tools/perf/builtin-diff.c

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