1 /* Implementation of the Patience Diff algorithm invented by Bram Cohen:
2  * Divide a diff problem into smaller chunks by an LCS of common-unique lines. */
3 /*
4  * Copyright (c) 2020 Neels Hofmeyr <neels@hofmeyr.de>
5  *
6  * Permission to use, copy, modify, and distribute this software for any
7  * purpose with or without fee is hereby granted, provided that the above
8  * copyright notice and this permission notice appear in all copies.
9  *
10  * THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES
11  * WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF
12  * MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR
13  * ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES
14  * WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN
15  * ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF
16  * OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.
17  */
18 
19 #include <assert.h>
20 #include <inttypes.h>
21 #include <errno.h>
22 #include <stdbool.h>
23 #include <stdio.h>
24 #include <stdlib.h>
25 
26 #include <arraylist.h>
27 #include <diff_main.h>
28 
29 #include "diff_internal.h"
30 #include "diff_debug.h"
31 
32 /* Per-atom state for the Patience Diff algorithm */
33 struct atom_patience {
34 	bool unique_in_both;
35 	struct diff_atom *pos_in_other;
36 	struct diff_atom *prev_stack;
37 	struct diff_range identical_lines;
38 };
39 
40 /* A diff_atom has a backpointer to the root diff_data. That points to the
41  * current diff_data, a possibly smaller section of the root. That current
42  * diff_data->algo_data is a pointer to an array of struct atom_patience. The
43  * atom's index in current diff_data gives the index in the atom_patience array.
44  */
45 #define PATIENCE(ATOM) \
46 	(((struct atom_patience*)((ATOM)->root->current->algo_data))\
47 	 [diff_atom_idx((ATOM)->root->current, ATOM)])
48 
49 int diff_atoms_qsort_compar(const void *_a, const void *_b)
50 {
51 	const struct diff_atom *a = *(struct diff_atom**)_a;
52 	const struct diff_atom *b = *(struct diff_atom**)_b;
53 	int cmp;
54 	int rc = 0;
55 
56 	/* If there's been an error (e.g. I/O error) in a previous compar, we
57 	 * have no way to abort the qsort but just report the rc and stop
58 	 * comparing. Make sure to catch errors on either side. If atoms are
59 	 * from more than one diff_data, make sure the error, if any, spreads
60 	 * to all of them, so we can cut short all future comparisons. */
61 	if (a->root->err)
62 		rc = a->root->err;
63 	if (b->root->err)
64 		rc = b->root->err;
65 	if (rc) {
66 		a->root->err = rc;
67 		b->root->err = rc;
68 		/* just return 'equal' to not swap more positions */
69 		return 0;
70 	}
71 
72 	/* Sort by the simplistic hash */
73 	if (a->hash < b->hash)
74 		return -1;
75 	if (a->hash > b->hash)
76 		return 1;
77 
78 	/* If hashes are the same, the lines may still differ. Do a full cmp. */
79 	rc = diff_atom_cmp(&cmp, a, b);
80 
81 	if (rc) {
82 		/* Mark the I/O error so that the caller can find out about it.
83 		 * For the case atoms are from more than one diff_data, mark in
84 		 * both. */
85 		a->root->err = rc;
86 		if (a->root != b->root)
87 			b->root->err = rc;
88 		return 0;
89 	}
90 
91 	return cmp;
92 }
93 
94 /* Sort an array of struct diff_atom* in-place. */
95 static int diff_atoms_qsort(struct diff_atom *atoms[],
96 			    size_t atoms_count)
97 {
98 	qsort(atoms, atoms_count, sizeof(struct diff_atom*),
99 	      diff_atoms_qsort_compar);
100 	return atoms[0]->root->err;
101 }
102 
103 static int
104 diff_atoms_mark_unique_in_both_by_qsort(struct diff_data *left,
105 					struct diff_data *right,
106 					unsigned int *unique_in_both_count_p)
107 {
108 	struct diff_atom *a;
109 	struct diff_atom *b;
110 	struct diff_atom **all_atoms =
111 		malloc((left->atoms.len + right->atoms.len)
112 		       * sizeof(struct diff_atom*));
113 	unsigned int len = 0;
114 	unsigned int i;
115 	unsigned int unique_in_both_count = 0;
116 	int rc;
117 	left->err = 0;
118 	right->err = 0;
119 	left->root->err = 0;
120 	right->root->err = 0;
121 	diff_data_foreach_atom(a, left) {
122 		all_atoms[len++] = a;
123 	}
124 	diff_data_foreach_atom(b, right) {
125 		all_atoms[len++] = b;
126 	}
127 
128 	rc = diff_atoms_qsort(all_atoms, len);
129 	if (rc)
130 		goto free_and_exit;
131 
132 	/* Now we have a sorted array of atom pointers. All similar lines are
133 	 * adjacent. Walk through the array and mark those that are unique on
134 	 * each side, but exist once in both sources. */
135 	for (i = 0; i < len; i++) {
136 		bool same;
137 		unsigned int j;
138 		unsigned int count_first_side = 1;
139 		unsigned int count_other_side = 0;
140 		a = all_atoms[i];
141 
142 		for (j = i+1; j < len; j++) {
143 			b = all_atoms[j];
144 			rc = diff_atom_same(&same, a, b);
145 			if (rc)
146 				goto free_and_exit;
147 			if (!same)
148 				break;
149 			/* A following atom is the same. See on which side the
150 			 * repetition counts. */
151 			if (a->root == b->root)
152 				count_first_side ++;
153 			else
154 				count_other_side ++;
155 		}
156 
157 		/* Counted a section of similar atoms, put the results back to
158 		 * the atoms. */
159 		if ((count_first_side == 1)
160 		    && (count_other_side == 1)) {
161 			b = all_atoms[i+1];
162 			PATIENCE(a).unique_in_both = true;
163 			PATIENCE(a).pos_in_other = b;
164 			PATIENCE(b).unique_in_both = true;
165 			PATIENCE(b).pos_in_other = a;
166 			unique_in_both_count++;
167 		}
168 	}
169 	*unique_in_both_count_p = unique_in_both_count;
170 	rc = 0;
171 free_and_exit:
172 	free(all_atoms);
173 	return rc;
174 }
175 
176 static int
177 diff_atoms_swallow_identical_neighbors(struct diff_data *left,
178 				       struct diff_data *right,
179 				       unsigned int *unique_in_both_count)
180 {
181 	debug("trivially combine identical lines"
182 	      " around unique_in_both lines\n");
183 
184 	unsigned int l_idx;
185 	unsigned int next_l_idx;
186 	unsigned int l_min = 0;
187 	unsigned int r_min = 0;
188 	for (l_idx = 0; l_idx < left->atoms.len; l_idx = next_l_idx) {
189 		next_l_idx = l_idx + 1;
190 		struct diff_atom *l = &left->atoms.head[l_idx];
191 
192 		if (!PATIENCE(l).unique_in_both)
193 			continue;
194 
195 		debug("check identical lines around ");
196 		debug_dump_atom(left, right, l);
197 
198 		unsigned int r_idx = diff_atom_idx(right, PATIENCE(l).pos_in_other);
199 
200 		struct diff_range identical_l;
201 		struct diff_range identical_r;
202 
203 		/* Swallow upwards.
204 		 * Each common-unique line swallows identical lines upwards and
205 		 * downwards.
206 		 * All common-unique lines that were part of the identical lines
207 		 * following below were already swallowed in the previous
208 		 * iteration, so we will never hit another common-unique line
209 		 * above. */
210 		for (identical_l.start = l_idx, identical_r.start = r_idx;
211 		     identical_l.start > l_min && identical_r.start > r_min;
212 		     identical_l.start--, identical_r.start--) {
213 			bool same;
214 			int r = diff_atom_same(&same,
215 				&left->atoms.head[identical_l.start - 1],
216 				&right->atoms.head[identical_r.start - 1]);
217 			if (r)
218 				return r;
219 			if (!same)
220 				break;
221 		}
222 
223 		/* Swallow downwards */
224 		for (identical_l.end = l_idx + 1, identical_r.end = r_idx + 1;
225 		     identical_l.end < left->atoms.len
226 			 && identical_r.end < right->atoms.len;
227 		     identical_l.end++, identical_r.end++,
228 		     next_l_idx++) {
229 			struct diff_atom *l_end;
230 			struct diff_atom *r_end;
231 			bool same;
232 			int r = diff_atom_same(&same,
233 					&left->atoms.head[identical_l.end],
234 					&right->atoms.head[identical_r.end]);
235 			if (r)
236 				return r;
237 			if (!same)
238 				break;
239 			l_end = &left->atoms.head[identical_l.end];
240 			r_end = &right->atoms.head[identical_r.end];
241 			if (!PATIENCE(l_end).unique_in_both)
242 				continue;
243 			/* Part of a chunk of identical lines, remove from
244 			 * listing of unique_in_both lines */
245 			PATIENCE(l_end).unique_in_both = false;
246 			PATIENCE(r_end).unique_in_both = false;
247 			(*unique_in_both_count)--;
248 		}
249 
250 		PATIENCE(l).identical_lines = identical_l;
251 		PATIENCE(PATIENCE(l).pos_in_other).identical_lines =
252 			identical_r;
253 
254 		l_min = identical_l.end;
255 		r_min = identical_r.end;
256 
257 		if (!diff_range_empty(&PATIENCE(l).identical_lines)) {
258 			debug("common-unique line at l=%u r=%u swallowed"
259 			      " identical lines l=%u-%u r=%u-%u\n",
260 			      l_idx, r_idx,
261 			      identical_l.start, identical_l.end,
262 			      identical_r.start, identical_r.end);
263 		}
264 		debug("next_l_idx = %u\n", next_l_idx);
265 	}
266 	return 0;
267 }
268 
269 /* binary search to find the stack to put this atom "card" on. */
270 static int
271 find_target_stack(struct diff_atom *atom,
272 		  struct diff_atom **patience_stacks,
273 		  unsigned int patience_stacks_count)
274 {
275 	unsigned int lo = 0;
276 	unsigned int hi = patience_stacks_count;
277 	while (lo < hi) {
278 		unsigned int mid = (lo + hi) >> 1;
279 
280 		if (PATIENCE(patience_stacks[mid]).pos_in_other
281 		    < PATIENCE(atom).pos_in_other)
282 			lo = mid + 1;
283 		else
284 			hi = mid;
285 	}
286 	return lo;
287 }
288 
289 /* Among the lines that appear exactly once in each side, find the longest
290  * streak that appear in both files in the same order (with other stuff allowed
291  * to interleave). Use patience sort for that, as in the Patience Diff
292  * algorithm.
293  * See https://bramcohen.livejournal.com/73318.html and, for a much more
294  * detailed explanation,
295  * https://blog.jcoglan.com/2017/09/19/the-patience-diff-algorithm/ */
296 int
297 diff_algo_patience(const struct diff_algo_config *algo_config,
298 		   struct diff_state *state)
299 {
300 	int rc;
301 	struct diff_data *left = &state->left;
302 	struct diff_data *right = &state->right;
303 	struct atom_patience *atom_patience_left =
304 		calloc(left->atoms.len, sizeof(struct atom_patience));
305 	struct atom_patience *atom_patience_right =
306 		calloc(right->atoms.len, sizeof(struct atom_patience));
307 	unsigned int unique_in_both_count;
308 	struct diff_atom **lcs = NULL;
309 
310 	debug("\n** %s\n", __func__);
311 
312 	left->root->current = left;
313 	right->root->current = right;
314 	left->algo_data = atom_patience_left;
315 	right->algo_data = atom_patience_right;
316 
317 	/* Find those lines that appear exactly once in 'left' and exactly once
318 	 * in 'right'. */
319 	rc = diff_atoms_mark_unique_in_both_by_qsort(left, right,
320 						     &unique_in_both_count);
321 	if (rc)
322 		goto free_and_exit;
323 
324 	debug("unique_in_both_count %u\n", unique_in_both_count);
325 	debug("left:\n");
326 	debug_dump(left);
327 	debug("right:\n");
328 	debug_dump(right);
329 
330 	if (!unique_in_both_count) {
331 		/* Cannot apply Patience, tell the caller to use fallback_algo
332 		 * instead. */
333 		rc = DIFF_RC_USE_DIFF_ALGO_FALLBACK;
334 		goto free_and_exit;
335 	}
336 
337 	rc = diff_atoms_swallow_identical_neighbors(left, right,
338 						    &unique_in_both_count);
339 	if (rc)
340 		goto free_and_exit;
341 	debug("After swallowing identical neighbors: unique_in_both = %u\n",
342 	      unique_in_both_count);
343 
344 	rc = ENOMEM;
345 
346 	/* An array of Longest Common Sequence is the result of the below
347 	 * subscope: */
348 	unsigned int lcs_count = 0;
349 	struct diff_atom *lcs_tail = NULL;
350 
351 	{
352 		/* This subscope marks the lifetime of the atom_pointers
353 		 * allocation */
354 
355 		/* One chunk of storage for atom pointers */
356 		struct diff_atom **atom_pointers;
357 		atom_pointers = recallocarray(NULL, 0, unique_in_both_count * 2,
358 					      sizeof(struct diff_atom*));
359 
360 		/* Half for the list of atoms that still need to be put on
361 		 * stacks */
362 		struct diff_atom **uniques = atom_pointers;
363 
364 		/* Half for the patience sort state's "card stacks" -- we
365 		 * remember only each stack's topmost "card" */
366 		struct diff_atom **patience_stacks;
367 		patience_stacks = atom_pointers + unique_in_both_count;
368 		unsigned int patience_stacks_count = 0;
369 
370 		/* Take all common, unique items from 'left' ... */
371 
372 		struct diff_atom *atom;
373 		struct diff_atom **uniques_end = uniques;
374 		diff_data_foreach_atom(atom, left) {
375 			if (!PATIENCE(atom).unique_in_both)
376 				continue;
377 			*uniques_end = atom;
378 			uniques_end++;
379 		}
380 
381 		/* ...and sort them to the order found in 'right'.
382 		 * The idea is to find the leftmost stack that has a higher line
383 		 * number and add it to the stack's top.
384 		 * If there is no such stack, open a new one on the right. The
385 		 * line number is derived from the atom*, which are array items
386 		 * and hence reflect the relative position in the source file.
387 		 * So we got the common-uniques from 'left' and sort them
388 		 * according to PATIENCE(atom).pos_in_other. */
389 		unsigned int i;
390 		for (i = 0; i < unique_in_both_count; i++) {
391 			atom = uniques[i];
392 			unsigned int target_stack;
393 			target_stack = find_target_stack(atom, patience_stacks,
394 							 patience_stacks_count);
395 			assert(target_stack <= patience_stacks_count);
396 			patience_stacks[target_stack] = atom;
397 			if (target_stack == patience_stacks_count)
398 				patience_stacks_count++;
399 
400 			/* Record a back reference to the next stack on the
401 			 * left, which will form the final longest sequence
402 			 * later. */
403 			PATIENCE(atom).prev_stack = target_stack ?
404 				patience_stacks[target_stack - 1] : NULL;
405 
406 		}
407 
408 		/* backtrace through prev_stack references to form the final
409 		 * longest common sequence */
410 		lcs_tail = patience_stacks[patience_stacks_count - 1];
411 		lcs_count = patience_stacks_count;
412 
413 		/* uniques and patience_stacks are no longer needed.
414 		 * Backpointers are in PATIENCE(atom).prev_stack */
415 		free(atom_pointers);
416 	}
417 
418 	lcs = recallocarray(NULL, 0, lcs_count, sizeof(struct diff_atom*));
419 	struct diff_atom **lcs_backtrace_pos = &lcs[lcs_count - 1];
420 	struct diff_atom *atom;
421 	for (atom = lcs_tail; atom; atom = PATIENCE(atom).prev_stack, lcs_backtrace_pos--) {
422 		assert(lcs_backtrace_pos >= lcs);
423 		*lcs_backtrace_pos = atom;
424 	}
425 
426 	unsigned int i;
427 	if (DEBUG) {
428 		debug("\npatience LCS:\n");
429 		for (i = 0; i < lcs_count; i++) {
430 			debug_dump_atom(left, right, lcs[i]);
431 		}
432 	}
433 
434 
435 	/* TODO: For each common-unique line found (now listed in lcs), swallow
436 	 * lines upwards and downwards that are identical on each side. Requires
437 	 * a way to represent atoms being glued to adjacent atoms. */
438 
439 	debug("\ntraverse LCS, possibly recursing:\n");
440 
441 	/* Now we have pinned positions in both files at which it makes sense to
442 	 * divide the diff problem into smaller chunks. Go into the next round:
443 	 * look at each section in turn, trying to again find common-unique
444 	 * lines in those smaller sections. As soon as no more are found, the
445 	 * remaining smaller sections are solved by Myers. */
446 	unsigned int left_pos = 0;
447 	unsigned int right_pos = 0;
448 	for (i = 0; i <= lcs_count; i++) {
449 		struct diff_atom *atom;
450 		struct diff_atom *atom_r;
451 		unsigned int left_idx;
452 		unsigned int right_idx;
453 
454 		if (i < lcs_count) {
455 			atom = lcs[i];
456 			atom_r = PATIENCE(atom).pos_in_other;
457 			debug("lcs[%u] = left[%u] = right[%u]\n", i,
458 			      diff_atom_idx(left, atom), diff_atom_idx(right, atom_r));
459 			left_idx = PATIENCE(atom).identical_lines.start;
460 			right_idx = PATIENCE(atom_r).identical_lines.start;
461 			debug(" identical lines l %u-%u  r %u-%u\n",
462 			      PATIENCE(atom).identical_lines.start, PATIENCE(atom).identical_lines.end,
463 			      PATIENCE(atom_r).identical_lines.start, PATIENCE(atom_r).identical_lines.end);
464 		} else {
465 			atom = NULL;
466 			atom_r = NULL;
467 			left_idx = left->atoms.len;
468 			right_idx = right->atoms.len;
469 		}
470 
471 		/* 'atom' now marks an atom that matches on both sides according
472 		 * to patience-diff (a common-unique identical atom in both
473 		 * files).
474 		 * Handle the section before and the atom itself; the section
475 		 * after will be handled by the next loop iteration -- note that
476 		 * i loops to last element + 1 ("i <= lcs_count"), so that there
477 		 * will be another final iteration to pick up the last remaining
478 		 * items after the last LCS atom.
479 		 * The sections before might also be empty on left and/or right.
480 		 * left_pos and right_pos mark the indexes of the first atoms
481 		 * that have not yet been handled in the previous loop
482 		 * iteration.  left_idx and right_idx mark the indexes of the
483 		 * matching atom on left and right, respectively. */
484 
485 		debug("iteration %u  left_pos %u  left_idx %u"
486 		      "  right_pos %u  right_idx %u\n",
487 		      i, left_pos, left_idx, right_pos, right_idx);
488 
489 		/* Section before the matching atom */
490 		struct diff_atom *left_atom = &left->atoms.head[left_pos];
491 		unsigned int left_section_len = left_idx - left_pos;
492 
493 		struct diff_atom *right_atom = &(right->atoms.head[right_pos]);
494 		unsigned int right_section_len = right_idx - right_pos;
495 
496 		if (left_section_len && right_section_len) {
497 			/* Record an unsolved chunk, the caller will apply
498 			 * inner_algo() on this chunk. */
499 			if (!diff_state_add_chunk(state, false,
500 						  left_atom, left_section_len,
501 						  right_atom,
502 						  right_section_len))
503 				goto free_and_exit;
504 		} else if (left_section_len && !right_section_len) {
505 			/* Only left atoms and none on the right, they form a
506 			 * "minus" chunk, then. */
507 			if (!diff_state_add_chunk(state, true,
508 						  left_atom, left_section_len,
509 						  right_atom, 0))
510 				goto free_and_exit;
511 		} else if (!left_section_len && right_section_len) {
512 			/* No left atoms, only atoms on the right, they form a
513 			 * "plus" chunk, then. */
514 			if (!diff_state_add_chunk(state, true,
515 						  left_atom, 0,
516 						  right_atom, right_section_len))
517 				goto free_and_exit;
518 		}
519 		/* else: left_section_len == 0 and right_section_len == 0, i.e.
520 		 * nothing here. */
521 
522 		/* The atom found to match on both sides forms a chunk of equals
523 		 * on each side. In the very last iteration of this loop, there
524 		 * is no matching atom, we were just cleaning out the remaining
525 		 * lines. */
526 		if (atom) {
527 			void *ok;
528 			ok = diff_state_add_chunk(state, true,
529 				left->atoms.head
530 				    + PATIENCE(atom).identical_lines.start,
531 				diff_range_len(&PATIENCE(atom).identical_lines),
532 				right->atoms.head
533 				    + PATIENCE(atom_r).identical_lines.start,
534 				diff_range_len(&PATIENCE(atom_r).identical_lines));
535 			if (!ok)
536 				goto free_and_exit;
537 			left_pos = PATIENCE(atom).identical_lines.end;
538 			right_pos = PATIENCE(atom_r).identical_lines.end;
539 		} else {
540 			left_pos = left_idx + 1;
541 			right_pos = right_idx + 1;
542 		}
543 		debug("end of iteration %u  left_pos %u  left_idx %u"
544 		      "  right_pos %u  right_idx %u\n",
545 		      i, left_pos, left_idx, right_pos, right_idx);
546 	}
547 	debug("** END %s\n", __func__);
548 
549 	rc = DIFF_RC_OK;
550 
551 free_and_exit:
552 	left->root->current = NULL;
553 	right->root->current = NULL;
554 	free(atom_patience_left);
555 	free(atom_patience_right);
556 	if (lcs)
557 		free(lcs);
558 	return rc;
559 }