1 /*
2 * Copyright (C) 2005 Junio C Hamano
3 */
4 #include "cache.h"
5 #include "diff.h"
6 #include "diffcore.h"
7 #include "delta.h"
9 struct diff_rename_pool {
10 struct diff_filespec **s;
11 int nr, alloc;
12 };
14 static void diff_rename_pool_clear(struct diff_rename_pool *pool)
15 {
16 pool->s = NULL; pool->nr = pool->alloc = 0;
17 }
19 static void diff_rename_pool_add(struct diff_rename_pool *pool,
20 struct diff_filespec *s)
21 {
22 if (S_ISDIR(s->mode))
23 return; /* no trees, please */
25 if (pool->alloc <= pool->nr) {
26 pool->alloc = alloc_nr(pool->alloc);
27 pool->s = xrealloc(pool->s,
28 sizeof(*(pool->s)) * pool->alloc);
29 }
30 pool->s[pool->nr] = s;
31 pool->nr++;
32 }
34 static int is_exact_match(struct diff_filespec *src, struct diff_filespec *dst)
35 {
36 if (src->sha1_valid && dst->sha1_valid &&
37 !memcmp(src->sha1, dst->sha1, 20))
38 return 1;
39 if (diff_populate_filespec(src) || diff_populate_filespec(dst))
40 /* this is an error but will be caught downstream */
41 return 0;
42 if (src->size == dst->size &&
43 !memcmp(src->data, dst->data, src->size))
44 return 1;
45 return 0;
46 }
48 struct diff_score {
49 struct diff_filespec *src;
50 struct diff_filespec *dst;
51 int score;
52 int rank;
53 };
55 static int estimate_similarity(struct diff_filespec *src,
56 struct diff_filespec *dst,
57 int minimum_score)
58 {
59 /* src points at a file that existed in the original tree (or
60 * optionally a file in the destination tree) and dst points
61 * at a newly created file. They may be quite similar, in which
62 * case we want to say src is renamed to dst or src is copied into
63 * dst, and then some edit has been applied to dst.
64 *
65 * Compare them and return how similar they are, representing
66 * the score as an integer between 0 and 10000, except
67 * where they match exactly it is considered better than anything
68 * else.
69 */
70 void *delta;
71 unsigned long delta_size, base_size;
72 int score;
74 /* We deal only with regular files. Symlink renames are handled
75 * only when they are exact matches --- in other words, no edits
76 * after renaming.
77 */
78 if (!S_ISREG(src->mode) || !S_ISREG(dst->mode))
79 return 0;
81 delta_size = ((src->size < dst->size) ?
82 (dst->size - src->size) : (src->size - dst->size));
83 base_size = ((src->size < dst->size) ? src->size : dst->size);
85 /* We would not consider edits that change the file size so
86 * drastically. delta_size must be smaller than
87 * (MAX_SCORE-minimum_score)/MAX_SCORE * min(src->size, dst->size).
88 * Note that base_size == 0 case is handled here already
89 * and the final score computation below would not have a
90 * divide-by-zero issue.
91 */
92 if (base_size * (MAX_SCORE-minimum_score) < delta_size * MAX_SCORE)
93 return 0;
95 delta = diff_delta(src->data, src->size,
96 dst->data, dst->size,
97 &delta_size);
98 /*
99 * We currently punt here, but we may later end up parsing the
100 * delta to really assess the extent of damage. A big consecutive
101 * remove would produce small delta_size that affects quite a
102 * big portion of the file.
103 */
104 free(delta);
106 /*
107 * Now we will give some score to it. 100% edit gets 0 points
108 * and 0% edit gets MAX_SCORE points.
109 */
110 score = MAX_SCORE - (MAX_SCORE * delta_size / base_size);
111 if (score < 0) return 0;
112 if (MAX_SCORE < score) return MAX_SCORE;
113 return score;
114 }
116 static void record_rename_pair(struct diff_queue_struct *outq,
117 struct diff_filespec *src,
118 struct diff_filespec *dst,
119 int rank,
120 int score)
121 {
122 /* The rank is used to sort the final output, because there
123 * are certain dependencies.
124 *
125 * - rank #0 depends on deleted ones.
126 * - rank #1 depends on kept files before they are modified.
127 * - rank #2 depends on kept files after they are modified;
128 * currently not used.
129 *
130 * Therefore, the final output order should be:
131 *
132 * 1. rank #0 rename/copy diffs.
133 * 2. deletions in the original.
134 * 3. rank #1 rename/copy diffs.
135 * 4. additions and modifications in the original.
136 * 5. rank #2 rename/copy diffs; currently not used.
137 *
138 * To achieve this sort order, we give xform_work the number
139 * above.
140 */
141 struct diff_filepair *dp = diff_queue(outq, src, dst);
142 dp->xfrm_work = (rank * 2 + 1) | (score<<RENAME_SCORE_SHIFT);
143 dst->xfrm_flags |= RENAME_DST_MATCHED;
144 }
146 #if 0
147 static void debug_filespec(struct diff_filespec *s, int x, const char *one)
148 {
149 fprintf(stderr, "queue[%d] %s (%s) %s %06o %s\n",
150 x, one,
151 s->path,
152 DIFF_FILE_VALID(s) ? "valid" : "invalid",
153 s->mode,
154 s->sha1_valid ? sha1_to_hex(s->sha1) : "");
155 fprintf(stderr, "queue[%d] %s size %lu flags %d\n",
156 x, one,
157 s->size, s->xfrm_flags);
158 }
160 static void debug_filepair(const struct diff_filepair *p, int i)
161 {
162 debug_filespec(p->one, i, "one");
163 debug_filespec(p->two, i, "two");
164 fprintf(stderr, "pair flags %d, orig order %d, score %d\n",
165 (p->xfrm_work & ((1<<RENAME_SCORE_SHIFT) - 1)),
166 p->orig_order,
167 (p->xfrm_work >> RENAME_SCORE_SHIFT));
168 }
170 static void debug_queue(const char *msg, struct diff_queue_struct *q)
171 {
172 int i;
173 if (msg)
174 fprintf(stderr, "%s\n", msg);
175 fprintf(stderr, "q->nr = %d\n", q->nr);
176 for (i = 0; i < q->nr; i++) {
177 struct diff_filepair *p = q->queue[i];
178 debug_filepair(p, i);
179 }
180 }
181 #else
182 #define debug_queue(a,b) do { ; /*nothing*/ } while(0)
183 #endif
185 /*
186 * We sort the outstanding diff entries according to the rank (see
187 * comment at the beginning of record_rename_pair) and tiebreak with
188 * the order in the original input.
189 */
190 static int rank_compare(const void *a_, const void *b_)
191 {
192 const struct diff_filepair *a = *(const struct diff_filepair **)a_;
193 const struct diff_filepair *b = *(const struct diff_filepair **)b_;
194 int a_rank = a->xfrm_work & ((1<<RENAME_SCORE_SHIFT) - 1);
195 int b_rank = b->xfrm_work & ((1<<RENAME_SCORE_SHIFT) - 1);
197 if (a_rank != b_rank)
198 return a_rank - b_rank;
199 return a->orig_order - b->orig_order;
200 }
202 /*
203 * We sort the rename similarity matrix with the score, in descending
204 * order (more similar first).
205 */
206 static int score_compare(const void *a_, const void *b_)
207 {
208 const struct diff_score *a = a_, *b = b_;
209 return b->score - a->score;
210 }
212 static int needs_to_stay(struct diff_queue_struct *q, int i,
213 struct diff_filespec *it)
214 {
215 /* If it will be used in later entry (either stay or used
216 * as the source of rename/copy), we need to copy, not rename.
217 */
218 while (i < q->nr) {
219 struct diff_filepair *p = q->queue[i++];
220 if (!DIFF_FILE_VALID(p->two))
221 continue; /* removed is fine */
222 if (strcmp(p->one->path, it->path))
223 continue; /* not relevant */
225 /* p has its src set to *it and it is not a delete;
226 * it will be used for in-place change or rename/copy,
227 * so we cannot rename it out.
228 */
229 return 1;
230 }
231 return 0;
232 }
234 int diff_scoreopt_parse(const char *opt)
235 {
236 int diglen, num, scale, i;
237 if (opt[0] != '-' || (opt[1] != 'M' && opt[1] != 'C'))
238 return -1; /* that is not a -M nor -C option */
239 diglen = strspn(opt+2, "0123456789");
240 if (diglen == 0 || strlen(opt+2) != diglen)
241 return 0; /* use default */
242 sscanf(opt+2, "%d", &num);
243 for (i = 0, scale = 1; i < diglen; i++)
244 scale *= 10;
246 /* user says num divided by scale and we say internally that
247 * is MAX_SCORE * num / scale.
248 */
249 return MAX_SCORE * num / scale;
250 }
252 void diffcore_rename(int detect_rename, int minimum_score)
253 {
254 struct diff_queue_struct *q = &diff_queued_diff;
255 struct diff_queue_struct outq;
256 struct diff_rename_pool created, deleted, stay;
257 struct diff_rename_pool *(srcs[2]);
258 struct diff_score *mx;
259 int h, i, j;
260 int num_create, num_src, dst_cnt, src_cnt;
262 if (!minimum_score)
263 minimum_score = DEFAULT_MINIMUM_SCORE;
264 outq.queue = NULL;
265 outq.nr = outq.alloc = 0;
267 diff_rename_pool_clear(&created);
268 diff_rename_pool_clear(&deleted);
269 diff_rename_pool_clear(&stay);
271 srcs[0] = &deleted;
272 srcs[1] = &stay;
274 for (i = 0; i < q->nr; i++) {
275 struct diff_filepair *p = q->queue[i];
276 if (!DIFF_FILE_VALID(p->one))
277 if (!DIFF_FILE_VALID(p->two))
278 continue; /* unmerged */
279 else
280 diff_rename_pool_add(&created, p->two);
281 else if (!DIFF_FILE_VALID(p->two))
282 diff_rename_pool_add(&deleted, p->one);
283 else if (1 < detect_rename) /* find copy, too */
284 diff_rename_pool_add(&stay, p->one);
285 }
286 if (created.nr == 0)
287 goto cleanup; /* nothing to do */
289 /* We really want to cull the candidates list early
290 * with cheap tests in order to avoid doing deltas.
291 *
292 * With the current callers, we should not have already
293 * matched entries at this point, but it is nonetheless
294 * checked for sanity.
295 */
296 for (i = 0; i < created.nr; i++) {
297 if (created.s[i]->xfrm_flags & RENAME_DST_MATCHED)
298 continue; /* we have matched exactly already */
299 for (h = 0; h < sizeof(srcs)/sizeof(srcs[0]); h++) {
300 struct diff_rename_pool *p = srcs[h];
301 for (j = 0; j < p->nr; j++) {
302 if (!is_exact_match(p->s[j], created.s[i]))
303 continue;
304 record_rename_pair(&outq,
305 p->s[j], created.s[i], h,
306 MAX_SCORE);
307 break; /* we are done with this entry */
308 }
309 }
310 }
311 debug_queue("done detecting exact", &outq);
313 /* Have we run out the created file pool? If so we can avoid
314 * doing the delta matrix altogether.
315 */
316 if (outq.nr == created.nr)
317 goto flush_rest;
319 num_create = (created.nr - outq.nr);
320 num_src = deleted.nr + stay.nr;
321 mx = xmalloc(sizeof(*mx) * num_create * num_src);
322 for (dst_cnt = i = 0; i < created.nr; i++) {
323 int base = dst_cnt * num_src;
324 if (created.s[i]->xfrm_flags & RENAME_DST_MATCHED)
325 continue; /* dealt with exact match already. */
326 for (src_cnt = h = 0; h < sizeof(srcs)/sizeof(srcs[0]); h++) {
327 struct diff_rename_pool *p = srcs[h];
328 for (j = 0; j < p->nr; j++, src_cnt++) {
329 struct diff_score *m = &mx[base + src_cnt];
330 m->src = p->s[j];
331 m->dst = created.s[i];
332 m->score = estimate_similarity(m->src, m->dst,
333 minimum_score);
334 m->rank = h;
335 }
336 }
337 dst_cnt++;
338 }
339 /* cost matrix sorted by most to least similar pair */
340 qsort(mx, num_create * num_src, sizeof(*mx), score_compare);
341 for (i = 0; i < num_create * num_src; i++) {
342 if (mx[i].dst->xfrm_flags & RENAME_DST_MATCHED)
343 continue; /* alreayd done, either exact or fuzzy. */
344 if (mx[i].score < minimum_score)
345 break; /* there is not any more diffs applicable. */
346 record_rename_pair(&outq,
347 mx[i].src, mx[i].dst, mx[i].rank,
348 mx[i].score);
349 }
350 free(mx);
351 debug_queue("done detecting fuzzy", &outq);
353 flush_rest:
354 /* At this point, we have found some renames and copies and they
355 * are kept in outq. The original list is still in *q.
356 *
357 * Scan the original list and move them into the outq; we will sort
358 * outq and swap it into the queue supplied to pass that to
359 * downstream, so we assign the sort keys in this loop.
360 *
361 * See comments at the top of record_rename_pair for numbers used
362 * to assign xfrm_work.
363 *
364 * Note that we have not annotated the diff_filepair with any comment
365 * so there is nothing other than p to free.
366 */
367 for (i = 0; i < q->nr; i++) {
368 struct diff_filepair *dp, *p = q->queue[i];
369 if (!DIFF_FILE_VALID(p->one)) {
370 /* creation or unmerged entries */
371 dp = diff_queue(&outq, p->one, p->two);
372 dp->xfrm_work = 4;
373 }
374 else if (!DIFF_FILE_VALID(p->two)) {
375 /* deletion */
376 dp = diff_queue(&outq, p->one, p->two);
377 dp->xfrm_work = 2;
378 }
379 else {
380 /* modification, or stay as is */
381 dp = diff_queue(&outq, p->one, p->two);
382 dp->xfrm_work = 4;
383 }
384 free(p);
385 }
386 debug_queue("done copying original", &outq);
388 /* Sort outq */
389 qsort(outq.queue, outq.nr, sizeof(outq.queue[0]), rank_compare);
391 debug_queue("done sorting", &outq);
393 free(q->queue);
394 q->nr = q->alloc = 0;
395 q->queue = NULL;
397 /* Copy it out to q, removing duplicates. */
398 for (i = 0; i < outq.nr; i++) {
399 struct diff_filepair *p = outq.queue[i];
400 if (!DIFF_FILE_VALID(p->one)) {
401 /* created or unmerged */
402 if (p->two->xfrm_flags & RENAME_DST_MATCHED)
403 ; /* rename/copy created it already */
404 else
405 diff_queue(q, p->one, p->two);
406 }
407 else if (!DIFF_FILE_VALID(p->two)) {
408 /* deleted */
409 if (p->one->xfrm_flags & RENAME_SRC_GONE)
410 ; /* rename/copy deleted it already */
411 else
412 diff_queue(q, p->one, p->two);
413 }
414 else if (strcmp(p->one->path, p->two->path)) {
415 /* rename or copy */
416 struct diff_filepair *dp =
417 diff_queue(q, p->one, p->two);
418 int msglen = (strlen(p->one->path) +
419 strlen(p->two->path) + 100);
420 int score = (p->xfrm_work >> RENAME_SCORE_SHIFT);
421 dp->xfrm_msg = xmalloc(msglen);
423 /* if we have a later entry that is a rename/copy
424 * that depends on p->one, then we copy here.
425 * otherwise we rename it.
426 */
427 if (needs_to_stay(&outq, i+1, p->one)) {
428 /* copy it */
429 sprintf(dp->xfrm_msg,
430 "similarity index %d%%\n"
431 "copy from %s\n"
432 "copy to %s\n",
433 (int)(0.5 + score * 100 / MAX_SCORE),
434 p->one->path, p->two->path);
435 }
436 else {
437 /* rename it, and mark it as gone. */
438 p->one->xfrm_flags |= RENAME_SRC_GONE;
439 sprintf(dp->xfrm_msg,
440 "similarity index %d%%\n"
441 "rename old %s\n"
442 "rename new %s\n",
443 (int)(0.5 + score * 100 / MAX_SCORE),
444 p->one->path, p->two->path);
445 }
446 }
447 else
448 /* otherwise it is a modified (or stayed) entry */
449 diff_queue(q, p->one, p->two);
450 diff_free_filepair(p);
451 }
453 free(outq.queue);
454 debug_queue("done collapsing", q);
456 cleanup:
457 free(created.s);
458 free(deleted.s);
459 free(stay.s);
460 return;
461 }