xref: /openbmc/linux/tools/perf/builtin-diff.c (revision 6279eb3d)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * builtin-diff.c
4  *
5  * Builtin diff command: Analyze two perf.data input files, look up and read
6  * DSOs and symbol information, sort them and produce a diff.
7  */
8 #include "builtin.h"
9 #include "perf.h"
10 
11 #include "util/debug.h"
12 #include "util/event.h"
13 #include "util/hist.h"
14 #include "util/evsel.h"
15 #include "util/evlist.h"
16 #include "util/session.h"
17 #include "util/tool.h"
18 #include "util/sort.h"
19 #include "util/srcline.h"
20 #include "util/symbol.h"
21 #include "util/data.h"
22 #include "util/config.h"
23 #include "util/time-utils.h"
24 #include "util/annotate.h"
25 #include "util/map.h"
26 #include <linux/zalloc.h>
27 #include <subcmd/pager.h>
28 #include <subcmd/parse-options.h>
29 
30 #include <errno.h>
31 #include <inttypes.h>
32 #include <stdlib.h>
33 #include <math.h>
34 
35 struct perf_diff {
36 	struct perf_tool		 tool;
37 	const char			*time_str;
38 	struct perf_time_interval	*ptime_range;
39 	int				 range_size;
40 	int				 range_num;
41 	bool				 has_br_stack;
42 };
43 
44 /* Diff command specific HPP columns. */
45 enum {
46 	PERF_HPP_DIFF__BASELINE,
47 	PERF_HPP_DIFF__PERIOD,
48 	PERF_HPP_DIFF__PERIOD_BASELINE,
49 	PERF_HPP_DIFF__DELTA,
50 	PERF_HPP_DIFF__RATIO,
51 	PERF_HPP_DIFF__WEIGHTED_DIFF,
52 	PERF_HPP_DIFF__FORMULA,
53 	PERF_HPP_DIFF__DELTA_ABS,
54 	PERF_HPP_DIFF__CYCLES,
55 
56 	PERF_HPP_DIFF__MAX_INDEX
57 };
58 
59 struct diff_hpp_fmt {
60 	struct perf_hpp_fmt	 fmt;
61 	int			 idx;
62 	char			*header;
63 	int			 header_width;
64 };
65 
66 struct data__file {
67 	struct perf_session	*session;
68 	struct perf_data	 data;
69 	int			 idx;
70 	struct hists		*hists;
71 	struct diff_hpp_fmt	 fmt[PERF_HPP_DIFF__MAX_INDEX];
72 };
73 
74 static struct data__file *data__files;
75 static int data__files_cnt;
76 
77 #define data__for_each_file_start(i, d, s)	\
78 	for (i = s, d = &data__files[s];	\
79 	     i < data__files_cnt;		\
80 	     i++, d = &data__files[i])
81 
82 #define data__for_each_file(i, d) data__for_each_file_start(i, d, 0)
83 #define data__for_each_file_new(i, d) data__for_each_file_start(i, d, 1)
84 
85 static bool force;
86 static bool show_period;
87 static bool show_formula;
88 static bool show_baseline_only;
89 static unsigned int sort_compute = 1;
90 
91 static s64 compute_wdiff_w1;
92 static s64 compute_wdiff_w2;
93 
94 static const char		*cpu_list;
95 static DECLARE_BITMAP(cpu_bitmap, MAX_NR_CPUS);
96 
97 static struct addr_location dummy_al;
98 
99 enum {
100 	COMPUTE_DELTA,
101 	COMPUTE_RATIO,
102 	COMPUTE_WEIGHTED_DIFF,
103 	COMPUTE_DELTA_ABS,
104 	COMPUTE_CYCLES,
105 	COMPUTE_MAX,
106 };
107 
108 const char *compute_names[COMPUTE_MAX] = {
109 	[COMPUTE_DELTA] = "delta",
110 	[COMPUTE_DELTA_ABS] = "delta-abs",
111 	[COMPUTE_RATIO] = "ratio",
112 	[COMPUTE_WEIGHTED_DIFF] = "wdiff",
113 	[COMPUTE_CYCLES] = "cycles",
114 };
115 
116 static int compute = COMPUTE_DELTA_ABS;
117 
118 static int compute_2_hpp[COMPUTE_MAX] = {
119 	[COMPUTE_DELTA]		= PERF_HPP_DIFF__DELTA,
120 	[COMPUTE_DELTA_ABS]	= PERF_HPP_DIFF__DELTA_ABS,
121 	[COMPUTE_RATIO]		= PERF_HPP_DIFF__RATIO,
122 	[COMPUTE_WEIGHTED_DIFF]	= PERF_HPP_DIFF__WEIGHTED_DIFF,
123 	[COMPUTE_CYCLES]	= PERF_HPP_DIFF__CYCLES,
124 };
125 
126 #define MAX_COL_WIDTH 70
127 
128 static struct header_column {
129 	const char *name;
130 	int width;
131 } columns[PERF_HPP_DIFF__MAX_INDEX] = {
132 	[PERF_HPP_DIFF__BASELINE] = {
133 		.name  = "Baseline",
134 	},
135 	[PERF_HPP_DIFF__PERIOD] = {
136 		.name  = "Period",
137 		.width = 14,
138 	},
139 	[PERF_HPP_DIFF__PERIOD_BASELINE] = {
140 		.name  = "Base period",
141 		.width = 14,
142 	},
143 	[PERF_HPP_DIFF__DELTA] = {
144 		.name  = "Delta",
145 		.width = 7,
146 	},
147 	[PERF_HPP_DIFF__DELTA_ABS] = {
148 		.name  = "Delta Abs",
149 		.width = 7,
150 	},
151 	[PERF_HPP_DIFF__RATIO] = {
152 		.name  = "Ratio",
153 		.width = 14,
154 	},
155 	[PERF_HPP_DIFF__WEIGHTED_DIFF] = {
156 		.name  = "Weighted diff",
157 		.width = 14,
158 	},
159 	[PERF_HPP_DIFF__FORMULA] = {
160 		.name  = "Formula",
161 		.width = MAX_COL_WIDTH,
162 	},
163 	[PERF_HPP_DIFF__CYCLES] = {
164 		.name  = "[Program Block Range] Cycles Diff",
165 		.width = 70,
166 	}
167 };
168 
169 static int setup_compute_opt_wdiff(char *opt)
170 {
171 	char *w1_str = opt;
172 	char *w2_str;
173 
174 	int ret = -EINVAL;
175 
176 	if (!opt)
177 		goto out;
178 
179 	w2_str = strchr(opt, ',');
180 	if (!w2_str)
181 		goto out;
182 
183 	*w2_str++ = 0x0;
184 	if (!*w2_str)
185 		goto out;
186 
187 	compute_wdiff_w1 = strtol(w1_str, NULL, 10);
188 	compute_wdiff_w2 = strtol(w2_str, NULL, 10);
189 
190 	if (!compute_wdiff_w1 || !compute_wdiff_w2)
191 		goto out;
192 
193 	pr_debug("compute wdiff w1(%" PRId64 ") w2(%" PRId64 ")\n",
194 		  compute_wdiff_w1, compute_wdiff_w2);
195 
196 	ret = 0;
197 
198  out:
199 	if (ret)
200 		pr_err("Failed: wrong weight data, use 'wdiff:w1,w2'\n");
201 
202 	return ret;
203 }
204 
205 static int setup_compute_opt(char *opt)
206 {
207 	if (compute == COMPUTE_WEIGHTED_DIFF)
208 		return setup_compute_opt_wdiff(opt);
209 
210 	if (opt) {
211 		pr_err("Failed: extra option specified '%s'", opt);
212 		return -EINVAL;
213 	}
214 
215 	return 0;
216 }
217 
218 static int setup_compute(const struct option *opt, const char *str,
219 			 int unset __maybe_unused)
220 {
221 	int *cp = (int *) opt->value;
222 	char *cstr = (char *) str;
223 	char buf[50];
224 	unsigned i;
225 	char *option;
226 
227 	if (!str) {
228 		*cp = COMPUTE_DELTA;
229 		return 0;
230 	}
231 
232 	option = strchr(str, ':');
233 	if (option) {
234 		unsigned len = option++ - str;
235 
236 		/*
237 		 * The str data are not writeable, so we need
238 		 * to use another buffer.
239 		 */
240 
241 		/* No option value is longer. */
242 		if (len >= sizeof(buf))
243 			return -EINVAL;
244 
245 		strncpy(buf, str, len);
246 		buf[len] = 0x0;
247 		cstr = buf;
248 	}
249 
250 	for (i = 0; i < COMPUTE_MAX; i++)
251 		if (!strcmp(cstr, compute_names[i])) {
252 			*cp = i;
253 			return setup_compute_opt(option);
254 		}
255 
256 	pr_err("Failed: '%s' is not computation method "
257 	       "(use 'delta','ratio' or 'wdiff')\n", str);
258 	return -EINVAL;
259 }
260 
261 static double period_percent(struct hist_entry *he, u64 period)
262 {
263 	u64 total = hists__total_period(he->hists);
264 
265 	return (period * 100.0) / total;
266 }
267 
268 static double compute_delta(struct hist_entry *he, struct hist_entry *pair)
269 {
270 	double old_percent = period_percent(he, he->stat.period);
271 	double new_percent = period_percent(pair, pair->stat.period);
272 
273 	pair->diff.period_ratio_delta = new_percent - old_percent;
274 	pair->diff.computed = true;
275 	return pair->diff.period_ratio_delta;
276 }
277 
278 static double compute_ratio(struct hist_entry *he, struct hist_entry *pair)
279 {
280 	double old_period = he->stat.period ?: 1;
281 	double new_period = pair->stat.period;
282 
283 	pair->diff.computed = true;
284 	pair->diff.period_ratio = new_period / old_period;
285 	return pair->diff.period_ratio;
286 }
287 
288 static s64 compute_wdiff(struct hist_entry *he, struct hist_entry *pair)
289 {
290 	u64 old_period = he->stat.period;
291 	u64 new_period = pair->stat.period;
292 
293 	pair->diff.computed = true;
294 	pair->diff.wdiff = new_period * compute_wdiff_w2 -
295 			   old_period * compute_wdiff_w1;
296 
297 	return pair->diff.wdiff;
298 }
299 
300 static int formula_delta(struct hist_entry *he, struct hist_entry *pair,
301 			 char *buf, size_t size)
302 {
303 	u64 he_total = he->hists->stats.total_period;
304 	u64 pair_total = pair->hists->stats.total_period;
305 
306 	if (symbol_conf.filter_relative) {
307 		he_total = he->hists->stats.total_non_filtered_period;
308 		pair_total = pair->hists->stats.total_non_filtered_period;
309 	}
310 	return scnprintf(buf, size,
311 			 "(%" PRIu64 " * 100 / %" PRIu64 ") - "
312 			 "(%" PRIu64 " * 100 / %" PRIu64 ")",
313 			 pair->stat.period, pair_total,
314 			 he->stat.period, he_total);
315 }
316 
317 static int formula_ratio(struct hist_entry *he, struct hist_entry *pair,
318 			 char *buf, size_t size)
319 {
320 	double old_period = he->stat.period;
321 	double new_period = pair->stat.period;
322 
323 	return scnprintf(buf, size, "%.0F / %.0F", new_period, old_period);
324 }
325 
326 static int formula_wdiff(struct hist_entry *he, struct hist_entry *pair,
327 			 char *buf, size_t size)
328 {
329 	u64 old_period = he->stat.period;
330 	u64 new_period = pair->stat.period;
331 
332 	return scnprintf(buf, size,
333 		  "(%" PRIu64 " * " "%" PRId64 ") - (%" PRIu64 " * " "%" PRId64 ")",
334 		  new_period, compute_wdiff_w2, old_period, compute_wdiff_w1);
335 }
336 
337 static int formula_fprintf(struct hist_entry *he, struct hist_entry *pair,
338 			   char *buf, size_t size)
339 {
340 	switch (compute) {
341 	case COMPUTE_DELTA:
342 	case COMPUTE_DELTA_ABS:
343 		return formula_delta(he, pair, buf, size);
344 	case COMPUTE_RATIO:
345 		return formula_ratio(he, pair, buf, size);
346 	case COMPUTE_WEIGHTED_DIFF:
347 		return formula_wdiff(he, pair, buf, size);
348 	default:
349 		BUG_ON(1);
350 	}
351 
352 	return -1;
353 }
354 
355 static void *block_hist_zalloc(size_t size)
356 {
357 	struct block_hist *bh;
358 
359 	bh = zalloc(size + sizeof(*bh));
360 	if (!bh)
361 		return NULL;
362 
363 	return &bh->he;
364 }
365 
366 static void block_hist_free(void *he)
367 {
368 	struct block_hist *bh;
369 
370 	bh = container_of(he, struct block_hist, he);
371 	hists__delete_entries(&bh->block_hists);
372 	free(bh);
373 }
374 
375 struct hist_entry_ops block_hist_ops = {
376 	.new    = block_hist_zalloc,
377 	.free   = block_hist_free,
378 };
379 
380 static int diff__process_sample_event(struct perf_tool *tool,
381 				      union perf_event *event,
382 				      struct perf_sample *sample,
383 				      struct evsel *evsel,
384 				      struct machine *machine)
385 {
386 	struct perf_diff *pdiff = container_of(tool, struct perf_diff, tool);
387 	struct addr_location al;
388 	struct hists *hists = evsel__hists(evsel);
389 	int ret = -1;
390 
391 	if (perf_time__ranges_skip_sample(pdiff->ptime_range, pdiff->range_num,
392 					  sample->time)) {
393 		return 0;
394 	}
395 
396 	if (machine__resolve(machine, &al, sample) < 0) {
397 		pr_warning("problem processing %d event, skipping it.\n",
398 			   event->header.type);
399 		return -1;
400 	}
401 
402 	if (cpu_list && !test_bit(sample->cpu, cpu_bitmap)) {
403 		ret = 0;
404 		goto out_put;
405 	}
406 
407 	if (compute != COMPUTE_CYCLES) {
408 		if (!hists__add_entry(hists, &al, NULL, NULL, NULL, sample,
409 				      true)) {
410 			pr_warning("problem incrementing symbol period, "
411 				   "skipping event\n");
412 			goto out_put;
413 		}
414 	} else {
415 		if (!hists__add_entry_ops(hists, &block_hist_ops, &al, NULL,
416 					  NULL, NULL, sample, true)) {
417 			pr_warning("problem incrementing symbol period, "
418 				   "skipping event\n");
419 			goto out_put;
420 		}
421 
422 		hist__account_cycles(sample->branch_stack, &al, sample, false);
423 	}
424 
425 	/*
426 	 * The total_period is updated here before going to the output
427 	 * tree since normally only the baseline hists will call
428 	 * hists__output_resort() and precompute needs the total
429 	 * period in order to sort entries by percentage delta.
430 	 */
431 	hists->stats.total_period += sample->period;
432 	if (!al.filtered)
433 		hists->stats.total_non_filtered_period += sample->period;
434 	ret = 0;
435 out_put:
436 	addr_location__put(&al);
437 	return ret;
438 }
439 
440 static struct perf_diff pdiff = {
441 	.tool = {
442 		.sample	= diff__process_sample_event,
443 		.mmap	= perf_event__process_mmap,
444 		.mmap2	= perf_event__process_mmap2,
445 		.comm	= perf_event__process_comm,
446 		.exit	= perf_event__process_exit,
447 		.fork	= perf_event__process_fork,
448 		.lost	= perf_event__process_lost,
449 		.namespaces = perf_event__process_namespaces,
450 		.ordered_events = true,
451 		.ordering_requires_timestamps = true,
452 	},
453 };
454 
455 static struct evsel *evsel_match(struct evsel *evsel,
456 				      struct evlist *evlist)
457 {
458 	struct evsel *e;
459 
460 	evlist__for_each_entry(evlist, e) {
461 		if (perf_evsel__match2(evsel, e))
462 			return e;
463 	}
464 
465 	return NULL;
466 }
467 
468 static void perf_evlist__collapse_resort(struct evlist *evlist)
469 {
470 	struct evsel *evsel;
471 
472 	evlist__for_each_entry(evlist, evsel) {
473 		struct hists *hists = evsel__hists(evsel);
474 
475 		hists__collapse_resort(hists, NULL);
476 	}
477 }
478 
479 static struct data__file *fmt_to_data_file(struct perf_hpp_fmt *fmt)
480 {
481 	struct diff_hpp_fmt *dfmt = container_of(fmt, struct diff_hpp_fmt, fmt);
482 	void *ptr = dfmt - dfmt->idx;
483 	struct data__file *d = container_of(ptr, struct data__file, fmt);
484 
485 	return d;
486 }
487 
488 static struct hist_entry*
489 get_pair_data(struct hist_entry *he, struct data__file *d)
490 {
491 	if (hist_entry__has_pairs(he)) {
492 		struct hist_entry *pair;
493 
494 		list_for_each_entry(pair, &he->pairs.head, pairs.node)
495 			if (pair->hists == d->hists)
496 				return pair;
497 	}
498 
499 	return NULL;
500 }
501 
502 static struct hist_entry*
503 get_pair_fmt(struct hist_entry *he, struct diff_hpp_fmt *dfmt)
504 {
505 	struct data__file *d = fmt_to_data_file(&dfmt->fmt);
506 
507 	return get_pair_data(he, d);
508 }
509 
510 static void hists__baseline_only(struct hists *hists)
511 {
512 	struct rb_root_cached *root;
513 	struct rb_node *next;
514 
515 	if (hists__has(hists, need_collapse))
516 		root = &hists->entries_collapsed;
517 	else
518 		root = hists->entries_in;
519 
520 	next = rb_first_cached(root);
521 	while (next != NULL) {
522 		struct hist_entry *he = rb_entry(next, struct hist_entry, rb_node_in);
523 
524 		next = rb_next(&he->rb_node_in);
525 		if (!hist_entry__next_pair(he)) {
526 			rb_erase_cached(&he->rb_node_in, root);
527 			hist_entry__delete(he);
528 		}
529 	}
530 }
531 
532 static int64_t block_cmp(struct perf_hpp_fmt *fmt __maybe_unused,
533 			 struct hist_entry *left, struct hist_entry *right)
534 {
535 	struct block_info *bi_l = left->block_info;
536 	struct block_info *bi_r = right->block_info;
537 	int cmp;
538 
539 	if (!bi_l->sym || !bi_r->sym) {
540 		if (!bi_l->sym && !bi_r->sym)
541 			return 0;
542 		else if (!bi_l->sym)
543 			return -1;
544 		else
545 			return 1;
546 	}
547 
548 	if (bi_l->sym == bi_r->sym) {
549 		if (bi_l->start == bi_r->start) {
550 			if (bi_l->end == bi_r->end)
551 				return 0;
552 			else
553 				return (int64_t)(bi_r->end - bi_l->end);
554 		} else
555 			return (int64_t)(bi_r->start - bi_l->start);
556 	} else {
557 		cmp = strcmp(bi_l->sym->name, bi_r->sym->name);
558 		return cmp;
559 	}
560 
561 	if (bi_l->sym->start != bi_r->sym->start)
562 		return (int64_t)(bi_r->sym->start - bi_l->sym->start);
563 
564 	return (int64_t)(bi_r->sym->end - bi_l->sym->end);
565 }
566 
567 static int64_t block_cycles_diff_cmp(struct hist_entry *left,
568 				     struct hist_entry *right)
569 {
570 	bool pairs_left  = hist_entry__has_pairs(left);
571 	bool pairs_right = hist_entry__has_pairs(right);
572 	s64 l, r;
573 
574 	if (!pairs_left && !pairs_right)
575 		return 0;
576 
577 	l = labs(left->diff.cycles);
578 	r = labs(right->diff.cycles);
579 	return r - l;
580 }
581 
582 static int64_t block_sort(struct perf_hpp_fmt *fmt __maybe_unused,
583 			  struct hist_entry *left, struct hist_entry *right)
584 {
585 	return block_cycles_diff_cmp(right, left);
586 }
587 
588 static void init_block_hist(struct block_hist *bh)
589 {
590 	__hists__init(&bh->block_hists, &bh->block_list);
591 	perf_hpp_list__init(&bh->block_list);
592 
593 	INIT_LIST_HEAD(&bh->block_fmt.list);
594 	INIT_LIST_HEAD(&bh->block_fmt.sort_list);
595 	bh->block_fmt.cmp = block_cmp;
596 	bh->block_fmt.sort = block_sort;
597 	perf_hpp_list__register_sort_field(&bh->block_list,
598 					   &bh->block_fmt);
599 	bh->valid = true;
600 }
601 
602 static void init_block_info(struct block_info *bi, struct symbol *sym,
603 			    struct cyc_hist *ch, int offset)
604 {
605 	bi->sym = sym;
606 	bi->start = ch->start;
607 	bi->end = offset;
608 	bi->cycles = ch->cycles;
609 	bi->cycles_aggr = ch->cycles_aggr;
610 	bi->num = ch->num;
611 	bi->num_aggr = ch->num_aggr;
612 }
613 
614 static int process_block_per_sym(struct hist_entry *he)
615 {
616 	struct annotation *notes;
617 	struct cyc_hist *ch;
618 	struct block_hist *bh;
619 
620 	if (!he->ms.map || !he->ms.sym)
621 		return 0;
622 
623 	notes = symbol__annotation(he->ms.sym);
624 	if (!notes || !notes->src || !notes->src->cycles_hist)
625 		return 0;
626 
627 	bh = container_of(he, struct block_hist, he);
628 	init_block_hist(bh);
629 
630 	ch = notes->src->cycles_hist;
631 	for (unsigned int i = 0; i < symbol__size(he->ms.sym); i++) {
632 		if (ch[i].num_aggr) {
633 			struct block_info *bi;
634 			struct hist_entry *he_block;
635 
636 			bi = block_info__new();
637 			if (!bi)
638 				return -1;
639 
640 			init_block_info(bi, he->ms.sym, &ch[i], i);
641 			he_block = hists__add_entry_block(&bh->block_hists,
642 							  &dummy_al, bi);
643 			if (!he_block) {
644 				block_info__put(bi);
645 				return -1;
646 			}
647 		}
648 	}
649 
650 	return 0;
651 }
652 
653 static int block_pair_cmp(struct hist_entry *a, struct hist_entry *b)
654 {
655 	struct block_info *bi_a = a->block_info;
656 	struct block_info *bi_b = b->block_info;
657 	int cmp;
658 
659 	if (!bi_a->sym || !bi_b->sym)
660 		return -1;
661 
662 	cmp = strcmp(bi_a->sym->name, bi_b->sym->name);
663 
664 	if ((!cmp) && (bi_a->start == bi_b->start) && (bi_a->end == bi_b->end))
665 		return 0;
666 
667 	return -1;
668 }
669 
670 static struct hist_entry *get_block_pair(struct hist_entry *he,
671 					 struct hists *hists_pair)
672 {
673 	struct rb_root_cached *root = hists_pair->entries_in;
674 	struct rb_node *next = rb_first_cached(root);
675 	int cmp;
676 
677 	while (next != NULL) {
678 		struct hist_entry *he_pair = rb_entry(next, struct hist_entry,
679 						      rb_node_in);
680 
681 		next = rb_next(&he_pair->rb_node_in);
682 
683 		cmp = block_pair_cmp(he_pair, he);
684 		if (!cmp)
685 			return he_pair;
686 	}
687 
688 	return NULL;
689 }
690 
691 static void compute_cycles_diff(struct hist_entry *he,
692 				struct hist_entry *pair)
693 {
694 	pair->diff.computed = true;
695 	if (pair->block_info->num && he->block_info->num) {
696 		pair->diff.cycles =
697 			pair->block_info->cycles_aggr / pair->block_info->num_aggr -
698 			he->block_info->cycles_aggr / he->block_info->num_aggr;
699 	}
700 }
701 
702 static void block_hists_match(struct hists *hists_base,
703 			      struct hists *hists_pair)
704 {
705 	struct rb_root_cached *root = hists_base->entries_in;
706 	struct rb_node *next = rb_first_cached(root);
707 
708 	while (next != NULL) {
709 		struct hist_entry *he = rb_entry(next, struct hist_entry,
710 						 rb_node_in);
711 		struct hist_entry *pair = get_block_pair(he, hists_pair);
712 
713 		next = rb_next(&he->rb_node_in);
714 
715 		if (pair) {
716 			hist_entry__add_pair(pair, he);
717 			compute_cycles_diff(he, pair);
718 		}
719 	}
720 }
721 
722 static int filter_cb(struct hist_entry *he, void *arg __maybe_unused)
723 {
724 	/* Skip the calculation of column length in output_resort */
725 	he->filtered = true;
726 	return 0;
727 }
728 
729 static void hists__precompute(struct hists *hists)
730 {
731 	struct rb_root_cached *root;
732 	struct rb_node *next;
733 
734 	if (hists__has(hists, need_collapse))
735 		root = &hists->entries_collapsed;
736 	else
737 		root = hists->entries_in;
738 
739 	next = rb_first_cached(root);
740 	while (next != NULL) {
741 		struct block_hist *bh, *pair_bh;
742 		struct hist_entry *he, *pair;
743 		struct data__file *d;
744 		int i;
745 
746 		he   = rb_entry(next, struct hist_entry, rb_node_in);
747 		next = rb_next(&he->rb_node_in);
748 
749 		if (compute == COMPUTE_CYCLES)
750 			process_block_per_sym(he);
751 
752 		data__for_each_file_new(i, d) {
753 			pair = get_pair_data(he, d);
754 			if (!pair)
755 				continue;
756 
757 			switch (compute) {
758 			case COMPUTE_DELTA:
759 			case COMPUTE_DELTA_ABS:
760 				compute_delta(he, pair);
761 				break;
762 			case COMPUTE_RATIO:
763 				compute_ratio(he, pair);
764 				break;
765 			case COMPUTE_WEIGHTED_DIFF:
766 				compute_wdiff(he, pair);
767 				break;
768 			case COMPUTE_CYCLES:
769 				process_block_per_sym(pair);
770 				bh = container_of(he, struct block_hist, he);
771 				pair_bh = container_of(pair, struct block_hist,
772 						       he);
773 
774 				if (bh->valid && pair_bh->valid) {
775 					block_hists_match(&bh->block_hists,
776 							  &pair_bh->block_hists);
777 					hists__output_resort_cb(&pair_bh->block_hists,
778 								NULL, filter_cb);
779 				}
780 				break;
781 			default:
782 				BUG_ON(1);
783 			}
784 		}
785 	}
786 }
787 
788 static int64_t cmp_doubles(double l, double r)
789 {
790 	if (l > r)
791 		return -1;
792 	else if (l < r)
793 		return 1;
794 	else
795 		return 0;
796 }
797 
798 static int64_t
799 __hist_entry__cmp_compute(struct hist_entry *left, struct hist_entry *right,
800 			int c)
801 {
802 	switch (c) {
803 	case COMPUTE_DELTA:
804 	{
805 		double l = left->diff.period_ratio_delta;
806 		double r = right->diff.period_ratio_delta;
807 
808 		return cmp_doubles(l, r);
809 	}
810 	case COMPUTE_DELTA_ABS:
811 	{
812 		double l = fabs(left->diff.period_ratio_delta);
813 		double r = fabs(right->diff.period_ratio_delta);
814 
815 		return cmp_doubles(l, r);
816 	}
817 	case COMPUTE_RATIO:
818 	{
819 		double l = left->diff.period_ratio;
820 		double r = right->diff.period_ratio;
821 
822 		return cmp_doubles(l, r);
823 	}
824 	case COMPUTE_WEIGHTED_DIFF:
825 	{
826 		s64 l = left->diff.wdiff;
827 		s64 r = right->diff.wdiff;
828 
829 		return r - l;
830 	}
831 	default:
832 		BUG_ON(1);
833 	}
834 
835 	return 0;
836 }
837 
838 static int64_t
839 hist_entry__cmp_compute(struct hist_entry *left, struct hist_entry *right,
840 			int c, int sort_idx)
841 {
842 	bool pairs_left  = hist_entry__has_pairs(left);
843 	bool pairs_right = hist_entry__has_pairs(right);
844 	struct hist_entry *p_right, *p_left;
845 
846 	if (!pairs_left && !pairs_right)
847 		return 0;
848 
849 	if (!pairs_left || !pairs_right)
850 		return pairs_left ? -1 : 1;
851 
852 	p_left  = get_pair_data(left,  &data__files[sort_idx]);
853 	p_right = get_pair_data(right, &data__files[sort_idx]);
854 
855 	if (!p_left && !p_right)
856 		return 0;
857 
858 	if (!p_left || !p_right)
859 		return p_left ? -1 : 1;
860 
861 	/*
862 	 * We have 2 entries of same kind, let's
863 	 * make the data comparison.
864 	 */
865 	return __hist_entry__cmp_compute(p_left, p_right, c);
866 }
867 
868 static int64_t
869 hist_entry__cmp_compute_idx(struct hist_entry *left, struct hist_entry *right,
870 			    int c, int sort_idx)
871 {
872 	struct hist_entry *p_right, *p_left;
873 
874 	p_left  = get_pair_data(left,  &data__files[sort_idx]);
875 	p_right = get_pair_data(right, &data__files[sort_idx]);
876 
877 	if (!p_left && !p_right)
878 		return 0;
879 
880 	if (!p_left || !p_right)
881 		return p_left ? -1 : 1;
882 
883 	if (c != COMPUTE_DELTA && c != COMPUTE_DELTA_ABS) {
884 		/*
885 		 * The delta can be computed without the baseline, but
886 		 * others are not.  Put those entries which have no
887 		 * values below.
888 		 */
889 		if (left->dummy && right->dummy)
890 			return 0;
891 
892 		if (left->dummy || right->dummy)
893 			return left->dummy ? 1 : -1;
894 	}
895 
896 	return __hist_entry__cmp_compute(p_left, p_right, c);
897 }
898 
899 static int64_t
900 hist_entry__cmp_nop(struct perf_hpp_fmt *fmt __maybe_unused,
901 		    struct hist_entry *left __maybe_unused,
902 		    struct hist_entry *right __maybe_unused)
903 {
904 	return 0;
905 }
906 
907 static int64_t
908 hist_entry__cmp_baseline(struct perf_hpp_fmt *fmt __maybe_unused,
909 			 struct hist_entry *left, struct hist_entry *right)
910 {
911 	if (left->stat.period == right->stat.period)
912 		return 0;
913 	return left->stat.period > right->stat.period ? 1 : -1;
914 }
915 
916 static int64_t
917 hist_entry__cmp_delta(struct perf_hpp_fmt *fmt,
918 		      struct hist_entry *left, struct hist_entry *right)
919 {
920 	struct data__file *d = fmt_to_data_file(fmt);
921 
922 	return hist_entry__cmp_compute(right, left, COMPUTE_DELTA, d->idx);
923 }
924 
925 static int64_t
926 hist_entry__cmp_delta_abs(struct perf_hpp_fmt *fmt,
927 		      struct hist_entry *left, struct hist_entry *right)
928 {
929 	struct data__file *d = fmt_to_data_file(fmt);
930 
931 	return hist_entry__cmp_compute(right, left, COMPUTE_DELTA_ABS, d->idx);
932 }
933 
934 static int64_t
935 hist_entry__cmp_ratio(struct perf_hpp_fmt *fmt,
936 		      struct hist_entry *left, struct hist_entry *right)
937 {
938 	struct data__file *d = fmt_to_data_file(fmt);
939 
940 	return hist_entry__cmp_compute(right, left, COMPUTE_RATIO, d->idx);
941 }
942 
943 static int64_t
944 hist_entry__cmp_wdiff(struct perf_hpp_fmt *fmt,
945 		      struct hist_entry *left, struct hist_entry *right)
946 {
947 	struct data__file *d = fmt_to_data_file(fmt);
948 
949 	return hist_entry__cmp_compute(right, left, COMPUTE_WEIGHTED_DIFF, d->idx);
950 }
951 
952 static int64_t
953 hist_entry__cmp_delta_idx(struct perf_hpp_fmt *fmt __maybe_unused,
954 			  struct hist_entry *left, struct hist_entry *right)
955 {
956 	return hist_entry__cmp_compute_idx(right, left, COMPUTE_DELTA,
957 					   sort_compute);
958 }
959 
960 static int64_t
961 hist_entry__cmp_delta_abs_idx(struct perf_hpp_fmt *fmt __maybe_unused,
962 			      struct hist_entry *left, struct hist_entry *right)
963 {
964 	return hist_entry__cmp_compute_idx(right, left, COMPUTE_DELTA_ABS,
965 					   sort_compute);
966 }
967 
968 static int64_t
969 hist_entry__cmp_ratio_idx(struct perf_hpp_fmt *fmt __maybe_unused,
970 			  struct hist_entry *left, struct hist_entry *right)
971 {
972 	return hist_entry__cmp_compute_idx(right, left, COMPUTE_RATIO,
973 					   sort_compute);
974 }
975 
976 static int64_t
977 hist_entry__cmp_wdiff_idx(struct perf_hpp_fmt *fmt __maybe_unused,
978 			  struct hist_entry *left, struct hist_entry *right)
979 {
980 	return hist_entry__cmp_compute_idx(right, left, COMPUTE_WEIGHTED_DIFF,
981 					   sort_compute);
982 }
983 
984 static void hists__process(struct hists *hists)
985 {
986 	if (show_baseline_only)
987 		hists__baseline_only(hists);
988 
989 	hists__precompute(hists);
990 	hists__output_resort(hists, NULL);
991 
992 	if (compute == COMPUTE_CYCLES)
993 		symbol_conf.report_block = true;
994 
995 	hists__fprintf(hists, !quiet, 0, 0, 0, stdout,
996 		       !symbol_conf.use_callchain);
997 }
998 
999 static void data__fprintf(void)
1000 {
1001 	struct data__file *d;
1002 	int i;
1003 
1004 	fprintf(stdout, "# Data files:\n");
1005 
1006 	data__for_each_file(i, d)
1007 		fprintf(stdout, "#  [%d] %s %s\n",
1008 			d->idx, d->data.path,
1009 			!d->idx ? "(Baseline)" : "");
1010 
1011 	fprintf(stdout, "#\n");
1012 }
1013 
1014 static void data_process(void)
1015 {
1016 	struct evlist *evlist_base = data__files[0].session->evlist;
1017 	struct evsel *evsel_base;
1018 	bool first = true;
1019 
1020 	evlist__for_each_entry(evlist_base, evsel_base) {
1021 		struct hists *hists_base = evsel__hists(evsel_base);
1022 		struct data__file *d;
1023 		int i;
1024 
1025 		data__for_each_file_new(i, d) {
1026 			struct evlist *evlist = d->session->evlist;
1027 			struct evsel *evsel;
1028 			struct hists *hists;
1029 
1030 			evsel = evsel_match(evsel_base, evlist);
1031 			if (!evsel)
1032 				continue;
1033 
1034 			hists = evsel__hists(evsel);
1035 			d->hists = hists;
1036 
1037 			hists__match(hists_base, hists);
1038 
1039 			if (!show_baseline_only)
1040 				hists__link(hists_base, hists);
1041 		}
1042 
1043 		if (!quiet) {
1044 			fprintf(stdout, "%s# Event '%s'\n#\n", first ? "" : "\n",
1045 				perf_evsel__name(evsel_base));
1046 		}
1047 
1048 		first = false;
1049 
1050 		if (verbose > 0 || ((data__files_cnt > 2) && !quiet))
1051 			data__fprintf();
1052 
1053 		/* Don't sort callchain for perf diff */
1054 		perf_evsel__reset_sample_bit(evsel_base, CALLCHAIN);
1055 
1056 		hists__process(hists_base);
1057 	}
1058 }
1059 
1060 static void data__free(struct data__file *d)
1061 {
1062 	int col;
1063 
1064 	for (col = 0; col < PERF_HPP_DIFF__MAX_INDEX; col++) {
1065 		struct diff_hpp_fmt *fmt = &d->fmt[col];
1066 
1067 		zfree(&fmt->header);
1068 	}
1069 }
1070 
1071 static int abstime_str_dup(char **pstr)
1072 {
1073 	char *str = NULL;
1074 
1075 	if (pdiff.time_str && strchr(pdiff.time_str, ':')) {
1076 		str = strdup(pdiff.time_str);
1077 		if (!str)
1078 			return -ENOMEM;
1079 	}
1080 
1081 	*pstr = str;
1082 	return 0;
1083 }
1084 
1085 static int parse_absolute_time(struct data__file *d, char **pstr)
1086 {
1087 	char *p = *pstr;
1088 	int ret;
1089 
1090 	/*
1091 	 * Absolute timestamp for one file has the format: a.b,c.d
1092 	 * For multiple files, the format is: a.b,c.d:a.b,c.d
1093 	 */
1094 	p = strchr(*pstr, ':');
1095 	if (p) {
1096 		if (p == *pstr) {
1097 			pr_err("Invalid time string\n");
1098 			return -EINVAL;
1099 		}
1100 
1101 		*p = 0;
1102 		p++;
1103 		if (*p == 0) {
1104 			pr_err("Invalid time string\n");
1105 			return -EINVAL;
1106 		}
1107 	}
1108 
1109 	ret = perf_time__parse_for_ranges(*pstr, d->session,
1110 					  &pdiff.ptime_range,
1111 					  &pdiff.range_size,
1112 					  &pdiff.range_num);
1113 	if (ret < 0)
1114 		return ret;
1115 
1116 	if (!p || *p == 0)
1117 		*pstr = NULL;
1118 	else
1119 		*pstr = p;
1120 
1121 	return ret;
1122 }
1123 
1124 static int parse_percent_time(struct data__file *d)
1125 {
1126 	int ret;
1127 
1128 	ret = perf_time__parse_for_ranges(pdiff.time_str, d->session,
1129 					  &pdiff.ptime_range,
1130 					  &pdiff.range_size,
1131 					  &pdiff.range_num);
1132 	return ret;
1133 }
1134 
1135 static int parse_time_str(struct data__file *d, char *abstime_ostr,
1136 			   char **pabstime_tmp)
1137 {
1138 	int ret = 0;
1139 
1140 	if (abstime_ostr)
1141 		ret = parse_absolute_time(d, pabstime_tmp);
1142 	else if (pdiff.time_str)
1143 		ret = parse_percent_time(d);
1144 
1145 	return ret;
1146 }
1147 
1148 static int check_file_brstack(void)
1149 {
1150 	struct data__file *d;
1151 	bool has_br_stack;
1152 	int i;
1153 
1154 	data__for_each_file(i, d) {
1155 		d->session = perf_session__new(&d->data, false, &pdiff.tool);
1156 		if (!d->session) {
1157 			pr_err("Failed to open %s\n", d->data.path);
1158 			return -1;
1159 		}
1160 
1161 		has_br_stack = perf_header__has_feat(&d->session->header,
1162 						     HEADER_BRANCH_STACK);
1163 		perf_session__delete(d->session);
1164 		if (!has_br_stack)
1165 			return 0;
1166 	}
1167 
1168 	/* Set only all files having branch stacks */
1169 	pdiff.has_br_stack = true;
1170 	return 0;
1171 }
1172 
1173 static int __cmd_diff(void)
1174 {
1175 	struct data__file *d;
1176 	int ret, i;
1177 	char *abstime_ostr, *abstime_tmp;
1178 
1179 	ret = abstime_str_dup(&abstime_ostr);
1180 	if (ret)
1181 		return ret;
1182 
1183 	abstime_tmp = abstime_ostr;
1184 	ret = -EINVAL;
1185 
1186 	data__for_each_file(i, d) {
1187 		d->session = perf_session__new(&d->data, false, &pdiff.tool);
1188 		if (!d->session) {
1189 			pr_err("Failed to open %s\n", d->data.path);
1190 			ret = -1;
1191 			goto out_delete;
1192 		}
1193 
1194 		if (pdiff.time_str) {
1195 			ret = parse_time_str(d, abstime_ostr, &abstime_tmp);
1196 			if (ret < 0)
1197 				goto out_delete;
1198 		}
1199 
1200 		if (cpu_list) {
1201 			ret = perf_session__cpu_bitmap(d->session, cpu_list,
1202 						       cpu_bitmap);
1203 			if (ret < 0)
1204 				goto out_delete;
1205 		}
1206 
1207 		ret = perf_session__process_events(d->session);
1208 		if (ret) {
1209 			pr_err("Failed to process %s\n", d->data.path);
1210 			goto out_delete;
1211 		}
1212 
1213 		perf_evlist__collapse_resort(d->session->evlist);
1214 
1215 		if (pdiff.ptime_range)
1216 			zfree(&pdiff.ptime_range);
1217 	}
1218 
1219 	data_process();
1220 
1221  out_delete:
1222 	data__for_each_file(i, d) {
1223 		perf_session__delete(d->session);
1224 		data__free(d);
1225 	}
1226 
1227 	free(data__files);
1228 
1229 	if (pdiff.ptime_range)
1230 		zfree(&pdiff.ptime_range);
1231 
1232 	if (abstime_ostr)
1233 		free(abstime_ostr);
1234 
1235 	return ret;
1236 }
1237 
1238 static const char * const diff_usage[] = {
1239 	"perf diff [<options>] [old_file] [new_file]",
1240 	NULL,
1241 };
1242 
1243 static const struct option options[] = {
1244 	OPT_INCR('v', "verbose", &verbose,
1245 		    "be more verbose (show symbol address, etc)"),
1246 	OPT_BOOLEAN('q', "quiet", &quiet, "Do not show any message"),
1247 	OPT_BOOLEAN('b', "baseline-only", &show_baseline_only,
1248 		    "Show only items with match in baseline"),
1249 	OPT_CALLBACK('c', "compute", &compute,
1250 		     "delta,delta-abs,ratio,wdiff:w1,w2 (default delta-abs),cycles",
1251 		     "Entries differential computation selection",
1252 		     setup_compute),
1253 	OPT_BOOLEAN('p', "period", &show_period,
1254 		    "Show period values."),
1255 	OPT_BOOLEAN('F', "formula", &show_formula,
1256 		    "Show formula."),
1257 	OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1258 		    "dump raw trace in ASCII"),
1259 	OPT_BOOLEAN('f', "force", &force, "don't complain, do it"),
1260 	OPT_STRING(0, "kallsyms", &symbol_conf.kallsyms_name,
1261 		   "file", "kallsyms pathname"),
1262 	OPT_BOOLEAN('m', "modules", &symbol_conf.use_modules,
1263 		    "load module symbols - WARNING: use only with -k and LIVE kernel"),
1264 	OPT_STRING('d', "dsos", &symbol_conf.dso_list_str, "dso[,dso...]",
1265 		   "only consider symbols in these dsos"),
1266 	OPT_STRING('C', "comms", &symbol_conf.comm_list_str, "comm[,comm...]",
1267 		   "only consider symbols in these comms"),
1268 	OPT_STRING('S', "symbols", &symbol_conf.sym_list_str, "symbol[,symbol...]",
1269 		   "only consider these symbols"),
1270 	OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
1271 		   "sort by key(s): pid, comm, dso, symbol, parent, cpu, srcline, ..."
1272 		   " Please refer the man page for the complete list."),
1273 	OPT_STRING_NOEMPTY('t', "field-separator", &symbol_conf.field_sep, "separator",
1274 		   "separator for columns, no spaces will be added between "
1275 		   "columns '.' is reserved."),
1276 	OPT_CALLBACK(0, "symfs", NULL, "directory",
1277 		     "Look for files with symbols relative to this directory",
1278 		     symbol__config_symfs),
1279 	OPT_UINTEGER('o', "order", &sort_compute, "Specify compute sorting."),
1280 	OPT_CALLBACK(0, "percentage", NULL, "relative|absolute",
1281 		     "How to display percentage of filtered entries", parse_filter_percentage),
1282 	OPT_STRING(0, "time", &pdiff.time_str, "str",
1283 		   "Time span (time percent or absolute timestamp)"),
1284 	OPT_STRING(0, "cpu", &cpu_list, "cpu", "list of cpus to profile"),
1285 	OPT_STRING(0, "pid", &symbol_conf.pid_list_str, "pid[,pid...]",
1286 		   "only consider symbols in these pids"),
1287 	OPT_STRING(0, "tid", &symbol_conf.tid_list_str, "tid[,tid...]",
1288 		   "only consider symbols in these tids"),
1289 	OPT_END()
1290 };
1291 
1292 static double baseline_percent(struct hist_entry *he)
1293 {
1294 	u64 total = hists__total_period(he->hists);
1295 
1296 	return 100.0 * he->stat.period / total;
1297 }
1298 
1299 static int hpp__color_baseline(struct perf_hpp_fmt *fmt,
1300 			       struct perf_hpp *hpp, struct hist_entry *he)
1301 {
1302 	struct diff_hpp_fmt *dfmt =
1303 		container_of(fmt, struct diff_hpp_fmt, fmt);
1304 	double percent = baseline_percent(he);
1305 	char pfmt[20] = " ";
1306 
1307 	if (!he->dummy) {
1308 		scnprintf(pfmt, 20, "%%%d.2f%%%%", dfmt->header_width - 1);
1309 		return percent_color_snprintf(hpp->buf, hpp->size,
1310 					      pfmt, percent);
1311 	} else
1312 		return scnprintf(hpp->buf, hpp->size, "%*s",
1313 				 dfmt->header_width, pfmt);
1314 }
1315 
1316 static int hpp__entry_baseline(struct hist_entry *he, char *buf, size_t size)
1317 {
1318 	double percent = baseline_percent(he);
1319 	const char *fmt = symbol_conf.field_sep ? "%.2f" : "%6.2f%%";
1320 	int ret = 0;
1321 
1322 	if (!he->dummy)
1323 		ret = scnprintf(buf, size, fmt, percent);
1324 
1325 	return ret;
1326 }
1327 
1328 static int cycles_printf(struct hist_entry *he, struct hist_entry *pair,
1329 			 struct perf_hpp *hpp, int width)
1330 {
1331 	struct block_hist *bh = container_of(he, struct block_hist, he);
1332 	struct block_hist *bh_pair = container_of(pair, struct block_hist, he);
1333 	struct hist_entry *block_he;
1334 	struct block_info *bi;
1335 	char buf[128];
1336 	char *start_line, *end_line;
1337 
1338 	block_he = hists__get_entry(&bh_pair->block_hists, bh->block_idx);
1339 	if (!block_he) {
1340 		hpp->skip = true;
1341 		return 0;
1342 	}
1343 
1344 	/*
1345 	 * Avoid printing the warning "addr2line_init failed for ..."
1346 	 */
1347 	symbol_conf.disable_add2line_warn = true;
1348 
1349 	bi = block_he->block_info;
1350 
1351 	start_line = map__srcline(he->ms.map, bi->sym->start + bi->start,
1352 				  he->ms.sym);
1353 
1354 	end_line = map__srcline(he->ms.map, bi->sym->start + bi->end,
1355 				he->ms.sym);
1356 
1357 	if ((start_line != SRCLINE_UNKNOWN) && (end_line != SRCLINE_UNKNOWN)) {
1358 		scnprintf(buf, sizeof(buf), "[%s -> %s] %4ld",
1359 			  start_line, end_line, block_he->diff.cycles);
1360 	} else {
1361 		scnprintf(buf, sizeof(buf), "[%7lx -> %7lx] %4ld",
1362 			  bi->start, bi->end, block_he->diff.cycles);
1363 	}
1364 
1365 	free_srcline(start_line);
1366 	free_srcline(end_line);
1367 
1368 	return scnprintf(hpp->buf, hpp->size, "%*s", width, buf);
1369 }
1370 
1371 static int __hpp__color_compare(struct perf_hpp_fmt *fmt,
1372 				struct perf_hpp *hpp, struct hist_entry *he,
1373 				int comparison_method)
1374 {
1375 	struct diff_hpp_fmt *dfmt =
1376 		container_of(fmt, struct diff_hpp_fmt, fmt);
1377 	struct hist_entry *pair = get_pair_fmt(he, dfmt);
1378 	double diff;
1379 	s64 wdiff;
1380 	char pfmt[20] = " ";
1381 
1382 	if (!pair) {
1383 		if (comparison_method == COMPUTE_CYCLES) {
1384 			struct block_hist *bh;
1385 
1386 			bh = container_of(he, struct block_hist, he);
1387 			if (bh->block_idx)
1388 				hpp->skip = true;
1389 		}
1390 
1391 		goto no_print;
1392 	}
1393 
1394 	switch (comparison_method) {
1395 	case COMPUTE_DELTA:
1396 		if (pair->diff.computed)
1397 			diff = pair->diff.period_ratio_delta;
1398 		else
1399 			diff = compute_delta(he, pair);
1400 
1401 		scnprintf(pfmt, 20, "%%%+d.2f%%%%", dfmt->header_width - 1);
1402 		return percent_color_snprintf(hpp->buf, hpp->size,
1403 					pfmt, diff);
1404 	case COMPUTE_RATIO:
1405 		if (he->dummy)
1406 			goto dummy_print;
1407 		if (pair->diff.computed)
1408 			diff = pair->diff.period_ratio;
1409 		else
1410 			diff = compute_ratio(he, pair);
1411 
1412 		scnprintf(pfmt, 20, "%%%d.6f", dfmt->header_width);
1413 		return value_color_snprintf(hpp->buf, hpp->size,
1414 					pfmt, diff);
1415 	case COMPUTE_WEIGHTED_DIFF:
1416 		if (he->dummy)
1417 			goto dummy_print;
1418 		if (pair->diff.computed)
1419 			wdiff = pair->diff.wdiff;
1420 		else
1421 			wdiff = compute_wdiff(he, pair);
1422 
1423 		scnprintf(pfmt, 20, "%%14ld", dfmt->header_width);
1424 		return color_snprintf(hpp->buf, hpp->size,
1425 				get_percent_color(wdiff),
1426 				pfmt, wdiff);
1427 	case COMPUTE_CYCLES:
1428 		return cycles_printf(he, pair, hpp, dfmt->header_width);
1429 	default:
1430 		BUG_ON(1);
1431 	}
1432 dummy_print:
1433 	return scnprintf(hpp->buf, hpp->size, "%*s",
1434 			dfmt->header_width, "N/A");
1435 no_print:
1436 	return scnprintf(hpp->buf, hpp->size, "%*s",
1437 			dfmt->header_width, pfmt);
1438 }
1439 
1440 static int hpp__color_delta(struct perf_hpp_fmt *fmt,
1441 			struct perf_hpp *hpp, struct hist_entry *he)
1442 {
1443 	return __hpp__color_compare(fmt, hpp, he, COMPUTE_DELTA);
1444 }
1445 
1446 static int hpp__color_ratio(struct perf_hpp_fmt *fmt,
1447 			struct perf_hpp *hpp, struct hist_entry *he)
1448 {
1449 	return __hpp__color_compare(fmt, hpp, he, COMPUTE_RATIO);
1450 }
1451 
1452 static int hpp__color_wdiff(struct perf_hpp_fmt *fmt,
1453 			struct perf_hpp *hpp, struct hist_entry *he)
1454 {
1455 	return __hpp__color_compare(fmt, hpp, he, COMPUTE_WEIGHTED_DIFF);
1456 }
1457 
1458 static int hpp__color_cycles(struct perf_hpp_fmt *fmt,
1459 			     struct perf_hpp *hpp, struct hist_entry *he)
1460 {
1461 	return __hpp__color_compare(fmt, hpp, he, COMPUTE_CYCLES);
1462 }
1463 
1464 static void
1465 hpp__entry_unpair(struct hist_entry *he, int idx, char *buf, size_t size)
1466 {
1467 	switch (idx) {
1468 	case PERF_HPP_DIFF__PERIOD_BASELINE:
1469 		scnprintf(buf, size, "%" PRIu64, he->stat.period);
1470 		break;
1471 
1472 	default:
1473 		break;
1474 	}
1475 }
1476 
1477 static void
1478 hpp__entry_pair(struct hist_entry *he, struct hist_entry *pair,
1479 		int idx, char *buf, size_t size)
1480 {
1481 	double diff;
1482 	double ratio;
1483 	s64 wdiff;
1484 
1485 	switch (idx) {
1486 	case PERF_HPP_DIFF__DELTA:
1487 	case PERF_HPP_DIFF__DELTA_ABS:
1488 		if (pair->diff.computed)
1489 			diff = pair->diff.period_ratio_delta;
1490 		else
1491 			diff = compute_delta(he, pair);
1492 
1493 		scnprintf(buf, size, "%+4.2F%%", diff);
1494 		break;
1495 
1496 	case PERF_HPP_DIFF__RATIO:
1497 		/* No point for ratio number if we are dummy.. */
1498 		if (he->dummy) {
1499 			scnprintf(buf, size, "N/A");
1500 			break;
1501 		}
1502 
1503 		if (pair->diff.computed)
1504 			ratio = pair->diff.period_ratio;
1505 		else
1506 			ratio = compute_ratio(he, pair);
1507 
1508 		if (ratio > 0.0)
1509 			scnprintf(buf, size, "%14.6F", ratio);
1510 		break;
1511 
1512 	case PERF_HPP_DIFF__WEIGHTED_DIFF:
1513 		/* No point for wdiff number if we are dummy.. */
1514 		if (he->dummy) {
1515 			scnprintf(buf, size, "N/A");
1516 			break;
1517 		}
1518 
1519 		if (pair->diff.computed)
1520 			wdiff = pair->diff.wdiff;
1521 		else
1522 			wdiff = compute_wdiff(he, pair);
1523 
1524 		if (wdiff != 0)
1525 			scnprintf(buf, size, "%14ld", wdiff);
1526 		break;
1527 
1528 	case PERF_HPP_DIFF__FORMULA:
1529 		formula_fprintf(he, pair, buf, size);
1530 		break;
1531 
1532 	case PERF_HPP_DIFF__PERIOD:
1533 		scnprintf(buf, size, "%" PRIu64, pair->stat.period);
1534 		break;
1535 
1536 	default:
1537 		BUG_ON(1);
1538 	};
1539 }
1540 
1541 static void
1542 __hpp__entry_global(struct hist_entry *he, struct diff_hpp_fmt *dfmt,
1543 		    char *buf, size_t size)
1544 {
1545 	struct hist_entry *pair = get_pair_fmt(he, dfmt);
1546 	int idx = dfmt->idx;
1547 
1548 	/* baseline is special */
1549 	if (idx == PERF_HPP_DIFF__BASELINE)
1550 		hpp__entry_baseline(he, buf, size);
1551 	else {
1552 		if (pair)
1553 			hpp__entry_pair(he, pair, idx, buf, size);
1554 		else
1555 			hpp__entry_unpair(he, idx, buf, size);
1556 	}
1557 }
1558 
1559 static int hpp__entry_global(struct perf_hpp_fmt *_fmt, struct perf_hpp *hpp,
1560 			     struct hist_entry *he)
1561 {
1562 	struct diff_hpp_fmt *dfmt =
1563 		container_of(_fmt, struct diff_hpp_fmt, fmt);
1564 	char buf[MAX_COL_WIDTH] = " ";
1565 
1566 	__hpp__entry_global(he, dfmt, buf, MAX_COL_WIDTH);
1567 
1568 	if (symbol_conf.field_sep)
1569 		return scnprintf(hpp->buf, hpp->size, "%s", buf);
1570 	else
1571 		return scnprintf(hpp->buf, hpp->size, "%*s",
1572 				 dfmt->header_width, buf);
1573 }
1574 
1575 static int hpp__header(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp,
1576 		       struct hists *hists __maybe_unused,
1577 		       int line __maybe_unused,
1578 		       int *span __maybe_unused)
1579 {
1580 	struct diff_hpp_fmt *dfmt =
1581 		container_of(fmt, struct diff_hpp_fmt, fmt);
1582 
1583 	BUG_ON(!dfmt->header);
1584 	return scnprintf(hpp->buf, hpp->size, dfmt->header);
1585 }
1586 
1587 static int hpp__width(struct perf_hpp_fmt *fmt,
1588 		      struct perf_hpp *hpp __maybe_unused,
1589 		      struct hists *hists __maybe_unused)
1590 {
1591 	struct diff_hpp_fmt *dfmt =
1592 		container_of(fmt, struct diff_hpp_fmt, fmt);
1593 
1594 	BUG_ON(dfmt->header_width <= 0);
1595 	return dfmt->header_width;
1596 }
1597 
1598 static void init_header(struct data__file *d, struct diff_hpp_fmt *dfmt)
1599 {
1600 #define MAX_HEADER_NAME 100
1601 	char buf_indent[MAX_HEADER_NAME];
1602 	char buf[MAX_HEADER_NAME];
1603 	const char *header = NULL;
1604 	int width = 0;
1605 
1606 	BUG_ON(dfmt->idx >= PERF_HPP_DIFF__MAX_INDEX);
1607 	header = columns[dfmt->idx].name;
1608 	width  = columns[dfmt->idx].width;
1609 
1610 	/* Only our defined HPP fmts should appear here. */
1611 	BUG_ON(!header);
1612 
1613 	if (data__files_cnt > 2)
1614 		scnprintf(buf, MAX_HEADER_NAME, "%s/%d", header, d->idx);
1615 
1616 #define NAME (data__files_cnt > 2 ? buf : header)
1617 	dfmt->header_width = width;
1618 	width = (int) strlen(NAME);
1619 	if (dfmt->header_width < width)
1620 		dfmt->header_width = width;
1621 
1622 	scnprintf(buf_indent, MAX_HEADER_NAME, "%*s",
1623 		  dfmt->header_width, NAME);
1624 
1625 	dfmt->header = strdup(buf_indent);
1626 #undef MAX_HEADER_NAME
1627 #undef NAME
1628 }
1629 
1630 static void data__hpp_register(struct data__file *d, int idx)
1631 {
1632 	struct diff_hpp_fmt *dfmt = &d->fmt[idx];
1633 	struct perf_hpp_fmt *fmt = &dfmt->fmt;
1634 
1635 	dfmt->idx = idx;
1636 
1637 	fmt->header = hpp__header;
1638 	fmt->width  = hpp__width;
1639 	fmt->entry  = hpp__entry_global;
1640 	fmt->cmp    = hist_entry__cmp_nop;
1641 	fmt->collapse = hist_entry__cmp_nop;
1642 
1643 	/* TODO more colors */
1644 	switch (idx) {
1645 	case PERF_HPP_DIFF__BASELINE:
1646 		fmt->color = hpp__color_baseline;
1647 		fmt->sort  = hist_entry__cmp_baseline;
1648 		break;
1649 	case PERF_HPP_DIFF__DELTA:
1650 		fmt->color = hpp__color_delta;
1651 		fmt->sort  = hist_entry__cmp_delta;
1652 		break;
1653 	case PERF_HPP_DIFF__RATIO:
1654 		fmt->color = hpp__color_ratio;
1655 		fmt->sort  = hist_entry__cmp_ratio;
1656 		break;
1657 	case PERF_HPP_DIFF__WEIGHTED_DIFF:
1658 		fmt->color = hpp__color_wdiff;
1659 		fmt->sort  = hist_entry__cmp_wdiff;
1660 		break;
1661 	case PERF_HPP_DIFF__DELTA_ABS:
1662 		fmt->color = hpp__color_delta;
1663 		fmt->sort  = hist_entry__cmp_delta_abs;
1664 		break;
1665 	case PERF_HPP_DIFF__CYCLES:
1666 		fmt->color = hpp__color_cycles;
1667 		fmt->sort  = hist_entry__cmp_nop;
1668 		break;
1669 	default:
1670 		fmt->sort  = hist_entry__cmp_nop;
1671 		break;
1672 	}
1673 
1674 	init_header(d, dfmt);
1675 	perf_hpp__column_register(fmt);
1676 	perf_hpp__register_sort_field(fmt);
1677 }
1678 
1679 static int ui_init(void)
1680 {
1681 	struct data__file *d;
1682 	struct perf_hpp_fmt *fmt;
1683 	int i;
1684 
1685 	data__for_each_file(i, d) {
1686 
1687 		/*
1688 		 * Baseline or compute realted columns:
1689 		 *
1690 		 *   PERF_HPP_DIFF__BASELINE
1691 		 *   PERF_HPP_DIFF__DELTA
1692 		 *   PERF_HPP_DIFF__RATIO
1693 		 *   PERF_HPP_DIFF__WEIGHTED_DIFF
1694 		 */
1695 		data__hpp_register(d, i ? compute_2_hpp[compute] :
1696 					  PERF_HPP_DIFF__BASELINE);
1697 
1698 		/*
1699 		 * And the rest:
1700 		 *
1701 		 * PERF_HPP_DIFF__FORMULA
1702 		 * PERF_HPP_DIFF__PERIOD
1703 		 * PERF_HPP_DIFF__PERIOD_BASELINE
1704 		 */
1705 		if (show_formula && i)
1706 			data__hpp_register(d, PERF_HPP_DIFF__FORMULA);
1707 
1708 		if (show_period)
1709 			data__hpp_register(d, i ? PERF_HPP_DIFF__PERIOD :
1710 						  PERF_HPP_DIFF__PERIOD_BASELINE);
1711 	}
1712 
1713 	if (!sort_compute)
1714 		return 0;
1715 
1716 	/*
1717 	 * Prepend an fmt to sort on columns at 'sort_compute' first.
1718 	 * This fmt is added only to the sort list but not to the
1719 	 * output fields list.
1720 	 *
1721 	 * Note that this column (data) can be compared twice - one
1722 	 * for this 'sort_compute' fmt and another for the normal
1723 	 * diff_hpp_fmt.  But it shouldn't a problem as most entries
1724 	 * will be sorted out by first try or baseline and comparing
1725 	 * is not a costly operation.
1726 	 */
1727 	fmt = zalloc(sizeof(*fmt));
1728 	if (fmt == NULL) {
1729 		pr_err("Memory allocation failed\n");
1730 		return -1;
1731 	}
1732 
1733 	fmt->cmp      = hist_entry__cmp_nop;
1734 	fmt->collapse = hist_entry__cmp_nop;
1735 
1736 	switch (compute) {
1737 	case COMPUTE_DELTA:
1738 		fmt->sort = hist_entry__cmp_delta_idx;
1739 		break;
1740 	case COMPUTE_RATIO:
1741 		fmt->sort = hist_entry__cmp_ratio_idx;
1742 		break;
1743 	case COMPUTE_WEIGHTED_DIFF:
1744 		fmt->sort = hist_entry__cmp_wdiff_idx;
1745 		break;
1746 	case COMPUTE_DELTA_ABS:
1747 		fmt->sort = hist_entry__cmp_delta_abs_idx;
1748 		break;
1749 	case COMPUTE_CYCLES:
1750 		/*
1751 		 * Should set since 'fmt->sort' is called without
1752 		 * checking valid during sorting
1753 		 */
1754 		fmt->sort = hist_entry__cmp_nop;
1755 		break;
1756 	default:
1757 		BUG_ON(1);
1758 	}
1759 
1760 	perf_hpp__prepend_sort_field(fmt);
1761 	return 0;
1762 }
1763 
1764 static int data_init(int argc, const char **argv)
1765 {
1766 	struct data__file *d;
1767 	static const char *defaults[] = {
1768 		"perf.data.old",
1769 		"perf.data",
1770 	};
1771 	bool use_default = true;
1772 	int i;
1773 
1774 	data__files_cnt = 2;
1775 
1776 	if (argc) {
1777 		if (argc == 1)
1778 			defaults[1] = argv[0];
1779 		else {
1780 			data__files_cnt = argc;
1781 			use_default = false;
1782 		}
1783 	} else if (perf_guest) {
1784 		defaults[0] = "perf.data.host";
1785 		defaults[1] = "perf.data.guest";
1786 	}
1787 
1788 	if (sort_compute >= (unsigned int) data__files_cnt) {
1789 		pr_err("Order option out of limit.\n");
1790 		return -EINVAL;
1791 	}
1792 
1793 	data__files = zalloc(sizeof(*data__files) * data__files_cnt);
1794 	if (!data__files)
1795 		return -ENOMEM;
1796 
1797 	data__for_each_file(i, d) {
1798 		struct perf_data *data = &d->data;
1799 
1800 		data->path  = use_default ? defaults[i] : argv[i];
1801 		data->mode  = PERF_DATA_MODE_READ,
1802 		data->force = force,
1803 
1804 		d->idx  = i;
1805 	}
1806 
1807 	return 0;
1808 }
1809 
1810 static int diff__config(const char *var, const char *value,
1811 			void *cb __maybe_unused)
1812 {
1813 	if (!strcmp(var, "diff.order")) {
1814 		int ret;
1815 		if (perf_config_int(&ret, var, value) < 0)
1816 			return -1;
1817 		sort_compute = ret;
1818 		return 0;
1819 	}
1820 	if (!strcmp(var, "diff.compute")) {
1821 		if (!strcmp(value, "delta")) {
1822 			compute = COMPUTE_DELTA;
1823 		} else if (!strcmp(value, "delta-abs")) {
1824 			compute = COMPUTE_DELTA_ABS;
1825 		} else if (!strcmp(value, "ratio")) {
1826 			compute = COMPUTE_RATIO;
1827 		} else if (!strcmp(value, "wdiff")) {
1828 			compute = COMPUTE_WEIGHTED_DIFF;
1829 		} else {
1830 			pr_err("Invalid compute method: %s\n", value);
1831 			return -1;
1832 		}
1833 	}
1834 
1835 	return 0;
1836 }
1837 
1838 int cmd_diff(int argc, const char **argv)
1839 {
1840 	int ret = hists__init();
1841 
1842 	if (ret < 0)
1843 		return ret;
1844 
1845 	perf_config(diff__config, NULL);
1846 
1847 	argc = parse_options(argc, argv, options, diff_usage, 0);
1848 
1849 	if (quiet)
1850 		perf_quiet_option();
1851 
1852 	symbol__annotation_init();
1853 
1854 	if (symbol__init(NULL) < 0)
1855 		return -1;
1856 
1857 	if (data_init(argc, argv) < 0)
1858 		return -1;
1859 
1860 	if (check_file_brstack() < 0)
1861 		return -1;
1862 
1863 	if (compute == COMPUTE_CYCLES && !pdiff.has_br_stack)
1864 		return -1;
1865 
1866 	if (ui_init() < 0)
1867 		return -1;
1868 
1869 	sort__mode = SORT_MODE__DIFF;
1870 
1871 	if (setup_sorting(NULL) < 0)
1872 		usage_with_options(diff_usage, options);
1873 
1874 	setup_pager();
1875 
1876 	sort__setup_elide(NULL);
1877 
1878 	return __cmd_diff();
1879 }
1880