perf stat: Define a structure for per-thread shadow stats
[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
12 /*
13  * AGGR_GLOBAL: Use CPU 0
14  * AGGR_SOCKET: Use first CPU of socket
15  * AGGR_CORE: Use first CPU of core
16  * AGGR_NONE: Use matching CPU
17  * AGGR_THREAD: Not supported?
18  */
19 static struct stats runtime_nsecs_stats[MAX_NR_CPUS];
20 static struct stats runtime_cycles_stats[NUM_CTX][MAX_NR_CPUS];
21 static struct stats runtime_stalled_cycles_front_stats[NUM_CTX][MAX_NR_CPUS];
22 static struct stats runtime_stalled_cycles_back_stats[NUM_CTX][MAX_NR_CPUS];
23 static struct stats runtime_branches_stats[NUM_CTX][MAX_NR_CPUS];
24 static struct stats runtime_cacherefs_stats[NUM_CTX][MAX_NR_CPUS];
25 static struct stats runtime_l1_dcache_stats[NUM_CTX][MAX_NR_CPUS];
26 static struct stats runtime_l1_icache_stats[NUM_CTX][MAX_NR_CPUS];
27 static struct stats runtime_ll_cache_stats[NUM_CTX][MAX_NR_CPUS];
28 static struct stats runtime_itlb_cache_stats[NUM_CTX][MAX_NR_CPUS];
29 static struct stats runtime_dtlb_cache_stats[NUM_CTX][MAX_NR_CPUS];
30 static struct stats runtime_cycles_in_tx_stats[NUM_CTX][MAX_NR_CPUS];
31 static struct stats runtime_transaction_stats[NUM_CTX][MAX_NR_CPUS];
32 static struct stats runtime_elision_stats[NUM_CTX][MAX_NR_CPUS];
33 static struct stats runtime_topdown_total_slots[NUM_CTX][MAX_NR_CPUS];
34 static struct stats runtime_topdown_slots_issued[NUM_CTX][MAX_NR_CPUS];
35 static struct stats runtime_topdown_slots_retired[NUM_CTX][MAX_NR_CPUS];
36 static struct stats runtime_topdown_fetch_bubbles[NUM_CTX][MAX_NR_CPUS];
37 static struct stats runtime_topdown_recovery_bubbles[NUM_CTX][MAX_NR_CPUS];
38 static struct stats runtime_smi_num_stats[NUM_CTX][MAX_NR_CPUS];
39 static struct stats runtime_aperf_stats[NUM_CTX][MAX_NR_CPUS];
40 static struct rblist runtime_saved_values;
41 static bool have_frontend_stalled;
42
43 struct stats walltime_nsecs_stats;
44
45 struct saved_value {
46         struct rb_node rb_node;
47         struct perf_evsel *evsel;
48         int cpu;
49         struct stats stats;
50 };
51
52 static int saved_value_cmp(struct rb_node *rb_node, const void *entry)
53 {
54         struct saved_value *a = container_of(rb_node,
55                                              struct saved_value,
56                                              rb_node);
57         const struct saved_value *b = entry;
58
59         if (a->cpu != b->cpu)
60                 return a->cpu - b->cpu;
61         if (a->evsel == b->evsel)
62                 return 0;
63         if ((char *)a->evsel < (char *)b->evsel)
64                 return -1;
65         return +1;
66 }
67
68 static struct rb_node *saved_value_new(struct rblist *rblist __maybe_unused,
69                                      const void *entry)
70 {
71         struct saved_value *nd = malloc(sizeof(struct saved_value));
72
73         if (!nd)
74                 return NULL;
75         memcpy(nd, entry, sizeof(struct saved_value));
76         return &nd->rb_node;
77 }
78
79 static void saved_value_delete(struct rblist *rblist __maybe_unused,
80                                struct rb_node *rb_node)
81 {
82         struct saved_value *v;
83
84         BUG_ON(!rb_node);
85         v = container_of(rb_node, struct saved_value, rb_node);
86         free(v);
87 }
88
89 static struct saved_value *saved_value_lookup(struct perf_evsel *evsel,
90                                               int cpu,
91                                               bool create)
92 {
93         struct rb_node *nd;
94         struct saved_value dm = {
95                 .cpu = cpu,
96                 .evsel = evsel,
97         };
98         nd = rblist__find(&runtime_saved_values, &dm);
99         if (nd)
100                 return container_of(nd, struct saved_value, rb_node);
101         if (create) {
102                 rblist__add_node(&runtime_saved_values, &dm);
103                 nd = rblist__find(&runtime_saved_values, &dm);
104                 if (nd)
105                         return container_of(nd, struct saved_value, rb_node);
106         }
107         return NULL;
108 }
109
110 void perf_stat__init_shadow_stats(void)
111 {
112         have_frontend_stalled = pmu_have_event("cpu", "stalled-cycles-frontend");
113         rblist__init(&runtime_saved_values);
114         runtime_saved_values.node_cmp = saved_value_cmp;
115         runtime_saved_values.node_new = saved_value_new;
116         runtime_saved_values.node_delete = saved_value_delete;
117 }
118
119 static int evsel_context(struct perf_evsel *evsel)
120 {
121         int ctx = 0;
122
123         if (evsel->attr.exclude_kernel)
124                 ctx |= CTX_BIT_KERNEL;
125         if (evsel->attr.exclude_user)
126                 ctx |= CTX_BIT_USER;
127         if (evsel->attr.exclude_hv)
128                 ctx |= CTX_BIT_HV;
129         if (evsel->attr.exclude_host)
130                 ctx |= CTX_BIT_HOST;
131         if (evsel->attr.exclude_idle)
132                 ctx |= CTX_BIT_IDLE;
133
134         return ctx;
135 }
136
137 void perf_stat__reset_shadow_stats(void)
138 {
139         struct rb_node *pos, *next;
140
141         memset(runtime_nsecs_stats, 0, sizeof(runtime_nsecs_stats));
142         memset(runtime_cycles_stats, 0, sizeof(runtime_cycles_stats));
143         memset(runtime_stalled_cycles_front_stats, 0, sizeof(runtime_stalled_cycles_front_stats));
144         memset(runtime_stalled_cycles_back_stats, 0, sizeof(runtime_stalled_cycles_back_stats));
145         memset(runtime_branches_stats, 0, sizeof(runtime_branches_stats));
146         memset(runtime_cacherefs_stats, 0, sizeof(runtime_cacherefs_stats));
147         memset(runtime_l1_dcache_stats, 0, sizeof(runtime_l1_dcache_stats));
148         memset(runtime_l1_icache_stats, 0, sizeof(runtime_l1_icache_stats));
149         memset(runtime_ll_cache_stats, 0, sizeof(runtime_ll_cache_stats));
150         memset(runtime_itlb_cache_stats, 0, sizeof(runtime_itlb_cache_stats));
151         memset(runtime_dtlb_cache_stats, 0, sizeof(runtime_dtlb_cache_stats));
152         memset(runtime_cycles_in_tx_stats, 0,
153                         sizeof(runtime_cycles_in_tx_stats));
154         memset(runtime_transaction_stats, 0,
155                 sizeof(runtime_transaction_stats));
156         memset(runtime_elision_stats, 0, sizeof(runtime_elision_stats));
157         memset(&walltime_nsecs_stats, 0, sizeof(walltime_nsecs_stats));
158         memset(runtime_topdown_total_slots, 0, sizeof(runtime_topdown_total_slots));
159         memset(runtime_topdown_slots_retired, 0, sizeof(runtime_topdown_slots_retired));
160         memset(runtime_topdown_slots_issued, 0, sizeof(runtime_topdown_slots_issued));
161         memset(runtime_topdown_fetch_bubbles, 0, sizeof(runtime_topdown_fetch_bubbles));
162         memset(runtime_topdown_recovery_bubbles, 0, sizeof(runtime_topdown_recovery_bubbles));
163         memset(runtime_smi_num_stats, 0, sizeof(runtime_smi_num_stats));
164         memset(runtime_aperf_stats, 0, sizeof(runtime_aperf_stats));
165
166         next = rb_first(&runtime_saved_values.entries);
167         while (next) {
168                 pos = next;
169                 next = rb_next(pos);
170                 memset(&container_of(pos, struct saved_value, rb_node)->stats,
171                        0,
172                        sizeof(struct stats));
173         }
174 }
175
176 /*
177  * Update various tracking values we maintain to print
178  * more semantic information such as miss/hit ratios,
179  * instruction rates, etc:
180  */
181 void perf_stat__update_shadow_stats(struct perf_evsel *counter, u64 count,
182                                     int cpu)
183 {
184         int ctx = evsel_context(counter);
185
186         count *= counter->scale;
187
188         if (perf_evsel__match(counter, SOFTWARE, SW_TASK_CLOCK) ||
189             perf_evsel__match(counter, SOFTWARE, SW_CPU_CLOCK))
190                 update_stats(&runtime_nsecs_stats[cpu], count);
191         else if (perf_evsel__match(counter, HARDWARE, HW_CPU_CYCLES))
192                 update_stats(&runtime_cycles_stats[ctx][cpu], count);
193         else if (perf_stat_evsel__is(counter, CYCLES_IN_TX))
194                 update_stats(&runtime_cycles_in_tx_stats[ctx][cpu], count);
195         else if (perf_stat_evsel__is(counter, TRANSACTION_START))
196                 update_stats(&runtime_transaction_stats[ctx][cpu], count);
197         else if (perf_stat_evsel__is(counter, ELISION_START))
198                 update_stats(&runtime_elision_stats[ctx][cpu], count);
199         else if (perf_stat_evsel__is(counter, TOPDOWN_TOTAL_SLOTS))
200                 update_stats(&runtime_topdown_total_slots[ctx][cpu], count);
201         else if (perf_stat_evsel__is(counter, TOPDOWN_SLOTS_ISSUED))
202                 update_stats(&runtime_topdown_slots_issued[ctx][cpu], count);
203         else if (perf_stat_evsel__is(counter, TOPDOWN_SLOTS_RETIRED))
204                 update_stats(&runtime_topdown_slots_retired[ctx][cpu], count);
205         else if (perf_stat_evsel__is(counter, TOPDOWN_FETCH_BUBBLES))
206                 update_stats(&runtime_topdown_fetch_bubbles[ctx][cpu], count);
207         else if (perf_stat_evsel__is(counter, TOPDOWN_RECOVERY_BUBBLES))
208                 update_stats(&runtime_topdown_recovery_bubbles[ctx][cpu], count);
209         else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_FRONTEND))
210                 update_stats(&runtime_stalled_cycles_front_stats[ctx][cpu], count);
211         else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_BACKEND))
212                 update_stats(&runtime_stalled_cycles_back_stats[ctx][cpu], count);
213         else if (perf_evsel__match(counter, HARDWARE, HW_BRANCH_INSTRUCTIONS))
214                 update_stats(&runtime_branches_stats[ctx][cpu], count);
215         else if (perf_evsel__match(counter, HARDWARE, HW_CACHE_REFERENCES))
216                 update_stats(&runtime_cacherefs_stats[ctx][cpu], count);
217         else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_L1D))
218                 update_stats(&runtime_l1_dcache_stats[ctx][cpu], count);
219         else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_L1I))
220                 update_stats(&runtime_ll_cache_stats[ctx][cpu], count);
221         else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_LL))
222                 update_stats(&runtime_ll_cache_stats[ctx][cpu], count);
223         else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_DTLB))
224                 update_stats(&runtime_dtlb_cache_stats[ctx][cpu], count);
225         else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_ITLB))
226                 update_stats(&runtime_itlb_cache_stats[ctx][cpu], count);
227         else if (perf_stat_evsel__is(counter, SMI_NUM))
228                 update_stats(&runtime_smi_num_stats[ctx][cpu], count);
229         else if (perf_stat_evsel__is(counter, APERF))
230                 update_stats(&runtime_aperf_stats[ctx][cpu], count);
231
232         if (counter->collect_stat) {
233                 struct saved_value *v = saved_value_lookup(counter, cpu, true);
234                 update_stats(&v->stats, count);
235         }
236 }
237
238 /* used for get_ratio_color() */
239 enum grc_type {
240         GRC_STALLED_CYCLES_FE,
241         GRC_STALLED_CYCLES_BE,
242         GRC_CACHE_MISSES,
243         GRC_MAX_NR
244 };
245
246 static const char *get_ratio_color(enum grc_type type, double ratio)
247 {
248         static const double grc_table[GRC_MAX_NR][3] = {
249                 [GRC_STALLED_CYCLES_FE] = { 50.0, 30.0, 10.0 },
250                 [GRC_STALLED_CYCLES_BE] = { 75.0, 50.0, 20.0 },
251                 [GRC_CACHE_MISSES]      = { 20.0, 10.0, 5.0 },
252         };
253         const char *color = PERF_COLOR_NORMAL;
254
255         if (ratio > grc_table[type][0])
256                 color = PERF_COLOR_RED;
257         else if (ratio > grc_table[type][1])
258                 color = PERF_COLOR_MAGENTA;
259         else if (ratio > grc_table[type][2])
260                 color = PERF_COLOR_YELLOW;
261
262         return color;
263 }
264
265 static struct perf_evsel *perf_stat__find_event(struct perf_evlist *evsel_list,
266                                                 const char *name)
267 {
268         struct perf_evsel *c2;
269
270         evlist__for_each_entry (evsel_list, c2) {
271                 if (!strcasecmp(c2->name, name))
272                         return c2;
273         }
274         return NULL;
275 }
276
277 /* Mark MetricExpr target events and link events using them to them. */
278 void perf_stat__collect_metric_expr(struct perf_evlist *evsel_list)
279 {
280         struct perf_evsel *counter, *leader, **metric_events, *oc;
281         bool found;
282         const char **metric_names;
283         int i;
284         int num_metric_names;
285
286         evlist__for_each_entry(evsel_list, counter) {
287                 bool invalid = false;
288
289                 leader = counter->leader;
290                 if (!counter->metric_expr)
291                         continue;
292                 metric_events = counter->metric_events;
293                 if (!metric_events) {
294                         if (expr__find_other(counter->metric_expr, counter->name,
295                                                 &metric_names, &num_metric_names) < 0)
296                                 continue;
297
298                         metric_events = calloc(sizeof(struct perf_evsel *),
299                                                num_metric_names + 1);
300                         if (!metric_events)
301                                 return;
302                         counter->metric_events = metric_events;
303                 }
304
305                 for (i = 0; i < num_metric_names; i++) {
306                         found = false;
307                         if (leader) {
308                                 /* Search in group */
309                                 for_each_group_member (oc, leader) {
310                                         if (!strcasecmp(oc->name, metric_names[i])) {
311                                                 found = true;
312                                                 break;
313                                         }
314                                 }
315                         }
316                         if (!found) {
317                                 /* Search ignoring groups */
318                                 oc = perf_stat__find_event(evsel_list, metric_names[i]);
319                         }
320                         if (!oc) {
321                                 /* Deduping one is good enough to handle duplicated PMUs. */
322                                 static char *printed;
323
324                                 /*
325                                  * Adding events automatically would be difficult, because
326                                  * it would risk creating groups that are not schedulable.
327                                  * perf stat doesn't understand all the scheduling constraints
328                                  * of events. So we ask the user instead to add the missing
329                                  * events.
330                                  */
331                                 if (!printed || strcasecmp(printed, metric_names[i])) {
332                                         fprintf(stderr,
333                                                 "Add %s event to groups to get metric expression for %s\n",
334                                                 metric_names[i],
335                                                 counter->name);
336                                         printed = strdup(metric_names[i]);
337                                 }
338                                 invalid = true;
339                                 continue;
340                         }
341                         metric_events[i] = oc;
342                         oc->collect_stat = true;
343                 }
344                 metric_events[i] = NULL;
345                 free(metric_names);
346                 if (invalid) {
347                         free(metric_events);
348                         counter->metric_events = NULL;
349                         counter->metric_expr = NULL;
350                 }
351         }
352 }
353
354 static void print_stalled_cycles_frontend(int cpu,
355                                           struct perf_evsel *evsel, double avg,
356                                           struct perf_stat_output_ctx *out)
357 {
358         double total, ratio = 0.0;
359         const char *color;
360         int ctx = evsel_context(evsel);
361
362         total = avg_stats(&runtime_cycles_stats[ctx][cpu]);
363
364         if (total)
365                 ratio = avg / total * 100.0;
366
367         color = get_ratio_color(GRC_STALLED_CYCLES_FE, ratio);
368
369         if (ratio)
370                 out->print_metric(out->ctx, color, "%7.2f%%", "frontend cycles idle",
371                                   ratio);
372         else
373                 out->print_metric(out->ctx, NULL, NULL, "frontend cycles idle", 0);
374 }
375
376 static void print_stalled_cycles_backend(int cpu,
377                                          struct perf_evsel *evsel, double avg,
378                                          struct perf_stat_output_ctx *out)
379 {
380         double total, ratio = 0.0;
381         const char *color;
382         int ctx = evsel_context(evsel);
383
384         total = avg_stats(&runtime_cycles_stats[ctx][cpu]);
385
386         if (total)
387                 ratio = avg / total * 100.0;
388
389         color = get_ratio_color(GRC_STALLED_CYCLES_BE, ratio);
390
391         out->print_metric(out->ctx, color, "%7.2f%%", "backend cycles idle", ratio);
392 }
393
394 static void print_branch_misses(int cpu,
395                                 struct perf_evsel *evsel,
396                                 double avg,
397                                 struct perf_stat_output_ctx *out)
398 {
399         double total, ratio = 0.0;
400         const char *color;
401         int ctx = evsel_context(evsel);
402
403         total = avg_stats(&runtime_branches_stats[ctx][cpu]);
404
405         if (total)
406                 ratio = avg / total * 100.0;
407
408         color = get_ratio_color(GRC_CACHE_MISSES, ratio);
409
410         out->print_metric(out->ctx, color, "%7.2f%%", "of all branches", ratio);
411 }
412
413 static void print_l1_dcache_misses(int cpu,
414                                    struct perf_evsel *evsel,
415                                    double avg,
416                                    struct perf_stat_output_ctx *out)
417 {
418         double total, ratio = 0.0;
419         const char *color;
420         int ctx = evsel_context(evsel);
421
422         total = avg_stats(&runtime_l1_dcache_stats[ctx][cpu]);
423
424         if (total)
425                 ratio = avg / total * 100.0;
426
427         color = get_ratio_color(GRC_CACHE_MISSES, ratio);
428
429         out->print_metric(out->ctx, color, "%7.2f%%", "of all L1-dcache hits", ratio);
430 }
431
432 static void print_l1_icache_misses(int cpu,
433                                    struct perf_evsel *evsel,
434                                    double avg,
435                                    struct perf_stat_output_ctx *out)
436 {
437         double total, ratio = 0.0;
438         const char *color;
439         int ctx = evsel_context(evsel);
440
441         total = avg_stats(&runtime_l1_icache_stats[ctx][cpu]);
442
443         if (total)
444                 ratio = avg / total * 100.0;
445
446         color = get_ratio_color(GRC_CACHE_MISSES, ratio);
447         out->print_metric(out->ctx, color, "%7.2f%%", "of all L1-icache hits", ratio);
448 }
449
450 static void print_dtlb_cache_misses(int cpu,
451                                     struct perf_evsel *evsel,
452                                     double avg,
453                                     struct perf_stat_output_ctx *out)
454 {
455         double total, ratio = 0.0;
456         const char *color;
457         int ctx = evsel_context(evsel);
458
459         total = avg_stats(&runtime_dtlb_cache_stats[ctx][cpu]);
460
461         if (total)
462                 ratio = avg / total * 100.0;
463
464         color = get_ratio_color(GRC_CACHE_MISSES, ratio);
465         out->print_metric(out->ctx, color, "%7.2f%%", "of all dTLB cache hits", ratio);
466 }
467
468 static void print_itlb_cache_misses(int cpu,
469                                     struct perf_evsel *evsel,
470                                     double avg,
471                                     struct perf_stat_output_ctx *out)
472 {
473         double total, ratio = 0.0;
474         const char *color;
475         int ctx = evsel_context(evsel);
476
477         total = avg_stats(&runtime_itlb_cache_stats[ctx][cpu]);
478
479         if (total)
480                 ratio = avg / total * 100.0;
481
482         color = get_ratio_color(GRC_CACHE_MISSES, ratio);
483         out->print_metric(out->ctx, color, "%7.2f%%", "of all iTLB cache hits", ratio);
484 }
485
486 static void print_ll_cache_misses(int cpu,
487                                   struct perf_evsel *evsel,
488                                   double avg,
489                                   struct perf_stat_output_ctx *out)
490 {
491         double total, ratio = 0.0;
492         const char *color;
493         int ctx = evsel_context(evsel);
494
495         total = avg_stats(&runtime_ll_cache_stats[ctx][cpu]);
496
497         if (total)
498                 ratio = avg / total * 100.0;
499
500         color = get_ratio_color(GRC_CACHE_MISSES, ratio);
501         out->print_metric(out->ctx, color, "%7.2f%%", "of all LL-cache hits", ratio);
502 }
503
504 /*
505  * High level "TopDown" CPU core pipe line bottleneck break down.
506  *
507  * Basic concept following
508  * Yasin, A Top Down Method for Performance analysis and Counter architecture
509  * ISPASS14
510  *
511  * The CPU pipeline is divided into 4 areas that can be bottlenecks:
512  *
513  * Frontend -> Backend -> Retiring
514  * BadSpeculation in addition means out of order execution that is thrown away
515  * (for example branch mispredictions)
516  * Frontend is instruction decoding.
517  * Backend is execution, like computation and accessing data in memory
518  * Retiring is good execution that is not directly bottlenecked
519  *
520  * The formulas are computed in slots.
521  * A slot is an entry in the pipeline each for the pipeline width
522  * (for example a 4-wide pipeline has 4 slots for each cycle)
523  *
524  * Formulas:
525  * BadSpeculation = ((SlotsIssued - SlotsRetired) + RecoveryBubbles) /
526  *                      TotalSlots
527  * Retiring = SlotsRetired / TotalSlots
528  * FrontendBound = FetchBubbles / TotalSlots
529  * BackendBound = 1.0 - BadSpeculation - Retiring - FrontendBound
530  *
531  * The kernel provides the mapping to the low level CPU events and any scaling
532  * needed for the CPU pipeline width, for example:
533  *
534  * TotalSlots = Cycles * 4
535  *
536  * The scaling factor is communicated in the sysfs unit.
537  *
538  * In some cases the CPU may not be able to measure all the formulas due to
539  * missing events. In this case multiple formulas are combined, as possible.
540  *
541  * Full TopDown supports more levels to sub-divide each area: for example
542  * BackendBound into computing bound and memory bound. For now we only
543  * support Level 1 TopDown.
544  */
545
546 static double sanitize_val(double x)
547 {
548         if (x < 0 && x >= -0.02)
549                 return 0.0;
550         return x;
551 }
552
553 static double td_total_slots(int ctx, int cpu)
554 {
555         return avg_stats(&runtime_topdown_total_slots[ctx][cpu]);
556 }
557
558 static double td_bad_spec(int ctx, int cpu)
559 {
560         double bad_spec = 0;
561         double total_slots;
562         double total;
563
564         total = avg_stats(&runtime_topdown_slots_issued[ctx][cpu]) -
565                 avg_stats(&runtime_topdown_slots_retired[ctx][cpu]) +
566                 avg_stats(&runtime_topdown_recovery_bubbles[ctx][cpu]);
567         total_slots = td_total_slots(ctx, cpu);
568         if (total_slots)
569                 bad_spec = total / total_slots;
570         return sanitize_val(bad_spec);
571 }
572
573 static double td_retiring(int ctx, int cpu)
574 {
575         double retiring = 0;
576         double total_slots = td_total_slots(ctx, cpu);
577         double ret_slots = avg_stats(&runtime_topdown_slots_retired[ctx][cpu]);
578
579         if (total_slots)
580                 retiring = ret_slots / total_slots;
581         return retiring;
582 }
583
584 static double td_fe_bound(int ctx, int cpu)
585 {
586         double fe_bound = 0;
587         double total_slots = td_total_slots(ctx, cpu);
588         double fetch_bub = avg_stats(&runtime_topdown_fetch_bubbles[ctx][cpu]);
589
590         if (total_slots)
591                 fe_bound = fetch_bub / total_slots;
592         return fe_bound;
593 }
594
595 static double td_be_bound(int ctx, int cpu)
596 {
597         double sum = (td_fe_bound(ctx, cpu) +
598                       td_bad_spec(ctx, cpu) +
599                       td_retiring(ctx, cpu));
600         if (sum == 0)
601                 return 0;
602         return sanitize_val(1.0 - sum);
603 }
604
605 static void print_smi_cost(int cpu, struct perf_evsel *evsel,
606                            struct perf_stat_output_ctx *out)
607 {
608         double smi_num, aperf, cycles, cost = 0.0;
609         int ctx = evsel_context(evsel);
610         const char *color = NULL;
611
612         smi_num = avg_stats(&runtime_smi_num_stats[ctx][cpu]);
613         aperf = avg_stats(&runtime_aperf_stats[ctx][cpu]);
614         cycles = avg_stats(&runtime_cycles_stats[ctx][cpu]);
615
616         if ((cycles == 0) || (aperf == 0))
617                 return;
618
619         if (smi_num)
620                 cost = (aperf - cycles) / aperf * 100.00;
621
622         if (cost > 10)
623                 color = PERF_COLOR_RED;
624         out->print_metric(out->ctx, color, "%8.1f%%", "SMI cycles%", cost);
625         out->print_metric(out->ctx, NULL, "%4.0f", "SMI#", smi_num);
626 }
627
628 static void generic_metric(const char *metric_expr,
629                            struct perf_evsel **metric_events,
630                            char *name,
631                            const char *metric_name,
632                            double avg,
633                            int cpu,
634                            struct perf_stat_output_ctx *out)
635 {
636         print_metric_t print_metric = out->print_metric;
637         struct parse_ctx pctx;
638         double ratio;
639         int i;
640         void *ctxp = out->ctx;
641
642         expr__ctx_init(&pctx);
643         expr__add_id(&pctx, name, avg);
644         for (i = 0; metric_events[i]; i++) {
645                 struct saved_value *v;
646                 struct stats *stats;
647                 double scale;
648
649                 if (!strcmp(metric_events[i]->name, "duration_time")) {
650                         stats = &walltime_nsecs_stats;
651                         scale = 1e-9;
652                 } else {
653                         v = saved_value_lookup(metric_events[i], cpu, false);
654                         if (!v)
655                                 break;
656                         stats = &v->stats;
657                         scale = 1.0;
658                 }
659                 expr__add_id(&pctx, metric_events[i]->name, avg_stats(stats)*scale);
660         }
661         if (!metric_events[i]) {
662                 const char *p = metric_expr;
663
664                 if (expr__parse(&ratio, &pctx, &p) == 0)
665                         print_metric(ctxp, NULL, "%8.1f",
666                                 metric_name ?
667                                 metric_name :
668                                 out->force_header ?  name : "",
669                                 ratio);
670                 else
671                         print_metric(ctxp, NULL, NULL,
672                                      out->force_header ?
673                                      (metric_name ? metric_name : name) : "", 0);
674         } else
675                 print_metric(ctxp, NULL, NULL, "", 0);
676 }
677
678 void perf_stat__print_shadow_stats(struct perf_evsel *evsel,
679                                    double avg, int cpu,
680                                    struct perf_stat_output_ctx *out,
681                                    struct rblist *metric_events)
682 {
683         void *ctxp = out->ctx;
684         print_metric_t print_metric = out->print_metric;
685         double total, ratio = 0.0, total2;
686         const char *color = NULL;
687         int ctx = evsel_context(evsel);
688         struct metric_event *me;
689         int num = 1;
690
691         if (perf_evsel__match(evsel, HARDWARE, HW_INSTRUCTIONS)) {
692                 total = avg_stats(&runtime_cycles_stats[ctx][cpu]);
693                 if (total) {
694                         ratio = avg / total;
695                         print_metric(ctxp, NULL, "%7.2f ",
696                                         "insn per cycle", ratio);
697                 } else {
698                         print_metric(ctxp, NULL, NULL, "insn per cycle", 0);
699                 }
700                 total = avg_stats(&runtime_stalled_cycles_front_stats[ctx][cpu]);
701                 total = max(total, avg_stats(&runtime_stalled_cycles_back_stats[ctx][cpu]));
702
703                 if (total && avg) {
704                         out->new_line(ctxp);
705                         ratio = total / avg;
706                         print_metric(ctxp, NULL, "%7.2f ",
707                                         "stalled cycles per insn",
708                                         ratio);
709                 } else if (have_frontend_stalled) {
710                         print_metric(ctxp, NULL, NULL,
711                                      "stalled cycles per insn", 0);
712                 }
713         } else if (perf_evsel__match(evsel, HARDWARE, HW_BRANCH_MISSES)) {
714                 if (runtime_branches_stats[ctx][cpu].n != 0)
715                         print_branch_misses(cpu, evsel, avg, out);
716                 else
717                         print_metric(ctxp, NULL, NULL, "of all branches", 0);
718         } else if (
719                 evsel->attr.type == PERF_TYPE_HW_CACHE &&
720                 evsel->attr.config ==  ( PERF_COUNT_HW_CACHE_L1D |
721                                         ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
722                                          ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) {
723                 if (runtime_l1_dcache_stats[ctx][cpu].n != 0)
724                         print_l1_dcache_misses(cpu, evsel, avg, out);
725                 else
726                         print_metric(ctxp, NULL, NULL, "of all L1-dcache hits", 0);
727         } else if (
728                 evsel->attr.type == PERF_TYPE_HW_CACHE &&
729                 evsel->attr.config ==  ( PERF_COUNT_HW_CACHE_L1I |
730                                         ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
731                                          ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) {
732                 if (runtime_l1_icache_stats[ctx][cpu].n != 0)
733                         print_l1_icache_misses(cpu, evsel, avg, out);
734                 else
735                         print_metric(ctxp, NULL, NULL, "of all L1-icache hits", 0);
736         } else if (
737                 evsel->attr.type == PERF_TYPE_HW_CACHE &&
738                 evsel->attr.config ==  ( PERF_COUNT_HW_CACHE_DTLB |
739                                         ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
740                                          ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) {
741                 if (runtime_dtlb_cache_stats[ctx][cpu].n != 0)
742                         print_dtlb_cache_misses(cpu, evsel, avg, out);
743                 else
744                         print_metric(ctxp, NULL, NULL, "of all dTLB cache hits", 0);
745         } else if (
746                 evsel->attr.type == PERF_TYPE_HW_CACHE &&
747                 evsel->attr.config ==  ( PERF_COUNT_HW_CACHE_ITLB |
748                                         ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
749                                          ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) {
750                 if (runtime_itlb_cache_stats[ctx][cpu].n != 0)
751                         print_itlb_cache_misses(cpu, evsel, avg, out);
752                 else
753                         print_metric(ctxp, NULL, NULL, "of all iTLB cache hits", 0);
754         } else if (
755                 evsel->attr.type == PERF_TYPE_HW_CACHE &&
756                 evsel->attr.config ==  ( PERF_COUNT_HW_CACHE_LL |
757                                         ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
758                                          ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) {
759                 if (runtime_ll_cache_stats[ctx][cpu].n != 0)
760                         print_ll_cache_misses(cpu, evsel, avg, out);
761                 else
762                         print_metric(ctxp, NULL, NULL, "of all LL-cache hits", 0);
763         } else if (perf_evsel__match(evsel, HARDWARE, HW_CACHE_MISSES)) {
764                 total = avg_stats(&runtime_cacherefs_stats[ctx][cpu]);
765
766                 if (total)
767                         ratio = avg * 100 / total;
768
769                 if (runtime_cacherefs_stats[ctx][cpu].n != 0)
770                         print_metric(ctxp, NULL, "%8.3f %%",
771                                      "of all cache refs", ratio);
772                 else
773                         print_metric(ctxp, NULL, NULL, "of all cache refs", 0);
774         } else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_FRONTEND)) {
775                 print_stalled_cycles_frontend(cpu, evsel, avg, out);
776         } else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_BACKEND)) {
777                 print_stalled_cycles_backend(cpu, evsel, avg, out);
778         } else if (perf_evsel__match(evsel, HARDWARE, HW_CPU_CYCLES)) {
779                 total = avg_stats(&runtime_nsecs_stats[cpu]);
780
781                 if (total) {
782                         ratio = avg / total;
783                         print_metric(ctxp, NULL, "%8.3f", "GHz", ratio);
784                 } else {
785                         print_metric(ctxp, NULL, NULL, "Ghz", 0);
786                 }
787         } else if (perf_stat_evsel__is(evsel, CYCLES_IN_TX)) {
788                 total = avg_stats(&runtime_cycles_stats[ctx][cpu]);
789                 if (total)
790                         print_metric(ctxp, NULL,
791                                         "%7.2f%%", "transactional cycles",
792                                         100.0 * (avg / total));
793                 else
794                         print_metric(ctxp, NULL, NULL, "transactional cycles",
795                                      0);
796         } else if (perf_stat_evsel__is(evsel, CYCLES_IN_TX_CP)) {
797                 total = avg_stats(&runtime_cycles_stats[ctx][cpu]);
798                 total2 = avg_stats(&runtime_cycles_in_tx_stats[ctx][cpu]);
799                 if (total2 < avg)
800                         total2 = avg;
801                 if (total)
802                         print_metric(ctxp, NULL, "%7.2f%%", "aborted cycles",
803                                 100.0 * ((total2-avg) / total));
804                 else
805                         print_metric(ctxp, NULL, NULL, "aborted cycles", 0);
806         } else if (perf_stat_evsel__is(evsel, TRANSACTION_START)) {
807                 total = avg_stats(&runtime_cycles_in_tx_stats[ctx][cpu]);
808
809                 if (avg)
810                         ratio = total / avg;
811
812                 if (runtime_cycles_in_tx_stats[ctx][cpu].n != 0)
813                         print_metric(ctxp, NULL, "%8.0f",
814                                      "cycles / transaction", ratio);
815                 else
816                         print_metric(ctxp, NULL, NULL, "cycles / transaction",
817                                      0);
818         } else if (perf_stat_evsel__is(evsel, ELISION_START)) {
819                 total = avg_stats(&runtime_cycles_in_tx_stats[ctx][cpu]);
820
821                 if (avg)
822                         ratio = total / avg;
823
824                 print_metric(ctxp, NULL, "%8.0f", "cycles / elision", ratio);
825         } else if (perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK) ||
826                    perf_evsel__match(evsel, SOFTWARE, SW_CPU_CLOCK)) {
827                 if ((ratio = avg_stats(&walltime_nsecs_stats)) != 0)
828                         print_metric(ctxp, NULL, "%8.3f", "CPUs utilized",
829                                      avg / ratio);
830                 else
831                         print_metric(ctxp, NULL, NULL, "CPUs utilized", 0);
832         } else if (perf_stat_evsel__is(evsel, TOPDOWN_FETCH_BUBBLES)) {
833                 double fe_bound = td_fe_bound(ctx, cpu);
834
835                 if (fe_bound > 0.2)
836                         color = PERF_COLOR_RED;
837                 print_metric(ctxp, color, "%8.1f%%", "frontend bound",
838                                 fe_bound * 100.);
839         } else if (perf_stat_evsel__is(evsel, TOPDOWN_SLOTS_RETIRED)) {
840                 double retiring = td_retiring(ctx, cpu);
841
842                 if (retiring > 0.7)
843                         color = PERF_COLOR_GREEN;
844                 print_metric(ctxp, color, "%8.1f%%", "retiring",
845                                 retiring * 100.);
846         } else if (perf_stat_evsel__is(evsel, TOPDOWN_RECOVERY_BUBBLES)) {
847                 double bad_spec = td_bad_spec(ctx, cpu);
848
849                 if (bad_spec > 0.1)
850                         color = PERF_COLOR_RED;
851                 print_metric(ctxp, color, "%8.1f%%", "bad speculation",
852                                 bad_spec * 100.);
853         } else if (perf_stat_evsel__is(evsel, TOPDOWN_SLOTS_ISSUED)) {
854                 double be_bound = td_be_bound(ctx, cpu);
855                 const char *name = "backend bound";
856                 static int have_recovery_bubbles = -1;
857
858                 /* In case the CPU does not support topdown-recovery-bubbles */
859                 if (have_recovery_bubbles < 0)
860                         have_recovery_bubbles = pmu_have_event("cpu",
861                                         "topdown-recovery-bubbles");
862                 if (!have_recovery_bubbles)
863                         name = "backend bound/bad spec";
864
865                 if (be_bound > 0.2)
866                         color = PERF_COLOR_RED;
867                 if (td_total_slots(ctx, cpu) > 0)
868                         print_metric(ctxp, color, "%8.1f%%", name,
869                                         be_bound * 100.);
870                 else
871                         print_metric(ctxp, NULL, NULL, name, 0);
872         } else if (evsel->metric_expr) {
873                 generic_metric(evsel->metric_expr, evsel->metric_events, evsel->name,
874                                 evsel->metric_name, avg, cpu, out);
875         } else if (runtime_nsecs_stats[cpu].n != 0) {
876                 char unit = 'M';
877                 char unit_buf[10];
878
879                 total = avg_stats(&runtime_nsecs_stats[cpu]);
880
881                 if (total)
882                         ratio = 1000.0 * avg / total;
883                 if (ratio < 0.001) {
884                         ratio *= 1000;
885                         unit = 'K';
886                 }
887                 snprintf(unit_buf, sizeof(unit_buf), "%c/sec", unit);
888                 print_metric(ctxp, NULL, "%8.3f", unit_buf, ratio);
889         } else if (perf_stat_evsel__is(evsel, SMI_NUM)) {
890                 print_smi_cost(cpu, evsel, out);
891         } else {
892                 num = 0;
893         }
894
895         if ((me = metricgroup__lookup(metric_events, evsel, false)) != NULL) {
896                 struct metric_expr *mexp;
897
898                 list_for_each_entry (mexp, &me->head, nd) {
899                         if (num++ > 0)
900                                 out->new_line(ctxp);
901                         generic_metric(mexp->metric_expr, mexp->metric_events,
902                                         evsel->name, mexp->metric_name,
903                                         avg, cpu, out);
904                 }
905         }
906         if (num == 0)
907                 print_metric(ctxp, NULL, NULL, NULL, 0);
908 }