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 "util/spark.h"
  27#include "util/block-info.h"
  28#include "util/stream.h"
  29#include <linux/err.h>
  30#include <linux/zalloc.h>
  31#include <subcmd/pager.h>
  32#include <subcmd/parse-options.h>
  33
  34#include <errno.h>
  35#include <inttypes.h>
  36#include <stdlib.h>
  37#include <math.h>
  38
  39struct perf_diff {
  40	struct perf_tool		 tool;
  41	const char			*time_str;
  42	struct perf_time_interval	*ptime_range;
  43	int				 range_size;
  44	int				 range_num;
  45	bool				 has_br_stack;
  46	bool				 stream;
  47};
  48
  49/* Diff command specific HPP columns. */
  50enum {
  51	PERF_HPP_DIFF__BASELINE,
  52	PERF_HPP_DIFF__PERIOD,
  53	PERF_HPP_DIFF__PERIOD_BASELINE,
  54	PERF_HPP_DIFF__DELTA,
  55	PERF_HPP_DIFF__RATIO,
  56	PERF_HPP_DIFF__WEIGHTED_DIFF,
  57	PERF_HPP_DIFF__FORMULA,
  58	PERF_HPP_DIFF__DELTA_ABS,
  59	PERF_HPP_DIFF__CYCLES,
  60	PERF_HPP_DIFF__CYCLES_HIST,
  61
  62	PERF_HPP_DIFF__MAX_INDEX
  63};
  64
  65struct diff_hpp_fmt {
  66	struct perf_hpp_fmt	 fmt;
  67	int			 idx;
  68	char			*header;
  69	int			 header_width;
  70};
  71
  72struct data__file {
  73	struct perf_session	*session;
  74	struct perf_data	 data;
  75	int			 idx;
  76	struct hists		*hists;
  77	struct evlist_streams	*evlist_streams;
  78	struct diff_hpp_fmt	 fmt[PERF_HPP_DIFF__MAX_INDEX];
  79};
  80
  81static struct data__file *data__files;
  82static int data__files_cnt;
  83
  84#define data__for_each_file_start(i, d, s)	\
  85	for (i = s, d = &data__files[s];	\
  86	     i < data__files_cnt;		\
  87	     i++, d = &data__files[i])
  88
  89#define data__for_each_file(i, d) data__for_each_file_start(i, d, 0)
  90#define data__for_each_file_new(i, d) data__for_each_file_start(i, d, 1)
  91
  92static bool force;
  93static bool show_period;
  94static bool show_formula;
  95static bool show_baseline_only;
  96static bool cycles_hist;
  97static unsigned int sort_compute = 1;
  98
  99static s64 compute_wdiff_w1;
 100static s64 compute_wdiff_w2;
 101
 102static const char		*cpu_list;
 103static DECLARE_BITMAP(cpu_bitmap, MAX_NR_CPUS);
 104
 105enum {
 106	COMPUTE_DELTA,
 107	COMPUTE_RATIO,
 108	COMPUTE_WEIGHTED_DIFF,
 109	COMPUTE_DELTA_ABS,
 110	COMPUTE_CYCLES,
 111	COMPUTE_MAX,
 112	COMPUTE_STREAM,	/* After COMPUTE_MAX to avoid use current compute arrays */
 113};
 114
 115const char *compute_names[COMPUTE_MAX] = {
 116	[COMPUTE_DELTA] = "delta",
 117	[COMPUTE_DELTA_ABS] = "delta-abs",
 118	[COMPUTE_RATIO] = "ratio",
 119	[COMPUTE_WEIGHTED_DIFF] = "wdiff",
 120	[COMPUTE_CYCLES] = "cycles",
 121};
 122
 123static int compute = COMPUTE_DELTA_ABS;
 124
 125static int compute_2_hpp[COMPUTE_MAX] = {
 126	[COMPUTE_DELTA]		= PERF_HPP_DIFF__DELTA,
 127	[COMPUTE_DELTA_ABS]	= PERF_HPP_DIFF__DELTA_ABS,
 128	[COMPUTE_RATIO]		= PERF_HPP_DIFF__RATIO,
 129	[COMPUTE_WEIGHTED_DIFF]	= PERF_HPP_DIFF__WEIGHTED_DIFF,
 130	[COMPUTE_CYCLES]	= PERF_HPP_DIFF__CYCLES,
 131};
 132
 133#define MAX_COL_WIDTH 70
 134
 135static struct header_column {
 136	const char *name;
 137	int width;
 138} columns[PERF_HPP_DIFF__MAX_INDEX] = {
 139	[PERF_HPP_DIFF__BASELINE] = {
 140		.name  = "Baseline",
 141	},
 142	[PERF_HPP_DIFF__PERIOD] = {
 143		.name  = "Period",
 144		.width = 14,
 145	},
 146	[PERF_HPP_DIFF__PERIOD_BASELINE] = {
 147		.name  = "Base period",
 148		.width = 14,
 149	},
 150	[PERF_HPP_DIFF__DELTA] = {
 151		.name  = "Delta",
 152		.width = 7,
 153	},
 154	[PERF_HPP_DIFF__DELTA_ABS] = {
 155		.name  = "Delta Abs",
 156		.width = 7,
 157	},
 158	[PERF_HPP_DIFF__RATIO] = {
 159		.name  = "Ratio",
 160		.width = 14,
 161	},
 162	[PERF_HPP_DIFF__WEIGHTED_DIFF] = {
 163		.name  = "Weighted diff",
 164		.width = 14,
 165	},
 166	[PERF_HPP_DIFF__FORMULA] = {
 167		.name  = "Formula",
 168		.width = MAX_COL_WIDTH,
 169	},
 170	[PERF_HPP_DIFF__CYCLES] = {
 171		.name  = "[Program Block Range] Cycles Diff",
 172		.width = 70,
 173	},
 174	[PERF_HPP_DIFF__CYCLES_HIST] = {
 175		.name  = "stddev/Hist",
 176		.width = NUM_SPARKS + 9,
 177	}
 178};
 179
 180static int setup_compute_opt_wdiff(char *opt)
 181{
 182	char *w1_str = opt;
 183	char *w2_str;
 184
 185	int ret = -EINVAL;
 186
 187	if (!opt)
 188		goto out;
 189
 190	w2_str = strchr(opt, ',');
 191	if (!w2_str)
 192		goto out;
 193
 194	*w2_str++ = 0x0;
 195	if (!*w2_str)
 196		goto out;
 197
 198	compute_wdiff_w1 = strtol(w1_str, NULL, 10);
 199	compute_wdiff_w2 = strtol(w2_str, NULL, 10);
 200
 201	if (!compute_wdiff_w1 || !compute_wdiff_w2)
 202		goto out;
 203
 204	pr_debug("compute wdiff w1(%" PRId64 ") w2(%" PRId64 ")\n",
 205		  compute_wdiff_w1, compute_wdiff_w2);
 206
 207	ret = 0;
 208
 209 out:
 210	if (ret)
 211		pr_err("Failed: wrong weight data, use 'wdiff:w1,w2'\n");
 212
 213	return ret;
 214}
 215
 216static int setup_compute_opt(char *opt)
 217{
 218	if (compute == COMPUTE_WEIGHTED_DIFF)
 219		return setup_compute_opt_wdiff(opt);
 220
 221	if (opt) {
 222		pr_err("Failed: extra option specified '%s'", opt);
 223		return -EINVAL;
 224	}
 225
 226	return 0;
 227}
 228
 229static int setup_compute(const struct option *opt, const char *str,
 230			 int unset __maybe_unused)
 231{
 232	int *cp = (int *) opt->value;
 233	char *cstr = (char *) str;
 234	char buf[50];
 235	unsigned i;
 236	char *option;
 237
 238	if (!str) {
 239		*cp = COMPUTE_DELTA;
 240		return 0;
 241	}
 242
 243	option = strchr(str, ':');
 244	if (option) {
 245		unsigned len = option++ - str;
 246
 247		/*
 248		 * The str data are not writeable, so we need
 249		 * to use another buffer.
 250		 */
 251
 252		/* No option value is longer. */
 253		if (len >= sizeof(buf))
 254			return -EINVAL;
 255
 256		strncpy(buf, str, len);
 257		buf[len] = 0x0;
 258		cstr = buf;
 259	}
 260
 261	for (i = 0; i < COMPUTE_MAX; i++)
 262		if (!strcmp(cstr, compute_names[i])) {
 263			*cp = i;
 264			return setup_compute_opt(option);
 265		}
 266
 267	pr_err("Failed: '%s' is not computation method "
 268	       "(use 'delta','ratio' or 'wdiff')\n", str);
 269	return -EINVAL;
 270}
 271
 272static double period_percent(struct hist_entry *he, u64 period)
 273{
 274	u64 total = hists__total_period(he->hists);
 275
 276	return (period * 100.0) / total;
 277}
 278
 279static double compute_delta(struct hist_entry *he, struct hist_entry *pair)
 280{
 281	double old_percent = period_percent(he, he->stat.period);
 282	double new_percent = period_percent(pair, pair->stat.period);
 283
 284	pair->diff.period_ratio_delta = new_percent - old_percent;
 285	pair->diff.computed = true;
 286	return pair->diff.period_ratio_delta;
 287}
 288
 289static double compute_ratio(struct hist_entry *he, struct hist_entry *pair)
 290{
 291	double old_period = he->stat.period ?: 1;
 292	double new_period = pair->stat.period;
 293
 294	pair->diff.computed = true;
 295	pair->diff.period_ratio = new_period / old_period;
 296	return pair->diff.period_ratio;
 297}
 298
 299static s64 compute_wdiff(struct hist_entry *he, struct hist_entry *pair)
 300{
 301	u64 old_period = he->stat.period;
 302	u64 new_period = pair->stat.period;
 303
 304	pair->diff.computed = true;
 305	pair->diff.wdiff = new_period * compute_wdiff_w2 -
 306			   old_period * compute_wdiff_w1;
 307
 308	return pair->diff.wdiff;
 309}
 310
 311static int formula_delta(struct hist_entry *he, struct hist_entry *pair,
 312			 char *buf, size_t size)
 313{
 314	u64 he_total = he->hists->stats.total_period;
 315	u64 pair_total = pair->hists->stats.total_period;
 316
 317	if (symbol_conf.filter_relative) {
 318		he_total = he->hists->stats.total_non_filtered_period;
 319		pair_total = pair->hists->stats.total_non_filtered_period;
 320	}
 321	return scnprintf(buf, size,
 322			 "(%" PRIu64 " * 100 / %" PRIu64 ") - "
 323			 "(%" PRIu64 " * 100 / %" PRIu64 ")",
 324			 pair->stat.period, pair_total,
 325			 he->stat.period, he_total);
 326}
 327
 328static int formula_ratio(struct hist_entry *he, struct hist_entry *pair,
 329			 char *buf, size_t size)
 330{
 331	double old_period = he->stat.period;
 332	double new_period = pair->stat.period;
 333
 334	return scnprintf(buf, size, "%.0F / %.0F", new_period, old_period);
 335}
 336
 337static int formula_wdiff(struct hist_entry *he, struct hist_entry *pair,
 338			 char *buf, size_t size)
 339{
 340	u64 old_period = he->stat.period;
 341	u64 new_period = pair->stat.period;
 342
 343	return scnprintf(buf, size,
 344		  "(%" PRIu64 " * " "%" PRId64 ") - (%" PRIu64 " * " "%" PRId64 ")",
 345		  new_period, compute_wdiff_w2, old_period, compute_wdiff_w1);
 346}
 347
 348static int formula_fprintf(struct hist_entry *he, struct hist_entry *pair,
 349			   char *buf, size_t size)
 350{
 351	switch (compute) {
 352	case COMPUTE_DELTA:
 353	case COMPUTE_DELTA_ABS:
 354		return formula_delta(he, pair, buf, size);
 355	case COMPUTE_RATIO:
 356		return formula_ratio(he, pair, buf, size);
 357	case COMPUTE_WEIGHTED_DIFF:
 358		return formula_wdiff(he, pair, buf, size);
 359	default:
 360		BUG_ON(1);
 361	}
 362
 363	return -1;
 364}
 365
 366static void *block_hist_zalloc(size_t size)
 367{
 368	struct block_hist *bh;
 369
 370	bh = zalloc(size + sizeof(*bh));
 371	if (!bh)
 372		return NULL;
 373
 374	return &bh->he;
 375}
 376
 377static void block_hist_free(void *he)
 378{
 379	struct block_hist *bh;
 380
 381	bh = container_of(he, struct block_hist, he);
 382	hists__delete_entries(&bh->block_hists);
 383	free(bh);
 384}
 385
 386struct hist_entry_ops block_hist_ops = {
 387	.new    = block_hist_zalloc,
 388	.free   = block_hist_free,
 389};
 390
 391static int diff__process_sample_event(struct perf_tool *tool,
 392				      union perf_event *event,
 393				      struct perf_sample *sample,
 394				      struct evsel *evsel,
 395				      struct machine *machine)
 396{
 397	struct perf_diff *pdiff = container_of(tool, struct perf_diff, tool);
 398	struct addr_location al;
 399	struct hists *hists = evsel__hists(evsel);
 400	struct hist_entry_iter iter = {
 401		.evsel	= evsel,
 402		.sample	= sample,
 403		.ops	= &hist_iter_normal,
 404	};
 405	int ret = -1;
 406
 407	if (perf_time__ranges_skip_sample(pdiff->ptime_range, pdiff->range_num,
 408					  sample->time)) {
 409		return 0;
 410	}
 411
 412	if (machine__resolve(machine, &al, sample) < 0) {
 413		pr_warning("problem processing %d event, skipping it.\n",
 414			   event->header.type);
 415		return -1;
 416	}
 417
 418	if (cpu_list && !test_bit(sample->cpu, cpu_bitmap)) {
 419		ret = 0;
 420		goto out_put;
 421	}
 422
 423	switch (compute) {
 424	case COMPUTE_CYCLES:
 
 
 
 
 
 
 425		if (!hists__add_entry_ops(hists, &block_hist_ops, &al, NULL,
 426					  NULL, NULL, sample, true)) {
 427			pr_warning("problem incrementing symbol period, "
 428				   "skipping event\n");
 429			goto out_put;
 430		}
 431
 432		hist__account_cycles(sample->branch_stack, &al, sample, false,
 433				     NULL);
 434		break;
 435
 436	case COMPUTE_STREAM:
 437		if (hist_entry_iter__add(&iter, &al, PERF_MAX_STACK_DEPTH,
 438					 NULL)) {
 439			pr_debug("problem adding hist entry, skipping event\n");
 440			goto out_put;
 441		}
 442		break;
 443
 444	default:
 445		if (!hists__add_entry(hists, &al, NULL, NULL, NULL, sample,
 446				      true)) {
 447			pr_warning("problem incrementing symbol period, "
 448				   "skipping event\n");
 449			goto out_put;
 450		}
 451	}
 452
 453	/*
 454	 * The total_period is updated here before going to the output
 455	 * tree since normally only the baseline hists will call
 456	 * hists__output_resort() and precompute needs the total
 457	 * period in order to sort entries by percentage delta.
 458	 */
 459	hists->stats.total_period += sample->period;
 460	if (!al.filtered)
 461		hists->stats.total_non_filtered_period += sample->period;
 462	ret = 0;
 463out_put:
 464	addr_location__put(&al);
 465	return ret;
 466}
 467
 468static struct perf_diff pdiff = {
 469	.tool = {
 470		.sample	= diff__process_sample_event,
 471		.mmap	= perf_event__process_mmap,
 472		.mmap2	= perf_event__process_mmap2,
 473		.comm	= perf_event__process_comm,
 474		.exit	= perf_event__process_exit,
 475		.fork	= perf_event__process_fork,
 476		.lost	= perf_event__process_lost,
 477		.namespaces = perf_event__process_namespaces,
 478		.cgroup = perf_event__process_cgroup,
 479		.ordered_events = true,
 480		.ordering_requires_timestamps = true,
 481	},
 482};
 483
 484static struct evsel *evsel_match(struct evsel *evsel,
 485				      struct evlist *evlist)
 486{
 487	struct evsel *e;
 488
 489	evlist__for_each_entry(evlist, e) {
 490		if (evsel__match2(evsel, e))
 491			return e;
 492	}
 493
 494	return NULL;
 495}
 496
 497static void evlist__collapse_resort(struct evlist *evlist)
 498{
 499	struct evsel *evsel;
 500
 501	evlist__for_each_entry(evlist, evsel) {
 502		struct hists *hists = evsel__hists(evsel);
 503
 504		hists__collapse_resort(hists, NULL);
 505	}
 506}
 507
 508static struct data__file *fmt_to_data_file(struct perf_hpp_fmt *fmt)
 509{
 510	struct diff_hpp_fmt *dfmt = container_of(fmt, struct diff_hpp_fmt, fmt);
 511	void *ptr = dfmt - dfmt->idx;
 512	struct data__file *d = container_of(ptr, struct data__file, fmt);
 513
 514	return d;
 515}
 516
 517static struct hist_entry*
 518get_pair_data(struct hist_entry *he, struct data__file *d)
 519{
 520	if (hist_entry__has_pairs(he)) {
 521		struct hist_entry *pair;
 522
 523		list_for_each_entry(pair, &he->pairs.head, pairs.node)
 524			if (pair->hists == d->hists)
 525				return pair;
 526	}
 527
 528	return NULL;
 529}
 530
 531static struct hist_entry*
 532get_pair_fmt(struct hist_entry *he, struct diff_hpp_fmt *dfmt)
 533{
 534	struct data__file *d = fmt_to_data_file(&dfmt->fmt);
 535
 536	return get_pair_data(he, d);
 537}
 538
 539static void hists__baseline_only(struct hists *hists)
 540{
 541	struct rb_root_cached *root;
 542	struct rb_node *next;
 543
 544	if (hists__has(hists, need_collapse))
 545		root = &hists->entries_collapsed;
 546	else
 547		root = hists->entries_in;
 548
 549	next = rb_first_cached(root);
 550	while (next != NULL) {
 551		struct hist_entry *he = rb_entry(next, struct hist_entry, rb_node_in);
 552
 553		next = rb_next(&he->rb_node_in);
 554		if (!hist_entry__next_pair(he)) {
 555			rb_erase_cached(&he->rb_node_in, root);
 556			hist_entry__delete(he);
 557		}
 558	}
 559}
 560
 561static int64_t block_cycles_diff_cmp(struct hist_entry *left,
 562				     struct hist_entry *right)
 563{
 564	bool pairs_left  = hist_entry__has_pairs(left);
 565	bool pairs_right = hist_entry__has_pairs(right);
 566	s64 l, r;
 567
 568	if (!pairs_left && !pairs_right)
 569		return 0;
 570
 571	l = llabs(left->diff.cycles);
 572	r = llabs(right->diff.cycles);
 573	return r - l;
 574}
 575
 576static int64_t block_sort(struct perf_hpp_fmt *fmt __maybe_unused,
 577			  struct hist_entry *left, struct hist_entry *right)
 578{
 579	return block_cycles_diff_cmp(right, left);
 580}
 581
 582static void init_block_hist(struct block_hist *bh)
 583{
 584	__hists__init(&bh->block_hists, &bh->block_list);
 585	perf_hpp_list__init(&bh->block_list);
 586
 587	INIT_LIST_HEAD(&bh->block_fmt.list);
 588	INIT_LIST_HEAD(&bh->block_fmt.sort_list);
 589	bh->block_fmt.cmp = block_info__cmp;
 590	bh->block_fmt.sort = block_sort;
 591	perf_hpp_list__register_sort_field(&bh->block_list,
 592					   &bh->block_fmt);
 593	bh->valid = true;
 594}
 595
 596static struct hist_entry *get_block_pair(struct hist_entry *he,
 597					 struct hists *hists_pair)
 598{
 599	struct rb_root_cached *root = hists_pair->entries_in;
 600	struct rb_node *next = rb_first_cached(root);
 601	int64_t cmp;
 602
 603	while (next != NULL) {
 604		struct hist_entry *he_pair = rb_entry(next, struct hist_entry,
 605						      rb_node_in);
 606
 607		next = rb_next(&he_pair->rb_node_in);
 608
 609		cmp = __block_info__cmp(he_pair, he);
 610		if (!cmp)
 611			return he_pair;
 612	}
 613
 614	return NULL;
 615}
 616
 617static void init_spark_values(unsigned long *svals, int num)
 618{
 619	for (int i = 0; i < num; i++)
 620		svals[i] = 0;
 621}
 622
 623static void update_spark_value(unsigned long *svals, int num,
 624			       struct stats *stats, u64 val)
 625{
 626	int n = stats->n;
 627
 628	if (n < num)
 629		svals[n] = val;
 630}
 631
 632static void compute_cycles_diff(struct hist_entry *he,
 633				struct hist_entry *pair)
 634{
 635	pair->diff.computed = true;
 636	if (pair->block_info->num && he->block_info->num) {
 637		pair->diff.cycles =
 638			pair->block_info->cycles_aggr / pair->block_info->num_aggr -
 639			he->block_info->cycles_aggr / he->block_info->num_aggr;
 640
 641		if (!cycles_hist)
 642			return;
 643
 644		init_stats(&pair->diff.stats);
 645		init_spark_values(pair->diff.svals, NUM_SPARKS);
 646
 647		for (int i = 0; i < pair->block_info->num; i++) {
 648			u64 val;
 649
 650			if (i >= he->block_info->num || i >= NUM_SPARKS)
 651				break;
 652
 653			val = llabs(pair->block_info->cycles_spark[i] -
 654				     he->block_info->cycles_spark[i]);
 655
 656			update_spark_value(pair->diff.svals, NUM_SPARKS,
 657					   &pair->diff.stats, val);
 658			update_stats(&pair->diff.stats, val);
 659		}
 660	}
 661}
 662
 663static void block_hists_match(struct hists *hists_base,
 664			      struct hists *hists_pair)
 665{
 666	struct rb_root_cached *root = hists_base->entries_in;
 667	struct rb_node *next = rb_first_cached(root);
 668
 669	while (next != NULL) {
 670		struct hist_entry *he = rb_entry(next, struct hist_entry,
 671						 rb_node_in);
 672		struct hist_entry *pair = get_block_pair(he, hists_pair);
 673
 674		next = rb_next(&he->rb_node_in);
 675
 676		if (pair) {
 677			hist_entry__add_pair(pair, he);
 678			compute_cycles_diff(he, pair);
 679		}
 680	}
 681}
 682
 683static void hists__precompute(struct hists *hists)
 684{
 685	struct rb_root_cached *root;
 686	struct rb_node *next;
 687
 688	if (hists__has(hists, need_collapse))
 689		root = &hists->entries_collapsed;
 690	else
 691		root = hists->entries_in;
 692
 693	next = rb_first_cached(root);
 694	while (next != NULL) {
 695		struct block_hist *bh, *pair_bh;
 696		struct hist_entry *he, *pair;
 697		struct data__file *d;
 698		int i;
 699
 700		he   = rb_entry(next, struct hist_entry, rb_node_in);
 701		next = rb_next(&he->rb_node_in);
 702
 703		if (compute == COMPUTE_CYCLES) {
 704			bh = container_of(he, struct block_hist, he);
 705			init_block_hist(bh);
 706			block_info__process_sym(he, bh, NULL, 0);
 707		}
 708
 709		data__for_each_file_new(i, d) {
 710			pair = get_pair_data(he, d);
 711			if (!pair)
 712				continue;
 713
 714			switch (compute) {
 715			case COMPUTE_DELTA:
 716			case COMPUTE_DELTA_ABS:
 717				compute_delta(he, pair);
 718				break;
 719			case COMPUTE_RATIO:
 720				compute_ratio(he, pair);
 721				break;
 722			case COMPUTE_WEIGHTED_DIFF:
 723				compute_wdiff(he, pair);
 724				break;
 725			case COMPUTE_CYCLES:
 726				pair_bh = container_of(pair, struct block_hist,
 727						       he);
 728				init_block_hist(pair_bh);
 729				block_info__process_sym(pair, pair_bh, NULL, 0);
 730
 731				bh = container_of(he, struct block_hist, he);
 732
 733				if (bh->valid && pair_bh->valid) {
 734					block_hists_match(&bh->block_hists,
 735							  &pair_bh->block_hists);
 736					hists__output_resort(&pair_bh->block_hists,
 737							     NULL);
 738				}
 739				break;
 740			default:
 741				BUG_ON(1);
 742			}
 743		}
 744	}
 745}
 746
 747static int64_t cmp_doubles(double l, double r)
 748{
 749	if (l > r)
 750		return -1;
 751	else if (l < r)
 752		return 1;
 753	else
 754		return 0;
 755}
 756
 757static int64_t
 758__hist_entry__cmp_compute(struct hist_entry *left, struct hist_entry *right,
 759			int c)
 760{
 761	switch (c) {
 762	case COMPUTE_DELTA:
 763	{
 764		double l = left->diff.period_ratio_delta;
 765		double r = right->diff.period_ratio_delta;
 766
 767		return cmp_doubles(l, r);
 768	}
 769	case COMPUTE_DELTA_ABS:
 770	{
 771		double l = fabs(left->diff.period_ratio_delta);
 772		double r = fabs(right->diff.period_ratio_delta);
 773
 774		return cmp_doubles(l, r);
 775	}
 776	case COMPUTE_RATIO:
 777	{
 778		double l = left->diff.period_ratio;
 779		double r = right->diff.period_ratio;
 780
 781		return cmp_doubles(l, r);
 782	}
 783	case COMPUTE_WEIGHTED_DIFF:
 784	{
 785		s64 l = left->diff.wdiff;
 786		s64 r = right->diff.wdiff;
 787
 788		return r - l;
 789	}
 790	default:
 791		BUG_ON(1);
 792	}
 793
 794	return 0;
 795}
 796
 797static int64_t
 798hist_entry__cmp_compute(struct hist_entry *left, struct hist_entry *right,
 799			int c, int sort_idx)
 800{
 801	bool pairs_left  = hist_entry__has_pairs(left);
 802	bool pairs_right = hist_entry__has_pairs(right);
 803	struct hist_entry *p_right, *p_left;
 804
 805	if (!pairs_left && !pairs_right)
 806		return 0;
 807
 808	if (!pairs_left || !pairs_right)
 809		return pairs_left ? -1 : 1;
 810
 811	p_left  = get_pair_data(left,  &data__files[sort_idx]);
 812	p_right = get_pair_data(right, &data__files[sort_idx]);
 813
 814	if (!p_left && !p_right)
 815		return 0;
 816
 817	if (!p_left || !p_right)
 818		return p_left ? -1 : 1;
 819
 820	/*
 821	 * We have 2 entries of same kind, let's
 822	 * make the data comparison.
 823	 */
 824	return __hist_entry__cmp_compute(p_left, p_right, c);
 825}
 826
 827static int64_t
 828hist_entry__cmp_compute_idx(struct hist_entry *left, struct hist_entry *right,
 829			    int c, int sort_idx)
 830{
 831	struct hist_entry *p_right, *p_left;
 832
 833	p_left  = get_pair_data(left,  &data__files[sort_idx]);
 834	p_right = get_pair_data(right, &data__files[sort_idx]);
 835
 836	if (!p_left && !p_right)
 837		return 0;
 838
 839	if (!p_left || !p_right)
 840		return p_left ? -1 : 1;
 841
 842	if (c != COMPUTE_DELTA && c != COMPUTE_DELTA_ABS) {
 843		/*
 844		 * The delta can be computed without the baseline, but
 845		 * others are not.  Put those entries which have no
 846		 * values below.
 847		 */
 848		if (left->dummy && right->dummy)
 849			return 0;
 850
 851		if (left->dummy || right->dummy)
 852			return left->dummy ? 1 : -1;
 853	}
 854
 855	return __hist_entry__cmp_compute(p_left, p_right, c);
 856}
 857
 858static int64_t
 859hist_entry__cmp_nop(struct perf_hpp_fmt *fmt __maybe_unused,
 860		    struct hist_entry *left __maybe_unused,
 861		    struct hist_entry *right __maybe_unused)
 862{
 863	return 0;
 864}
 865
 866static int64_t
 867hist_entry__cmp_baseline(struct perf_hpp_fmt *fmt __maybe_unused,
 868			 struct hist_entry *left, struct hist_entry *right)
 869{
 870	if (left->stat.period == right->stat.period)
 871		return 0;
 872	return left->stat.period > right->stat.period ? 1 : -1;
 873}
 874
 875static int64_t
 876hist_entry__cmp_delta(struct perf_hpp_fmt *fmt,
 877		      struct hist_entry *left, struct hist_entry *right)
 878{
 879	struct data__file *d = fmt_to_data_file(fmt);
 880
 881	return hist_entry__cmp_compute(right, left, COMPUTE_DELTA, d->idx);
 882}
 883
 884static int64_t
 885hist_entry__cmp_delta_abs(struct perf_hpp_fmt *fmt,
 886		      struct hist_entry *left, struct hist_entry *right)
 887{
 888	struct data__file *d = fmt_to_data_file(fmt);
 889
 890	return hist_entry__cmp_compute(right, left, COMPUTE_DELTA_ABS, d->idx);
 891}
 892
 893static int64_t
 894hist_entry__cmp_ratio(struct perf_hpp_fmt *fmt,
 895		      struct hist_entry *left, struct hist_entry *right)
 896{
 897	struct data__file *d = fmt_to_data_file(fmt);
 898
 899	return hist_entry__cmp_compute(right, left, COMPUTE_RATIO, d->idx);
 900}
 901
 902static int64_t
 903hist_entry__cmp_wdiff(struct perf_hpp_fmt *fmt,
 904		      struct hist_entry *left, struct hist_entry *right)
 905{
 906	struct data__file *d = fmt_to_data_file(fmt);
 907
 908	return hist_entry__cmp_compute(right, left, COMPUTE_WEIGHTED_DIFF, d->idx);
 909}
 910
 911static int64_t
 912hist_entry__cmp_delta_idx(struct perf_hpp_fmt *fmt __maybe_unused,
 913			  struct hist_entry *left, struct hist_entry *right)
 914{
 915	return hist_entry__cmp_compute_idx(right, left, COMPUTE_DELTA,
 916					   sort_compute);
 917}
 918
 919static int64_t
 920hist_entry__cmp_delta_abs_idx(struct perf_hpp_fmt *fmt __maybe_unused,
 921			      struct hist_entry *left, struct hist_entry *right)
 922{
 923	return hist_entry__cmp_compute_idx(right, left, COMPUTE_DELTA_ABS,
 924					   sort_compute);
 925}
 926
 927static int64_t
 928hist_entry__cmp_ratio_idx(struct perf_hpp_fmt *fmt __maybe_unused,
 929			  struct hist_entry *left, struct hist_entry *right)
 930{
 931	return hist_entry__cmp_compute_idx(right, left, COMPUTE_RATIO,
 932					   sort_compute);
 933}
 934
 935static int64_t
 936hist_entry__cmp_wdiff_idx(struct perf_hpp_fmt *fmt __maybe_unused,
 937			  struct hist_entry *left, struct hist_entry *right)
 938{
 939	return hist_entry__cmp_compute_idx(right, left, COMPUTE_WEIGHTED_DIFF,
 940					   sort_compute);
 941}
 942
 943static void hists__process(struct hists *hists)
 944{
 945	if (show_baseline_only)
 946		hists__baseline_only(hists);
 947
 948	hists__precompute(hists);
 949	hists__output_resort(hists, NULL);
 950
 951	if (compute == COMPUTE_CYCLES)
 952		symbol_conf.report_block = true;
 953
 954	hists__fprintf(hists, !quiet, 0, 0, 0, stdout,
 955		       !symbol_conf.use_callchain);
 956}
 957
 958static void data__fprintf(void)
 959{
 960	struct data__file *d;
 961	int i;
 962
 963	fprintf(stdout, "# Data files:\n");
 964
 965	data__for_each_file(i, d)
 966		fprintf(stdout, "#  [%d] %s %s\n",
 967			d->idx, d->data.path,
 968			!d->idx ? "(Baseline)" : "");
 969
 970	fprintf(stdout, "#\n");
 971}
 972
 973static void data_process(void)
 974{
 975	struct evlist *evlist_base = data__files[0].session->evlist;
 976	struct evsel *evsel_base;
 977	bool first = true;
 978
 979	evlist__for_each_entry(evlist_base, evsel_base) {
 980		struct hists *hists_base = evsel__hists(evsel_base);
 981		struct data__file *d;
 982		int i;
 983
 984		data__for_each_file_new(i, d) {
 985			struct evlist *evlist = d->session->evlist;
 986			struct evsel *evsel;
 987			struct hists *hists;
 988
 989			evsel = evsel_match(evsel_base, evlist);
 990			if (!evsel)
 991				continue;
 992
 993			hists = evsel__hists(evsel);
 994			d->hists = hists;
 995
 996			hists__match(hists_base, hists);
 997
 998			if (!show_baseline_only)
 999				hists__link(hists_base, hists);
1000		}
1001
1002		if (!quiet) {
1003			fprintf(stdout, "%s# Event '%s'\n#\n", first ? "" : "\n",
1004				evsel__name(evsel_base));
1005		}
1006
1007		first = false;
1008
1009		if (verbose > 0 || ((data__files_cnt > 2) && !quiet))
1010			data__fprintf();
1011
1012		/* Don't sort callchain for perf diff */
1013		evsel__reset_sample_bit(evsel_base, CALLCHAIN);
1014
1015		hists__process(hists_base);
1016	}
1017}
1018
1019static int process_base_stream(struct data__file *data_base,
1020			       struct data__file *data_pair,
1021			       const char *title __maybe_unused)
1022{
1023	struct evlist *evlist_base = data_base->session->evlist;
1024	struct evlist *evlist_pair = data_pair->session->evlist;
1025	struct evsel *evsel_base, *evsel_pair;
1026	struct evsel_streams *es_base, *es_pair;
1027
1028	evlist__for_each_entry(evlist_base, evsel_base) {
1029		evsel_pair = evsel_match(evsel_base, evlist_pair);
1030		if (!evsel_pair)
1031			continue;
1032
1033		es_base = evsel_streams__entry(data_base->evlist_streams,
1034					       evsel_base->core.idx);
1035		if (!es_base)
1036			return -1;
1037
1038		es_pair = evsel_streams__entry(data_pair->evlist_streams,
1039					       evsel_pair->core.idx);
1040		if (!es_pair)
1041			return -1;
1042
1043		evsel_streams__match(es_base, es_pair);
1044		evsel_streams__report(es_base, es_pair);
1045	}
1046
1047	return 0;
1048}
1049
1050static void stream_process(void)
1051{
1052	/*
1053	 * Stream comparison only supports two data files.
1054	 * perf.data.old and perf.data. data__files[0] is perf.data.old,
1055	 * data__files[1] is perf.data.
1056	 */
1057	process_base_stream(&data__files[0], &data__files[1],
1058			    "# Output based on old perf data:\n#\n");
1059}
1060
1061static void data__free(struct data__file *d)
1062{
1063	int col;
1064
1065	if (d->evlist_streams)
1066		evlist_streams__delete(d->evlist_streams);
1067
1068	for (col = 0; col < PERF_HPP_DIFF__MAX_INDEX; col++) {
1069		struct diff_hpp_fmt *fmt = &d->fmt[col];
1070
1071		zfree(&fmt->header);
1072	}
1073}
1074
1075static int abstime_str_dup(char **pstr)
1076{
1077	char *str = NULL;
1078
1079	if (pdiff.time_str && strchr(pdiff.time_str, ':')) {
1080		str = strdup(pdiff.time_str);
1081		if (!str)
1082			return -ENOMEM;
1083	}
1084
1085	*pstr = str;
1086	return 0;
1087}
1088
1089static int parse_absolute_time(struct data__file *d, char **pstr)
1090{
1091	char *p = *pstr;
1092	int ret;
1093
1094	/*
1095	 * Absolute timestamp for one file has the format: a.b,c.d
1096	 * For multiple files, the format is: a.b,c.d:a.b,c.d
1097	 */
1098	p = strchr(*pstr, ':');
1099	if (p) {
1100		if (p == *pstr) {
1101			pr_err("Invalid time string\n");
1102			return -EINVAL;
1103		}
1104
1105		*p = 0;
1106		p++;
1107		if (*p == 0) {
1108			pr_err("Invalid time string\n");
1109			return -EINVAL;
1110		}
1111	}
1112
1113	ret = perf_time__parse_for_ranges(*pstr, d->session,
1114					  &pdiff.ptime_range,
1115					  &pdiff.range_size,
1116					  &pdiff.range_num);
1117	if (ret < 0)
1118		return ret;
1119
1120	if (!p || *p == 0)
1121		*pstr = NULL;
1122	else
1123		*pstr = p;
1124
1125	return ret;
1126}
1127
1128static int parse_percent_time(struct data__file *d)
1129{
1130	int ret;
1131
1132	ret = perf_time__parse_for_ranges(pdiff.time_str, d->session,
1133					  &pdiff.ptime_range,
1134					  &pdiff.range_size,
1135					  &pdiff.range_num);
1136	return ret;
1137}
1138
1139static int parse_time_str(struct data__file *d, char *abstime_ostr,
1140			   char **pabstime_tmp)
1141{
1142	int ret = 0;
1143
1144	if (abstime_ostr)
1145		ret = parse_absolute_time(d, pabstime_tmp);
1146	else if (pdiff.time_str)
1147		ret = parse_percent_time(d);
1148
1149	return ret;
1150}
1151
1152static int check_file_brstack(void)
1153{
1154	struct data__file *d;
1155	bool has_br_stack;
1156	int i;
1157
1158	data__for_each_file(i, d) {
1159		d->session = perf_session__new(&d->data, &pdiff.tool);
1160		if (IS_ERR(d->session)) {
1161			pr_err("Failed to open %s\n", d->data.path);
1162			return PTR_ERR(d->session);
1163		}
1164
1165		has_br_stack = perf_header__has_feat(&d->session->header,
1166						     HEADER_BRANCH_STACK);
1167		perf_session__delete(d->session);
1168		if (!has_br_stack)
1169			return 0;
1170	}
1171
1172	/* Set only all files having branch stacks */
1173	pdiff.has_br_stack = true;
1174	return 0;
1175}
1176
1177static int __cmd_diff(void)
1178{
1179	struct data__file *d;
1180	int ret, i;
1181	char *abstime_ostr, *abstime_tmp;
1182
1183	ret = abstime_str_dup(&abstime_ostr);
1184	if (ret)
1185		return ret;
1186
1187	abstime_tmp = abstime_ostr;
1188	ret = -EINVAL;
1189
1190	data__for_each_file(i, d) {
1191		d->session = perf_session__new(&d->data, &pdiff.tool);
1192		if (IS_ERR(d->session)) {
1193			ret = PTR_ERR(d->session);
1194			pr_err("Failed to open %s\n", d->data.path);
1195			goto out_delete;
1196		}
1197
1198		if (pdiff.time_str) {
1199			ret = parse_time_str(d, abstime_ostr, &abstime_tmp);
1200			if (ret < 0)
1201				goto out_delete;
1202		}
1203
1204		if (cpu_list) {
1205			ret = perf_session__cpu_bitmap(d->session, cpu_list,
1206						       cpu_bitmap);
1207			if (ret < 0)
1208				goto out_delete;
1209		}
1210
1211		ret = perf_session__process_events(d->session);
1212		if (ret) {
1213			pr_err("Failed to process %s\n", d->data.path);
1214			goto out_delete;
1215		}
1216
1217		evlist__collapse_resort(d->session->evlist);
1218
1219		if (pdiff.ptime_range)
1220			zfree(&pdiff.ptime_range);
1221
1222		if (compute == COMPUTE_STREAM) {
1223			d->evlist_streams = evlist__create_streams(
1224						d->session->evlist, 5);
1225			if (!d->evlist_streams) {
1226				ret = -ENOMEM;
1227				goto out_delete;
1228			}
1229		}
1230	}
1231
1232	if (compute == COMPUTE_STREAM)
1233		stream_process();
1234	else
1235		data_process();
1236
1237 out_delete:
1238	data__for_each_file(i, d) {
1239		if (!IS_ERR(d->session))
1240			perf_session__delete(d->session);
1241		data__free(d);
1242	}
1243
1244	free(data__files);
1245
1246	if (pdiff.ptime_range)
1247		zfree(&pdiff.ptime_range);
1248
1249	if (abstime_ostr)
1250		free(abstime_ostr);
1251
1252	return ret;
1253}
1254
1255static const char * const diff_usage[] = {
1256	"perf diff [<options>] [old_file] [new_file]",
1257	NULL,
1258};
1259
1260static const struct option options[] = {
1261	OPT_INCR('v', "verbose", &verbose,
1262		    "be more verbose (show symbol address, etc)"),
1263	OPT_BOOLEAN('q', "quiet", &quiet, "Do not show any warnings or messages"),
1264	OPT_BOOLEAN('b', "baseline-only", &show_baseline_only,
1265		    "Show only items with match in baseline"),
1266	OPT_CALLBACK('c', "compute", &compute,
1267		     "delta,delta-abs,ratio,wdiff:w1,w2 (default delta-abs),cycles",
1268		     "Entries differential computation selection",
1269		     setup_compute),
1270	OPT_BOOLEAN('p', "period", &show_period,
1271		    "Show period values."),
1272	OPT_BOOLEAN('F', "formula", &show_formula,
1273		    "Show formula."),
1274	OPT_BOOLEAN(0, "cycles-hist", &cycles_hist,
1275		    "Show cycles histogram and standard deviation "
1276		    "- WARNING: use only with -c cycles."),
1277	OPT_BOOLEAN('D', "dump-raw-trace", &dump_trace,
1278		    "dump raw trace in ASCII"),
1279	OPT_BOOLEAN('f', "force", &force, "don't complain, do it"),
1280	OPT_STRING(0, "kallsyms", &symbol_conf.kallsyms_name,
1281		   "file", "kallsyms pathname"),
1282	OPT_BOOLEAN('m', "modules", &symbol_conf.use_modules,
1283		    "load module symbols - WARNING: use only with -k and LIVE kernel"),
1284	OPT_STRING('d', "dsos", &symbol_conf.dso_list_str, "dso[,dso...]",
1285		   "only consider symbols in these dsos"),
1286	OPT_STRING('C', "comms", &symbol_conf.comm_list_str, "comm[,comm...]",
1287		   "only consider symbols in these comms"),
1288	OPT_STRING('S', "symbols", &symbol_conf.sym_list_str, "symbol[,symbol...]",
1289		   "only consider these symbols"),
1290	OPT_STRING('s', "sort", &sort_order, "key[,key2...]",
1291		   "sort by key(s): pid, comm, dso, symbol, parent, cpu, srcline, ..."
1292		   " Please refer the man page for the complete list."),
1293	OPT_STRING_NOEMPTY('t', "field-separator", &symbol_conf.field_sep, "separator",
1294		   "separator for columns, no spaces will be added between "
1295		   "columns '.' is reserved."),
1296	OPT_CALLBACK(0, "symfs", NULL, "directory",
1297		     "Look for files with symbols relative to this directory",
1298		     symbol__config_symfs),
1299	OPT_UINTEGER('o', "order", &sort_compute, "Specify compute sorting."),
1300	OPT_CALLBACK(0, "percentage", NULL, "relative|absolute",
1301		     "How to display percentage of filtered entries", parse_filter_percentage),
1302	OPT_STRING(0, "time", &pdiff.time_str, "str",
1303		   "Time span (time percent or absolute timestamp)"),
1304	OPT_STRING(0, "cpu", &cpu_list, "cpu", "list of cpus to profile"),
1305	OPT_STRING(0, "pid", &symbol_conf.pid_list_str, "pid[,pid...]",
1306		   "only consider symbols in these pids"),
1307	OPT_STRING(0, "tid", &symbol_conf.tid_list_str, "tid[,tid...]",
1308		   "only consider symbols in these tids"),
1309	OPT_BOOLEAN(0, "stream", &pdiff.stream,
1310		    "Enable hot streams comparison."),
1311	OPT_END()
1312};
1313
1314static double baseline_percent(struct hist_entry *he)
1315{
1316	u64 total = hists__total_period(he->hists);
1317
1318	return 100.0 * he->stat.period / total;
1319}
1320
1321static int hpp__color_baseline(struct perf_hpp_fmt *fmt,
1322			       struct perf_hpp *hpp, struct hist_entry *he)
1323{
1324	struct diff_hpp_fmt *dfmt =
1325		container_of(fmt, struct diff_hpp_fmt, fmt);
1326	double percent = baseline_percent(he);
1327	char pfmt[20] = " ";
1328
1329	if (!he->dummy) {
1330		scnprintf(pfmt, 20, "%%%d.2f%%%%", dfmt->header_width - 1);
1331		return percent_color_snprintf(hpp->buf, hpp->size,
1332					      pfmt, percent);
1333	} else
1334		return scnprintf(hpp->buf, hpp->size, "%*s",
1335				 dfmt->header_width, pfmt);
1336}
1337
1338static int hpp__entry_baseline(struct hist_entry *he, char *buf, size_t size)
1339{
1340	double percent = baseline_percent(he);
1341	const char *fmt = symbol_conf.field_sep ? "%.2f" : "%6.2f%%";
1342	int ret = 0;
1343
1344	if (!he->dummy)
1345		ret = scnprintf(buf, size, fmt, percent);
1346
1347	return ret;
1348}
1349
1350static int cycles_printf(struct hist_entry *he, struct hist_entry *pair,
1351			 struct perf_hpp *hpp, int width)
1352{
1353	struct block_hist *bh = container_of(he, struct block_hist, he);
1354	struct block_hist *bh_pair = container_of(pair, struct block_hist, he);
1355	struct hist_entry *block_he;
1356	struct block_info *bi;
1357	char buf[128];
1358	char *start_line, *end_line;
1359
1360	block_he = hists__get_entry(&bh_pair->block_hists, bh->block_idx);
1361	if (!block_he) {
1362		hpp->skip = true;
1363		return 0;
1364	}
1365
1366	/*
1367	 * Avoid printing the warning "addr2line_init failed for ..."
1368	 */
1369	symbol_conf.disable_add2line_warn = true;
1370
1371	bi = block_he->block_info;
1372
1373	start_line = map__srcline(he->ms.map, bi->sym->start + bi->start,
1374				  he->ms.sym);
1375
1376	end_line = map__srcline(he->ms.map, bi->sym->start + bi->end,
1377				he->ms.sym);
1378
1379	if ((strncmp(start_line, SRCLINE_UNKNOWN, strlen(SRCLINE_UNKNOWN)) != 0) &&
1380	    (strncmp(end_line, SRCLINE_UNKNOWN, strlen(SRCLINE_UNKNOWN)) != 0)) {
1381		scnprintf(buf, sizeof(buf), "[%s -> %s] %4ld",
1382			  start_line, end_line, block_he->diff.cycles);
1383	} else {
1384		scnprintf(buf, sizeof(buf), "[%7lx -> %7lx] %4ld",
1385			  bi->start, bi->end, block_he->diff.cycles);
1386	}
1387
1388	free_srcline(start_line);
1389	free_srcline(end_line);
1390
1391	return scnprintf(hpp->buf, hpp->size, "%*s", width, buf);
1392}
1393
1394static int __hpp__color_compare(struct perf_hpp_fmt *fmt,
1395				struct perf_hpp *hpp, struct hist_entry *he,
1396				int comparison_method)
1397{
1398	struct diff_hpp_fmt *dfmt =
1399		container_of(fmt, struct diff_hpp_fmt, fmt);
1400	struct hist_entry *pair = get_pair_fmt(he, dfmt);
1401	double diff;
1402	s64 wdiff;
1403	char pfmt[20] = " ";
1404
1405	if (!pair) {
1406		if (comparison_method == COMPUTE_CYCLES) {
1407			struct block_hist *bh;
1408
1409			bh = container_of(he, struct block_hist, he);
1410			if (bh->block_idx)
1411				hpp->skip = true;
1412		}
1413
1414		goto no_print;
1415	}
1416
1417	switch (comparison_method) {
1418	case COMPUTE_DELTA:
1419		if (pair->diff.computed)
1420			diff = pair->diff.period_ratio_delta;
1421		else
1422			diff = compute_delta(he, pair);
1423
1424		scnprintf(pfmt, 20, "%%%+d.2f%%%%", dfmt->header_width - 1);
1425		return percent_color_snprintf(hpp->buf, hpp->size,
1426					pfmt, diff);
1427	case COMPUTE_RATIO:
1428		if (he->dummy)
1429			goto dummy_print;
1430		if (pair->diff.computed)
1431			diff = pair->diff.period_ratio;
1432		else
1433			diff = compute_ratio(he, pair);
1434
1435		scnprintf(pfmt, 20, "%%%d.6f", dfmt->header_width);
1436		return value_color_snprintf(hpp->buf, hpp->size,
1437					pfmt, diff);
1438	case COMPUTE_WEIGHTED_DIFF:
1439		if (he->dummy)
1440			goto dummy_print;
1441		if (pair->diff.computed)
1442			wdiff = pair->diff.wdiff;
1443		else
1444			wdiff = compute_wdiff(he, pair);
1445
1446		scnprintf(pfmt, 20, "%%14ld", dfmt->header_width);
1447		return color_snprintf(hpp->buf, hpp->size,
1448				get_percent_color(wdiff),
1449				pfmt, wdiff);
1450	case COMPUTE_CYCLES:
1451		return cycles_printf(he, pair, hpp, dfmt->header_width);
1452	default:
1453		BUG_ON(1);
1454	}
1455dummy_print:
1456	return scnprintf(hpp->buf, hpp->size, "%*s",
1457			dfmt->header_width, "N/A");
1458no_print:
1459	return scnprintf(hpp->buf, hpp->size, "%*s",
1460			dfmt->header_width, pfmt);
1461}
1462
1463static int hpp__color_delta(struct perf_hpp_fmt *fmt,
1464			struct perf_hpp *hpp, struct hist_entry *he)
1465{
1466	return __hpp__color_compare(fmt, hpp, he, COMPUTE_DELTA);
1467}
1468
1469static int hpp__color_ratio(struct perf_hpp_fmt *fmt,
1470			struct perf_hpp *hpp, struct hist_entry *he)
1471{
1472	return __hpp__color_compare(fmt, hpp, he, COMPUTE_RATIO);
1473}
1474
1475static int hpp__color_wdiff(struct perf_hpp_fmt *fmt,
1476			struct perf_hpp *hpp, struct hist_entry *he)
1477{
1478	return __hpp__color_compare(fmt, hpp, he, COMPUTE_WEIGHTED_DIFF);
1479}
1480
1481static int hpp__color_cycles(struct perf_hpp_fmt *fmt,
1482			     struct perf_hpp *hpp, struct hist_entry *he)
1483{
1484	return __hpp__color_compare(fmt, hpp, he, COMPUTE_CYCLES);
1485}
1486
1487static int all_zero(unsigned long *vals, int len)
1488{
1489	int i;
1490
1491	for (i = 0; i < len; i++)
1492		if (vals[i] != 0)
1493			return 0;
1494	return 1;
1495}
1496
1497static int print_cycles_spark(char *bf, int size, unsigned long *svals, u64 n)
1498{
1499	int printed;
1500
1501	if (n <= 1)
1502		return 0;
1503
1504	if (n > NUM_SPARKS)
1505		n = NUM_SPARKS;
1506	if (all_zero(svals, n))
1507		return 0;
1508
1509	printed = print_spark(bf, size, svals, n);
1510	printed += scnprintf(bf + printed, size - printed, " ");
1511	return printed;
1512}
1513
1514static int hpp__color_cycles_hist(struct perf_hpp_fmt *fmt,
1515			    struct perf_hpp *hpp, struct hist_entry *he)
1516{
1517	struct diff_hpp_fmt *dfmt =
1518		container_of(fmt, struct diff_hpp_fmt, fmt);
1519	struct hist_entry *pair = get_pair_fmt(he, dfmt);
1520	struct block_hist *bh = container_of(he, struct block_hist, he);
1521	struct block_hist *bh_pair;
1522	struct hist_entry *block_he;
1523	char spark[32], buf[128];
1524	double r;
1525	int ret, pad;
1526
1527	if (!pair) {
1528		if (bh->block_idx)
1529			hpp->skip = true;
1530
1531		goto no_print;
1532	}
1533
1534	bh_pair = container_of(pair, struct block_hist, he);
1535
1536	block_he = hists__get_entry(&bh_pair->block_hists, bh->block_idx);
1537	if (!block_he) {
1538		hpp->skip = true;
1539		goto no_print;
1540	}
1541
1542	ret = print_cycles_spark(spark, sizeof(spark), block_he->diff.svals,
1543				 block_he->diff.stats.n);
1544
1545	r = rel_stddev_stats(stddev_stats(&block_he->diff.stats),
1546			     avg_stats(&block_he->diff.stats));
1547
1548	if (ret) {
1549		/*
1550		 * Padding spaces if number of sparks less than NUM_SPARKS
1551		 * otherwise the output is not aligned.
1552		 */
1553		pad = NUM_SPARKS - ((ret - 1) / 3);
1554		scnprintf(buf, sizeof(buf), "%s%5.1f%% %s", "\u00B1", r, spark);
1555		ret = scnprintf(hpp->buf, hpp->size, "%*s",
1556				dfmt->header_width, buf);
1557
1558		if (pad) {
1559			ret += scnprintf(hpp->buf + ret, hpp->size - ret,
1560					 "%-*s", pad, " ");
1561		}
1562
1563		return ret;
1564	}
1565
1566no_print:
1567	return scnprintf(hpp->buf, hpp->size, "%*s",
1568			dfmt->header_width, " ");
1569}
1570
1571static void
1572hpp__entry_unpair(struct hist_entry *he, int idx, char *buf, size_t size)
1573{
1574	switch (idx) {
1575	case PERF_HPP_DIFF__PERIOD_BASELINE:
1576		scnprintf(buf, size, "%" PRIu64, he->stat.period);
1577		break;
1578
1579	default:
1580		break;
1581	}
1582}
1583
1584static void
1585hpp__entry_pair(struct hist_entry *he, struct hist_entry *pair,
1586		int idx, char *buf, size_t size)
1587{
1588	double diff;
1589	double ratio;
1590	s64 wdiff;
1591
1592	switch (idx) {
1593	case PERF_HPP_DIFF__DELTA:
1594	case PERF_HPP_DIFF__DELTA_ABS:
1595		if (pair->diff.computed)
1596			diff = pair->diff.period_ratio_delta;
1597		else
1598			diff = compute_delta(he, pair);
1599
1600		scnprintf(buf, size, "%+4.2F%%", diff);
1601		break;
1602
1603	case PERF_HPP_DIFF__RATIO:
1604		/* No point for ratio number if we are dummy.. */
1605		if (he->dummy) {
1606			scnprintf(buf, size, "N/A");
1607			break;
1608		}
1609
1610		if (pair->diff.computed)
1611			ratio = pair->diff.period_ratio;
1612		else
1613			ratio = compute_ratio(he, pair);
1614
1615		if (ratio > 0.0)
1616			scnprintf(buf, size, "%14.6F", ratio);
1617		break;
1618
1619	case PERF_HPP_DIFF__WEIGHTED_DIFF:
1620		/* No point for wdiff number if we are dummy.. */
1621		if (he->dummy) {
1622			scnprintf(buf, size, "N/A");
1623			break;
1624		}
1625
1626		if (pair->diff.computed)
1627			wdiff = pair->diff.wdiff;
1628		else
1629			wdiff = compute_wdiff(he, pair);
1630
1631		if (wdiff != 0)
1632			scnprintf(buf, size, "%14ld", wdiff);
1633		break;
1634
1635	case PERF_HPP_DIFF__FORMULA:
1636		formula_fprintf(he, pair, buf, size);
1637		break;
1638
1639	case PERF_HPP_DIFF__PERIOD:
1640		scnprintf(buf, size, "%" PRIu64, pair->stat.period);
1641		break;
1642
1643	default:
1644		BUG_ON(1);
1645	}
1646}
1647
1648static void
1649__hpp__entry_global(struct hist_entry *he, struct diff_hpp_fmt *dfmt,
1650		    char *buf, size_t size)
1651{
1652	struct hist_entry *pair = get_pair_fmt(he, dfmt);
1653	int idx = dfmt->idx;
1654
1655	/* baseline is special */
1656	if (idx == PERF_HPP_DIFF__BASELINE)
1657		hpp__entry_baseline(he, buf, size);
1658	else {
1659		if (pair)
1660			hpp__entry_pair(he, pair, idx, buf, size);
1661		else
1662			hpp__entry_unpair(he, idx, buf, size);
1663	}
1664}
1665
1666static int hpp__entry_global(struct perf_hpp_fmt *_fmt, struct perf_hpp *hpp,
1667			     struct hist_entry *he)
1668{
1669	struct diff_hpp_fmt *dfmt =
1670		container_of(_fmt, struct diff_hpp_fmt, fmt);
1671	char buf[MAX_COL_WIDTH] = " ";
1672
1673	__hpp__entry_global(he, dfmt, buf, MAX_COL_WIDTH);
1674
1675	if (symbol_conf.field_sep)
1676		return scnprintf(hpp->buf, hpp->size, "%s", buf);
1677	else
1678		return scnprintf(hpp->buf, hpp->size, "%*s",
1679				 dfmt->header_width, buf);
1680}
1681
1682static int hpp__header(struct perf_hpp_fmt *fmt, struct perf_hpp *hpp,
1683		       struct hists *hists __maybe_unused,
1684		       int line __maybe_unused,
1685		       int *span __maybe_unused)
1686{
1687	struct diff_hpp_fmt *dfmt =
1688		container_of(fmt, struct diff_hpp_fmt, fmt);
1689
1690	BUG_ON(!dfmt->header);
1691	return scnprintf(hpp->buf, hpp->size, dfmt->header);
1692}
1693
1694static int hpp__width(struct perf_hpp_fmt *fmt,
1695		      struct perf_hpp *hpp __maybe_unused,
1696		      struct hists *hists __maybe_unused)
1697{
1698	struct diff_hpp_fmt *dfmt =
1699		container_of(fmt, struct diff_hpp_fmt, fmt);
1700
1701	BUG_ON(dfmt->header_width <= 0);
1702	return dfmt->header_width;
1703}
1704
1705static void init_header(struct data__file *d, struct diff_hpp_fmt *dfmt)
1706{
1707#define MAX_HEADER_NAME 100
1708	char buf_indent[MAX_HEADER_NAME];
1709	char buf[MAX_HEADER_NAME];
1710	const char *header = NULL;
1711	int width = 0;
1712
1713	BUG_ON(dfmt->idx >= PERF_HPP_DIFF__MAX_INDEX);
1714	header = columns[dfmt->idx].name;
1715	width  = columns[dfmt->idx].width;
1716
1717	/* Only our defined HPP fmts should appear here. */
1718	BUG_ON(!header);
1719
1720	if (data__files_cnt > 2)
1721		scnprintf(buf, MAX_HEADER_NAME, "%s/%d", header, d->idx);
1722
1723#define NAME (data__files_cnt > 2 ? buf : header)
1724	dfmt->header_width = width;
1725	width = (int) strlen(NAME);
1726	if (dfmt->header_width < width)
1727		dfmt->header_width = width;
1728
1729	scnprintf(buf_indent, MAX_HEADER_NAME, "%*s",
1730		  dfmt->header_width, NAME);
1731
1732	dfmt->header = strdup(buf_indent);
1733#undef MAX_HEADER_NAME
1734#undef NAME
1735}
1736
1737static void data__hpp_register(struct data__file *d, int idx)
1738{
1739	struct diff_hpp_fmt *dfmt = &d->fmt[idx];
1740	struct perf_hpp_fmt *fmt = &dfmt->fmt;
1741
1742	dfmt->idx = idx;
1743
1744	fmt->header = hpp__header;
1745	fmt->width  = hpp__width;
1746	fmt->entry  = hpp__entry_global;
1747	fmt->cmp    = hist_entry__cmp_nop;
1748	fmt->collapse = hist_entry__cmp_nop;
1749
1750	/* TODO more colors */
1751	switch (idx) {
1752	case PERF_HPP_DIFF__BASELINE:
1753		fmt->color = hpp__color_baseline;
1754		fmt->sort  = hist_entry__cmp_baseline;
1755		break;
1756	case PERF_HPP_DIFF__DELTA:
1757		fmt->color = hpp__color_delta;
1758		fmt->sort  = hist_entry__cmp_delta;
1759		break;
1760	case PERF_HPP_DIFF__RATIO:
1761		fmt->color = hpp__color_ratio;
1762		fmt->sort  = hist_entry__cmp_ratio;
1763		break;
1764	case PERF_HPP_DIFF__WEIGHTED_DIFF:
1765		fmt->color = hpp__color_wdiff;
1766		fmt->sort  = hist_entry__cmp_wdiff;
1767		break;
1768	case PERF_HPP_DIFF__DELTA_ABS:
1769		fmt->color = hpp__color_delta;
1770		fmt->sort  = hist_entry__cmp_delta_abs;
1771		break;
1772	case PERF_HPP_DIFF__CYCLES:
1773		fmt->color = hpp__color_cycles;
1774		fmt->sort  = hist_entry__cmp_nop;
1775		break;
1776	case PERF_HPP_DIFF__CYCLES_HIST:
1777		fmt->color = hpp__color_cycles_hist;
1778		fmt->sort  = hist_entry__cmp_nop;
1779		break;
1780	default:
1781		fmt->sort  = hist_entry__cmp_nop;
1782		break;
1783	}
1784
1785	init_header(d, dfmt);
1786	perf_hpp__column_register(fmt);
1787	perf_hpp__register_sort_field(fmt);
1788}
1789
1790static int ui_init(void)
1791{
1792	struct data__file *d;
1793	struct perf_hpp_fmt *fmt;
1794	int i;
1795
1796	data__for_each_file(i, d) {
1797
1798		/*
1799		 * Baseline or compute related columns:
1800		 *
1801		 *   PERF_HPP_DIFF__BASELINE
1802		 *   PERF_HPP_DIFF__DELTA
1803		 *   PERF_HPP_DIFF__RATIO
1804		 *   PERF_HPP_DIFF__WEIGHTED_DIFF
1805		 *   PERF_HPP_DIFF__CYCLES
1806		 */
1807		data__hpp_register(d, i ? compute_2_hpp[compute] :
1808					  PERF_HPP_DIFF__BASELINE);
1809
1810		if (cycles_hist && i)
1811			data__hpp_register(d, PERF_HPP_DIFF__CYCLES_HIST);
1812
1813		/*
1814		 * And the rest:
1815		 *
1816		 * PERF_HPP_DIFF__FORMULA
1817		 * PERF_HPP_DIFF__PERIOD
1818		 * PERF_HPP_DIFF__PERIOD_BASELINE
1819		 */
1820		if (show_formula && i)
1821			data__hpp_register(d, PERF_HPP_DIFF__FORMULA);
1822
1823		if (show_period)
1824			data__hpp_register(d, i ? PERF_HPP_DIFF__PERIOD :
1825						  PERF_HPP_DIFF__PERIOD_BASELINE);
1826	}
1827
1828	if (!sort_compute)
1829		return 0;
1830
1831	/*
1832	 * Prepend an fmt to sort on columns at 'sort_compute' first.
1833	 * This fmt is added only to the sort list but not to the
1834	 * output fields list.
1835	 *
1836	 * Note that this column (data) can be compared twice - one
1837	 * for this 'sort_compute' fmt and another for the normal
1838	 * diff_hpp_fmt.  But it shouldn't a problem as most entries
1839	 * will be sorted out by first try or baseline and comparing
1840	 * is not a costly operation.
1841	 */
1842	fmt = zalloc(sizeof(*fmt));
1843	if (fmt == NULL) {
1844		pr_err("Memory allocation failed\n");
1845		return -1;
1846	}
1847
1848	fmt->cmp      = hist_entry__cmp_nop;
1849	fmt->collapse = hist_entry__cmp_nop;
1850
1851	switch (compute) {
1852	case COMPUTE_DELTA:
1853		fmt->sort = hist_entry__cmp_delta_idx;
1854		break;
1855	case COMPUTE_RATIO:
1856		fmt->sort = hist_entry__cmp_ratio_idx;
1857		break;
1858	case COMPUTE_WEIGHTED_DIFF:
1859		fmt->sort = hist_entry__cmp_wdiff_idx;
1860		break;
1861	case COMPUTE_DELTA_ABS:
1862		fmt->sort = hist_entry__cmp_delta_abs_idx;
1863		break;
1864	case COMPUTE_CYCLES:
1865		/*
1866		 * Should set since 'fmt->sort' is called without
1867		 * checking valid during sorting
1868		 */
1869		fmt->sort = hist_entry__cmp_nop;
1870		break;
1871	default:
1872		BUG_ON(1);
1873	}
1874
1875	perf_hpp__prepend_sort_field(fmt);
1876	return 0;
1877}
1878
1879static int data_init(int argc, const char **argv)
1880{
1881	struct data__file *d;
1882	static const char *defaults[] = {
1883		"perf.data.old",
1884		"perf.data",
1885	};
1886	bool use_default = true;
1887	int i;
1888
1889	data__files_cnt = 2;
1890
1891	if (argc) {
1892		if (argc == 1)
1893			defaults[1] = argv[0];
1894		else {
1895			data__files_cnt = argc;
1896			use_default = false;
1897		}
1898	} else if (perf_guest) {
1899		defaults[0] = "perf.data.host";
1900		defaults[1] = "perf.data.guest";
1901	}
1902
1903	if (sort_compute >= (unsigned int) data__files_cnt) {
1904		pr_err("Order option out of limit.\n");
1905		return -EINVAL;
1906	}
1907
1908	data__files = zalloc(sizeof(*data__files) * data__files_cnt);
1909	if (!data__files)
1910		return -ENOMEM;
1911
1912	data__for_each_file(i, d) {
1913		struct perf_data *data = &d->data;
1914
1915		data->path  = use_default ? defaults[i] : argv[i];
1916		data->mode  = PERF_DATA_MODE_READ,
1917		data->force = force,
1918
1919		d->idx  = i;
1920	}
1921
1922	return 0;
1923}
1924
1925static int diff__config(const char *var, const char *value,
1926			void *cb __maybe_unused)
1927{
1928	if (!strcmp(var, "diff.order")) {
1929		int ret;
1930		if (perf_config_int(&ret, var, value) < 0)
1931			return -1;
1932		sort_compute = ret;
1933		return 0;
1934	}
1935	if (!strcmp(var, "diff.compute")) {
1936		if (!strcmp(value, "delta")) {
1937			compute = COMPUTE_DELTA;
1938		} else if (!strcmp(value, "delta-abs")) {
1939			compute = COMPUTE_DELTA_ABS;
1940		} else if (!strcmp(value, "ratio")) {
1941			compute = COMPUTE_RATIO;
1942		} else if (!strcmp(value, "wdiff")) {
1943			compute = COMPUTE_WEIGHTED_DIFF;
1944		} else {
1945			pr_err("Invalid compute method: %s\n", value);
1946			return -1;
1947		}
1948	}
1949
1950	return 0;
1951}
1952
1953int cmd_diff(int argc, const char **argv)
1954{
1955	int ret = hists__init();
1956
1957	if (ret < 0)
1958		return ret;
1959
1960	perf_config(diff__config, NULL);
1961
1962	argc = parse_options(argc, argv, options, diff_usage, 0);
1963
1964	if (quiet)
1965		perf_quiet_option();
1966
1967	if (cycles_hist && (compute != COMPUTE_CYCLES))
1968		usage_with_options(diff_usage, options);
1969
1970	if (pdiff.stream)
1971		compute = COMPUTE_STREAM;
1972
1973	symbol__annotation_init();
1974
1975	if (symbol__init(NULL) < 0)
1976		return -1;
1977
1978	if (data_init(argc, argv) < 0)
1979		return -1;
1980
1981	if (check_file_brstack() < 0)
1982		return -1;
1983
1984	if ((compute == COMPUTE_CYCLES || compute == COMPUTE_STREAM)
1985	    && !pdiff.has_br_stack) {
1986		return -1;
1987	}
1988
1989	if (compute == COMPUTE_STREAM) {
1990		symbol_conf.show_branchflag_count = true;
1991		symbol_conf.disable_add2line_warn = true;
1992		callchain_param.mode = CHAIN_FLAT;
1993		callchain_param.key = CCKEY_SRCLINE;
1994		callchain_param.branch_callstack = 1;
1995		symbol_conf.use_callchain = true;
1996		callchain_register_param(&callchain_param);
1997		sort_order = "srcline,symbol,dso";
1998	} else {
1999		if (ui_init() < 0)
2000			return -1;
2001
2002		sort__mode = SORT_MODE__DIFF;
2003	}
2004
2005	if (setup_sorting(NULL) < 0)
2006		usage_with_options(diff_usage, options);
2007
2008	setup_pager();
2009
2010	sort__setup_elide(NULL);
2011
2012	return __cmd_diff();
2013}