xref: /openbmc/linux/tools/perf/util/stat-shadow.c (revision f7af616c)
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 "cgroup.h"
12 #include "units.h"
13 #include <linux/zalloc.h>
14 #include "iostat.h"
15 
16 /*
17  * AGGR_GLOBAL: Use CPU 0
18  * AGGR_SOCKET: Use first CPU of socket
19  * AGGR_DIE: Use first CPU of die
20  * AGGR_CORE: Use first CPU of core
21  * AGGR_NONE: Use matching CPU
22  * AGGR_THREAD: Not supported?
23  */
24 
25 struct runtime_stat rt_stat;
26 struct stats walltime_nsecs_stats;
27 
28 struct saved_value {
29 	struct rb_node rb_node;
30 	struct evsel *evsel;
31 	enum stat_type type;
32 	int ctx;
33 	int cpu;
34 	struct cgroup *cgrp;
35 	struct runtime_stat *stat;
36 	struct stats stats;
37 	u64 metric_total;
38 	int metric_other;
39 };
40 
41 static int saved_value_cmp(struct rb_node *rb_node, const void *entry)
42 {
43 	struct saved_value *a = container_of(rb_node,
44 					     struct saved_value,
45 					     rb_node);
46 	const struct saved_value *b = entry;
47 
48 	if (a->cpu != b->cpu)
49 		return a->cpu - b->cpu;
50 
51 	/*
52 	 * Previously the rbtree was used to link generic metrics.
53 	 * The keys were evsel/cpu. Now the rbtree is extended to support
54 	 * per-thread shadow stats. For shadow stats case, the keys
55 	 * are cpu/type/ctx/stat (evsel is NULL). For generic metrics
56 	 * case, the keys are still evsel/cpu (type/ctx/stat are 0 or NULL).
57 	 */
58 	if (a->type != b->type)
59 		return a->type - b->type;
60 
61 	if (a->ctx != b->ctx)
62 		return a->ctx - b->ctx;
63 
64 	if (a->cgrp != b->cgrp)
65 		return (char *)a->cgrp < (char *)b->cgrp ? -1 : +1;
66 
67 	if (a->evsel == NULL && b->evsel == NULL) {
68 		if (a->stat == b->stat)
69 			return 0;
70 
71 		if ((char *)a->stat < (char *)b->stat)
72 			return -1;
73 
74 		return 1;
75 	}
76 
77 	if (a->evsel == b->evsel)
78 		return 0;
79 	if ((char *)a->evsel < (char *)b->evsel)
80 		return -1;
81 	return +1;
82 }
83 
84 static struct rb_node *saved_value_new(struct rblist *rblist __maybe_unused,
85 				     const void *entry)
86 {
87 	struct saved_value *nd = malloc(sizeof(struct saved_value));
88 
89 	if (!nd)
90 		return NULL;
91 	memcpy(nd, entry, sizeof(struct saved_value));
92 	return &nd->rb_node;
93 }
94 
95 static void saved_value_delete(struct rblist *rblist __maybe_unused,
96 			       struct rb_node *rb_node)
97 {
98 	struct saved_value *v;
99 
100 	BUG_ON(!rb_node);
101 	v = container_of(rb_node, struct saved_value, rb_node);
102 	free(v);
103 }
104 
105 static struct saved_value *saved_value_lookup(struct evsel *evsel,
106 					      int cpu,
107 					      bool create,
108 					      enum stat_type type,
109 					      int ctx,
110 					      struct runtime_stat *st,
111 					      struct cgroup *cgrp)
112 {
113 	struct rblist *rblist;
114 	struct rb_node *nd;
115 	struct saved_value dm = {
116 		.cpu = cpu,
117 		.evsel = evsel,
118 		.type = type,
119 		.ctx = ctx,
120 		.stat = st,
121 		.cgrp = cgrp,
122 	};
123 
124 	rblist = &st->value_list;
125 
126 	/* don't use context info for clock events */
127 	if (type == STAT_NSECS)
128 		dm.ctx = 0;
129 
130 	nd = rblist__find(rblist, &dm);
131 	if (nd)
132 		return container_of(nd, struct saved_value, rb_node);
133 	if (create) {
134 		rblist__add_node(rblist, &dm);
135 		nd = rblist__find(rblist, &dm);
136 		if (nd)
137 			return container_of(nd, struct saved_value, rb_node);
138 	}
139 	return NULL;
140 }
141 
142 void runtime_stat__init(struct runtime_stat *st)
143 {
144 	struct rblist *rblist = &st->value_list;
145 
146 	rblist__init(rblist);
147 	rblist->node_cmp = saved_value_cmp;
148 	rblist->node_new = saved_value_new;
149 	rblist->node_delete = saved_value_delete;
150 }
151 
152 void runtime_stat__exit(struct runtime_stat *st)
153 {
154 	rblist__exit(&st->value_list);
155 }
156 
157 void perf_stat__init_shadow_stats(void)
158 {
159 	runtime_stat__init(&rt_stat);
160 }
161 
162 static int evsel_context(struct evsel *evsel)
163 {
164 	int ctx = 0;
165 
166 	if (evsel->core.attr.exclude_kernel)
167 		ctx |= CTX_BIT_KERNEL;
168 	if (evsel->core.attr.exclude_user)
169 		ctx |= CTX_BIT_USER;
170 	if (evsel->core.attr.exclude_hv)
171 		ctx |= CTX_BIT_HV;
172 	if (evsel->core.attr.exclude_host)
173 		ctx |= CTX_BIT_HOST;
174 	if (evsel->core.attr.exclude_idle)
175 		ctx |= CTX_BIT_IDLE;
176 
177 	return ctx;
178 }
179 
180 static void reset_stat(struct runtime_stat *st)
181 {
182 	struct rblist *rblist;
183 	struct rb_node *pos, *next;
184 
185 	rblist = &st->value_list;
186 	next = rb_first_cached(&rblist->entries);
187 	while (next) {
188 		pos = next;
189 		next = rb_next(pos);
190 		memset(&container_of(pos, struct saved_value, rb_node)->stats,
191 		       0,
192 		       sizeof(struct stats));
193 	}
194 }
195 
196 void perf_stat__reset_shadow_stats(void)
197 {
198 	reset_stat(&rt_stat);
199 	memset(&walltime_nsecs_stats, 0, sizeof(walltime_nsecs_stats));
200 }
201 
202 void perf_stat__reset_shadow_per_stat(struct runtime_stat *st)
203 {
204 	reset_stat(st);
205 }
206 
207 struct runtime_stat_data {
208 	int ctx;
209 	struct cgroup *cgrp;
210 };
211 
212 static void update_runtime_stat(struct runtime_stat *st,
213 				enum stat_type type,
214 				int cpu, u64 count,
215 				struct runtime_stat_data *rsd)
216 {
217 	struct saved_value *v = saved_value_lookup(NULL, cpu, true, type,
218 						   rsd->ctx, st, rsd->cgrp);
219 
220 	if (v)
221 		update_stats(&v->stats, count);
222 }
223 
224 /*
225  * Update various tracking values we maintain to print
226  * more semantic information such as miss/hit ratios,
227  * instruction rates, etc:
228  */
229 void perf_stat__update_shadow_stats(struct evsel *counter, u64 count,
230 				    int cpu, struct runtime_stat *st)
231 {
232 	u64 count_ns = count;
233 	struct saved_value *v;
234 	struct runtime_stat_data rsd = {
235 		.ctx = evsel_context(counter),
236 		.cgrp = counter->cgrp,
237 	};
238 
239 	count *= counter->scale;
240 
241 	if (evsel__is_clock(counter))
242 		update_runtime_stat(st, STAT_NSECS, cpu, count_ns, &rsd);
243 	else if (evsel__match(counter, HARDWARE, HW_CPU_CYCLES))
244 		update_runtime_stat(st, STAT_CYCLES, cpu, count, &rsd);
245 	else if (perf_stat_evsel__is(counter, CYCLES_IN_TX))
246 		update_runtime_stat(st, STAT_CYCLES_IN_TX, cpu, count, &rsd);
247 	else if (perf_stat_evsel__is(counter, TRANSACTION_START))
248 		update_runtime_stat(st, STAT_TRANSACTION, cpu, count, &rsd);
249 	else if (perf_stat_evsel__is(counter, ELISION_START))
250 		update_runtime_stat(st, STAT_ELISION, cpu, count, &rsd);
251 	else if (perf_stat_evsel__is(counter, TOPDOWN_TOTAL_SLOTS))
252 		update_runtime_stat(st, STAT_TOPDOWN_TOTAL_SLOTS,
253 				    cpu, count, &rsd);
254 	else if (perf_stat_evsel__is(counter, TOPDOWN_SLOTS_ISSUED))
255 		update_runtime_stat(st, STAT_TOPDOWN_SLOTS_ISSUED,
256 				    cpu, count, &rsd);
257 	else if (perf_stat_evsel__is(counter, TOPDOWN_SLOTS_RETIRED))
258 		update_runtime_stat(st, STAT_TOPDOWN_SLOTS_RETIRED,
259 				    cpu, count, &rsd);
260 	else if (perf_stat_evsel__is(counter, TOPDOWN_FETCH_BUBBLES))
261 		update_runtime_stat(st, STAT_TOPDOWN_FETCH_BUBBLES,
262 				    cpu, count, &rsd);
263 	else if (perf_stat_evsel__is(counter, TOPDOWN_RECOVERY_BUBBLES))
264 		update_runtime_stat(st, STAT_TOPDOWN_RECOVERY_BUBBLES,
265 				    cpu, count, &rsd);
266 	else if (perf_stat_evsel__is(counter, TOPDOWN_RETIRING))
267 		update_runtime_stat(st, STAT_TOPDOWN_RETIRING,
268 				    cpu, count, &rsd);
269 	else if (perf_stat_evsel__is(counter, TOPDOWN_BAD_SPEC))
270 		update_runtime_stat(st, STAT_TOPDOWN_BAD_SPEC,
271 				    cpu, count, &rsd);
272 	else if (perf_stat_evsel__is(counter, TOPDOWN_FE_BOUND))
273 		update_runtime_stat(st, STAT_TOPDOWN_FE_BOUND,
274 				    cpu, count, &rsd);
275 	else if (perf_stat_evsel__is(counter, TOPDOWN_BE_BOUND))
276 		update_runtime_stat(st, STAT_TOPDOWN_BE_BOUND,
277 				    cpu, count, &rsd);
278 	else if (perf_stat_evsel__is(counter, TOPDOWN_HEAVY_OPS))
279 		update_runtime_stat(st, STAT_TOPDOWN_HEAVY_OPS,
280 				    cpu, count, &rsd);
281 	else if (perf_stat_evsel__is(counter, TOPDOWN_BR_MISPREDICT))
282 		update_runtime_stat(st, STAT_TOPDOWN_BR_MISPREDICT,
283 				    cpu, count, &rsd);
284 	else if (perf_stat_evsel__is(counter, TOPDOWN_FETCH_LAT))
285 		update_runtime_stat(st, STAT_TOPDOWN_FETCH_LAT,
286 				    cpu, count, &rsd);
287 	else if (perf_stat_evsel__is(counter, TOPDOWN_MEM_BOUND))
288 		update_runtime_stat(st, STAT_TOPDOWN_MEM_BOUND,
289 				    cpu, count, &rsd);
290 	else if (evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_FRONTEND))
291 		update_runtime_stat(st, STAT_STALLED_CYCLES_FRONT,
292 				    cpu, count, &rsd);
293 	else if (evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_BACKEND))
294 		update_runtime_stat(st, STAT_STALLED_CYCLES_BACK,
295 				    cpu, count, &rsd);
296 	else if (evsel__match(counter, HARDWARE, HW_BRANCH_INSTRUCTIONS))
297 		update_runtime_stat(st, STAT_BRANCHES, cpu, count, &rsd);
298 	else if (evsel__match(counter, HARDWARE, HW_CACHE_REFERENCES))
299 		update_runtime_stat(st, STAT_CACHEREFS, cpu, count, &rsd);
300 	else if (evsel__match(counter, HW_CACHE, HW_CACHE_L1D))
301 		update_runtime_stat(st, STAT_L1_DCACHE, cpu, count, &rsd);
302 	else if (evsel__match(counter, HW_CACHE, HW_CACHE_L1I))
303 		update_runtime_stat(st, STAT_L1_ICACHE, cpu, count, &rsd);
304 	else if (evsel__match(counter, HW_CACHE, HW_CACHE_LL))
305 		update_runtime_stat(st, STAT_LL_CACHE, cpu, count, &rsd);
306 	else if (evsel__match(counter, HW_CACHE, HW_CACHE_DTLB))
307 		update_runtime_stat(st, STAT_DTLB_CACHE, cpu, count, &rsd);
308 	else if (evsel__match(counter, HW_CACHE, HW_CACHE_ITLB))
309 		update_runtime_stat(st, STAT_ITLB_CACHE, cpu, count, &rsd);
310 	else if (perf_stat_evsel__is(counter, SMI_NUM))
311 		update_runtime_stat(st, STAT_SMI_NUM, cpu, count, &rsd);
312 	else if (perf_stat_evsel__is(counter, APERF))
313 		update_runtime_stat(st, STAT_APERF, cpu, count, &rsd);
314 
315 	if (counter->collect_stat) {
316 		v = saved_value_lookup(counter, cpu, true, STAT_NONE, 0, st,
317 				       rsd.cgrp);
318 		update_stats(&v->stats, count);
319 		if (counter->metric_leader)
320 			v->metric_total += count;
321 	} else if (counter->metric_leader) {
322 		v = saved_value_lookup(counter->metric_leader,
323 				       cpu, true, STAT_NONE, 0, st, rsd.cgrp);
324 		v->metric_total += count;
325 		v->metric_other++;
326 	}
327 }
328 
329 /* used for get_ratio_color() */
330 enum grc_type {
331 	GRC_STALLED_CYCLES_FE,
332 	GRC_STALLED_CYCLES_BE,
333 	GRC_CACHE_MISSES,
334 	GRC_MAX_NR
335 };
336 
337 static const char *get_ratio_color(enum grc_type type, double ratio)
338 {
339 	static const double grc_table[GRC_MAX_NR][3] = {
340 		[GRC_STALLED_CYCLES_FE] = { 50.0, 30.0, 10.0 },
341 		[GRC_STALLED_CYCLES_BE] = { 75.0, 50.0, 20.0 },
342 		[GRC_CACHE_MISSES] 	= { 20.0, 10.0, 5.0 },
343 	};
344 	const char *color = PERF_COLOR_NORMAL;
345 
346 	if (ratio > grc_table[type][0])
347 		color = PERF_COLOR_RED;
348 	else if (ratio > grc_table[type][1])
349 		color = PERF_COLOR_MAGENTA;
350 	else if (ratio > grc_table[type][2])
351 		color = PERF_COLOR_YELLOW;
352 
353 	return color;
354 }
355 
356 static struct evsel *perf_stat__find_event(struct evlist *evsel_list,
357 						const char *name)
358 {
359 	struct evsel *c2;
360 
361 	evlist__for_each_entry (evsel_list, c2) {
362 		if (!strcasecmp(c2->name, name) && !c2->collect_stat)
363 			return c2;
364 	}
365 	return NULL;
366 }
367 
368 /* Mark MetricExpr target events and link events using them to them. */
369 void perf_stat__collect_metric_expr(struct evlist *evsel_list)
370 {
371 	struct evsel *counter, *leader, **metric_events, *oc;
372 	bool found;
373 	struct expr_parse_ctx ctx;
374 	struct hashmap_entry *cur;
375 	size_t bkt;
376 	int i;
377 
378 	expr__ctx_init(&ctx);
379 	evlist__for_each_entry(evsel_list, counter) {
380 		bool invalid = false;
381 
382 		leader = counter->leader;
383 		if (!counter->metric_expr)
384 			continue;
385 
386 		expr__ctx_clear(&ctx);
387 		metric_events = counter->metric_events;
388 		if (!metric_events) {
389 			if (expr__find_other(counter->metric_expr,
390 					     counter->name,
391 					     &ctx, 1) < 0)
392 				continue;
393 
394 			metric_events = calloc(sizeof(struct evsel *),
395 					       hashmap__size(&ctx.ids) + 1);
396 			if (!metric_events) {
397 				expr__ctx_clear(&ctx);
398 				return;
399 			}
400 			counter->metric_events = metric_events;
401 		}
402 
403 		i = 0;
404 		hashmap__for_each_entry((&ctx.ids), cur, bkt) {
405 			const char *metric_name = (const char *)cur->key;
406 
407 			found = false;
408 			if (leader) {
409 				/* Search in group */
410 				for_each_group_member (oc, leader) {
411 					if (!strcasecmp(oc->name,
412 							metric_name) &&
413 						!oc->collect_stat) {
414 						found = true;
415 						break;
416 					}
417 				}
418 			}
419 			if (!found) {
420 				/* Search ignoring groups */
421 				oc = perf_stat__find_event(evsel_list,
422 							   metric_name);
423 			}
424 			if (!oc) {
425 				/* Deduping one is good enough to handle duplicated PMUs. */
426 				static char *printed;
427 
428 				/*
429 				 * Adding events automatically would be difficult, because
430 				 * it would risk creating groups that are not schedulable.
431 				 * perf stat doesn't understand all the scheduling constraints
432 				 * of events. So we ask the user instead to add the missing
433 				 * events.
434 				 */
435 				if (!printed ||
436 				    strcasecmp(printed, metric_name)) {
437 					fprintf(stderr,
438 						"Add %s event to groups to get metric expression for %s\n",
439 						metric_name,
440 						counter->name);
441 					printed = strdup(metric_name);
442 				}
443 				invalid = true;
444 				continue;
445 			}
446 			metric_events[i++] = oc;
447 			oc->collect_stat = true;
448 		}
449 		metric_events[i] = NULL;
450 		if (invalid) {
451 			free(metric_events);
452 			counter->metric_events = NULL;
453 			counter->metric_expr = NULL;
454 		}
455 	}
456 	expr__ctx_clear(&ctx);
457 }
458 
459 static double runtime_stat_avg(struct runtime_stat *st,
460 			       enum stat_type type, int cpu,
461 			       struct runtime_stat_data *rsd)
462 {
463 	struct saved_value *v;
464 
465 	v = saved_value_lookup(NULL, cpu, false, type, rsd->ctx, st, rsd->cgrp);
466 	if (!v)
467 		return 0.0;
468 
469 	return avg_stats(&v->stats);
470 }
471 
472 static double runtime_stat_n(struct runtime_stat *st,
473 			     enum stat_type type, int cpu,
474 			     struct runtime_stat_data *rsd)
475 {
476 	struct saved_value *v;
477 
478 	v = saved_value_lookup(NULL, cpu, false, type, rsd->ctx, st, rsd->cgrp);
479 	if (!v)
480 		return 0.0;
481 
482 	return v->stats.n;
483 }
484 
485 static void print_stalled_cycles_frontend(struct perf_stat_config *config,
486 					  int cpu, double avg,
487 					  struct perf_stat_output_ctx *out,
488 					  struct runtime_stat *st,
489 					  struct runtime_stat_data *rsd)
490 {
491 	double total, ratio = 0.0;
492 	const char *color;
493 
494 	total = runtime_stat_avg(st, STAT_CYCLES, cpu, rsd);
495 
496 	if (total)
497 		ratio = avg / total * 100.0;
498 
499 	color = get_ratio_color(GRC_STALLED_CYCLES_FE, ratio);
500 
501 	if (ratio)
502 		out->print_metric(config, out->ctx, color, "%7.2f%%", "frontend cycles idle",
503 				  ratio);
504 	else
505 		out->print_metric(config, out->ctx, NULL, NULL, "frontend cycles idle", 0);
506 }
507 
508 static void print_stalled_cycles_backend(struct perf_stat_config *config,
509 					 int cpu, double avg,
510 					 struct perf_stat_output_ctx *out,
511 					 struct runtime_stat *st,
512 					 struct runtime_stat_data *rsd)
513 {
514 	double total, ratio = 0.0;
515 	const char *color;
516 
517 	total = runtime_stat_avg(st, STAT_CYCLES, cpu, rsd);
518 
519 	if (total)
520 		ratio = avg / total * 100.0;
521 
522 	color = get_ratio_color(GRC_STALLED_CYCLES_BE, ratio);
523 
524 	out->print_metric(config, out->ctx, color, "%7.2f%%", "backend cycles idle", ratio);
525 }
526 
527 static void print_branch_misses(struct perf_stat_config *config,
528 				int cpu, double avg,
529 				struct perf_stat_output_ctx *out,
530 				struct runtime_stat *st,
531 				struct runtime_stat_data *rsd)
532 {
533 	double total, ratio = 0.0;
534 	const char *color;
535 
536 	total = runtime_stat_avg(st, STAT_BRANCHES, cpu, rsd);
537 
538 	if (total)
539 		ratio = avg / total * 100.0;
540 
541 	color = get_ratio_color(GRC_CACHE_MISSES, ratio);
542 
543 	out->print_metric(config, out->ctx, color, "%7.2f%%", "of all branches", ratio);
544 }
545 
546 static void print_l1_dcache_misses(struct perf_stat_config *config,
547 				   int cpu, double avg,
548 				   struct perf_stat_output_ctx *out,
549 				   struct runtime_stat *st,
550 				   struct runtime_stat_data *rsd)
551 {
552 	double total, ratio = 0.0;
553 	const char *color;
554 
555 	total = runtime_stat_avg(st, STAT_L1_DCACHE, cpu, rsd);
556 
557 	if (total)
558 		ratio = avg / total * 100.0;
559 
560 	color = get_ratio_color(GRC_CACHE_MISSES, ratio);
561 
562 	out->print_metric(config, out->ctx, color, "%7.2f%%", "of all L1-dcache accesses", ratio);
563 }
564 
565 static void print_l1_icache_misses(struct perf_stat_config *config,
566 				   int cpu, double avg,
567 				   struct perf_stat_output_ctx *out,
568 				   struct runtime_stat *st,
569 				   struct runtime_stat_data *rsd)
570 {
571 	double total, ratio = 0.0;
572 	const char *color;
573 
574 	total = runtime_stat_avg(st, STAT_L1_ICACHE, cpu, rsd);
575 
576 	if (total)
577 		ratio = avg / total * 100.0;
578 
579 	color = get_ratio_color(GRC_CACHE_MISSES, ratio);
580 	out->print_metric(config, out->ctx, color, "%7.2f%%", "of all L1-icache accesses", ratio);
581 }
582 
583 static void print_dtlb_cache_misses(struct perf_stat_config *config,
584 				    int cpu, double avg,
585 				    struct perf_stat_output_ctx *out,
586 				    struct runtime_stat *st,
587 				    struct runtime_stat_data *rsd)
588 {
589 	double total, ratio = 0.0;
590 	const char *color;
591 
592 	total = runtime_stat_avg(st, STAT_DTLB_CACHE, cpu, rsd);
593 
594 	if (total)
595 		ratio = avg / total * 100.0;
596 
597 	color = get_ratio_color(GRC_CACHE_MISSES, ratio);
598 	out->print_metric(config, out->ctx, color, "%7.2f%%", "of all dTLB cache accesses", ratio);
599 }
600 
601 static void print_itlb_cache_misses(struct perf_stat_config *config,
602 				    int cpu, double avg,
603 				    struct perf_stat_output_ctx *out,
604 				    struct runtime_stat *st,
605 				    struct runtime_stat_data *rsd)
606 {
607 	double total, ratio = 0.0;
608 	const char *color;
609 
610 	total = runtime_stat_avg(st, STAT_ITLB_CACHE, cpu, rsd);
611 
612 	if (total)
613 		ratio = avg / total * 100.0;
614 
615 	color = get_ratio_color(GRC_CACHE_MISSES, ratio);
616 	out->print_metric(config, out->ctx, color, "%7.2f%%", "of all iTLB cache accesses", ratio);
617 }
618 
619 static void print_ll_cache_misses(struct perf_stat_config *config,
620 				  int cpu, double avg,
621 				  struct perf_stat_output_ctx *out,
622 				  struct runtime_stat *st,
623 				  struct runtime_stat_data *rsd)
624 {
625 	double total, ratio = 0.0;
626 	const char *color;
627 
628 	total = runtime_stat_avg(st, STAT_LL_CACHE, cpu, rsd);
629 
630 	if (total)
631 		ratio = avg / total * 100.0;
632 
633 	color = get_ratio_color(GRC_CACHE_MISSES, ratio);
634 	out->print_metric(config, out->ctx, color, "%7.2f%%", "of all LL-cache accesses", ratio);
635 }
636 
637 /*
638  * High level "TopDown" CPU core pipe line bottleneck break down.
639  *
640  * Basic concept following
641  * Yasin, A Top Down Method for Performance analysis and Counter architecture
642  * ISPASS14
643  *
644  * The CPU pipeline is divided into 4 areas that can be bottlenecks:
645  *
646  * Frontend -> Backend -> Retiring
647  * BadSpeculation in addition means out of order execution that is thrown away
648  * (for example branch mispredictions)
649  * Frontend is instruction decoding.
650  * Backend is execution, like computation and accessing data in memory
651  * Retiring is good execution that is not directly bottlenecked
652  *
653  * The formulas are computed in slots.
654  * A slot is an entry in the pipeline each for the pipeline width
655  * (for example a 4-wide pipeline has 4 slots for each cycle)
656  *
657  * Formulas:
658  * BadSpeculation = ((SlotsIssued - SlotsRetired) + RecoveryBubbles) /
659  *			TotalSlots
660  * Retiring = SlotsRetired / TotalSlots
661  * FrontendBound = FetchBubbles / TotalSlots
662  * BackendBound = 1.0 - BadSpeculation - Retiring - FrontendBound
663  *
664  * The kernel provides the mapping to the low level CPU events and any scaling
665  * needed for the CPU pipeline width, for example:
666  *
667  * TotalSlots = Cycles * 4
668  *
669  * The scaling factor is communicated in the sysfs unit.
670  *
671  * In some cases the CPU may not be able to measure all the formulas due to
672  * missing events. In this case multiple formulas are combined, as possible.
673  *
674  * Full TopDown supports more levels to sub-divide each area: for example
675  * BackendBound into computing bound and memory bound. For now we only
676  * support Level 1 TopDown.
677  */
678 
679 static double sanitize_val(double x)
680 {
681 	if (x < 0 && x >= -0.02)
682 		return 0.0;
683 	return x;
684 }
685 
686 static double td_total_slots(int cpu, struct runtime_stat *st,
687 			     struct runtime_stat_data *rsd)
688 {
689 	return runtime_stat_avg(st, STAT_TOPDOWN_TOTAL_SLOTS, cpu, rsd);
690 }
691 
692 static double td_bad_spec(int cpu, struct runtime_stat *st,
693 			  struct runtime_stat_data *rsd)
694 {
695 	double bad_spec = 0;
696 	double total_slots;
697 	double total;
698 
699 	total = runtime_stat_avg(st, STAT_TOPDOWN_SLOTS_ISSUED, cpu, rsd) -
700 		runtime_stat_avg(st, STAT_TOPDOWN_SLOTS_RETIRED, cpu, rsd) +
701 		runtime_stat_avg(st, STAT_TOPDOWN_RECOVERY_BUBBLES, cpu, rsd);
702 
703 	total_slots = td_total_slots(cpu, st, rsd);
704 	if (total_slots)
705 		bad_spec = total / total_slots;
706 	return sanitize_val(bad_spec);
707 }
708 
709 static double td_retiring(int cpu, struct runtime_stat *st,
710 			  struct runtime_stat_data *rsd)
711 {
712 	double retiring = 0;
713 	double total_slots = td_total_slots(cpu, st, rsd);
714 	double ret_slots = runtime_stat_avg(st, STAT_TOPDOWN_SLOTS_RETIRED,
715 					    cpu, rsd);
716 
717 	if (total_slots)
718 		retiring = ret_slots / total_slots;
719 	return retiring;
720 }
721 
722 static double td_fe_bound(int cpu, struct runtime_stat *st,
723 			  struct runtime_stat_data *rsd)
724 {
725 	double fe_bound = 0;
726 	double total_slots = td_total_slots(cpu, st, rsd);
727 	double fetch_bub = runtime_stat_avg(st, STAT_TOPDOWN_FETCH_BUBBLES,
728 					    cpu, rsd);
729 
730 	if (total_slots)
731 		fe_bound = fetch_bub / total_slots;
732 	return fe_bound;
733 }
734 
735 static double td_be_bound(int cpu, struct runtime_stat *st,
736 			  struct runtime_stat_data *rsd)
737 {
738 	double sum = (td_fe_bound(cpu, st, rsd) +
739 		      td_bad_spec(cpu, st, rsd) +
740 		      td_retiring(cpu, st, rsd));
741 	if (sum == 0)
742 		return 0;
743 	return sanitize_val(1.0 - sum);
744 }
745 
746 /*
747  * Kernel reports metrics multiplied with slots. To get back
748  * the ratios we need to recreate the sum.
749  */
750 
751 static double td_metric_ratio(int cpu, enum stat_type type,
752 			      struct runtime_stat *stat,
753 			      struct runtime_stat_data *rsd)
754 {
755 	double sum = runtime_stat_avg(stat, STAT_TOPDOWN_RETIRING, cpu, rsd) +
756 		runtime_stat_avg(stat, STAT_TOPDOWN_FE_BOUND, cpu, rsd) +
757 		runtime_stat_avg(stat, STAT_TOPDOWN_BE_BOUND, cpu, rsd) +
758 		runtime_stat_avg(stat, STAT_TOPDOWN_BAD_SPEC, cpu, rsd);
759 	double d = runtime_stat_avg(stat, type, cpu, rsd);
760 
761 	if (sum)
762 		return d / sum;
763 	return 0;
764 }
765 
766 /*
767  * ... but only if most of the values are actually available.
768  * We allow two missing.
769  */
770 
771 static bool full_td(int cpu, struct runtime_stat *stat,
772 		    struct runtime_stat_data *rsd)
773 {
774 	int c = 0;
775 
776 	if (runtime_stat_avg(stat, STAT_TOPDOWN_RETIRING, cpu, rsd) > 0)
777 		c++;
778 	if (runtime_stat_avg(stat, STAT_TOPDOWN_BE_BOUND, cpu, rsd) > 0)
779 		c++;
780 	if (runtime_stat_avg(stat, STAT_TOPDOWN_FE_BOUND, cpu, rsd) > 0)
781 		c++;
782 	if (runtime_stat_avg(stat, STAT_TOPDOWN_BAD_SPEC, cpu, rsd) > 0)
783 		c++;
784 	return c >= 2;
785 }
786 
787 static void print_smi_cost(struct perf_stat_config *config, int cpu,
788 			   struct perf_stat_output_ctx *out,
789 			   struct runtime_stat *st,
790 			   struct runtime_stat_data *rsd)
791 {
792 	double smi_num, aperf, cycles, cost = 0.0;
793 	const char *color = NULL;
794 
795 	smi_num = runtime_stat_avg(st, STAT_SMI_NUM, cpu, rsd);
796 	aperf = runtime_stat_avg(st, STAT_APERF, cpu, rsd);
797 	cycles = runtime_stat_avg(st, STAT_CYCLES, cpu, rsd);
798 
799 	if ((cycles == 0) || (aperf == 0))
800 		return;
801 
802 	if (smi_num)
803 		cost = (aperf - cycles) / aperf * 100.00;
804 
805 	if (cost > 10)
806 		color = PERF_COLOR_RED;
807 	out->print_metric(config, out->ctx, color, "%8.1f%%", "SMI cycles%", cost);
808 	out->print_metric(config, out->ctx, NULL, "%4.0f", "SMI#", smi_num);
809 }
810 
811 static int prepare_metric(struct evsel **metric_events,
812 			  struct metric_ref *metric_refs,
813 			  struct expr_parse_ctx *pctx,
814 			  int cpu,
815 			  struct runtime_stat *st)
816 {
817 	double scale;
818 	char *n, *pn;
819 	int i, j, ret;
820 
821 	expr__ctx_init(pctx);
822 	for (i = 0; metric_events[i]; i++) {
823 		struct saved_value *v;
824 		struct stats *stats;
825 		u64 metric_total = 0;
826 
827 		if (!strcmp(metric_events[i]->name, "duration_time")) {
828 			stats = &walltime_nsecs_stats;
829 			scale = 1e-9;
830 		} else {
831 			v = saved_value_lookup(metric_events[i], cpu, false,
832 					       STAT_NONE, 0, st,
833 					       metric_events[i]->cgrp);
834 			if (!v)
835 				break;
836 			stats = &v->stats;
837 			scale = 1.0;
838 
839 			if (v->metric_other)
840 				metric_total = v->metric_total;
841 		}
842 
843 		n = strdup(metric_events[i]->name);
844 		if (!n)
845 			return -ENOMEM;
846 		/*
847 		 * This display code with --no-merge adds [cpu] postfixes.
848 		 * These are not supported by the parser. Remove everything
849 		 * after the space.
850 		 */
851 		pn = strchr(n, ' ');
852 		if (pn)
853 			*pn = 0;
854 
855 		if (metric_total)
856 			expr__add_id_val(pctx, n, metric_total);
857 		else
858 			expr__add_id_val(pctx, n, avg_stats(stats)*scale);
859 	}
860 
861 	for (j = 0; metric_refs && metric_refs[j].metric_name; j++) {
862 		ret = expr__add_ref(pctx, &metric_refs[j]);
863 		if (ret)
864 			return ret;
865 	}
866 
867 	return i;
868 }
869 
870 static void generic_metric(struct perf_stat_config *config,
871 			   const char *metric_expr,
872 			   struct evsel **metric_events,
873 			   struct metric_ref *metric_refs,
874 			   char *name,
875 			   const char *metric_name,
876 			   const char *metric_unit,
877 			   int runtime,
878 			   int cpu,
879 			   struct perf_stat_output_ctx *out,
880 			   struct runtime_stat *st)
881 {
882 	print_metric_t print_metric = out->print_metric;
883 	struct expr_parse_ctx pctx;
884 	double ratio, scale;
885 	int i;
886 	void *ctxp = out->ctx;
887 
888 	i = prepare_metric(metric_events, metric_refs, &pctx, cpu, st);
889 	if (i < 0)
890 		return;
891 
892 	if (!metric_events[i]) {
893 		if (expr__parse(&ratio, &pctx, metric_expr, runtime) == 0) {
894 			char *unit;
895 			char metric_bf[64];
896 
897 			if (metric_unit && metric_name) {
898 				if (perf_pmu__convert_scale(metric_unit,
899 					&unit, &scale) >= 0) {
900 					ratio *= scale;
901 				}
902 				if (strstr(metric_expr, "?"))
903 					scnprintf(metric_bf, sizeof(metric_bf),
904 					  "%s  %s_%d", unit, metric_name, runtime);
905 				else
906 					scnprintf(metric_bf, sizeof(metric_bf),
907 					  "%s  %s", unit, metric_name);
908 
909 				print_metric(config, ctxp, NULL, "%8.1f",
910 					     metric_bf, ratio);
911 			} else {
912 				print_metric(config, ctxp, NULL, "%8.2f",
913 					metric_name ?
914 					metric_name :
915 					out->force_header ?  name : "",
916 					ratio);
917 			}
918 		} else {
919 			print_metric(config, ctxp, NULL, NULL,
920 				     out->force_header ?
921 				     (metric_name ? metric_name : name) : "", 0);
922 		}
923 	} else {
924 		print_metric(config, ctxp, NULL, NULL,
925 			     out->force_header ?
926 			     (metric_name ? metric_name : name) : "", 0);
927 	}
928 
929 	expr__ctx_clear(&pctx);
930 }
931 
932 double test_generic_metric(struct metric_expr *mexp, int cpu, struct runtime_stat *st)
933 {
934 	struct expr_parse_ctx pctx;
935 	double ratio = 0.0;
936 
937 	if (prepare_metric(mexp->metric_events, mexp->metric_refs, &pctx, cpu, st) < 0)
938 		goto out;
939 
940 	if (expr__parse(&ratio, &pctx, mexp->metric_expr, 1))
941 		ratio = 0.0;
942 
943 out:
944 	expr__ctx_clear(&pctx);
945 	return ratio;
946 }
947 
948 void perf_stat__print_shadow_stats(struct perf_stat_config *config,
949 				   struct evsel *evsel,
950 				   double avg, int cpu,
951 				   struct perf_stat_output_ctx *out,
952 				   struct rblist *metric_events,
953 				   struct runtime_stat *st)
954 {
955 	void *ctxp = out->ctx;
956 	print_metric_t print_metric = out->print_metric;
957 	double total, ratio = 0.0, total2;
958 	const char *color = NULL;
959 	struct runtime_stat_data rsd = {
960 		.ctx = evsel_context(evsel),
961 		.cgrp = evsel->cgrp,
962 	};
963 	struct metric_event *me;
964 	int num = 1;
965 
966 	if (config->iostat_run) {
967 		iostat_print_metric(config, evsel, out);
968 	} else if (evsel__match(evsel, HARDWARE, HW_INSTRUCTIONS)) {
969 		total = runtime_stat_avg(st, STAT_CYCLES, cpu, &rsd);
970 
971 		if (total) {
972 			ratio = avg / total;
973 			print_metric(config, ctxp, NULL, "%7.2f ",
974 					"insn per cycle", ratio);
975 		} else {
976 			print_metric(config, ctxp, NULL, NULL, "insn per cycle", 0);
977 		}
978 
979 		total = runtime_stat_avg(st, STAT_STALLED_CYCLES_FRONT, cpu, &rsd);
980 
981 		total = max(total, runtime_stat_avg(st,
982 						    STAT_STALLED_CYCLES_BACK,
983 						    cpu, &rsd));
984 
985 		if (total && avg) {
986 			out->new_line(config, ctxp);
987 			ratio = total / avg;
988 			print_metric(config, ctxp, NULL, "%7.2f ",
989 					"stalled cycles per insn",
990 					ratio);
991 		}
992 	} else if (evsel__match(evsel, HARDWARE, HW_BRANCH_MISSES)) {
993 		if (runtime_stat_n(st, STAT_BRANCHES, cpu, &rsd) != 0)
994 			print_branch_misses(config, cpu, avg, out, st, &rsd);
995 		else
996 			print_metric(config, ctxp, NULL, NULL, "of all branches", 0);
997 	} else if (
998 		evsel->core.attr.type == PERF_TYPE_HW_CACHE &&
999 		evsel->core.attr.config ==  ( PERF_COUNT_HW_CACHE_L1D |
1000 					((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
1001 					 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) {
1002 
1003 		if (runtime_stat_n(st, STAT_L1_DCACHE, cpu, &rsd) != 0)
1004 			print_l1_dcache_misses(config, cpu, avg, out, st, &rsd);
1005 		else
1006 			print_metric(config, ctxp, NULL, NULL, "of all L1-dcache accesses", 0);
1007 	} else if (
1008 		evsel->core.attr.type == PERF_TYPE_HW_CACHE &&
1009 		evsel->core.attr.config ==  ( PERF_COUNT_HW_CACHE_L1I |
1010 					((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
1011 					 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) {
1012 
1013 		if (runtime_stat_n(st, STAT_L1_ICACHE, cpu, &rsd) != 0)
1014 			print_l1_icache_misses(config, cpu, avg, out, st, &rsd);
1015 		else
1016 			print_metric(config, ctxp, NULL, NULL, "of all L1-icache accesses", 0);
1017 	} else if (
1018 		evsel->core.attr.type == PERF_TYPE_HW_CACHE &&
1019 		evsel->core.attr.config ==  ( PERF_COUNT_HW_CACHE_DTLB |
1020 					((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
1021 					 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) {
1022 
1023 		if (runtime_stat_n(st, STAT_DTLB_CACHE, cpu, &rsd) != 0)
1024 			print_dtlb_cache_misses(config, cpu, avg, out, st, &rsd);
1025 		else
1026 			print_metric(config, ctxp, NULL, NULL, "of all dTLB cache accesses", 0);
1027 	} else if (
1028 		evsel->core.attr.type == PERF_TYPE_HW_CACHE &&
1029 		evsel->core.attr.config ==  ( PERF_COUNT_HW_CACHE_ITLB |
1030 					((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
1031 					 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) {
1032 
1033 		if (runtime_stat_n(st, STAT_ITLB_CACHE, cpu, &rsd) != 0)
1034 			print_itlb_cache_misses(config, cpu, avg, out, st, &rsd);
1035 		else
1036 			print_metric(config, ctxp, NULL, NULL, "of all iTLB cache accesses", 0);
1037 	} else if (
1038 		evsel->core.attr.type == PERF_TYPE_HW_CACHE &&
1039 		evsel->core.attr.config ==  ( PERF_COUNT_HW_CACHE_LL |
1040 					((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
1041 					 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) {
1042 
1043 		if (runtime_stat_n(st, STAT_LL_CACHE, cpu, &rsd) != 0)
1044 			print_ll_cache_misses(config, cpu, avg, out, st, &rsd);
1045 		else
1046 			print_metric(config, ctxp, NULL, NULL, "of all LL-cache accesses", 0);
1047 	} else if (evsel__match(evsel, HARDWARE, HW_CACHE_MISSES)) {
1048 		total = runtime_stat_avg(st, STAT_CACHEREFS, cpu, &rsd);
1049 
1050 		if (total)
1051 			ratio = avg * 100 / total;
1052 
1053 		if (runtime_stat_n(st, STAT_CACHEREFS, cpu, &rsd) != 0)
1054 			print_metric(config, ctxp, NULL, "%8.3f %%",
1055 				     "of all cache refs", ratio);
1056 		else
1057 			print_metric(config, ctxp, NULL, NULL, "of all cache refs", 0);
1058 	} else if (evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_FRONTEND)) {
1059 		print_stalled_cycles_frontend(config, cpu, avg, out, st, &rsd);
1060 	} else if (evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_BACKEND)) {
1061 		print_stalled_cycles_backend(config, cpu, avg, out, st, &rsd);
1062 	} else if (evsel__match(evsel, HARDWARE, HW_CPU_CYCLES)) {
1063 		total = runtime_stat_avg(st, STAT_NSECS, cpu, &rsd);
1064 
1065 		if (total) {
1066 			ratio = avg / total;
1067 			print_metric(config, ctxp, NULL, "%8.3f", "GHz", ratio);
1068 		} else {
1069 			print_metric(config, ctxp, NULL, NULL, "Ghz", 0);
1070 		}
1071 	} else if (perf_stat_evsel__is(evsel, CYCLES_IN_TX)) {
1072 		total = runtime_stat_avg(st, STAT_CYCLES, cpu, &rsd);
1073 
1074 		if (total)
1075 			print_metric(config, ctxp, NULL,
1076 					"%7.2f%%", "transactional cycles",
1077 					100.0 * (avg / total));
1078 		else
1079 			print_metric(config, ctxp, NULL, NULL, "transactional cycles",
1080 				     0);
1081 	} else if (perf_stat_evsel__is(evsel, CYCLES_IN_TX_CP)) {
1082 		total = runtime_stat_avg(st, STAT_CYCLES, cpu, &rsd);
1083 		total2 = runtime_stat_avg(st, STAT_CYCLES_IN_TX, cpu, &rsd);
1084 
1085 		if (total2 < avg)
1086 			total2 = avg;
1087 		if (total)
1088 			print_metric(config, ctxp, NULL, "%7.2f%%", "aborted cycles",
1089 				100.0 * ((total2-avg) / total));
1090 		else
1091 			print_metric(config, ctxp, NULL, NULL, "aborted cycles", 0);
1092 	} else if (perf_stat_evsel__is(evsel, TRANSACTION_START)) {
1093 		total = runtime_stat_avg(st, STAT_CYCLES_IN_TX, cpu, &rsd);
1094 
1095 		if (avg)
1096 			ratio = total / avg;
1097 
1098 		if (runtime_stat_n(st, STAT_CYCLES_IN_TX, cpu, &rsd) != 0)
1099 			print_metric(config, ctxp, NULL, "%8.0f",
1100 				     "cycles / transaction", ratio);
1101 		else
1102 			print_metric(config, ctxp, NULL, NULL, "cycles / transaction",
1103 				      0);
1104 	} else if (perf_stat_evsel__is(evsel, ELISION_START)) {
1105 		total = runtime_stat_avg(st, STAT_CYCLES_IN_TX, cpu, &rsd);
1106 
1107 		if (avg)
1108 			ratio = total / avg;
1109 
1110 		print_metric(config, ctxp, NULL, "%8.0f", "cycles / elision", ratio);
1111 	} else if (evsel__is_clock(evsel)) {
1112 		if ((ratio = avg_stats(&walltime_nsecs_stats)) != 0)
1113 			print_metric(config, ctxp, NULL, "%8.3f", "CPUs utilized",
1114 				     avg / (ratio * evsel->scale));
1115 		else
1116 			print_metric(config, ctxp, NULL, NULL, "CPUs utilized", 0);
1117 	} else if (perf_stat_evsel__is(evsel, TOPDOWN_FETCH_BUBBLES)) {
1118 		double fe_bound = td_fe_bound(cpu, st, &rsd);
1119 
1120 		if (fe_bound > 0.2)
1121 			color = PERF_COLOR_RED;
1122 		print_metric(config, ctxp, color, "%8.1f%%", "frontend bound",
1123 				fe_bound * 100.);
1124 	} else if (perf_stat_evsel__is(evsel, TOPDOWN_SLOTS_RETIRED)) {
1125 		double retiring = td_retiring(cpu, st, &rsd);
1126 
1127 		if (retiring > 0.7)
1128 			color = PERF_COLOR_GREEN;
1129 		print_metric(config, ctxp, color, "%8.1f%%", "retiring",
1130 				retiring * 100.);
1131 	} else if (perf_stat_evsel__is(evsel, TOPDOWN_RECOVERY_BUBBLES)) {
1132 		double bad_spec = td_bad_spec(cpu, st, &rsd);
1133 
1134 		if (bad_spec > 0.1)
1135 			color = PERF_COLOR_RED;
1136 		print_metric(config, ctxp, color, "%8.1f%%", "bad speculation",
1137 				bad_spec * 100.);
1138 	} else if (perf_stat_evsel__is(evsel, TOPDOWN_SLOTS_ISSUED)) {
1139 		double be_bound = td_be_bound(cpu, st, &rsd);
1140 		const char *name = "backend bound";
1141 		static int have_recovery_bubbles = -1;
1142 
1143 		/* In case the CPU does not support topdown-recovery-bubbles */
1144 		if (have_recovery_bubbles < 0)
1145 			have_recovery_bubbles = pmu_have_event("cpu",
1146 					"topdown-recovery-bubbles");
1147 		if (!have_recovery_bubbles)
1148 			name = "backend bound/bad spec";
1149 
1150 		if (be_bound > 0.2)
1151 			color = PERF_COLOR_RED;
1152 		if (td_total_slots(cpu, st, &rsd) > 0)
1153 			print_metric(config, ctxp, color, "%8.1f%%", name,
1154 					be_bound * 100.);
1155 		else
1156 			print_metric(config, ctxp, NULL, NULL, name, 0);
1157 	} else if (perf_stat_evsel__is(evsel, TOPDOWN_RETIRING) &&
1158 		   full_td(cpu, st, &rsd)) {
1159 		double retiring = td_metric_ratio(cpu,
1160 						  STAT_TOPDOWN_RETIRING, st,
1161 						  &rsd);
1162 		if (retiring > 0.7)
1163 			color = PERF_COLOR_GREEN;
1164 		print_metric(config, ctxp, color, "%8.1f%%", "retiring",
1165 				retiring * 100.);
1166 	} else if (perf_stat_evsel__is(evsel, TOPDOWN_FE_BOUND) &&
1167 		   full_td(cpu, st, &rsd)) {
1168 		double fe_bound = td_metric_ratio(cpu,
1169 						  STAT_TOPDOWN_FE_BOUND, st,
1170 						  &rsd);
1171 		if (fe_bound > 0.2)
1172 			color = PERF_COLOR_RED;
1173 		print_metric(config, ctxp, color, "%8.1f%%", "frontend bound",
1174 				fe_bound * 100.);
1175 	} else if (perf_stat_evsel__is(evsel, TOPDOWN_BE_BOUND) &&
1176 		   full_td(cpu, st, &rsd)) {
1177 		double be_bound = td_metric_ratio(cpu,
1178 						  STAT_TOPDOWN_BE_BOUND, st,
1179 						  &rsd);
1180 		if (be_bound > 0.2)
1181 			color = PERF_COLOR_RED;
1182 		print_metric(config, ctxp, color, "%8.1f%%", "backend bound",
1183 				be_bound * 100.);
1184 	} else if (perf_stat_evsel__is(evsel, TOPDOWN_BAD_SPEC) &&
1185 		   full_td(cpu, st, &rsd)) {
1186 		double bad_spec = td_metric_ratio(cpu,
1187 						  STAT_TOPDOWN_BAD_SPEC, st,
1188 						  &rsd);
1189 		if (bad_spec > 0.1)
1190 			color = PERF_COLOR_RED;
1191 		print_metric(config, ctxp, color, "%8.1f%%", "bad speculation",
1192 				bad_spec * 100.);
1193 	} else if (perf_stat_evsel__is(evsel, TOPDOWN_HEAVY_OPS) &&
1194 			full_td(cpu, st, &rsd) && (config->topdown_level > 1)) {
1195 		double retiring = td_metric_ratio(cpu,
1196 						  STAT_TOPDOWN_RETIRING, st,
1197 						  &rsd);
1198 		double heavy_ops = td_metric_ratio(cpu,
1199 						   STAT_TOPDOWN_HEAVY_OPS, st,
1200 						   &rsd);
1201 		double light_ops = retiring - heavy_ops;
1202 
1203 		if (retiring > 0.7 && heavy_ops > 0.1)
1204 			color = PERF_COLOR_GREEN;
1205 		print_metric(config, ctxp, color, "%8.1f%%", "heavy operations",
1206 				heavy_ops * 100.);
1207 		if (retiring > 0.7 && light_ops > 0.6)
1208 			color = PERF_COLOR_GREEN;
1209 		else
1210 			color = NULL;
1211 		print_metric(config, ctxp, color, "%8.1f%%", "light operations",
1212 				light_ops * 100.);
1213 	} else if (perf_stat_evsel__is(evsel, TOPDOWN_BR_MISPREDICT) &&
1214 			full_td(cpu, st, &rsd) && (config->topdown_level > 1)) {
1215 		double bad_spec = td_metric_ratio(cpu,
1216 						  STAT_TOPDOWN_BAD_SPEC, st,
1217 						  &rsd);
1218 		double br_mis = td_metric_ratio(cpu,
1219 						STAT_TOPDOWN_BR_MISPREDICT, st,
1220 						&rsd);
1221 		double m_clears = bad_spec - br_mis;
1222 
1223 		if (bad_spec > 0.1 && br_mis > 0.05)
1224 			color = PERF_COLOR_RED;
1225 		print_metric(config, ctxp, color, "%8.1f%%", "branch mispredict",
1226 				br_mis * 100.);
1227 		if (bad_spec > 0.1 && m_clears > 0.05)
1228 			color = PERF_COLOR_RED;
1229 		else
1230 			color = NULL;
1231 		print_metric(config, ctxp, color, "%8.1f%%", "machine clears",
1232 				m_clears * 100.);
1233 	} else if (perf_stat_evsel__is(evsel, TOPDOWN_FETCH_LAT) &&
1234 			full_td(cpu, st, &rsd) && (config->topdown_level > 1)) {
1235 		double fe_bound = td_metric_ratio(cpu,
1236 						  STAT_TOPDOWN_FE_BOUND, st,
1237 						  &rsd);
1238 		double fetch_lat = td_metric_ratio(cpu,
1239 						   STAT_TOPDOWN_FETCH_LAT, st,
1240 						   &rsd);
1241 		double fetch_bw = fe_bound - fetch_lat;
1242 
1243 		if (fe_bound > 0.2 && fetch_lat > 0.15)
1244 			color = PERF_COLOR_RED;
1245 		print_metric(config, ctxp, color, "%8.1f%%", "fetch latency",
1246 				fetch_lat * 100.);
1247 		if (fe_bound > 0.2 && fetch_bw > 0.1)
1248 			color = PERF_COLOR_RED;
1249 		else
1250 			color = NULL;
1251 		print_metric(config, ctxp, color, "%8.1f%%", "fetch bandwidth",
1252 				fetch_bw * 100.);
1253 	} else if (perf_stat_evsel__is(evsel, TOPDOWN_MEM_BOUND) &&
1254 			full_td(cpu, st, &rsd) && (config->topdown_level > 1)) {
1255 		double be_bound = td_metric_ratio(cpu,
1256 						  STAT_TOPDOWN_BE_BOUND, st,
1257 						  &rsd);
1258 		double mem_bound = td_metric_ratio(cpu,
1259 						   STAT_TOPDOWN_MEM_BOUND, st,
1260 						   &rsd);
1261 		double core_bound = be_bound - mem_bound;
1262 
1263 		if (be_bound > 0.2 && mem_bound > 0.2)
1264 			color = PERF_COLOR_RED;
1265 		print_metric(config, ctxp, color, "%8.1f%%", "memory bound",
1266 				mem_bound * 100.);
1267 		if (be_bound > 0.2 && core_bound > 0.1)
1268 			color = PERF_COLOR_RED;
1269 		else
1270 			color = NULL;
1271 		print_metric(config, ctxp, color, "%8.1f%%", "Core bound",
1272 				core_bound * 100.);
1273 	} else if (evsel->metric_expr) {
1274 		generic_metric(config, evsel->metric_expr, evsel->metric_events, NULL,
1275 				evsel->name, evsel->metric_name, NULL, 1, cpu, out, st);
1276 	} else if (runtime_stat_n(st, STAT_NSECS, cpu, &rsd) != 0) {
1277 		char unit = ' ';
1278 		char unit_buf[10] = "/sec";
1279 
1280 		total = runtime_stat_avg(st, STAT_NSECS, cpu, &rsd);
1281 		if (total)
1282 			ratio = convert_unit_double(1000000000.0 * avg / total, &unit);
1283 
1284 		if (unit != ' ')
1285 			snprintf(unit_buf, sizeof(unit_buf), "%c/sec", unit);
1286 		print_metric(config, ctxp, NULL, "%8.3f", unit_buf, ratio);
1287 	} else if (perf_stat_evsel__is(evsel, SMI_NUM)) {
1288 		print_smi_cost(config, cpu, out, st, &rsd);
1289 	} else {
1290 		num = 0;
1291 	}
1292 
1293 	if ((me = metricgroup__lookup(metric_events, evsel, false)) != NULL) {
1294 		struct metric_expr *mexp;
1295 
1296 		list_for_each_entry (mexp, &me->head, nd) {
1297 			if (num++ > 0)
1298 				out->new_line(config, ctxp);
1299 			generic_metric(config, mexp->metric_expr, mexp->metric_events,
1300 					mexp->metric_refs, evsel->name, mexp->metric_name,
1301 					mexp->metric_unit, mexp->runtime, cpu, out, st);
1302 		}
1303 	}
1304 	if (num == 0)
1305 		print_metric(config, ctxp, NULL, NULL, NULL, 0);
1306 }
1307