perf expr: Migrate expr ids table to a hashmap
[linux-2.6-microblaze.git] / tools / perf / util / stat-shadow.c
1 // SPDX-License-Identifier: GPL-2.0
2 #include <stdio.h>
3 #include "evsel.h"
4 #include "stat.h"
5 #include "color.h"
6 #include "pmu.h"
7 #include "rblist.h"
8 #include "evlist.h"
9 #include "expr.h"
10 #include "metricgroup.h"
11 #include <linux/zalloc.h>
12
13 /*
14  * AGGR_GLOBAL: Use CPU 0
15  * AGGR_SOCKET: Use first CPU of socket
16  * AGGR_DIE: Use first CPU of die
17  * AGGR_CORE: Use first CPU of core
18  * AGGR_NONE: Use matching CPU
19  * AGGR_THREAD: Not supported?
20  */
21
22 struct runtime_stat rt_stat;
23 struct stats walltime_nsecs_stats;
24
25 struct saved_value {
26         struct rb_node rb_node;
27         struct evsel *evsel;
28         enum stat_type type;
29         int ctx;
30         int cpu;
31         struct runtime_stat *stat;
32         struct stats stats;
33         u64 metric_total;
34         int metric_other;
35 };
36
37 static int saved_value_cmp(struct rb_node *rb_node, const void *entry)
38 {
39         struct saved_value *a = container_of(rb_node,
40                                              struct saved_value,
41                                              rb_node);
42         const struct saved_value *b = entry;
43
44         if (a->cpu != b->cpu)
45                 return a->cpu - b->cpu;
46
47         /*
48          * Previously the rbtree was used to link generic metrics.
49          * The keys were evsel/cpu. Now the rbtree is extended to support
50          * per-thread shadow stats. For shadow stats case, the keys
51          * are cpu/type/ctx/stat (evsel is NULL). For generic metrics
52          * case, the keys are still evsel/cpu (type/ctx/stat are 0 or NULL).
53          */
54         if (a->type != b->type)
55                 return a->type - b->type;
56
57         if (a->ctx != b->ctx)
58                 return a->ctx - b->ctx;
59
60         if (a->evsel == NULL && b->evsel == NULL) {
61                 if (a->stat == b->stat)
62                         return 0;
63
64                 if ((char *)a->stat < (char *)b->stat)
65                         return -1;
66
67                 return 1;
68         }
69
70         if (a->evsel == b->evsel)
71                 return 0;
72         if ((char *)a->evsel < (char *)b->evsel)
73                 return -1;
74         return +1;
75 }
76
77 static struct rb_node *saved_value_new(struct rblist *rblist __maybe_unused,
78                                      const void *entry)
79 {
80         struct saved_value *nd = malloc(sizeof(struct saved_value));
81
82         if (!nd)
83                 return NULL;
84         memcpy(nd, entry, sizeof(struct saved_value));
85         return &nd->rb_node;
86 }
87
88 static void saved_value_delete(struct rblist *rblist __maybe_unused,
89                                struct rb_node *rb_node)
90 {
91         struct saved_value *v;
92
93         BUG_ON(!rb_node);
94         v = container_of(rb_node, struct saved_value, rb_node);
95         free(v);
96 }
97
98 static struct saved_value *saved_value_lookup(struct evsel *evsel,
99                                               int cpu,
100                                               bool create,
101                                               enum stat_type type,
102                                               int ctx,
103                                               struct runtime_stat *st)
104 {
105         struct rblist *rblist;
106         struct rb_node *nd;
107         struct saved_value dm = {
108                 .cpu = cpu,
109                 .evsel = evsel,
110                 .type = type,
111                 .ctx = ctx,
112                 .stat = st,
113         };
114
115         rblist = &st->value_list;
116
117         nd = rblist__find(rblist, &dm);
118         if (nd)
119                 return container_of(nd, struct saved_value, rb_node);
120         if (create) {
121                 rblist__add_node(rblist, &dm);
122                 nd = rblist__find(rblist, &dm);
123                 if (nd)
124                         return container_of(nd, struct saved_value, rb_node);
125         }
126         return NULL;
127 }
128
129 void runtime_stat__init(struct runtime_stat *st)
130 {
131         struct rblist *rblist = &st->value_list;
132
133         rblist__init(rblist);
134         rblist->node_cmp = saved_value_cmp;
135         rblist->node_new = saved_value_new;
136         rblist->node_delete = saved_value_delete;
137 }
138
139 void runtime_stat__exit(struct runtime_stat *st)
140 {
141         rblist__exit(&st->value_list);
142 }
143
144 void perf_stat__init_shadow_stats(void)
145 {
146         runtime_stat__init(&rt_stat);
147 }
148
149 static int evsel_context(struct evsel *evsel)
150 {
151         int ctx = 0;
152
153         if (evsel->core.attr.exclude_kernel)
154                 ctx |= CTX_BIT_KERNEL;
155         if (evsel->core.attr.exclude_user)
156                 ctx |= CTX_BIT_USER;
157         if (evsel->core.attr.exclude_hv)
158                 ctx |= CTX_BIT_HV;
159         if (evsel->core.attr.exclude_host)
160                 ctx |= CTX_BIT_HOST;
161         if (evsel->core.attr.exclude_idle)
162                 ctx |= CTX_BIT_IDLE;
163
164         return ctx;
165 }
166
167 static void reset_stat(struct runtime_stat *st)
168 {
169         struct rblist *rblist;
170         struct rb_node *pos, *next;
171
172         rblist = &st->value_list;
173         next = rb_first_cached(&rblist->entries);
174         while (next) {
175                 pos = next;
176                 next = rb_next(pos);
177                 memset(&container_of(pos, struct saved_value, rb_node)->stats,
178                        0,
179                        sizeof(struct stats));
180         }
181 }
182
183 void perf_stat__reset_shadow_stats(void)
184 {
185         reset_stat(&rt_stat);
186         memset(&walltime_nsecs_stats, 0, sizeof(walltime_nsecs_stats));
187 }
188
189 void perf_stat__reset_shadow_per_stat(struct runtime_stat *st)
190 {
191         reset_stat(st);
192 }
193
194 static void update_runtime_stat(struct runtime_stat *st,
195                                 enum stat_type type,
196                                 int ctx, int cpu, u64 count)
197 {
198         struct saved_value *v = saved_value_lookup(NULL, cpu, true,
199                                                    type, ctx, st);
200
201         if (v)
202                 update_stats(&v->stats, count);
203 }
204
205 /*
206  * Update various tracking values we maintain to print
207  * more semantic information such as miss/hit ratios,
208  * instruction rates, etc:
209  */
210 void perf_stat__update_shadow_stats(struct evsel *counter, u64 count,
211                                     int cpu, struct runtime_stat *st)
212 {
213         int ctx = evsel_context(counter);
214         u64 count_ns = count;
215         struct saved_value *v;
216
217         count *= counter->scale;
218
219         if (evsel__is_clock(counter))
220                 update_runtime_stat(st, STAT_NSECS, 0, cpu, count_ns);
221         else if (evsel__match(counter, HARDWARE, HW_CPU_CYCLES))
222                 update_runtime_stat(st, STAT_CYCLES, ctx, cpu, count);
223         else if (perf_stat_evsel__is(counter, CYCLES_IN_TX))
224                 update_runtime_stat(st, STAT_CYCLES_IN_TX, ctx, cpu, count);
225         else if (perf_stat_evsel__is(counter, TRANSACTION_START))
226                 update_runtime_stat(st, STAT_TRANSACTION, ctx, cpu, count);
227         else if (perf_stat_evsel__is(counter, ELISION_START))
228                 update_runtime_stat(st, STAT_ELISION, ctx, cpu, count);
229         else if (perf_stat_evsel__is(counter, TOPDOWN_TOTAL_SLOTS))
230                 update_runtime_stat(st, STAT_TOPDOWN_TOTAL_SLOTS,
231                                     ctx, cpu, count);
232         else if (perf_stat_evsel__is(counter, TOPDOWN_SLOTS_ISSUED))
233                 update_runtime_stat(st, STAT_TOPDOWN_SLOTS_ISSUED,
234                                     ctx, cpu, count);
235         else if (perf_stat_evsel__is(counter, TOPDOWN_SLOTS_RETIRED))
236                 update_runtime_stat(st, STAT_TOPDOWN_SLOTS_RETIRED,
237                                     ctx, cpu, count);
238         else if (perf_stat_evsel__is(counter, TOPDOWN_FETCH_BUBBLES))
239                 update_runtime_stat(st, STAT_TOPDOWN_FETCH_BUBBLES,
240                                     ctx, cpu, count);
241         else if (perf_stat_evsel__is(counter, TOPDOWN_RECOVERY_BUBBLES))
242                 update_runtime_stat(st, STAT_TOPDOWN_RECOVERY_BUBBLES,
243                                     ctx, cpu, count);
244         else if (evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_FRONTEND))
245                 update_runtime_stat(st, STAT_STALLED_CYCLES_FRONT,
246                                     ctx, cpu, count);
247         else if (evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_BACKEND))
248                 update_runtime_stat(st, STAT_STALLED_CYCLES_BACK,
249                                     ctx, cpu, count);
250         else if (evsel__match(counter, HARDWARE, HW_BRANCH_INSTRUCTIONS))
251                 update_runtime_stat(st, STAT_BRANCHES, ctx, cpu, count);
252         else if (evsel__match(counter, HARDWARE, HW_CACHE_REFERENCES))
253                 update_runtime_stat(st, STAT_CACHEREFS, ctx, cpu, count);
254         else if (evsel__match(counter, HW_CACHE, HW_CACHE_L1D))
255                 update_runtime_stat(st, STAT_L1_DCACHE, ctx, cpu, count);
256         else if (evsel__match(counter, HW_CACHE, HW_CACHE_L1I))
257                 update_runtime_stat(st, STAT_L1_ICACHE, ctx, cpu, count);
258         else if (evsel__match(counter, HW_CACHE, HW_CACHE_LL))
259                 update_runtime_stat(st, STAT_LL_CACHE, ctx, cpu, count);
260         else if (evsel__match(counter, HW_CACHE, HW_CACHE_DTLB))
261                 update_runtime_stat(st, STAT_DTLB_CACHE, ctx, cpu, count);
262         else if (evsel__match(counter, HW_CACHE, HW_CACHE_ITLB))
263                 update_runtime_stat(st, STAT_ITLB_CACHE, ctx, cpu, count);
264         else if (perf_stat_evsel__is(counter, SMI_NUM))
265                 update_runtime_stat(st, STAT_SMI_NUM, ctx, cpu, count);
266         else if (perf_stat_evsel__is(counter, APERF))
267                 update_runtime_stat(st, STAT_APERF, ctx, cpu, count);
268
269         if (counter->collect_stat) {
270                 v = saved_value_lookup(counter, cpu, true, STAT_NONE, 0, st);
271                 update_stats(&v->stats, count);
272                 if (counter->metric_leader)
273                         v->metric_total += count;
274         } else if (counter->metric_leader) {
275                 v = saved_value_lookup(counter->metric_leader,
276                                        cpu, true, STAT_NONE, 0, st);
277                 v->metric_total += count;
278                 v->metric_other++;
279         }
280 }
281
282 /* used for get_ratio_color() */
283 enum grc_type {
284         GRC_STALLED_CYCLES_FE,
285         GRC_STALLED_CYCLES_BE,
286         GRC_CACHE_MISSES,
287         GRC_MAX_NR
288 };
289
290 static const char *get_ratio_color(enum grc_type type, double ratio)
291 {
292         static const double grc_table[GRC_MAX_NR][3] = {
293                 [GRC_STALLED_CYCLES_FE] = { 50.0, 30.0, 10.0 },
294                 [GRC_STALLED_CYCLES_BE] = { 75.0, 50.0, 20.0 },
295                 [GRC_CACHE_MISSES]      = { 20.0, 10.0, 5.0 },
296         };
297         const char *color = PERF_COLOR_NORMAL;
298
299         if (ratio > grc_table[type][0])
300                 color = PERF_COLOR_RED;
301         else if (ratio > grc_table[type][1])
302                 color = PERF_COLOR_MAGENTA;
303         else if (ratio > grc_table[type][2])
304                 color = PERF_COLOR_YELLOW;
305
306         return color;
307 }
308
309 static struct evsel *perf_stat__find_event(struct evlist *evsel_list,
310                                                 const char *name)
311 {
312         struct evsel *c2;
313
314         evlist__for_each_entry (evsel_list, c2) {
315                 if (!strcasecmp(c2->name, name) && !c2->collect_stat)
316                         return c2;
317         }
318         return NULL;
319 }
320
321 /* Mark MetricExpr target events and link events using them to them. */
322 void perf_stat__collect_metric_expr(struct evlist *evsel_list)
323 {
324         struct evsel *counter, *leader, **metric_events, *oc;
325         bool found;
326         struct expr_parse_ctx ctx;
327         struct hashmap_entry *cur;
328         size_t bkt;
329         int i;
330
331         expr__ctx_init(&ctx);
332         evlist__for_each_entry(evsel_list, counter) {
333                 bool invalid = false;
334
335                 leader = counter->leader;
336                 if (!counter->metric_expr)
337                         continue;
338
339                 expr__ctx_clear(&ctx);
340                 metric_events = counter->metric_events;
341                 if (!metric_events) {
342                         if (expr__find_other(counter->metric_expr,
343                                              counter->name,
344                                              &ctx, 1) < 0)
345                                 continue;
346
347                         metric_events = calloc(sizeof(struct evsel *),
348                                                hashmap__size(&ctx.ids) + 1);
349                         if (!metric_events) {
350                                 expr__ctx_clear(&ctx);
351                                 return;
352                         }
353                         counter->metric_events = metric_events;
354                 }
355
356                 i = 0;
357                 hashmap__for_each_entry((&ctx.ids), cur, bkt) {
358                         const char *metric_name = (const char *)cur->key;
359
360                         found = false;
361                         if (leader) {
362                                 /* Search in group */
363                                 for_each_group_member (oc, leader) {
364                                         if (!strcasecmp(oc->name,
365                                                         metric_name) &&
366                                                 !oc->collect_stat) {
367                                                 found = true;
368                                                 break;
369                                         }
370                                 }
371                         }
372                         if (!found) {
373                                 /* Search ignoring groups */
374                                 oc = perf_stat__find_event(evsel_list,
375                                                            metric_name);
376                         }
377                         if (!oc) {
378                                 /* Deduping one is good enough to handle duplicated PMUs. */
379                                 static char *printed;
380
381                                 /*
382                                  * Adding events automatically would be difficult, because
383                                  * it would risk creating groups that are not schedulable.
384                                  * perf stat doesn't understand all the scheduling constraints
385                                  * of events. So we ask the user instead to add the missing
386                                  * events.
387                                  */
388                                 if (!printed ||
389                                     strcasecmp(printed, metric_name)) {
390                                         fprintf(stderr,
391                                                 "Add %s event to groups to get metric expression for %s\n",
392                                                 metric_name,
393                                                 counter->name);
394                                         printed = strdup(metric_name);
395                                 }
396                                 invalid = true;
397                                 continue;
398                         }
399                         metric_events[i++] = oc;
400                         oc->collect_stat = true;
401                 }
402                 metric_events[i] = NULL;
403                 if (invalid) {
404                         free(metric_events);
405                         counter->metric_events = NULL;
406                         counter->metric_expr = NULL;
407                 }
408         }
409         expr__ctx_clear(&ctx);
410 }
411
412 static double runtime_stat_avg(struct runtime_stat *st,
413                                enum stat_type type, int ctx, int cpu)
414 {
415         struct saved_value *v;
416
417         v = saved_value_lookup(NULL, cpu, false, type, ctx, st);
418         if (!v)
419                 return 0.0;
420
421         return avg_stats(&v->stats);
422 }
423
424 static double runtime_stat_n(struct runtime_stat *st,
425                              enum stat_type type, int ctx, int cpu)
426 {
427         struct saved_value *v;
428
429         v = saved_value_lookup(NULL, cpu, false, type, ctx, st);
430         if (!v)
431                 return 0.0;
432
433         return v->stats.n;
434 }
435
436 static void print_stalled_cycles_frontend(struct perf_stat_config *config,
437                                           int cpu,
438                                           struct evsel *evsel, double avg,
439                                           struct perf_stat_output_ctx *out,
440                                           struct runtime_stat *st)
441 {
442         double total, ratio = 0.0;
443         const char *color;
444         int ctx = evsel_context(evsel);
445
446         total = runtime_stat_avg(st, STAT_CYCLES, ctx, cpu);
447
448         if (total)
449                 ratio = avg / total * 100.0;
450
451         color = get_ratio_color(GRC_STALLED_CYCLES_FE, ratio);
452
453         if (ratio)
454                 out->print_metric(config, out->ctx, color, "%7.2f%%", "frontend cycles idle",
455                                   ratio);
456         else
457                 out->print_metric(config, out->ctx, NULL, NULL, "frontend cycles idle", 0);
458 }
459
460 static void print_stalled_cycles_backend(struct perf_stat_config *config,
461                                          int cpu,
462                                          struct evsel *evsel, double avg,
463                                          struct perf_stat_output_ctx *out,
464                                          struct runtime_stat *st)
465 {
466         double total, ratio = 0.0;
467         const char *color;
468         int ctx = evsel_context(evsel);
469
470         total = runtime_stat_avg(st, STAT_CYCLES, ctx, cpu);
471
472         if (total)
473                 ratio = avg / total * 100.0;
474
475         color = get_ratio_color(GRC_STALLED_CYCLES_BE, ratio);
476
477         out->print_metric(config, out->ctx, color, "%7.2f%%", "backend cycles idle", ratio);
478 }
479
480 static void print_branch_misses(struct perf_stat_config *config,
481                                 int cpu,
482                                 struct evsel *evsel,
483                                 double avg,
484                                 struct perf_stat_output_ctx *out,
485                                 struct runtime_stat *st)
486 {
487         double total, ratio = 0.0;
488         const char *color;
489         int ctx = evsel_context(evsel);
490
491         total = runtime_stat_avg(st, STAT_BRANCHES, ctx, cpu);
492
493         if (total)
494                 ratio = avg / total * 100.0;
495
496         color = get_ratio_color(GRC_CACHE_MISSES, ratio);
497
498         out->print_metric(config, out->ctx, color, "%7.2f%%", "of all branches", ratio);
499 }
500
501 static void print_l1_dcache_misses(struct perf_stat_config *config,
502                                    int cpu,
503                                    struct evsel *evsel,
504                                    double avg,
505                                    struct perf_stat_output_ctx *out,
506                                    struct runtime_stat *st)
507
508 {
509         double total, ratio = 0.0;
510         const char *color;
511         int ctx = evsel_context(evsel);
512
513         total = runtime_stat_avg(st, STAT_L1_DCACHE, ctx, cpu);
514
515         if (total)
516                 ratio = avg / total * 100.0;
517
518         color = get_ratio_color(GRC_CACHE_MISSES, ratio);
519
520         out->print_metric(config, out->ctx, color, "%7.2f%%", "of all L1-dcache hits", ratio);
521 }
522
523 static void print_l1_icache_misses(struct perf_stat_config *config,
524                                    int cpu,
525                                    struct evsel *evsel,
526                                    double avg,
527                                    struct perf_stat_output_ctx *out,
528                                    struct runtime_stat *st)
529
530 {
531         double total, ratio = 0.0;
532         const char *color;
533         int ctx = evsel_context(evsel);
534
535         total = runtime_stat_avg(st, STAT_L1_ICACHE, ctx, cpu);
536
537         if (total)
538                 ratio = avg / total * 100.0;
539
540         color = get_ratio_color(GRC_CACHE_MISSES, ratio);
541         out->print_metric(config, out->ctx, color, "%7.2f%%", "of all L1-icache hits", ratio);
542 }
543
544 static void print_dtlb_cache_misses(struct perf_stat_config *config,
545                                     int cpu,
546                                     struct evsel *evsel,
547                                     double avg,
548                                     struct perf_stat_output_ctx *out,
549                                     struct runtime_stat *st)
550 {
551         double total, ratio = 0.0;
552         const char *color;
553         int ctx = evsel_context(evsel);
554
555         total = runtime_stat_avg(st, STAT_DTLB_CACHE, ctx, cpu);
556
557         if (total)
558                 ratio = avg / total * 100.0;
559
560         color = get_ratio_color(GRC_CACHE_MISSES, ratio);
561         out->print_metric(config, out->ctx, color, "%7.2f%%", "of all dTLB cache hits", ratio);
562 }
563
564 static void print_itlb_cache_misses(struct perf_stat_config *config,
565                                     int cpu,
566                                     struct evsel *evsel,
567                                     double avg,
568                                     struct perf_stat_output_ctx *out,
569                                     struct runtime_stat *st)
570 {
571         double total, ratio = 0.0;
572         const char *color;
573         int ctx = evsel_context(evsel);
574
575         total = runtime_stat_avg(st, STAT_ITLB_CACHE, ctx, cpu);
576
577         if (total)
578                 ratio = avg / total * 100.0;
579
580         color = get_ratio_color(GRC_CACHE_MISSES, ratio);
581         out->print_metric(config, out->ctx, color, "%7.2f%%", "of all iTLB cache hits", ratio);
582 }
583
584 static void print_ll_cache_misses(struct perf_stat_config *config,
585                                   int cpu,
586                                   struct evsel *evsel,
587                                   double avg,
588                                   struct perf_stat_output_ctx *out,
589                                   struct runtime_stat *st)
590 {
591         double total, ratio = 0.0;
592         const char *color;
593         int ctx = evsel_context(evsel);
594
595         total = runtime_stat_avg(st, STAT_LL_CACHE, ctx, cpu);
596
597         if (total)
598                 ratio = avg / total * 100.0;
599
600         color = get_ratio_color(GRC_CACHE_MISSES, ratio);
601         out->print_metric(config, out->ctx, color, "%7.2f%%", "of all LL-cache hits", ratio);
602 }
603
604 /*
605  * High level "TopDown" CPU core pipe line bottleneck break down.
606  *
607  * Basic concept following
608  * Yasin, A Top Down Method for Performance analysis and Counter architecture
609  * ISPASS14
610  *
611  * The CPU pipeline is divided into 4 areas that can be bottlenecks:
612  *
613  * Frontend -> Backend -> Retiring
614  * BadSpeculation in addition means out of order execution that is thrown away
615  * (for example branch mispredictions)
616  * Frontend is instruction decoding.
617  * Backend is execution, like computation and accessing data in memory
618  * Retiring is good execution that is not directly bottlenecked
619  *
620  * The formulas are computed in slots.
621  * A slot is an entry in the pipeline each for the pipeline width
622  * (for example a 4-wide pipeline has 4 slots for each cycle)
623  *
624  * Formulas:
625  * BadSpeculation = ((SlotsIssued - SlotsRetired) + RecoveryBubbles) /
626  *                      TotalSlots
627  * Retiring = SlotsRetired / TotalSlots
628  * FrontendBound = FetchBubbles / TotalSlots
629  * BackendBound = 1.0 - BadSpeculation - Retiring - FrontendBound
630  *
631  * The kernel provides the mapping to the low level CPU events and any scaling
632  * needed for the CPU pipeline width, for example:
633  *
634  * TotalSlots = Cycles * 4
635  *
636  * The scaling factor is communicated in the sysfs unit.
637  *
638  * In some cases the CPU may not be able to measure all the formulas due to
639  * missing events. In this case multiple formulas are combined, as possible.
640  *
641  * Full TopDown supports more levels to sub-divide each area: for example
642  * BackendBound into computing bound and memory bound. For now we only
643  * support Level 1 TopDown.
644  */
645
646 static double sanitize_val(double x)
647 {
648         if (x < 0 && x >= -0.02)
649                 return 0.0;
650         return x;
651 }
652
653 static double td_total_slots(int ctx, int cpu, struct runtime_stat *st)
654 {
655         return runtime_stat_avg(st, STAT_TOPDOWN_TOTAL_SLOTS, ctx, cpu);
656 }
657
658 static double td_bad_spec(int ctx, int cpu, struct runtime_stat *st)
659 {
660         double bad_spec = 0;
661         double total_slots;
662         double total;
663
664         total = runtime_stat_avg(st, STAT_TOPDOWN_SLOTS_ISSUED, ctx, cpu) -
665                 runtime_stat_avg(st, STAT_TOPDOWN_SLOTS_RETIRED, ctx, cpu) +
666                 runtime_stat_avg(st, STAT_TOPDOWN_RECOVERY_BUBBLES, ctx, cpu);
667
668         total_slots = td_total_slots(ctx, cpu, st);
669         if (total_slots)
670                 bad_spec = total / total_slots;
671         return sanitize_val(bad_spec);
672 }
673
674 static double td_retiring(int ctx, int cpu, struct runtime_stat *st)
675 {
676         double retiring = 0;
677         double total_slots = td_total_slots(ctx, cpu, st);
678         double ret_slots = runtime_stat_avg(st, STAT_TOPDOWN_SLOTS_RETIRED,
679                                             ctx, cpu);
680
681         if (total_slots)
682                 retiring = ret_slots / total_slots;
683         return retiring;
684 }
685
686 static double td_fe_bound(int ctx, int cpu, struct runtime_stat *st)
687 {
688         double fe_bound = 0;
689         double total_slots = td_total_slots(ctx, cpu, st);
690         double fetch_bub = runtime_stat_avg(st, STAT_TOPDOWN_FETCH_BUBBLES,
691                                             ctx, cpu);
692
693         if (total_slots)
694                 fe_bound = fetch_bub / total_slots;
695         return fe_bound;
696 }
697
698 static double td_be_bound(int ctx, int cpu, struct runtime_stat *st)
699 {
700         double sum = (td_fe_bound(ctx, cpu, st) +
701                       td_bad_spec(ctx, cpu, st) +
702                       td_retiring(ctx, cpu, st));
703         if (sum == 0)
704                 return 0;
705         return sanitize_val(1.0 - sum);
706 }
707
708 static void print_smi_cost(struct perf_stat_config *config,
709                            int cpu, struct evsel *evsel,
710                            struct perf_stat_output_ctx *out,
711                            struct runtime_stat *st)
712 {
713         double smi_num, aperf, cycles, cost = 0.0;
714         int ctx = evsel_context(evsel);
715         const char *color = NULL;
716
717         smi_num = runtime_stat_avg(st, STAT_SMI_NUM, ctx, cpu);
718         aperf = runtime_stat_avg(st, STAT_APERF, ctx, cpu);
719         cycles = runtime_stat_avg(st, STAT_CYCLES, ctx, cpu);
720
721         if ((cycles == 0) || (aperf == 0))
722                 return;
723
724         if (smi_num)
725                 cost = (aperf - cycles) / aperf * 100.00;
726
727         if (cost > 10)
728                 color = PERF_COLOR_RED;
729         out->print_metric(config, out->ctx, color, "%8.1f%%", "SMI cycles%", cost);
730         out->print_metric(config, out->ctx, NULL, "%4.0f", "SMI#", smi_num);
731 }
732
733 static void generic_metric(struct perf_stat_config *config,
734                            const char *metric_expr,
735                            struct evsel **metric_events,
736                            char *name,
737                            const char *metric_name,
738                            const char *metric_unit,
739                            int runtime,
740                            double avg,
741                            int cpu,
742                            struct perf_stat_output_ctx *out,
743                            struct runtime_stat *st)
744 {
745         print_metric_t print_metric = out->print_metric;
746         struct expr_parse_ctx pctx;
747         double ratio, scale;
748         int i;
749         void *ctxp = out->ctx;
750         char *n, *pn;
751
752         expr__ctx_init(&pctx);
753         /* Must be first id entry */
754         n = strdup(name);
755         if (!n)
756                 return;
757         expr__add_id(&pctx, n, avg);
758         for (i = 0; metric_events[i]; i++) {
759                 struct saved_value *v;
760                 struct stats *stats;
761                 u64 metric_total = 0;
762
763                 if (!strcmp(metric_events[i]->name, "duration_time")) {
764                         stats = &walltime_nsecs_stats;
765                         scale = 1e-9;
766                 } else {
767                         v = saved_value_lookup(metric_events[i], cpu, false,
768                                                STAT_NONE, 0, st);
769                         if (!v)
770                                 break;
771                         stats = &v->stats;
772                         scale = 1.0;
773
774                         if (v->metric_other)
775                                 metric_total = v->metric_total;
776                 }
777
778                 n = strdup(metric_events[i]->name);
779                 if (!n)
780                         return;
781                 /*
782                  * This display code with --no-merge adds [cpu] postfixes.
783                  * These are not supported by the parser. Remove everything
784                  * after the space.
785                  */
786                 pn = strchr(n, ' ');
787                 if (pn)
788                         *pn = 0;
789
790                 if (metric_total)
791                         expr__add_id(&pctx, n, metric_total);
792                 else
793                         expr__add_id(&pctx, n, avg_stats(stats)*scale);
794         }
795
796         if (!metric_events[i]) {
797                 if (expr__parse(&ratio, &pctx, metric_expr, runtime) == 0) {
798                         char *unit;
799                         char metric_bf[64];
800
801                         if (metric_unit && metric_name) {
802                                 if (perf_pmu__convert_scale(metric_unit,
803                                         &unit, &scale) >= 0) {
804                                         ratio *= scale;
805                                 }
806                                 if (strstr(metric_expr, "?"))
807                                         scnprintf(metric_bf, sizeof(metric_bf),
808                                           "%s  %s_%d", unit, metric_name, runtime);
809                                 else
810                                         scnprintf(metric_bf, sizeof(metric_bf),
811                                           "%s  %s", unit, metric_name);
812
813                                 print_metric(config, ctxp, NULL, "%8.1f",
814                                              metric_bf, ratio);
815                         } else {
816                                 print_metric(config, ctxp, NULL, "%8.2f",
817                                         metric_name ?
818                                         metric_name :
819                                         out->force_header ?  name : "",
820                                         ratio);
821                         }
822                 } else {
823                         print_metric(config, ctxp, NULL, NULL,
824                                      out->force_header ?
825                                      (metric_name ? metric_name : name) : "", 0);
826                 }
827         } else {
828                 print_metric(config, ctxp, NULL, NULL,
829                              out->force_header ?
830                              (metric_name ? metric_name : name) : "", 0);
831         }
832
833         expr__ctx_clear(&pctx);
834 }
835
836 void perf_stat__print_shadow_stats(struct perf_stat_config *config,
837                                    struct evsel *evsel,
838                                    double avg, int cpu,
839                                    struct perf_stat_output_ctx *out,
840                                    struct rblist *metric_events,
841                                    struct runtime_stat *st)
842 {
843         void *ctxp = out->ctx;
844         print_metric_t print_metric = out->print_metric;
845         double total, ratio = 0.0, total2;
846         const char *color = NULL;
847         int ctx = evsel_context(evsel);
848         struct metric_event *me;
849         int num = 1;
850
851         if (evsel__match(evsel, HARDWARE, HW_INSTRUCTIONS)) {
852                 total = runtime_stat_avg(st, STAT_CYCLES, ctx, cpu);
853
854                 if (total) {
855                         ratio = avg / total;
856                         print_metric(config, ctxp, NULL, "%7.2f ",
857                                         "insn per cycle", ratio);
858                 } else {
859                         print_metric(config, ctxp, NULL, NULL, "insn per cycle", 0);
860                 }
861
862                 total = runtime_stat_avg(st, STAT_STALLED_CYCLES_FRONT,
863                                          ctx, cpu);
864
865                 total = max(total, runtime_stat_avg(st,
866                                                     STAT_STALLED_CYCLES_BACK,
867                                                     ctx, cpu));
868
869                 if (total && avg) {
870                         out->new_line(config, ctxp);
871                         ratio = total / avg;
872                         print_metric(config, ctxp, NULL, "%7.2f ",
873                                         "stalled cycles per insn",
874                                         ratio);
875                 }
876         } else if (evsel__match(evsel, HARDWARE, HW_BRANCH_MISSES)) {
877                 if (runtime_stat_n(st, STAT_BRANCHES, ctx, cpu) != 0)
878                         print_branch_misses(config, cpu, evsel, avg, out, st);
879                 else
880                         print_metric(config, ctxp, NULL, NULL, "of all branches", 0);
881         } else if (
882                 evsel->core.attr.type == PERF_TYPE_HW_CACHE &&
883                 evsel->core.attr.config ==  ( PERF_COUNT_HW_CACHE_L1D |
884                                         ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
885                                          ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) {
886
887                 if (runtime_stat_n(st, STAT_L1_DCACHE, ctx, cpu) != 0)
888                         print_l1_dcache_misses(config, cpu, evsel, avg, out, st);
889                 else
890                         print_metric(config, ctxp, NULL, NULL, "of all L1-dcache hits", 0);
891         } else if (
892                 evsel->core.attr.type == PERF_TYPE_HW_CACHE &&
893                 evsel->core.attr.config ==  ( PERF_COUNT_HW_CACHE_L1I |
894                                         ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
895                                          ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) {
896
897                 if (runtime_stat_n(st, STAT_L1_ICACHE, ctx, cpu) != 0)
898                         print_l1_icache_misses(config, cpu, evsel, avg, out, st);
899                 else
900                         print_metric(config, ctxp, NULL, NULL, "of all L1-icache hits", 0);
901         } else if (
902                 evsel->core.attr.type == PERF_TYPE_HW_CACHE &&
903                 evsel->core.attr.config ==  ( PERF_COUNT_HW_CACHE_DTLB |
904                                         ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
905                                          ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) {
906
907                 if (runtime_stat_n(st, STAT_DTLB_CACHE, ctx, cpu) != 0)
908                         print_dtlb_cache_misses(config, cpu, evsel, avg, out, st);
909                 else
910                         print_metric(config, ctxp, NULL, NULL, "of all dTLB cache hits", 0);
911         } else if (
912                 evsel->core.attr.type == PERF_TYPE_HW_CACHE &&
913                 evsel->core.attr.config ==  ( PERF_COUNT_HW_CACHE_ITLB |
914                                         ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
915                                          ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) {
916
917                 if (runtime_stat_n(st, STAT_ITLB_CACHE, ctx, cpu) != 0)
918                         print_itlb_cache_misses(config, cpu, evsel, avg, out, st);
919                 else
920                         print_metric(config, ctxp, NULL, NULL, "of all iTLB cache hits", 0);
921         } else if (
922                 evsel->core.attr.type == PERF_TYPE_HW_CACHE &&
923                 evsel->core.attr.config ==  ( PERF_COUNT_HW_CACHE_LL |
924                                         ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
925                                          ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) {
926
927                 if (runtime_stat_n(st, STAT_LL_CACHE, ctx, cpu) != 0)
928                         print_ll_cache_misses(config, cpu, evsel, avg, out, st);
929                 else
930                         print_metric(config, ctxp, NULL, NULL, "of all LL-cache hits", 0);
931         } else if (evsel__match(evsel, HARDWARE, HW_CACHE_MISSES)) {
932                 total = runtime_stat_avg(st, STAT_CACHEREFS, ctx, cpu);
933
934                 if (total)
935                         ratio = avg * 100 / total;
936
937                 if (runtime_stat_n(st, STAT_CACHEREFS, ctx, cpu) != 0)
938                         print_metric(config, ctxp, NULL, "%8.3f %%",
939                                      "of all cache refs", ratio);
940                 else
941                         print_metric(config, ctxp, NULL, NULL, "of all cache refs", 0);
942         } else if (evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_FRONTEND)) {
943                 print_stalled_cycles_frontend(config, cpu, evsel, avg, out, st);
944         } else if (evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_BACKEND)) {
945                 print_stalled_cycles_backend(config, cpu, evsel, avg, out, st);
946         } else if (evsel__match(evsel, HARDWARE, HW_CPU_CYCLES)) {
947                 total = runtime_stat_avg(st, STAT_NSECS, 0, cpu);
948
949                 if (total) {
950                         ratio = avg / total;
951                         print_metric(config, ctxp, NULL, "%8.3f", "GHz", ratio);
952                 } else {
953                         print_metric(config, ctxp, NULL, NULL, "Ghz", 0);
954                 }
955         } else if (perf_stat_evsel__is(evsel, CYCLES_IN_TX)) {
956                 total = runtime_stat_avg(st, STAT_CYCLES, ctx, cpu);
957
958                 if (total)
959                         print_metric(config, ctxp, NULL,
960                                         "%7.2f%%", "transactional cycles",
961                                         100.0 * (avg / total));
962                 else
963                         print_metric(config, ctxp, NULL, NULL, "transactional cycles",
964                                      0);
965         } else if (perf_stat_evsel__is(evsel, CYCLES_IN_TX_CP)) {
966                 total = runtime_stat_avg(st, STAT_CYCLES, ctx, cpu);
967                 total2 = runtime_stat_avg(st, STAT_CYCLES_IN_TX, ctx, cpu);
968
969                 if (total2 < avg)
970                         total2 = avg;
971                 if (total)
972                         print_metric(config, ctxp, NULL, "%7.2f%%", "aborted cycles",
973                                 100.0 * ((total2-avg) / total));
974                 else
975                         print_metric(config, ctxp, NULL, NULL, "aborted cycles", 0);
976         } else if (perf_stat_evsel__is(evsel, TRANSACTION_START)) {
977                 total = runtime_stat_avg(st, STAT_CYCLES_IN_TX,
978                                          ctx, cpu);
979
980                 if (avg)
981                         ratio = total / avg;
982
983                 if (runtime_stat_n(st, STAT_CYCLES_IN_TX, ctx, cpu) != 0)
984                         print_metric(config, ctxp, NULL, "%8.0f",
985                                      "cycles / transaction", ratio);
986                 else
987                         print_metric(config, ctxp, NULL, NULL, "cycles / transaction",
988                                       0);
989         } else if (perf_stat_evsel__is(evsel, ELISION_START)) {
990                 total = runtime_stat_avg(st, STAT_CYCLES_IN_TX,
991                                          ctx, cpu);
992
993                 if (avg)
994                         ratio = total / avg;
995
996                 print_metric(config, ctxp, NULL, "%8.0f", "cycles / elision", ratio);
997         } else if (evsel__is_clock(evsel)) {
998                 if ((ratio = avg_stats(&walltime_nsecs_stats)) != 0)
999                         print_metric(config, ctxp, NULL, "%8.3f", "CPUs utilized",
1000                                      avg / (ratio * evsel->scale));
1001                 else
1002                         print_metric(config, ctxp, NULL, NULL, "CPUs utilized", 0);
1003         } else if (perf_stat_evsel__is(evsel, TOPDOWN_FETCH_BUBBLES)) {
1004                 double fe_bound = td_fe_bound(ctx, cpu, st);
1005
1006                 if (fe_bound > 0.2)
1007                         color = PERF_COLOR_RED;
1008                 print_metric(config, ctxp, color, "%8.1f%%", "frontend bound",
1009                                 fe_bound * 100.);
1010         } else if (perf_stat_evsel__is(evsel, TOPDOWN_SLOTS_RETIRED)) {
1011                 double retiring = td_retiring(ctx, cpu, st);
1012
1013                 if (retiring > 0.7)
1014                         color = PERF_COLOR_GREEN;
1015                 print_metric(config, ctxp, color, "%8.1f%%", "retiring",
1016                                 retiring * 100.);
1017         } else if (perf_stat_evsel__is(evsel, TOPDOWN_RECOVERY_BUBBLES)) {
1018                 double bad_spec = td_bad_spec(ctx, cpu, st);
1019
1020                 if (bad_spec > 0.1)
1021                         color = PERF_COLOR_RED;
1022                 print_metric(config, ctxp, color, "%8.1f%%", "bad speculation",
1023                                 bad_spec * 100.);
1024         } else if (perf_stat_evsel__is(evsel, TOPDOWN_SLOTS_ISSUED)) {
1025                 double be_bound = td_be_bound(ctx, cpu, st);
1026                 const char *name = "backend bound";
1027                 static int have_recovery_bubbles = -1;
1028
1029                 /* In case the CPU does not support topdown-recovery-bubbles */
1030                 if (have_recovery_bubbles < 0)
1031                         have_recovery_bubbles = pmu_have_event("cpu",
1032                                         "topdown-recovery-bubbles");
1033                 if (!have_recovery_bubbles)
1034                         name = "backend bound/bad spec";
1035
1036                 if (be_bound > 0.2)
1037                         color = PERF_COLOR_RED;
1038                 if (td_total_slots(ctx, cpu, st) > 0)
1039                         print_metric(config, ctxp, color, "%8.1f%%", name,
1040                                         be_bound * 100.);
1041                 else
1042                         print_metric(config, ctxp, NULL, NULL, name, 0);
1043         } else if (evsel->metric_expr) {
1044                 generic_metric(config, evsel->metric_expr, evsel->metric_events, evsel->name,
1045                                 evsel->metric_name, NULL, 1, avg, cpu, out, st);
1046         } else if (runtime_stat_n(st, STAT_NSECS, 0, cpu) != 0) {
1047                 char unit = 'M';
1048                 char unit_buf[10];
1049
1050                 total = runtime_stat_avg(st, STAT_NSECS, 0, cpu);
1051
1052                 if (total)
1053                         ratio = 1000.0 * avg / total;
1054                 if (ratio < 0.001) {
1055                         ratio *= 1000;
1056                         unit = 'K';
1057                 }
1058                 snprintf(unit_buf, sizeof(unit_buf), "%c/sec", unit);
1059                 print_metric(config, ctxp, NULL, "%8.3f", unit_buf, ratio);
1060         } else if (perf_stat_evsel__is(evsel, SMI_NUM)) {
1061                 print_smi_cost(config, cpu, evsel, out, st);
1062         } else {
1063                 num = 0;
1064         }
1065
1066         if ((me = metricgroup__lookup(metric_events, evsel, false)) != NULL) {
1067                 struct metric_expr *mexp;
1068
1069                 list_for_each_entry (mexp, &me->head, nd) {
1070                         if (num++ > 0)
1071                                 out->new_line(config, ctxp);
1072                         generic_metric(config, mexp->metric_expr, mexp->metric_events,
1073                                         evsel->name, mexp->metric_name,
1074                                         mexp->metric_unit, mexp->runtime, avg, cpu, out, st);
1075                 }
1076         }
1077         if (num == 0)
1078                 print_metric(config, ctxp, NULL, NULL, NULL, 0);
1079 }