xref: /linux/tools/perf/util/stat-shadow.c (revision 37932c188ef1b471eae29249df045c8e567772d0)
1 #include <stdio.h>
2 #include "evsel.h"
3 #include "stat.h"
4 #include "color.h"
5 #include "pmu.h"
6 #include "rblist.h"
7 #include "evlist.h"
8 #include "expr.h"
9 
10 enum {
11 	CTX_BIT_USER	= 1 << 0,
12 	CTX_BIT_KERNEL	= 1 << 1,
13 	CTX_BIT_HV	= 1 << 2,
14 	CTX_BIT_HOST	= 1 << 3,
15 	CTX_BIT_IDLE	= 1 << 4,
16 	CTX_BIT_MAX	= 1 << 5,
17 };
18 
19 #define NUM_CTX CTX_BIT_MAX
20 
21 /*
22  * AGGR_GLOBAL: Use CPU 0
23  * AGGR_SOCKET: Use first CPU of socket
24  * AGGR_CORE: Use first CPU of core
25  * AGGR_NONE: Use matching CPU
26  * AGGR_THREAD: Not supported?
27  */
28 static struct stats runtime_nsecs_stats[MAX_NR_CPUS];
29 static struct stats runtime_cycles_stats[NUM_CTX][MAX_NR_CPUS];
30 static struct stats runtime_stalled_cycles_front_stats[NUM_CTX][MAX_NR_CPUS];
31 static struct stats runtime_stalled_cycles_back_stats[NUM_CTX][MAX_NR_CPUS];
32 static struct stats runtime_branches_stats[NUM_CTX][MAX_NR_CPUS];
33 static struct stats runtime_cacherefs_stats[NUM_CTX][MAX_NR_CPUS];
34 static struct stats runtime_l1_dcache_stats[NUM_CTX][MAX_NR_CPUS];
35 static struct stats runtime_l1_icache_stats[NUM_CTX][MAX_NR_CPUS];
36 static struct stats runtime_ll_cache_stats[NUM_CTX][MAX_NR_CPUS];
37 static struct stats runtime_itlb_cache_stats[NUM_CTX][MAX_NR_CPUS];
38 static struct stats runtime_dtlb_cache_stats[NUM_CTX][MAX_NR_CPUS];
39 static struct stats runtime_cycles_in_tx_stats[NUM_CTX][MAX_NR_CPUS];
40 static struct stats runtime_transaction_stats[NUM_CTX][MAX_NR_CPUS];
41 static struct stats runtime_elision_stats[NUM_CTX][MAX_NR_CPUS];
42 static struct stats runtime_topdown_total_slots[NUM_CTX][MAX_NR_CPUS];
43 static struct stats runtime_topdown_slots_issued[NUM_CTX][MAX_NR_CPUS];
44 static struct stats runtime_topdown_slots_retired[NUM_CTX][MAX_NR_CPUS];
45 static struct stats runtime_topdown_fetch_bubbles[NUM_CTX][MAX_NR_CPUS];
46 static struct stats runtime_topdown_recovery_bubbles[NUM_CTX][MAX_NR_CPUS];
47 static struct rblist runtime_saved_values;
48 static bool have_frontend_stalled;
49 
50 struct stats walltime_nsecs_stats;
51 
52 struct saved_value {
53 	struct rb_node rb_node;
54 	struct perf_evsel *evsel;
55 	int cpu;
56 	int ctx;
57 	struct stats stats;
58 };
59 
60 static int saved_value_cmp(struct rb_node *rb_node, const void *entry)
61 {
62 	struct saved_value *a = container_of(rb_node,
63 					     struct saved_value,
64 					     rb_node);
65 	const struct saved_value *b = entry;
66 
67 	if (a->ctx != b->ctx)
68 		return a->ctx - b->ctx;
69 	if (a->cpu != b->cpu)
70 		return a->cpu - b->cpu;
71 	return a->evsel - b->evsel;
72 }
73 
74 static struct rb_node *saved_value_new(struct rblist *rblist __maybe_unused,
75 				     const void *entry)
76 {
77 	struct saved_value *nd = malloc(sizeof(struct saved_value));
78 
79 	if (!nd)
80 		return NULL;
81 	memcpy(nd, entry, sizeof(struct saved_value));
82 	return &nd->rb_node;
83 }
84 
85 static struct saved_value *saved_value_lookup(struct perf_evsel *evsel,
86 					      int cpu, int ctx,
87 					      bool create)
88 {
89 	struct rb_node *nd;
90 	struct saved_value dm = {
91 		.cpu = cpu,
92 		.ctx = ctx,
93 		.evsel = evsel,
94 	};
95 	nd = rblist__find(&runtime_saved_values, &dm);
96 	if (nd)
97 		return container_of(nd, struct saved_value, rb_node);
98 	if (create) {
99 		rblist__add_node(&runtime_saved_values, &dm);
100 		nd = rblist__find(&runtime_saved_values, &dm);
101 		if (nd)
102 			return container_of(nd, struct saved_value, rb_node);
103 	}
104 	return NULL;
105 }
106 
107 void perf_stat__init_shadow_stats(void)
108 {
109 	have_frontend_stalled = pmu_have_event("cpu", "stalled-cycles-frontend");
110 	rblist__init(&runtime_saved_values);
111 	runtime_saved_values.node_cmp = saved_value_cmp;
112 	runtime_saved_values.node_new = saved_value_new;
113 	/* No delete for now */
114 }
115 
116 static int evsel_context(struct perf_evsel *evsel)
117 {
118 	int ctx = 0;
119 
120 	if (evsel->attr.exclude_kernel)
121 		ctx |= CTX_BIT_KERNEL;
122 	if (evsel->attr.exclude_user)
123 		ctx |= CTX_BIT_USER;
124 	if (evsel->attr.exclude_hv)
125 		ctx |= CTX_BIT_HV;
126 	if (evsel->attr.exclude_host)
127 		ctx |= CTX_BIT_HOST;
128 	if (evsel->attr.exclude_idle)
129 		ctx |= CTX_BIT_IDLE;
130 
131 	return ctx;
132 }
133 
134 void perf_stat__reset_shadow_stats(void)
135 {
136 	struct rb_node *pos, *next;
137 
138 	memset(runtime_nsecs_stats, 0, sizeof(runtime_nsecs_stats));
139 	memset(runtime_cycles_stats, 0, sizeof(runtime_cycles_stats));
140 	memset(runtime_stalled_cycles_front_stats, 0, sizeof(runtime_stalled_cycles_front_stats));
141 	memset(runtime_stalled_cycles_back_stats, 0, sizeof(runtime_stalled_cycles_back_stats));
142 	memset(runtime_branches_stats, 0, sizeof(runtime_branches_stats));
143 	memset(runtime_cacherefs_stats, 0, sizeof(runtime_cacherefs_stats));
144 	memset(runtime_l1_dcache_stats, 0, sizeof(runtime_l1_dcache_stats));
145 	memset(runtime_l1_icache_stats, 0, sizeof(runtime_l1_icache_stats));
146 	memset(runtime_ll_cache_stats, 0, sizeof(runtime_ll_cache_stats));
147 	memset(runtime_itlb_cache_stats, 0, sizeof(runtime_itlb_cache_stats));
148 	memset(runtime_dtlb_cache_stats, 0, sizeof(runtime_dtlb_cache_stats));
149 	memset(runtime_cycles_in_tx_stats, 0,
150 			sizeof(runtime_cycles_in_tx_stats));
151 	memset(runtime_transaction_stats, 0,
152 		sizeof(runtime_transaction_stats));
153 	memset(runtime_elision_stats, 0, sizeof(runtime_elision_stats));
154 	memset(&walltime_nsecs_stats, 0, sizeof(walltime_nsecs_stats));
155 	memset(runtime_topdown_total_slots, 0, sizeof(runtime_topdown_total_slots));
156 	memset(runtime_topdown_slots_retired, 0, sizeof(runtime_topdown_slots_retired));
157 	memset(runtime_topdown_slots_issued, 0, sizeof(runtime_topdown_slots_issued));
158 	memset(runtime_topdown_fetch_bubbles, 0, sizeof(runtime_topdown_fetch_bubbles));
159 	memset(runtime_topdown_recovery_bubbles, 0, sizeof(runtime_topdown_recovery_bubbles));
160 
161 	next = rb_first(&runtime_saved_values.entries);
162 	while (next) {
163 		pos = next;
164 		next = rb_next(pos);
165 		memset(&container_of(pos, struct saved_value, rb_node)->stats,
166 		       0,
167 		       sizeof(struct stats));
168 	}
169 }
170 
171 /*
172  * Update various tracking values we maintain to print
173  * more semantic information such as miss/hit ratios,
174  * instruction rates, etc:
175  */
176 void perf_stat__update_shadow_stats(struct perf_evsel *counter, u64 *count,
177 				    int cpu)
178 {
179 	int ctx = evsel_context(counter);
180 
181 	if (perf_evsel__match(counter, SOFTWARE, SW_TASK_CLOCK) ||
182 	    perf_evsel__match(counter, SOFTWARE, SW_CPU_CLOCK))
183 		update_stats(&runtime_nsecs_stats[cpu], count[0]);
184 	else if (perf_evsel__match(counter, HARDWARE, HW_CPU_CYCLES))
185 		update_stats(&runtime_cycles_stats[ctx][cpu], count[0]);
186 	else if (perf_stat_evsel__is(counter, CYCLES_IN_TX))
187 		update_stats(&runtime_cycles_in_tx_stats[ctx][cpu], count[0]);
188 	else if (perf_stat_evsel__is(counter, TRANSACTION_START))
189 		update_stats(&runtime_transaction_stats[ctx][cpu], count[0]);
190 	else if (perf_stat_evsel__is(counter, ELISION_START))
191 		update_stats(&runtime_elision_stats[ctx][cpu], count[0]);
192 	else if (perf_stat_evsel__is(counter, TOPDOWN_TOTAL_SLOTS))
193 		update_stats(&runtime_topdown_total_slots[ctx][cpu], count[0]);
194 	else if (perf_stat_evsel__is(counter, TOPDOWN_SLOTS_ISSUED))
195 		update_stats(&runtime_topdown_slots_issued[ctx][cpu], count[0]);
196 	else if (perf_stat_evsel__is(counter, TOPDOWN_SLOTS_RETIRED))
197 		update_stats(&runtime_topdown_slots_retired[ctx][cpu], count[0]);
198 	else if (perf_stat_evsel__is(counter, TOPDOWN_FETCH_BUBBLES))
199 		update_stats(&runtime_topdown_fetch_bubbles[ctx][cpu],count[0]);
200 	else if (perf_stat_evsel__is(counter, TOPDOWN_RECOVERY_BUBBLES))
201 		update_stats(&runtime_topdown_recovery_bubbles[ctx][cpu], count[0]);
202 	else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_FRONTEND))
203 		update_stats(&runtime_stalled_cycles_front_stats[ctx][cpu], count[0]);
204 	else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_BACKEND))
205 		update_stats(&runtime_stalled_cycles_back_stats[ctx][cpu], count[0]);
206 	else if (perf_evsel__match(counter, HARDWARE, HW_BRANCH_INSTRUCTIONS))
207 		update_stats(&runtime_branches_stats[ctx][cpu], count[0]);
208 	else if (perf_evsel__match(counter, HARDWARE, HW_CACHE_REFERENCES))
209 		update_stats(&runtime_cacherefs_stats[ctx][cpu], count[0]);
210 	else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_L1D))
211 		update_stats(&runtime_l1_dcache_stats[ctx][cpu], count[0]);
212 	else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_L1I))
213 		update_stats(&runtime_ll_cache_stats[ctx][cpu], count[0]);
214 	else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_LL))
215 		update_stats(&runtime_ll_cache_stats[ctx][cpu], count[0]);
216 	else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_DTLB))
217 		update_stats(&runtime_dtlb_cache_stats[ctx][cpu], count[0]);
218 	else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_ITLB))
219 		update_stats(&runtime_itlb_cache_stats[ctx][cpu], count[0]);
220 
221 	if (counter->collect_stat) {
222 		struct saved_value *v = saved_value_lookup(counter, cpu, ctx,
223 							   true);
224 		update_stats(&v->stats, count[0]);
225 	}
226 }
227 
228 /* used for get_ratio_color() */
229 enum grc_type {
230 	GRC_STALLED_CYCLES_FE,
231 	GRC_STALLED_CYCLES_BE,
232 	GRC_CACHE_MISSES,
233 	GRC_MAX_NR
234 };
235 
236 static const char *get_ratio_color(enum grc_type type, double ratio)
237 {
238 	static const double grc_table[GRC_MAX_NR][3] = {
239 		[GRC_STALLED_CYCLES_FE] = { 50.0, 30.0, 10.0 },
240 		[GRC_STALLED_CYCLES_BE] = { 75.0, 50.0, 20.0 },
241 		[GRC_CACHE_MISSES] 	= { 20.0, 10.0, 5.0 },
242 	};
243 	const char *color = PERF_COLOR_NORMAL;
244 
245 	if (ratio > grc_table[type][0])
246 		color = PERF_COLOR_RED;
247 	else if (ratio > grc_table[type][1])
248 		color = PERF_COLOR_MAGENTA;
249 	else if (ratio > grc_table[type][2])
250 		color = PERF_COLOR_YELLOW;
251 
252 	return color;
253 }
254 
255 static struct perf_evsel *perf_stat__find_event(struct perf_evlist *evsel_list,
256 						const char *name)
257 {
258 	struct perf_evsel *c2;
259 
260 	evlist__for_each_entry (evsel_list, c2) {
261 		if (!strcasecmp(c2->name, name))
262 			return c2;
263 	}
264 	return NULL;
265 }
266 
267 /* Mark MetricExpr target events and link events using them to them. */
268 void perf_stat__collect_metric_expr(struct perf_evlist *evsel_list)
269 {
270 	struct perf_evsel *counter, *leader, **metric_events, *oc;
271 	bool found;
272 	const char **metric_names;
273 	int i;
274 	int num_metric_names;
275 
276 	evlist__for_each_entry(evsel_list, counter) {
277 		bool invalid = false;
278 
279 		leader = counter->leader;
280 		if (!counter->metric_expr)
281 			continue;
282 		metric_events = counter->metric_events;
283 		if (!metric_events) {
284 			if (expr__find_other(counter->metric_expr, counter->name,
285 						&metric_names, &num_metric_names) < 0)
286 				continue;
287 
288 			metric_events = calloc(sizeof(struct perf_evsel *),
289 					       num_metric_names + 1);
290 			if (!metric_events)
291 				return;
292 			counter->metric_events = metric_events;
293 		}
294 
295 		for (i = 0; i < num_metric_names; i++) {
296 			found = false;
297 			if (leader) {
298 				/* Search in group */
299 				for_each_group_member (oc, leader) {
300 					if (!strcasecmp(oc->name, metric_names[i])) {
301 						found = true;
302 						break;
303 					}
304 				}
305 			}
306 			if (!found) {
307 				/* Search ignoring groups */
308 				oc = perf_stat__find_event(evsel_list, metric_names[i]);
309 			}
310 			if (!oc) {
311 				/* Deduping one is good enough to handle duplicated PMUs. */
312 				static char *printed;
313 
314 				/*
315 				 * Adding events automatically would be difficult, because
316 				 * it would risk creating groups that are not schedulable.
317 				 * perf stat doesn't understand all the scheduling constraints
318 				 * of events. So we ask the user instead to add the missing
319 				 * events.
320 				 */
321 				if (!printed || strcasecmp(printed, metric_names[i])) {
322 					fprintf(stderr,
323 						"Add %s event to groups to get metric expression for %s\n",
324 						metric_names[i],
325 						counter->name);
326 					printed = strdup(metric_names[i]);
327 				}
328 				invalid = true;
329 				continue;
330 			}
331 			metric_events[i] = oc;
332 			oc->collect_stat = true;
333 		}
334 		metric_events[i] = NULL;
335 		free(metric_names);
336 		if (invalid) {
337 			free(metric_events);
338 			counter->metric_events = NULL;
339 			counter->metric_expr = NULL;
340 		}
341 	}
342 }
343 
344 static void print_stalled_cycles_frontend(int cpu,
345 					  struct perf_evsel *evsel, double avg,
346 					  struct perf_stat_output_ctx *out)
347 {
348 	double total, ratio = 0.0;
349 	const char *color;
350 	int ctx = evsel_context(evsel);
351 
352 	total = avg_stats(&runtime_cycles_stats[ctx][cpu]);
353 
354 	if (total)
355 		ratio = avg / total * 100.0;
356 
357 	color = get_ratio_color(GRC_STALLED_CYCLES_FE, ratio);
358 
359 	if (ratio)
360 		out->print_metric(out->ctx, color, "%7.2f%%", "frontend cycles idle",
361 				  ratio);
362 	else
363 		out->print_metric(out->ctx, NULL, NULL, "frontend cycles idle", 0);
364 }
365 
366 static void print_stalled_cycles_backend(int cpu,
367 					 struct perf_evsel *evsel, double avg,
368 					 struct perf_stat_output_ctx *out)
369 {
370 	double total, ratio = 0.0;
371 	const char *color;
372 	int ctx = evsel_context(evsel);
373 
374 	total = avg_stats(&runtime_cycles_stats[ctx][cpu]);
375 
376 	if (total)
377 		ratio = avg / total * 100.0;
378 
379 	color = get_ratio_color(GRC_STALLED_CYCLES_BE, ratio);
380 
381 	out->print_metric(out->ctx, color, "%7.2f%%", "backend cycles idle", ratio);
382 }
383 
384 static void print_branch_misses(int cpu,
385 				struct perf_evsel *evsel,
386 				double avg,
387 				struct perf_stat_output_ctx *out)
388 {
389 	double total, ratio = 0.0;
390 	const char *color;
391 	int ctx = evsel_context(evsel);
392 
393 	total = avg_stats(&runtime_branches_stats[ctx][cpu]);
394 
395 	if (total)
396 		ratio = avg / total * 100.0;
397 
398 	color = get_ratio_color(GRC_CACHE_MISSES, ratio);
399 
400 	out->print_metric(out->ctx, color, "%7.2f%%", "of all branches", ratio);
401 }
402 
403 static void print_l1_dcache_misses(int cpu,
404 				   struct perf_evsel *evsel,
405 				   double avg,
406 				   struct perf_stat_output_ctx *out)
407 {
408 	double total, ratio = 0.0;
409 	const char *color;
410 	int ctx = evsel_context(evsel);
411 
412 	total = avg_stats(&runtime_l1_dcache_stats[ctx][cpu]);
413 
414 	if (total)
415 		ratio = avg / total * 100.0;
416 
417 	color = get_ratio_color(GRC_CACHE_MISSES, ratio);
418 
419 	out->print_metric(out->ctx, color, "%7.2f%%", "of all L1-dcache hits", ratio);
420 }
421 
422 static void print_l1_icache_misses(int cpu,
423 				   struct perf_evsel *evsel,
424 				   double avg,
425 				   struct perf_stat_output_ctx *out)
426 {
427 	double total, ratio = 0.0;
428 	const char *color;
429 	int ctx = evsel_context(evsel);
430 
431 	total = avg_stats(&runtime_l1_icache_stats[ctx][cpu]);
432 
433 	if (total)
434 		ratio = avg / total * 100.0;
435 
436 	color = get_ratio_color(GRC_CACHE_MISSES, ratio);
437 	out->print_metric(out->ctx, color, "%7.2f%%", "of all L1-icache hits", ratio);
438 }
439 
440 static void print_dtlb_cache_misses(int cpu,
441 				    struct perf_evsel *evsel,
442 				    double avg,
443 				    struct perf_stat_output_ctx *out)
444 {
445 	double total, ratio = 0.0;
446 	const char *color;
447 	int ctx = evsel_context(evsel);
448 
449 	total = avg_stats(&runtime_dtlb_cache_stats[ctx][cpu]);
450 
451 	if (total)
452 		ratio = avg / total * 100.0;
453 
454 	color = get_ratio_color(GRC_CACHE_MISSES, ratio);
455 	out->print_metric(out->ctx, color, "%7.2f%%", "of all dTLB cache hits", ratio);
456 }
457 
458 static void print_itlb_cache_misses(int cpu,
459 				    struct perf_evsel *evsel,
460 				    double avg,
461 				    struct perf_stat_output_ctx *out)
462 {
463 	double total, ratio = 0.0;
464 	const char *color;
465 	int ctx = evsel_context(evsel);
466 
467 	total = avg_stats(&runtime_itlb_cache_stats[ctx][cpu]);
468 
469 	if (total)
470 		ratio = avg / total * 100.0;
471 
472 	color = get_ratio_color(GRC_CACHE_MISSES, ratio);
473 	out->print_metric(out->ctx, color, "%7.2f%%", "of all iTLB cache hits", ratio);
474 }
475 
476 static void print_ll_cache_misses(int cpu,
477 				  struct perf_evsel *evsel,
478 				  double avg,
479 				  struct perf_stat_output_ctx *out)
480 {
481 	double total, ratio = 0.0;
482 	const char *color;
483 	int ctx = evsel_context(evsel);
484 
485 	total = avg_stats(&runtime_ll_cache_stats[ctx][cpu]);
486 
487 	if (total)
488 		ratio = avg / total * 100.0;
489 
490 	color = get_ratio_color(GRC_CACHE_MISSES, ratio);
491 	out->print_metric(out->ctx, color, "%7.2f%%", "of all LL-cache hits", ratio);
492 }
493 
494 /*
495  * High level "TopDown" CPU core pipe line bottleneck break down.
496  *
497  * Basic concept following
498  * Yasin, A Top Down Method for Performance analysis and Counter architecture
499  * ISPASS14
500  *
501  * The CPU pipeline is divided into 4 areas that can be bottlenecks:
502  *
503  * Frontend -> Backend -> Retiring
504  * BadSpeculation in addition means out of order execution that is thrown away
505  * (for example branch mispredictions)
506  * Frontend is instruction decoding.
507  * Backend is execution, like computation and accessing data in memory
508  * Retiring is good execution that is not directly bottlenecked
509  *
510  * The formulas are computed in slots.
511  * A slot is an entry in the pipeline each for the pipeline width
512  * (for example a 4-wide pipeline has 4 slots for each cycle)
513  *
514  * Formulas:
515  * BadSpeculation = ((SlotsIssued - SlotsRetired) + RecoveryBubbles) /
516  *			TotalSlots
517  * Retiring = SlotsRetired / TotalSlots
518  * FrontendBound = FetchBubbles / TotalSlots
519  * BackendBound = 1.0 - BadSpeculation - Retiring - FrontendBound
520  *
521  * The kernel provides the mapping to the low level CPU events and any scaling
522  * needed for the CPU pipeline width, for example:
523  *
524  * TotalSlots = Cycles * 4
525  *
526  * The scaling factor is communicated in the sysfs unit.
527  *
528  * In some cases the CPU may not be able to measure all the formulas due to
529  * missing events. In this case multiple formulas are combined, as possible.
530  *
531  * Full TopDown supports more levels to sub-divide each area: for example
532  * BackendBound into computing bound and memory bound. For now we only
533  * support Level 1 TopDown.
534  */
535 
536 static double sanitize_val(double x)
537 {
538 	if (x < 0 && x >= -0.02)
539 		return 0.0;
540 	return x;
541 }
542 
543 static double td_total_slots(int ctx, int cpu)
544 {
545 	return avg_stats(&runtime_topdown_total_slots[ctx][cpu]);
546 }
547 
548 static double td_bad_spec(int ctx, int cpu)
549 {
550 	double bad_spec = 0;
551 	double total_slots;
552 	double total;
553 
554 	total = avg_stats(&runtime_topdown_slots_issued[ctx][cpu]) -
555 		avg_stats(&runtime_topdown_slots_retired[ctx][cpu]) +
556 		avg_stats(&runtime_topdown_recovery_bubbles[ctx][cpu]);
557 	total_slots = td_total_slots(ctx, cpu);
558 	if (total_slots)
559 		bad_spec = total / total_slots;
560 	return sanitize_val(bad_spec);
561 }
562 
563 static double td_retiring(int ctx, int cpu)
564 {
565 	double retiring = 0;
566 	double total_slots = td_total_slots(ctx, cpu);
567 	double ret_slots = avg_stats(&runtime_topdown_slots_retired[ctx][cpu]);
568 
569 	if (total_slots)
570 		retiring = ret_slots / total_slots;
571 	return retiring;
572 }
573 
574 static double td_fe_bound(int ctx, int cpu)
575 {
576 	double fe_bound = 0;
577 	double total_slots = td_total_slots(ctx, cpu);
578 	double fetch_bub = avg_stats(&runtime_topdown_fetch_bubbles[ctx][cpu]);
579 
580 	if (total_slots)
581 		fe_bound = fetch_bub / total_slots;
582 	return fe_bound;
583 }
584 
585 static double td_be_bound(int ctx, int cpu)
586 {
587 	double sum = (td_fe_bound(ctx, cpu) +
588 		      td_bad_spec(ctx, cpu) +
589 		      td_retiring(ctx, cpu));
590 	if (sum == 0)
591 		return 0;
592 	return sanitize_val(1.0 - sum);
593 }
594 
595 void perf_stat__print_shadow_stats(struct perf_evsel *evsel,
596 				   double avg, int cpu,
597 				   struct perf_stat_output_ctx *out)
598 {
599 	void *ctxp = out->ctx;
600 	print_metric_t print_metric = out->print_metric;
601 	double total, ratio = 0.0, total2;
602 	const char *color = NULL;
603 	int ctx = evsel_context(evsel);
604 
605 	if (perf_evsel__match(evsel, HARDWARE, HW_INSTRUCTIONS)) {
606 		total = avg_stats(&runtime_cycles_stats[ctx][cpu]);
607 		if (total) {
608 			ratio = avg / total;
609 			print_metric(ctxp, NULL, "%7.2f ",
610 					"insn per cycle", ratio);
611 		} else {
612 			print_metric(ctxp, NULL, NULL, "insn per cycle", 0);
613 		}
614 		total = avg_stats(&runtime_stalled_cycles_front_stats[ctx][cpu]);
615 		total = max(total, avg_stats(&runtime_stalled_cycles_back_stats[ctx][cpu]));
616 
617 		if (total && avg) {
618 			out->new_line(ctxp);
619 			ratio = total / avg;
620 			print_metric(ctxp, NULL, "%7.2f ",
621 					"stalled cycles per insn",
622 					ratio);
623 		} else if (have_frontend_stalled) {
624 			print_metric(ctxp, NULL, NULL,
625 				     "stalled cycles per insn", 0);
626 		}
627 	} else if (perf_evsel__match(evsel, HARDWARE, HW_BRANCH_MISSES)) {
628 		if (runtime_branches_stats[ctx][cpu].n != 0)
629 			print_branch_misses(cpu, evsel, avg, out);
630 		else
631 			print_metric(ctxp, NULL, NULL, "of all branches", 0);
632 	} else if (
633 		evsel->attr.type == PERF_TYPE_HW_CACHE &&
634 		evsel->attr.config ==  ( PERF_COUNT_HW_CACHE_L1D |
635 					((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
636 					 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) {
637 		if (runtime_l1_dcache_stats[ctx][cpu].n != 0)
638 			print_l1_dcache_misses(cpu, evsel, avg, out);
639 		else
640 			print_metric(ctxp, NULL, NULL, "of all L1-dcache hits", 0);
641 	} else if (
642 		evsel->attr.type == PERF_TYPE_HW_CACHE &&
643 		evsel->attr.config ==  ( PERF_COUNT_HW_CACHE_L1I |
644 					((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
645 					 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) {
646 		if (runtime_l1_icache_stats[ctx][cpu].n != 0)
647 			print_l1_icache_misses(cpu, evsel, avg, out);
648 		else
649 			print_metric(ctxp, NULL, NULL, "of all L1-icache hits", 0);
650 	} else if (
651 		evsel->attr.type == PERF_TYPE_HW_CACHE &&
652 		evsel->attr.config ==  ( PERF_COUNT_HW_CACHE_DTLB |
653 					((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
654 					 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) {
655 		if (runtime_dtlb_cache_stats[ctx][cpu].n != 0)
656 			print_dtlb_cache_misses(cpu, evsel, avg, out);
657 		else
658 			print_metric(ctxp, NULL, NULL, "of all dTLB cache hits", 0);
659 	} else if (
660 		evsel->attr.type == PERF_TYPE_HW_CACHE &&
661 		evsel->attr.config ==  ( PERF_COUNT_HW_CACHE_ITLB |
662 					((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
663 					 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) {
664 		if (runtime_itlb_cache_stats[ctx][cpu].n != 0)
665 			print_itlb_cache_misses(cpu, evsel, avg, out);
666 		else
667 			print_metric(ctxp, NULL, NULL, "of all iTLB cache hits", 0);
668 	} else if (
669 		evsel->attr.type == PERF_TYPE_HW_CACHE &&
670 		evsel->attr.config ==  ( PERF_COUNT_HW_CACHE_LL |
671 					((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
672 					 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16))) {
673 		if (runtime_ll_cache_stats[ctx][cpu].n != 0)
674 			print_ll_cache_misses(cpu, evsel, avg, out);
675 		else
676 			print_metric(ctxp, NULL, NULL, "of all LL-cache hits", 0);
677 	} else if (perf_evsel__match(evsel, HARDWARE, HW_CACHE_MISSES)) {
678 		total = avg_stats(&runtime_cacherefs_stats[ctx][cpu]);
679 
680 		if (total)
681 			ratio = avg * 100 / total;
682 
683 		if (runtime_cacherefs_stats[ctx][cpu].n != 0)
684 			print_metric(ctxp, NULL, "%8.3f %%",
685 				     "of all cache refs", ratio);
686 		else
687 			print_metric(ctxp, NULL, NULL, "of all cache refs", 0);
688 	} else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_FRONTEND)) {
689 		print_stalled_cycles_frontend(cpu, evsel, avg, out);
690 	} else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_BACKEND)) {
691 		print_stalled_cycles_backend(cpu, evsel, avg, out);
692 	} else if (perf_evsel__match(evsel, HARDWARE, HW_CPU_CYCLES)) {
693 		total = avg_stats(&runtime_nsecs_stats[cpu]);
694 
695 		if (total) {
696 			ratio = avg / total;
697 			print_metric(ctxp, NULL, "%8.3f", "GHz", ratio);
698 		} else {
699 			print_metric(ctxp, NULL, NULL, "Ghz", 0);
700 		}
701 	} else if (perf_stat_evsel__is(evsel, CYCLES_IN_TX)) {
702 		total = avg_stats(&runtime_cycles_stats[ctx][cpu]);
703 		if (total)
704 			print_metric(ctxp, NULL,
705 					"%7.2f%%", "transactional cycles",
706 					100.0 * (avg / total));
707 		else
708 			print_metric(ctxp, NULL, NULL, "transactional cycles",
709 				     0);
710 	} else if (perf_stat_evsel__is(evsel, CYCLES_IN_TX_CP)) {
711 		total = avg_stats(&runtime_cycles_stats[ctx][cpu]);
712 		total2 = avg_stats(&runtime_cycles_in_tx_stats[ctx][cpu]);
713 		if (total2 < avg)
714 			total2 = avg;
715 		if (total)
716 			print_metric(ctxp, NULL, "%7.2f%%", "aborted cycles",
717 				100.0 * ((total2-avg) / total));
718 		else
719 			print_metric(ctxp, NULL, NULL, "aborted cycles", 0);
720 	} else if (perf_stat_evsel__is(evsel, TRANSACTION_START)) {
721 		total = avg_stats(&runtime_cycles_in_tx_stats[ctx][cpu]);
722 
723 		if (avg)
724 			ratio = total / avg;
725 
726 		if (runtime_cycles_in_tx_stats[ctx][cpu].n != 0)
727 			print_metric(ctxp, NULL, "%8.0f",
728 				     "cycles / transaction", ratio);
729 		else
730 			print_metric(ctxp, NULL, NULL, "cycles / transaction",
731 				     0);
732 	} else if (perf_stat_evsel__is(evsel, ELISION_START)) {
733 		total = avg_stats(&runtime_cycles_in_tx_stats[ctx][cpu]);
734 
735 		if (avg)
736 			ratio = total / avg;
737 
738 		print_metric(ctxp, NULL, "%8.0f", "cycles / elision", ratio);
739 	} else if (perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK) ||
740 		   perf_evsel__match(evsel, SOFTWARE, SW_CPU_CLOCK)) {
741 		if ((ratio = avg_stats(&walltime_nsecs_stats)) != 0)
742 			print_metric(ctxp, NULL, "%8.3f", "CPUs utilized",
743 				     avg / ratio);
744 		else
745 			print_metric(ctxp, NULL, NULL, "CPUs utilized", 0);
746 	} else if (perf_stat_evsel__is(evsel, TOPDOWN_FETCH_BUBBLES)) {
747 		double fe_bound = td_fe_bound(ctx, cpu);
748 
749 		if (fe_bound > 0.2)
750 			color = PERF_COLOR_RED;
751 		print_metric(ctxp, color, "%8.1f%%", "frontend bound",
752 				fe_bound * 100.);
753 	} else if (perf_stat_evsel__is(evsel, TOPDOWN_SLOTS_RETIRED)) {
754 		double retiring = td_retiring(ctx, cpu);
755 
756 		if (retiring > 0.7)
757 			color = PERF_COLOR_GREEN;
758 		print_metric(ctxp, color, "%8.1f%%", "retiring",
759 				retiring * 100.);
760 	} else if (perf_stat_evsel__is(evsel, TOPDOWN_RECOVERY_BUBBLES)) {
761 		double bad_spec = td_bad_spec(ctx, cpu);
762 
763 		if (bad_spec > 0.1)
764 			color = PERF_COLOR_RED;
765 		print_metric(ctxp, color, "%8.1f%%", "bad speculation",
766 				bad_spec * 100.);
767 	} else if (perf_stat_evsel__is(evsel, TOPDOWN_SLOTS_ISSUED)) {
768 		double be_bound = td_be_bound(ctx, cpu);
769 		const char *name = "backend bound";
770 		static int have_recovery_bubbles = -1;
771 
772 		/* In case the CPU does not support topdown-recovery-bubbles */
773 		if (have_recovery_bubbles < 0)
774 			have_recovery_bubbles = pmu_have_event("cpu",
775 					"topdown-recovery-bubbles");
776 		if (!have_recovery_bubbles)
777 			name = "backend bound/bad spec";
778 
779 		if (be_bound > 0.2)
780 			color = PERF_COLOR_RED;
781 		if (td_total_slots(ctx, cpu) > 0)
782 			print_metric(ctxp, color, "%8.1f%%", name,
783 					be_bound * 100.);
784 		else
785 			print_metric(ctxp, NULL, NULL, name, 0);
786 	} else if (evsel->metric_expr) {
787 		struct parse_ctx pctx;
788 		int i;
789 
790 		expr__ctx_init(&pctx);
791 		expr__add_id(&pctx, evsel->name, avg);
792 		for (i = 0; evsel->metric_events[i]; i++) {
793 			struct saved_value *v;
794 
795 			v = saved_value_lookup(evsel->metric_events[i], cpu, ctx, false);
796 			if (!v)
797 				break;
798 			expr__add_id(&pctx, evsel->metric_events[i]->name,
799 					     avg_stats(&v->stats));
800 		}
801 		if (!evsel->metric_events[i]) {
802 			const char *p = evsel->metric_expr;
803 
804 			if (expr__parse(&ratio, &pctx, &p) == 0)
805 				print_metric(ctxp, NULL, "%8.1f",
806 					out->force_header ? evsel->name : "",
807 					ratio);
808 			else
809 				print_metric(ctxp, NULL, NULL, "", 0);
810 		} else
811 			print_metric(ctxp, NULL, NULL, "", 0);
812 	} else if (runtime_nsecs_stats[cpu].n != 0) {
813 		char unit = 'M';
814 		char unit_buf[10];
815 
816 		total = avg_stats(&runtime_nsecs_stats[cpu]);
817 
818 		if (total)
819 			ratio = 1000.0 * avg / total;
820 		if (ratio < 0.001) {
821 			ratio *= 1000;
822 			unit = 'K';
823 		}
824 		snprintf(unit_buf, sizeof(unit_buf), "%c/sec", unit);
825 		print_metric(ctxp, NULL, "%8.3f", unit_buf, ratio);
826 	} else {
827 		print_metric(ctxp, NULL, NULL, NULL, 0);
828 	}
829 }
830