xref: /openbmc/linux/tools/perf/util/stat-shadow.c (revision d9fd5a71)
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 <linux/zalloc.h>
13 
14 /*
15  * AGGR_GLOBAL: Use CPU 0
16  * AGGR_SOCKET: Use first CPU of socket
17  * AGGR_DIE: Use first CPU of die
18  * AGGR_CORE: Use first CPU of core
19  * AGGR_NONE: Use matching CPU
20  * AGGR_THREAD: Not supported?
21  */
22 
23 struct runtime_stat rt_stat;
24 struct stats walltime_nsecs_stats;
25 
26 struct saved_value {
27 	struct rb_node rb_node;
28 	struct evsel *evsel;
29 	enum stat_type type;
30 	int ctx;
31 	int cpu;
32 	struct cgroup *cgrp;
33 	struct runtime_stat *stat;
34 	struct stats stats;
35 	u64 metric_total;
36 	int metric_other;
37 };
38 
39 static int saved_value_cmp(struct rb_node *rb_node, const void *entry)
40 {
41 	struct saved_value *a = container_of(rb_node,
42 					     struct saved_value,
43 					     rb_node);
44 	const struct saved_value *b = entry;
45 
46 	if (a->cpu != b->cpu)
47 		return a->cpu - b->cpu;
48 
49 	/*
50 	 * Previously the rbtree was used to link generic metrics.
51 	 * The keys were evsel/cpu. Now the rbtree is extended to support
52 	 * per-thread shadow stats. For shadow stats case, the keys
53 	 * are cpu/type/ctx/stat (evsel is NULL). For generic metrics
54 	 * case, the keys are still evsel/cpu (type/ctx/stat are 0 or NULL).
55 	 */
56 	if (a->type != b->type)
57 		return a->type - b->type;
58 
59 	if (a->ctx != b->ctx)
60 		return a->ctx - b->ctx;
61 
62 	if (a->cgrp != b->cgrp)
63 		return (char *)a->cgrp < (char *)b->cgrp ? -1 : +1;
64 
65 	if (a->evsel == NULL && b->evsel == NULL) {
66 		if (a->stat == b->stat)
67 			return 0;
68 
69 		if ((char *)a->stat < (char *)b->stat)
70 			return -1;
71 
72 		return 1;
73 	}
74 
75 	if (a->evsel == b->evsel)
76 		return 0;
77 	if ((char *)a->evsel < (char *)b->evsel)
78 		return -1;
79 	return +1;
80 }
81 
82 static struct rb_node *saved_value_new(struct rblist *rblist __maybe_unused,
83 				     const void *entry)
84 {
85 	struct saved_value *nd = malloc(sizeof(struct saved_value));
86 
87 	if (!nd)
88 		return NULL;
89 	memcpy(nd, entry, sizeof(struct saved_value));
90 	return &nd->rb_node;
91 }
92 
93 static void saved_value_delete(struct rblist *rblist __maybe_unused,
94 			       struct rb_node *rb_node)
95 {
96 	struct saved_value *v;
97 
98 	BUG_ON(!rb_node);
99 	v = container_of(rb_node, struct saved_value, rb_node);
100 	free(v);
101 }
102 
103 static struct saved_value *saved_value_lookup(struct evsel *evsel,
104 					      int cpu,
105 					      bool create,
106 					      enum stat_type type,
107 					      int ctx,
108 					      struct runtime_stat *st,
109 					      struct cgroup *cgrp)
110 {
111 	struct rblist *rblist;
112 	struct rb_node *nd;
113 	struct saved_value dm = {
114 		.cpu = cpu,
115 		.evsel = evsel,
116 		.type = type,
117 		.ctx = ctx,
118 		.stat = st,
119 		.cgrp = cgrp,
120 	};
121 
122 	rblist = &st->value_list;
123 
124 	/* don't use context info for clock events */
125 	if (type == STAT_NSECS)
126 		dm.ctx = 0;
127 
128 	nd = rblist__find(rblist, &dm);
129 	if (nd)
130 		return container_of(nd, struct saved_value, rb_node);
131 	if (create) {
132 		rblist__add_node(rblist, &dm);
133 		nd = rblist__find(rblist, &dm);
134 		if (nd)
135 			return container_of(nd, struct saved_value, rb_node);
136 	}
137 	return NULL;
138 }
139 
140 void runtime_stat__init(struct runtime_stat *st)
141 {
142 	struct rblist *rblist = &st->value_list;
143 
144 	rblist__init(rblist);
145 	rblist->node_cmp = saved_value_cmp;
146 	rblist->node_new = saved_value_new;
147 	rblist->node_delete = saved_value_delete;
148 }
149 
150 void runtime_stat__exit(struct runtime_stat *st)
151 {
152 	rblist__exit(&st->value_list);
153 }
154 
155 void perf_stat__init_shadow_stats(void)
156 {
157 	runtime_stat__init(&rt_stat);
158 }
159 
160 static int evsel_context(struct evsel *evsel)
161 {
162 	int ctx = 0;
163 
164 	if (evsel->core.attr.exclude_kernel)
165 		ctx |= CTX_BIT_KERNEL;
166 	if (evsel->core.attr.exclude_user)
167 		ctx |= CTX_BIT_USER;
168 	if (evsel->core.attr.exclude_hv)
169 		ctx |= CTX_BIT_HV;
170 	if (evsel->core.attr.exclude_host)
171 		ctx |= CTX_BIT_HOST;
172 	if (evsel->core.attr.exclude_idle)
173 		ctx |= CTX_BIT_IDLE;
174 
175 	return ctx;
176 }
177 
178 static void reset_stat(struct runtime_stat *st)
179 {
180 	struct rblist *rblist;
181 	struct rb_node *pos, *next;
182 
183 	rblist = &st->value_list;
184 	next = rb_first_cached(&rblist->entries);
185 	while (next) {
186 		pos = next;
187 		next = rb_next(pos);
188 		memset(&container_of(pos, struct saved_value, rb_node)->stats,
189 		       0,
190 		       sizeof(struct stats));
191 	}
192 }
193 
194 void perf_stat__reset_shadow_stats(void)
195 {
196 	reset_stat(&rt_stat);
197 	memset(&walltime_nsecs_stats, 0, sizeof(walltime_nsecs_stats));
198 }
199 
200 void perf_stat__reset_shadow_per_stat(struct runtime_stat *st)
201 {
202 	reset_stat(st);
203 }
204 
205 struct runtime_stat_data {
206 	int ctx;
207 	struct cgroup *cgrp;
208 };
209 
210 static void update_runtime_stat(struct runtime_stat *st,
211 				enum stat_type type,
212 				int cpu, u64 count,
213 				struct runtime_stat_data *rsd)
214 {
215 	struct saved_value *v = saved_value_lookup(NULL, cpu, true, type,
216 						   rsd->ctx, st, rsd->cgrp);
217 
218 	if (v)
219 		update_stats(&v->stats, count);
220 }
221 
222 /*
223  * Update various tracking values we maintain to print
224  * more semantic information such as miss/hit ratios,
225  * instruction rates, etc:
226  */
227 void perf_stat__update_shadow_stats(struct evsel *counter, u64 count,
228 				    int cpu, struct runtime_stat *st)
229 {
230 	u64 count_ns = count;
231 	struct saved_value *v;
232 	struct runtime_stat_data rsd = {
233 		.ctx = evsel_context(counter),
234 		.cgrp = counter->cgrp,
235 	};
236 
237 	count *= counter->scale;
238 
239 	if (evsel__is_clock(counter))
240 		update_runtime_stat(st, STAT_NSECS, cpu, count_ns, &rsd);
241 	else if (evsel__match(counter, HARDWARE, HW_CPU_CYCLES))
242 		update_runtime_stat(st, STAT_CYCLES, cpu, count, &rsd);
243 	else if (perf_stat_evsel__is(counter, CYCLES_IN_TX))
244 		update_runtime_stat(st, STAT_CYCLES_IN_TX, cpu, count, &rsd);
245 	else if (perf_stat_evsel__is(counter, TRANSACTION_START))
246 		update_runtime_stat(st, STAT_TRANSACTION, cpu, count, &rsd);
247 	else if (perf_stat_evsel__is(counter, ELISION_START))
248 		update_runtime_stat(st, STAT_ELISION, cpu, count, &rsd);
249 	else if (perf_stat_evsel__is(counter, TOPDOWN_TOTAL_SLOTS))
250 		update_runtime_stat(st, STAT_TOPDOWN_TOTAL_SLOTS,
251 				    cpu, count, &rsd);
252 	else if (perf_stat_evsel__is(counter, TOPDOWN_SLOTS_ISSUED))
253 		update_runtime_stat(st, STAT_TOPDOWN_SLOTS_ISSUED,
254 				    cpu, count, &rsd);
255 	else if (perf_stat_evsel__is(counter, TOPDOWN_SLOTS_RETIRED))
256 		update_runtime_stat(st, STAT_TOPDOWN_SLOTS_RETIRED,
257 				    cpu, count, &rsd);
258 	else if (perf_stat_evsel__is(counter, TOPDOWN_FETCH_BUBBLES))
259 		update_runtime_stat(st, STAT_TOPDOWN_FETCH_BUBBLES,
260 				    cpu, count, &rsd);
261 	else if (perf_stat_evsel__is(counter, TOPDOWN_RECOVERY_BUBBLES))
262 		update_runtime_stat(st, STAT_TOPDOWN_RECOVERY_BUBBLES,
263 				    cpu, count, &rsd);
264 	else if (perf_stat_evsel__is(counter, TOPDOWN_RETIRING))
265 		update_runtime_stat(st, STAT_TOPDOWN_RETIRING,
266 				    cpu, count, &rsd);
267 	else if (perf_stat_evsel__is(counter, TOPDOWN_BAD_SPEC))
268 		update_runtime_stat(st, STAT_TOPDOWN_BAD_SPEC,
269 				    cpu, count, &rsd);
270 	else if (perf_stat_evsel__is(counter, TOPDOWN_FE_BOUND))
271 		update_runtime_stat(st, STAT_TOPDOWN_FE_BOUND,
272 				    cpu, count, &rsd);
273 	else if (perf_stat_evsel__is(counter, TOPDOWN_BE_BOUND))
274 		update_runtime_stat(st, STAT_TOPDOWN_BE_BOUND,
275 				    cpu, count, &rsd);
276 	else if (perf_stat_evsel__is(counter, TOPDOWN_HEAVY_OPS))
277 		update_runtime_stat(st, STAT_TOPDOWN_HEAVY_OPS,
278 				    cpu, count, &rsd);
279 	else if (perf_stat_evsel__is(counter, TOPDOWN_BR_MISPREDICT))
280 		update_runtime_stat(st, STAT_TOPDOWN_BR_MISPREDICT,
281 				    cpu, count, &rsd);
282 	else if (perf_stat_evsel__is(counter, TOPDOWN_FETCH_LAT))
283 		update_runtime_stat(st, STAT_TOPDOWN_FETCH_LAT,
284 				    cpu, count, &rsd);
285 	else if (perf_stat_evsel__is(counter, TOPDOWN_MEM_BOUND))
286 		update_runtime_stat(st, STAT_TOPDOWN_MEM_BOUND,
287 				    cpu, count, &rsd);
288 	else if (evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_FRONTEND))
289 		update_runtime_stat(st, STAT_STALLED_CYCLES_FRONT,
290 				    cpu, count, &rsd);
291 	else if (evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_BACKEND))
292 		update_runtime_stat(st, STAT_STALLED_CYCLES_BACK,
293 				    cpu, count, &rsd);
294 	else if (evsel__match(counter, HARDWARE, HW_BRANCH_INSTRUCTIONS))
295 		update_runtime_stat(st, STAT_BRANCHES, cpu, count, &rsd);
296 	else if (evsel__match(counter, HARDWARE, HW_CACHE_REFERENCES))
297 		update_runtime_stat(st, STAT_CACHEREFS, cpu, count, &rsd);
298 	else if (evsel__match(counter, HW_CACHE, HW_CACHE_L1D))
299 		update_runtime_stat(st, STAT_L1_DCACHE, cpu, count, &rsd);
300 	else if (evsel__match(counter, HW_CACHE, HW_CACHE_L1I))
301 		update_runtime_stat(st, STAT_L1_ICACHE, cpu, count, &rsd);
302 	else if (evsel__match(counter, HW_CACHE, HW_CACHE_LL))
303 		update_runtime_stat(st, STAT_LL_CACHE, cpu, count, &rsd);
304 	else if (evsel__match(counter, HW_CACHE, HW_CACHE_DTLB))
305 		update_runtime_stat(st, STAT_DTLB_CACHE, cpu, count, &rsd);
306 	else if (evsel__match(counter, HW_CACHE, HW_CACHE_ITLB))
307 		update_runtime_stat(st, STAT_ITLB_CACHE, cpu, count, &rsd);
308 	else if (perf_stat_evsel__is(counter, SMI_NUM))
309 		update_runtime_stat(st, STAT_SMI_NUM, cpu, count, &rsd);
310 	else if (perf_stat_evsel__is(counter, APERF))
311 		update_runtime_stat(st, STAT_APERF, cpu, count, &rsd);
312 
313 	if (counter->collect_stat) {
314 		v = saved_value_lookup(counter, cpu, true, STAT_NONE, 0, st,
315 				       rsd.cgrp);
316 		update_stats(&v->stats, count);
317 		if (counter->metric_leader)
318 			v->metric_total += count;
319 	} else if (counter->metric_leader) {
320 		v = saved_value_lookup(counter->metric_leader,
321 				       cpu, true, STAT_NONE, 0, st, rsd.cgrp);
322 		v->metric_total += count;
323 		v->metric_other++;
324 	}
325 }
326 
327 /* used for get_ratio_color() */
328 enum grc_type {
329 	GRC_STALLED_CYCLES_FE,
330 	GRC_STALLED_CYCLES_BE,
331 	GRC_CACHE_MISSES,
332 	GRC_MAX_NR
333 };
334 
335 static const char *get_ratio_color(enum grc_type type, double ratio)
336 {
337 	static const double grc_table[GRC_MAX_NR][3] = {
338 		[GRC_STALLED_CYCLES_FE] = { 50.0, 30.0, 10.0 },
339 		[GRC_STALLED_CYCLES_BE] = { 75.0, 50.0, 20.0 },
340 		[GRC_CACHE_MISSES] 	= { 20.0, 10.0, 5.0 },
341 	};
342 	const char *color = PERF_COLOR_NORMAL;
343 
344 	if (ratio > grc_table[type][0])
345 		color = PERF_COLOR_RED;
346 	else if (ratio > grc_table[type][1])
347 		color = PERF_COLOR_MAGENTA;
348 	else if (ratio > grc_table[type][2])
349 		color = PERF_COLOR_YELLOW;
350 
351 	return color;
352 }
353 
354 static struct evsel *perf_stat__find_event(struct evlist *evsel_list,
355 						const char *name)
356 {
357 	struct evsel *c2;
358 
359 	evlist__for_each_entry (evsel_list, c2) {
360 		if (!strcasecmp(c2->name, name) && !c2->collect_stat)
361 			return c2;
362 	}
363 	return NULL;
364 }
365 
366 /* Mark MetricExpr target events and link events using them to them. */
367 void perf_stat__collect_metric_expr(struct evlist *evsel_list)
368 {
369 	struct evsel *counter, *leader, **metric_events, *oc;
370 	bool found;
371 	struct expr_parse_ctx ctx;
372 	struct hashmap_entry *cur;
373 	size_t bkt;
374 	int i;
375 
376 	expr__ctx_init(&ctx);
377 	evlist__for_each_entry(evsel_list, counter) {
378 		bool invalid = false;
379 
380 		leader = counter->leader;
381 		if (!counter->metric_expr)
382 			continue;
383 
384 		expr__ctx_clear(&ctx);
385 		metric_events = counter->metric_events;
386 		if (!metric_events) {
387 			if (expr__find_other(counter->metric_expr,
388 					     counter->name,
389 					     &ctx, 1) < 0)
390 				continue;
391 
392 			metric_events = calloc(sizeof(struct evsel *),
393 					       hashmap__size(&ctx.ids) + 1);
394 			if (!metric_events) {
395 				expr__ctx_clear(&ctx);
396 				return;
397 			}
398 			counter->metric_events = metric_events;
399 		}
400 
401 		i = 0;
402 		hashmap__for_each_entry((&ctx.ids), cur, bkt) {
403 			const char *metric_name = (const char *)cur->key;
404 
405 			found = false;
406 			if (leader) {
407 				/* Search in group */
408 				for_each_group_member (oc, leader) {
409 					if (!strcasecmp(oc->name,
410 							metric_name) &&
411 						!oc->collect_stat) {
412 						found = true;
413 						break;
414 					}
415 				}
416 			}
417 			if (!found) {
418 				/* Search ignoring groups */
419 				oc = perf_stat__find_event(evsel_list,
420 							   metric_name);
421 			}
422 			if (!oc) {
423 				/* Deduping one is good enough to handle duplicated PMUs. */
424 				static char *printed;
425 
426 				/*
427 				 * Adding events automatically would be difficult, because
428 				 * it would risk creating groups that are not schedulable.
429 				 * perf stat doesn't understand all the scheduling constraints
430 				 * of events. So we ask the user instead to add the missing
431 				 * events.
432 				 */
433 				if (!printed ||
434 				    strcasecmp(printed, metric_name)) {
435 					fprintf(stderr,
436 						"Add %s event to groups to get metric expression for %s\n",
437 						metric_name,
438 						counter->name);
439 					printed = strdup(metric_name);
440 				}
441 				invalid = true;
442 				continue;
443 			}
444 			metric_events[i++] = oc;
445 			oc->collect_stat = true;
446 		}
447 		metric_events[i] = NULL;
448 		if (invalid) {
449 			free(metric_events);
450 			counter->metric_events = NULL;
451 			counter->metric_expr = NULL;
452 		}
453 	}
454 	expr__ctx_clear(&ctx);
455 }
456 
457 static double runtime_stat_avg(struct runtime_stat *st,
458 			       enum stat_type type, int cpu,
459 			       struct runtime_stat_data *rsd)
460 {
461 	struct saved_value *v;
462 
463 	v = saved_value_lookup(NULL, cpu, false, type, rsd->ctx, st, rsd->cgrp);
464 	if (!v)
465 		return 0.0;
466 
467 	return avg_stats(&v->stats);
468 }
469 
470 static double runtime_stat_n(struct runtime_stat *st,
471 			     enum stat_type type, int cpu,
472 			     struct runtime_stat_data *rsd)
473 {
474 	struct saved_value *v;
475 
476 	v = saved_value_lookup(NULL, cpu, false, type, rsd->ctx, st, rsd->cgrp);
477 	if (!v)
478 		return 0.0;
479 
480 	return v->stats.n;
481 }
482 
483 static void print_stalled_cycles_frontend(struct perf_stat_config *config,
484 					  int cpu, double avg,
485 					  struct perf_stat_output_ctx *out,
486 					  struct runtime_stat *st,
487 					  struct runtime_stat_data *rsd)
488 {
489 	double total, ratio = 0.0;
490 	const char *color;
491 
492 	total = runtime_stat_avg(st, STAT_CYCLES, cpu, rsd);
493 
494 	if (total)
495 		ratio = avg / total * 100.0;
496 
497 	color = get_ratio_color(GRC_STALLED_CYCLES_FE, ratio);
498 
499 	if (ratio)
500 		out->print_metric(config, out->ctx, color, "%7.2f%%", "frontend cycles idle",
501 				  ratio);
502 	else
503 		out->print_metric(config, out->ctx, NULL, NULL, "frontend cycles idle", 0);
504 }
505 
506 static void print_stalled_cycles_backend(struct perf_stat_config *config,
507 					 int cpu, double avg,
508 					 struct perf_stat_output_ctx *out,
509 					 struct runtime_stat *st,
510 					 struct runtime_stat_data *rsd)
511 {
512 	double total, ratio = 0.0;
513 	const char *color;
514 
515 	total = runtime_stat_avg(st, STAT_CYCLES, cpu, rsd);
516 
517 	if (total)
518 		ratio = avg / total * 100.0;
519 
520 	color = get_ratio_color(GRC_STALLED_CYCLES_BE, ratio);
521 
522 	out->print_metric(config, out->ctx, color, "%7.2f%%", "backend cycles idle", ratio);
523 }
524 
525 static void print_branch_misses(struct perf_stat_config *config,
526 				int cpu, double avg,
527 				struct perf_stat_output_ctx *out,
528 				struct runtime_stat *st,
529 				struct runtime_stat_data *rsd)
530 {
531 	double total, ratio = 0.0;
532 	const char *color;
533 
534 	total = runtime_stat_avg(st, STAT_BRANCHES, cpu, rsd);
535 
536 	if (total)
537 		ratio = avg / total * 100.0;
538 
539 	color = get_ratio_color(GRC_CACHE_MISSES, ratio);
540 
541 	out->print_metric(config, out->ctx, color, "%7.2f%%", "of all branches", ratio);
542 }
543 
544 static void print_l1_dcache_misses(struct perf_stat_config *config,
545 				   int cpu, double avg,
546 				   struct perf_stat_output_ctx *out,
547 				   struct runtime_stat *st,
548 				   struct runtime_stat_data *rsd)
549 {
550 	double total, ratio = 0.0;
551 	const char *color;
552 
553 	total = runtime_stat_avg(st, STAT_L1_DCACHE, cpu, rsd);
554 
555 	if (total)
556 		ratio = avg / total * 100.0;
557 
558 	color = get_ratio_color(GRC_CACHE_MISSES, ratio);
559 
560 	out->print_metric(config, out->ctx, color, "%7.2f%%", "of all L1-dcache accesses", ratio);
561 }
562 
563 static void print_l1_icache_misses(struct perf_stat_config *config,
564 				   int cpu, double avg,
565 				   struct perf_stat_output_ctx *out,
566 				   struct runtime_stat *st,
567 				   struct runtime_stat_data *rsd)
568 {
569 	double total, ratio = 0.0;
570 	const char *color;
571 
572 	total = runtime_stat_avg(st, STAT_L1_ICACHE, cpu, rsd);
573 
574 	if (total)
575 		ratio = avg / total * 100.0;
576 
577 	color = get_ratio_color(GRC_CACHE_MISSES, ratio);
578 	out->print_metric(config, out->ctx, color, "%7.2f%%", "of all L1-icache accesses", ratio);
579 }
580 
581 static void print_dtlb_cache_misses(struct perf_stat_config *config,
582 				    int cpu, double avg,
583 				    struct perf_stat_output_ctx *out,
584 				    struct runtime_stat *st,
585 				    struct runtime_stat_data *rsd)
586 {
587 	double total, ratio = 0.0;
588 	const char *color;
589 
590 	total = runtime_stat_avg(st, STAT_DTLB_CACHE, cpu, rsd);
591 
592 	if (total)
593 		ratio = avg / total * 100.0;
594 
595 	color = get_ratio_color(GRC_CACHE_MISSES, ratio);
596 	out->print_metric(config, out->ctx, color, "%7.2f%%", "of all dTLB cache accesses", ratio);
597 }
598 
599 static void print_itlb_cache_misses(struct perf_stat_config *config,
600 				    int cpu, double avg,
601 				    struct perf_stat_output_ctx *out,
602 				    struct runtime_stat *st,
603 				    struct runtime_stat_data *rsd)
604 {
605 	double total, ratio = 0.0;
606 	const char *color;
607 
608 	total = runtime_stat_avg(st, STAT_ITLB_CACHE, cpu, rsd);
609 
610 	if (total)
611 		ratio = avg / total * 100.0;
612 
613 	color = get_ratio_color(GRC_CACHE_MISSES, ratio);
614 	out->print_metric(config, out->ctx, color, "%7.2f%%", "of all iTLB cache accesses", ratio);
615 }
616 
617 static void print_ll_cache_misses(struct perf_stat_config *config,
618 				  int cpu, double avg,
619 				  struct perf_stat_output_ctx *out,
620 				  struct runtime_stat *st,
621 				  struct runtime_stat_data *rsd)
622 {
623 	double total, ratio = 0.0;
624 	const char *color;
625 
626 	total = runtime_stat_avg(st, STAT_LL_CACHE, cpu, rsd);
627 
628 	if (total)
629 		ratio = avg / total * 100.0;
630 
631 	color = get_ratio_color(GRC_CACHE_MISSES, ratio);
632 	out->print_metric(config, out->ctx, color, "%7.2f%%", "of all LL-cache accesses", ratio);
633 }
634 
635 /*
636  * High level "TopDown" CPU core pipe line bottleneck break down.
637  *
638  * Basic concept following
639  * Yasin, A Top Down Method for Performance analysis and Counter architecture
640  * ISPASS14
641  *
642  * The CPU pipeline is divided into 4 areas that can be bottlenecks:
643  *
644  * Frontend -> Backend -> Retiring
645  * BadSpeculation in addition means out of order execution that is thrown away
646  * (for example branch mispredictions)
647  * Frontend is instruction decoding.
648  * Backend is execution, like computation and accessing data in memory
649  * Retiring is good execution that is not directly bottlenecked
650  *
651  * The formulas are computed in slots.
652  * A slot is an entry in the pipeline each for the pipeline width
653  * (for example a 4-wide pipeline has 4 slots for each cycle)
654  *
655  * Formulas:
656  * BadSpeculation = ((SlotsIssued - SlotsRetired) + RecoveryBubbles) /
657  *			TotalSlots
658  * Retiring = SlotsRetired / TotalSlots
659  * FrontendBound = FetchBubbles / TotalSlots
660  * BackendBound = 1.0 - BadSpeculation - Retiring - FrontendBound
661  *
662  * The kernel provides the mapping to the low level CPU events and any scaling
663  * needed for the CPU pipeline width, for example:
664  *
665  * TotalSlots = Cycles * 4
666  *
667  * The scaling factor is communicated in the sysfs unit.
668  *
669  * In some cases the CPU may not be able to measure all the formulas due to
670  * missing events. In this case multiple formulas are combined, as possible.
671  *
672  * Full TopDown supports more levels to sub-divide each area: for example
673  * BackendBound into computing bound and memory bound. For now we only
674  * support Level 1 TopDown.
675  */
676 
677 static double sanitize_val(double x)
678 {
679 	if (x < 0 && x >= -0.02)
680 		return 0.0;
681 	return x;
682 }
683 
684 static double td_total_slots(int cpu, struct runtime_stat *st,
685 			     struct runtime_stat_data *rsd)
686 {
687 	return runtime_stat_avg(st, STAT_TOPDOWN_TOTAL_SLOTS, cpu, rsd);
688 }
689 
690 static double td_bad_spec(int cpu, struct runtime_stat *st,
691 			  struct runtime_stat_data *rsd)
692 {
693 	double bad_spec = 0;
694 	double total_slots;
695 	double total;
696 
697 	total = runtime_stat_avg(st, STAT_TOPDOWN_SLOTS_ISSUED, cpu, rsd) -
698 		runtime_stat_avg(st, STAT_TOPDOWN_SLOTS_RETIRED, cpu, rsd) +
699 		runtime_stat_avg(st, STAT_TOPDOWN_RECOVERY_BUBBLES, cpu, rsd);
700 
701 	total_slots = td_total_slots(cpu, st, rsd);
702 	if (total_slots)
703 		bad_spec = total / total_slots;
704 	return sanitize_val(bad_spec);
705 }
706 
707 static double td_retiring(int cpu, struct runtime_stat *st,
708 			  struct runtime_stat_data *rsd)
709 {
710 	double retiring = 0;
711 	double total_slots = td_total_slots(cpu, st, rsd);
712 	double ret_slots = runtime_stat_avg(st, STAT_TOPDOWN_SLOTS_RETIRED,
713 					    cpu, rsd);
714 
715 	if (total_slots)
716 		retiring = ret_slots / total_slots;
717 	return retiring;
718 }
719 
720 static double td_fe_bound(int cpu, struct runtime_stat *st,
721 			  struct runtime_stat_data *rsd)
722 {
723 	double fe_bound = 0;
724 	double total_slots = td_total_slots(cpu, st, rsd);
725 	double fetch_bub = runtime_stat_avg(st, STAT_TOPDOWN_FETCH_BUBBLES,
726 					    cpu, rsd);
727 
728 	if (total_slots)
729 		fe_bound = fetch_bub / total_slots;
730 	return fe_bound;
731 }
732 
733 static double td_be_bound(int cpu, struct runtime_stat *st,
734 			  struct runtime_stat_data *rsd)
735 {
736 	double sum = (td_fe_bound(cpu, st, rsd) +
737 		      td_bad_spec(cpu, st, rsd) +
738 		      td_retiring(cpu, st, rsd));
739 	if (sum == 0)
740 		return 0;
741 	return sanitize_val(1.0 - sum);
742 }
743 
744 /*
745  * Kernel reports metrics multiplied with slots. To get back
746  * the ratios we need to recreate the sum.
747  */
748 
749 static double td_metric_ratio(int cpu, enum stat_type type,
750 			      struct runtime_stat *stat,
751 			      struct runtime_stat_data *rsd)
752 {
753 	double sum = runtime_stat_avg(stat, STAT_TOPDOWN_RETIRING, cpu, rsd) +
754 		runtime_stat_avg(stat, STAT_TOPDOWN_FE_BOUND, cpu, rsd) +
755 		runtime_stat_avg(stat, STAT_TOPDOWN_BE_BOUND, cpu, rsd) +
756 		runtime_stat_avg(stat, STAT_TOPDOWN_BAD_SPEC, cpu, rsd);
757 	double d = runtime_stat_avg(stat, type, cpu, rsd);
758 
759 	if (sum)
760 		return d / sum;
761 	return 0;
762 }
763 
764 /*
765  * ... but only if most of the values are actually available.
766  * We allow two missing.
767  */
768 
769 static bool full_td(int cpu, struct runtime_stat *stat,
770 		    struct runtime_stat_data *rsd)
771 {
772 	int c = 0;
773 
774 	if (runtime_stat_avg(stat, STAT_TOPDOWN_RETIRING, cpu, rsd) > 0)
775 		c++;
776 	if (runtime_stat_avg(stat, STAT_TOPDOWN_BE_BOUND, cpu, rsd) > 0)
777 		c++;
778 	if (runtime_stat_avg(stat, STAT_TOPDOWN_FE_BOUND, cpu, rsd) > 0)
779 		c++;
780 	if (runtime_stat_avg(stat, STAT_TOPDOWN_BAD_SPEC, cpu, rsd) > 0)
781 		c++;
782 	return c >= 2;
783 }
784 
785 static void print_smi_cost(struct perf_stat_config *config, int cpu,
786 			   struct perf_stat_output_ctx *out,
787 			   struct runtime_stat *st,
788 			   struct runtime_stat_data *rsd)
789 {
790 	double smi_num, aperf, cycles, cost = 0.0;
791 	const char *color = NULL;
792 
793 	smi_num = runtime_stat_avg(st, STAT_SMI_NUM, cpu, rsd);
794 	aperf = runtime_stat_avg(st, STAT_APERF, cpu, rsd);
795 	cycles = runtime_stat_avg(st, STAT_CYCLES, cpu, rsd);
796 
797 	if ((cycles == 0) || (aperf == 0))
798 		return;
799 
800 	if (smi_num)
801 		cost = (aperf - cycles) / aperf * 100.00;
802 
803 	if (cost > 10)
804 		color = PERF_COLOR_RED;
805 	out->print_metric(config, out->ctx, color, "%8.1f%%", "SMI cycles%", cost);
806 	out->print_metric(config, out->ctx, NULL, "%4.0f", "SMI#", smi_num);
807 }
808 
809 static int prepare_metric(struct evsel **metric_events,
810 			  struct metric_ref *metric_refs,
811 			  struct expr_parse_ctx *pctx,
812 			  int cpu,
813 			  struct runtime_stat *st)
814 {
815 	double scale;
816 	char *n, *pn;
817 	int i, j, ret;
818 
819 	expr__ctx_init(pctx);
820 	for (i = 0; metric_events[i]; i++) {
821 		struct saved_value *v;
822 		struct stats *stats;
823 		u64 metric_total = 0;
824 
825 		if (!strcmp(metric_events[i]->name, "duration_time")) {
826 			stats = &walltime_nsecs_stats;
827 			scale = 1e-9;
828 		} else {
829 			v = saved_value_lookup(metric_events[i], cpu, false,
830 					       STAT_NONE, 0, st,
831 					       metric_events[i]->cgrp);
832 			if (!v)
833 				break;
834 			stats = &v->stats;
835 			scale = 1.0;
836 
837 			if (v->metric_other)
838 				metric_total = v->metric_total;
839 		}
840 
841 		n = strdup(metric_events[i]->name);
842 		if (!n)
843 			return -ENOMEM;
844 		/*
845 		 * This display code with --no-merge adds [cpu] postfixes.
846 		 * These are not supported by the parser. Remove everything
847 		 * after the space.
848 		 */
849 		pn = strchr(n, ' ');
850 		if (pn)
851 			*pn = 0;
852 
853 		if (metric_total)
854 			expr__add_id_val(pctx, n, metric_total);
855 		else
856 			expr__add_id_val(pctx, n, avg_stats(stats)*scale);
857 	}
858 
859 	for (j = 0; metric_refs && metric_refs[j].metric_name; j++) {
860 		ret = expr__add_ref(pctx, &metric_refs[j]);
861 		if (ret)
862 			return ret;
863 	}
864 
865 	return i;
866 }
867 
868 static void generic_metric(struct perf_stat_config *config,
869 			   const char *metric_expr,
870 			   struct evsel **metric_events,
871 			   struct metric_ref *metric_refs,
872 			   char *name,
873 			   const char *metric_name,
874 			   const char *metric_unit,
875 			   int runtime,
876 			   int cpu,
877 			   struct perf_stat_output_ctx *out,
878 			   struct runtime_stat *st)
879 {
880 	print_metric_t print_metric = out->print_metric;
881 	struct expr_parse_ctx pctx;
882 	double ratio, scale;
883 	int i;
884 	void *ctxp = out->ctx;
885 
886 	i = prepare_metric(metric_events, metric_refs, &pctx, cpu, st);
887 	if (i < 0)
888 		return;
889 
890 	if (!metric_events[i]) {
891 		if (expr__parse(&ratio, &pctx, metric_expr, runtime) == 0) {
892 			char *unit;
893 			char metric_bf[64];
894 
895 			if (metric_unit && metric_name) {
896 				if (perf_pmu__convert_scale(metric_unit,
897 					&unit, &scale) >= 0) {
898 					ratio *= scale;
899 				}
900 				if (strstr(metric_expr, "?"))
901 					scnprintf(metric_bf, sizeof(metric_bf),
902 					  "%s  %s_%d", unit, metric_name, runtime);
903 				else
904 					scnprintf(metric_bf, sizeof(metric_bf),
905 					  "%s  %s", unit, metric_name);
906 
907 				print_metric(config, ctxp, NULL, "%8.1f",
908 					     metric_bf, ratio);
909 			} else {
910 				print_metric(config, ctxp, NULL, "%8.2f",
911 					metric_name ?
912 					metric_name :
913 					out->force_header ?  name : "",
914 					ratio);
915 			}
916 		} else {
917 			print_metric(config, ctxp, NULL, NULL,
918 				     out->force_header ?
919 				     (metric_name ? metric_name : name) : "", 0);
920 		}
921 	} else {
922 		print_metric(config, ctxp, NULL, NULL,
923 			     out->force_header ?
924 			     (metric_name ? metric_name : name) : "", 0);
925 	}
926 
927 	expr__ctx_clear(&pctx);
928 }
929 
930 double test_generic_metric(struct metric_expr *mexp, int cpu, struct runtime_stat *st)
931 {
932 	struct expr_parse_ctx pctx;
933 	double ratio = 0.0;
934 
935 	if (prepare_metric(mexp->metric_events, mexp->metric_refs, &pctx, cpu, st) < 0)
936 		goto out;
937 
938 	if (expr__parse(&ratio, &pctx, mexp->metric_expr, 1))
939 		ratio = 0.0;
940 
941 out:
942 	expr__ctx_clear(&pctx);
943 	return ratio;
944 }
945 
946 void perf_stat__print_shadow_stats(struct perf_stat_config *config,
947 				   struct evsel *evsel,
948 				   double avg, int cpu,
949 				   struct perf_stat_output_ctx *out,
950 				   struct rblist *metric_events,
951 				   struct runtime_stat *st)
952 {
953 	void *ctxp = out->ctx;
954 	print_metric_t print_metric = out->print_metric;
955 	double total, ratio = 0.0, total2;
956 	const char *color = NULL;
957 	struct runtime_stat_data rsd = {
958 		.ctx = evsel_context(evsel),
959 		.cgrp = evsel->cgrp,
960 	};
961 	struct metric_event *me;
962 	int num = 1;
963 
964 	if (evsel__match(evsel, HARDWARE, HW_INSTRUCTIONS)) {
965 		total = runtime_stat_avg(st, STAT_CYCLES, cpu, &rsd);
966 
967 		if (total) {
968 			ratio = avg / total;
969 			print_metric(config, ctxp, NULL, "%7.2f ",
970 					"insn per cycle", ratio);
971 		} else {
972 			print_metric(config, ctxp, NULL, NULL, "insn per cycle", 0);
973 		}
974 
975 		total = runtime_stat_avg(st, STAT_STALLED_CYCLES_FRONT, cpu, &rsd);
976 
977 		total = max(total, runtime_stat_avg(st,
978 						    STAT_STALLED_CYCLES_BACK,
979 						    cpu, &rsd));
980 
981 		if (total && avg) {
982 			out->new_line(config, ctxp);
983 			ratio = total / avg;
984 			print_metric(config, ctxp, NULL, "%7.2f ",
985 					"stalled cycles per insn",
986 					ratio);
987 		}
988 	} else if (evsel__match(evsel, HARDWARE, HW_BRANCH_MISSES)) {
989 		if (runtime_stat_n(st, STAT_BRANCHES, cpu, &rsd) != 0)
990 			print_branch_misses(config, cpu, avg, out, st, &rsd);
991 		else
992 			print_metric(config, ctxp, NULL, NULL, "of all branches", 0);
993 	} else if (
994 		evsel->core.attr.type == PERF_TYPE_HW_CACHE &&
995 		evsel->core.attr.config ==  ( PERF_COUNT_HW_CACHE_L1D |
996 					((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
997 					 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) {
998 
999 		if (runtime_stat_n(st, STAT_L1_DCACHE, cpu, &rsd) != 0)
1000 			print_l1_dcache_misses(config, cpu, avg, out, st, &rsd);
1001 		else
1002 			print_metric(config, ctxp, NULL, NULL, "of all L1-dcache accesses", 0);
1003 	} else if (
1004 		evsel->core.attr.type == PERF_TYPE_HW_CACHE &&
1005 		evsel->core.attr.config ==  ( PERF_COUNT_HW_CACHE_L1I |
1006 					((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
1007 					 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) {
1008 
1009 		if (runtime_stat_n(st, STAT_L1_ICACHE, cpu, &rsd) != 0)
1010 			print_l1_icache_misses(config, cpu, avg, out, st, &rsd);
1011 		else
1012 			print_metric(config, ctxp, NULL, NULL, "of all L1-icache accesses", 0);
1013 	} else if (
1014 		evsel->core.attr.type == PERF_TYPE_HW_CACHE &&
1015 		evsel->core.attr.config ==  ( PERF_COUNT_HW_CACHE_DTLB |
1016 					((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
1017 					 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) {
1018 
1019 		if (runtime_stat_n(st, STAT_DTLB_CACHE, cpu, &rsd) != 0)
1020 			print_dtlb_cache_misses(config, cpu, avg, out, st, &rsd);
1021 		else
1022 			print_metric(config, ctxp, NULL, NULL, "of all dTLB cache accesses", 0);
1023 	} else if (
1024 		evsel->core.attr.type == PERF_TYPE_HW_CACHE &&
1025 		evsel->core.attr.config ==  ( PERF_COUNT_HW_CACHE_ITLB |
1026 					((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
1027 					 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) {
1028 
1029 		if (runtime_stat_n(st, STAT_ITLB_CACHE, cpu, &rsd) != 0)
1030 			print_itlb_cache_misses(config, cpu, avg, out, st, &rsd);
1031 		else
1032 			print_metric(config, ctxp, NULL, NULL, "of all iTLB cache accesses", 0);
1033 	} else if (
1034 		evsel->core.attr.type == PERF_TYPE_HW_CACHE &&
1035 		evsel->core.attr.config ==  ( PERF_COUNT_HW_CACHE_LL |
1036 					((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
1037 					 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) {
1038 
1039 		if (runtime_stat_n(st, STAT_LL_CACHE, cpu, &rsd) != 0)
1040 			print_ll_cache_misses(config, cpu, avg, out, st, &rsd);
1041 		else
1042 			print_metric(config, ctxp, NULL, NULL, "of all LL-cache accesses", 0);
1043 	} else if (evsel__match(evsel, HARDWARE, HW_CACHE_MISSES)) {
1044 		total = runtime_stat_avg(st, STAT_CACHEREFS, cpu, &rsd);
1045 
1046 		if (total)
1047 			ratio = avg * 100 / total;
1048 
1049 		if (runtime_stat_n(st, STAT_CACHEREFS, cpu, &rsd) != 0)
1050 			print_metric(config, ctxp, NULL, "%8.3f %%",
1051 				     "of all cache refs", ratio);
1052 		else
1053 			print_metric(config, ctxp, NULL, NULL, "of all cache refs", 0);
1054 	} else if (evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_FRONTEND)) {
1055 		print_stalled_cycles_frontend(config, cpu, avg, out, st, &rsd);
1056 	} else if (evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_BACKEND)) {
1057 		print_stalled_cycles_backend(config, cpu, avg, out, st, &rsd);
1058 	} else if (evsel__match(evsel, HARDWARE, HW_CPU_CYCLES)) {
1059 		total = runtime_stat_avg(st, STAT_NSECS, cpu, &rsd);
1060 
1061 		if (total) {
1062 			ratio = avg / total;
1063 			print_metric(config, ctxp, NULL, "%8.3f", "GHz", ratio);
1064 		} else {
1065 			print_metric(config, ctxp, NULL, NULL, "Ghz", 0);
1066 		}
1067 	} else if (perf_stat_evsel__is(evsel, CYCLES_IN_TX)) {
1068 		total = runtime_stat_avg(st, STAT_CYCLES, cpu, &rsd);
1069 
1070 		if (total)
1071 			print_metric(config, ctxp, NULL,
1072 					"%7.2f%%", "transactional cycles",
1073 					100.0 * (avg / total));
1074 		else
1075 			print_metric(config, ctxp, NULL, NULL, "transactional cycles",
1076 				     0);
1077 	} else if (perf_stat_evsel__is(evsel, CYCLES_IN_TX_CP)) {
1078 		total = runtime_stat_avg(st, STAT_CYCLES, cpu, &rsd);
1079 		total2 = runtime_stat_avg(st, STAT_CYCLES_IN_TX, cpu, &rsd);
1080 
1081 		if (total2 < avg)
1082 			total2 = avg;
1083 		if (total)
1084 			print_metric(config, ctxp, NULL, "%7.2f%%", "aborted cycles",
1085 				100.0 * ((total2-avg) / total));
1086 		else
1087 			print_metric(config, ctxp, NULL, NULL, "aborted cycles", 0);
1088 	} else if (perf_stat_evsel__is(evsel, TRANSACTION_START)) {
1089 		total = runtime_stat_avg(st, STAT_CYCLES_IN_TX, cpu, &rsd);
1090 
1091 		if (avg)
1092 			ratio = total / avg;
1093 
1094 		if (runtime_stat_n(st, STAT_CYCLES_IN_TX, cpu, &rsd) != 0)
1095 			print_metric(config, ctxp, NULL, "%8.0f",
1096 				     "cycles / transaction", ratio);
1097 		else
1098 			print_metric(config, ctxp, NULL, NULL, "cycles / transaction",
1099 				      0);
1100 	} else if (perf_stat_evsel__is(evsel, ELISION_START)) {
1101 		total = runtime_stat_avg(st, STAT_CYCLES_IN_TX, cpu, &rsd);
1102 
1103 		if (avg)
1104 			ratio = total / avg;
1105 
1106 		print_metric(config, ctxp, NULL, "%8.0f", "cycles / elision", ratio);
1107 	} else if (evsel__is_clock(evsel)) {
1108 		if ((ratio = avg_stats(&walltime_nsecs_stats)) != 0)
1109 			print_metric(config, ctxp, NULL, "%8.3f", "CPUs utilized",
1110 				     avg / (ratio * evsel->scale));
1111 		else
1112 			print_metric(config, ctxp, NULL, NULL, "CPUs utilized", 0);
1113 	} else if (perf_stat_evsel__is(evsel, TOPDOWN_FETCH_BUBBLES)) {
1114 		double fe_bound = td_fe_bound(cpu, st, &rsd);
1115 
1116 		if (fe_bound > 0.2)
1117 			color = PERF_COLOR_RED;
1118 		print_metric(config, ctxp, color, "%8.1f%%", "frontend bound",
1119 				fe_bound * 100.);
1120 	} else if (perf_stat_evsel__is(evsel, TOPDOWN_SLOTS_RETIRED)) {
1121 		double retiring = td_retiring(cpu, st, &rsd);
1122 
1123 		if (retiring > 0.7)
1124 			color = PERF_COLOR_GREEN;
1125 		print_metric(config, ctxp, color, "%8.1f%%", "retiring",
1126 				retiring * 100.);
1127 	} else if (perf_stat_evsel__is(evsel, TOPDOWN_RECOVERY_BUBBLES)) {
1128 		double bad_spec = td_bad_spec(cpu, st, &rsd);
1129 
1130 		if (bad_spec > 0.1)
1131 			color = PERF_COLOR_RED;
1132 		print_metric(config, ctxp, color, "%8.1f%%", "bad speculation",
1133 				bad_spec * 100.);
1134 	} else if (perf_stat_evsel__is(evsel, TOPDOWN_SLOTS_ISSUED)) {
1135 		double be_bound = td_be_bound(cpu, st, &rsd);
1136 		const char *name = "backend bound";
1137 		static int have_recovery_bubbles = -1;
1138 
1139 		/* In case the CPU does not support topdown-recovery-bubbles */
1140 		if (have_recovery_bubbles < 0)
1141 			have_recovery_bubbles = pmu_have_event("cpu",
1142 					"topdown-recovery-bubbles");
1143 		if (!have_recovery_bubbles)
1144 			name = "backend bound/bad spec";
1145 
1146 		if (be_bound > 0.2)
1147 			color = PERF_COLOR_RED;
1148 		if (td_total_slots(cpu, st, &rsd) > 0)
1149 			print_metric(config, ctxp, color, "%8.1f%%", name,
1150 					be_bound * 100.);
1151 		else
1152 			print_metric(config, ctxp, NULL, NULL, name, 0);
1153 	} else if (perf_stat_evsel__is(evsel, TOPDOWN_RETIRING) &&
1154 		   full_td(cpu, st, &rsd)) {
1155 		double retiring = td_metric_ratio(cpu,
1156 						  STAT_TOPDOWN_RETIRING, st,
1157 						  &rsd);
1158 		if (retiring > 0.7)
1159 			color = PERF_COLOR_GREEN;
1160 		print_metric(config, ctxp, color, "%8.1f%%", "retiring",
1161 				retiring * 100.);
1162 	} else if (perf_stat_evsel__is(evsel, TOPDOWN_FE_BOUND) &&
1163 		   full_td(cpu, st, &rsd)) {
1164 		double fe_bound = td_metric_ratio(cpu,
1165 						  STAT_TOPDOWN_FE_BOUND, st,
1166 						  &rsd);
1167 		if (fe_bound > 0.2)
1168 			color = PERF_COLOR_RED;
1169 		print_metric(config, ctxp, color, "%8.1f%%", "frontend bound",
1170 				fe_bound * 100.);
1171 	} else if (perf_stat_evsel__is(evsel, TOPDOWN_BE_BOUND) &&
1172 		   full_td(cpu, st, &rsd)) {
1173 		double be_bound = td_metric_ratio(cpu,
1174 						  STAT_TOPDOWN_BE_BOUND, st,
1175 						  &rsd);
1176 		if (be_bound > 0.2)
1177 			color = PERF_COLOR_RED;
1178 		print_metric(config, ctxp, color, "%8.1f%%", "backend bound",
1179 				be_bound * 100.);
1180 	} else if (perf_stat_evsel__is(evsel, TOPDOWN_BAD_SPEC) &&
1181 		   full_td(cpu, st, &rsd)) {
1182 		double bad_spec = td_metric_ratio(cpu,
1183 						  STAT_TOPDOWN_BAD_SPEC, st,
1184 						  &rsd);
1185 		if (bad_spec > 0.1)
1186 			color = PERF_COLOR_RED;
1187 		print_metric(config, ctxp, color, "%8.1f%%", "bad speculation",
1188 				bad_spec * 100.);
1189 	} else if (perf_stat_evsel__is(evsel, TOPDOWN_HEAVY_OPS) &&
1190 			full_td(cpu, st, &rsd) && (config->topdown_level > 1)) {
1191 		double retiring = td_metric_ratio(cpu,
1192 						  STAT_TOPDOWN_RETIRING, st,
1193 						  &rsd);
1194 		double heavy_ops = td_metric_ratio(cpu,
1195 						   STAT_TOPDOWN_HEAVY_OPS, st,
1196 						   &rsd);
1197 		double light_ops = retiring - heavy_ops;
1198 
1199 		if (retiring > 0.7 && heavy_ops > 0.1)
1200 			color = PERF_COLOR_GREEN;
1201 		print_metric(config, ctxp, color, "%8.1f%%", "heavy operations",
1202 				heavy_ops * 100.);
1203 		if (retiring > 0.7 && light_ops > 0.6)
1204 			color = PERF_COLOR_GREEN;
1205 		else
1206 			color = NULL;
1207 		print_metric(config, ctxp, color, "%8.1f%%", "light operations",
1208 				light_ops * 100.);
1209 	} else if (perf_stat_evsel__is(evsel, TOPDOWN_BR_MISPREDICT) &&
1210 			full_td(cpu, st, &rsd) && (config->topdown_level > 1)) {
1211 		double bad_spec = td_metric_ratio(cpu,
1212 						  STAT_TOPDOWN_BAD_SPEC, st,
1213 						  &rsd);
1214 		double br_mis = td_metric_ratio(cpu,
1215 						STAT_TOPDOWN_BR_MISPREDICT, st,
1216 						&rsd);
1217 		double m_clears = bad_spec - br_mis;
1218 
1219 		if (bad_spec > 0.1 && br_mis > 0.05)
1220 			color = PERF_COLOR_RED;
1221 		print_metric(config, ctxp, color, "%8.1f%%", "branch mispredict",
1222 				br_mis * 100.);
1223 		if (bad_spec > 0.1 && m_clears > 0.05)
1224 			color = PERF_COLOR_RED;
1225 		else
1226 			color = NULL;
1227 		print_metric(config, ctxp, color, "%8.1f%%", "machine clears",
1228 				m_clears * 100.);
1229 	} else if (perf_stat_evsel__is(evsel, TOPDOWN_FETCH_LAT) &&
1230 			full_td(cpu, st, &rsd) && (config->topdown_level > 1)) {
1231 		double fe_bound = td_metric_ratio(cpu,
1232 						  STAT_TOPDOWN_FE_BOUND, st,
1233 						  &rsd);
1234 		double fetch_lat = td_metric_ratio(cpu,
1235 						   STAT_TOPDOWN_FETCH_LAT, st,
1236 						   &rsd);
1237 		double fetch_bw = fe_bound - fetch_lat;
1238 
1239 		if (fe_bound > 0.2 && fetch_lat > 0.15)
1240 			color = PERF_COLOR_RED;
1241 		print_metric(config, ctxp, color, "%8.1f%%", "fetch latency",
1242 				fetch_lat * 100.);
1243 		if (fe_bound > 0.2 && fetch_bw > 0.1)
1244 			color = PERF_COLOR_RED;
1245 		else
1246 			color = NULL;
1247 		print_metric(config, ctxp, color, "%8.1f%%", "fetch bandwidth",
1248 				fetch_bw * 100.);
1249 	} else if (perf_stat_evsel__is(evsel, TOPDOWN_MEM_BOUND) &&
1250 			full_td(cpu, st, &rsd) && (config->topdown_level > 1)) {
1251 		double be_bound = td_metric_ratio(cpu,
1252 						  STAT_TOPDOWN_BE_BOUND, st,
1253 						  &rsd);
1254 		double mem_bound = td_metric_ratio(cpu,
1255 						   STAT_TOPDOWN_MEM_BOUND, st,
1256 						   &rsd);
1257 		double core_bound = be_bound - mem_bound;
1258 
1259 		if (be_bound > 0.2 && mem_bound > 0.2)
1260 			color = PERF_COLOR_RED;
1261 		print_metric(config, ctxp, color, "%8.1f%%", "memory bound",
1262 				mem_bound * 100.);
1263 		if (be_bound > 0.2 && core_bound > 0.1)
1264 			color = PERF_COLOR_RED;
1265 		else
1266 			color = NULL;
1267 		print_metric(config, ctxp, color, "%8.1f%%", "Core bound",
1268 				core_bound * 100.);
1269 	} else if (evsel->metric_expr) {
1270 		generic_metric(config, evsel->metric_expr, evsel->metric_events, NULL,
1271 				evsel->name, evsel->metric_name, NULL, 1, cpu, out, st);
1272 	} else if (runtime_stat_n(st, STAT_NSECS, cpu, &rsd) != 0) {
1273 		char unit = 'M';
1274 		char unit_buf[10];
1275 
1276 		total = runtime_stat_avg(st, STAT_NSECS, cpu, &rsd);
1277 
1278 		if (total)
1279 			ratio = 1000.0 * avg / total;
1280 		if (ratio < 0.001) {
1281 			ratio *= 1000;
1282 			unit = 'K';
1283 		}
1284 		snprintf(unit_buf, sizeof(unit_buf), "%c/sec", unit);
1285 		print_metric(config, ctxp, NULL, "%8.3f", unit_buf, ratio);
1286 	} else if (perf_stat_evsel__is(evsel, SMI_NUM)) {
1287 		print_smi_cost(config, cpu, out, st, &rsd);
1288 	} else {
1289 		num = 0;
1290 	}
1291 
1292 	if ((me = metricgroup__lookup(metric_events, evsel, false)) != NULL) {
1293 		struct metric_expr *mexp;
1294 
1295 		list_for_each_entry (mexp, &me->head, nd) {
1296 			if (num++ > 0)
1297 				out->new_line(config, ctxp);
1298 			generic_metric(config, mexp->metric_expr, mexp->metric_events,
1299 					mexp->metric_refs, evsel->name, mexp->metric_name,
1300 					mexp->metric_unit, mexp->runtime, cpu, out, st);
1301 		}
1302 	}
1303 	if (num == 0)
1304 		print_metric(config, ctxp, NULL, NULL, NULL, 0);
1305 }
1306