff(path, NULL, NULL, NULL);