aboutsummaryrefslogtreecommitdiffstats
path: root/tools/perf/builtin-stat.c
diff options
context:
space:
mode:
Diffstat (limited to 'tools/perf/builtin-stat.c')
-rw-r--r--tools/perf/builtin-stat.c580
1 files changed, 514 insertions, 66 deletions
diff --git a/tools/perf/builtin-stat.c b/tools/perf/builtin-stat.c
index e2109f9b43e..a9f06715e44 100644
--- a/tools/perf/builtin-stat.c
+++ b/tools/perf/builtin-stat.c
@@ -6,24 +6,28 @@
6 * 6 *
7 * Sample output: 7 * Sample output:
8 8
9 $ perf stat ~/hackbench 10 9 $ perf stat ./hackbench 10
10 Time: 0.104
11 10
12 Performance counter stats for '/home/mingo/hackbench': 11 Time: 0.118
13 12
14 1255.538611 task clock ticks # 10.143 CPU utilization factor 13 Performance counter stats for './hackbench 10':
15 54011 context switches # 0.043 M/sec
16 385 CPU migrations # 0.000 M/sec
17 17755 pagefaults # 0.014 M/sec
18 3808323185 CPU cycles # 3033.219 M/sec
19 1575111190 instructions # 1254.530 M/sec
20 17367895 cache references # 13.833 M/sec
21 7674421 cache misses # 6.112 M/sec
22 14
23 Wall-clock time elapsed: 123.786620 msecs 15 1708.761321 task-clock # 11.037 CPUs utilized
16 41,190 context-switches # 0.024 M/sec
17 6,735 CPU-migrations # 0.004 M/sec
18 17,318 page-faults # 0.010 M/sec
19 5,205,202,243 cycles # 3.046 GHz
20 3,856,436,920 stalled-cycles-frontend # 74.09% frontend cycles idle
21 1,600,790,871 stalled-cycles-backend # 30.75% backend cycles idle
22 2,603,501,247 instructions # 0.50 insns per cycle
23 # 1.48 stalled cycles per insn
24 484,357,498 branches # 283.455 M/sec
25 6,388,934 branch-misses # 1.32% of all branches
26
27 0.154822978 seconds time elapsed
24 28
25 * 29 *
26 * Copyright (C) 2008, Red Hat Inc, Ingo Molnar <mingo@redhat.com> 30 * Copyright (C) 2008-2011, Red Hat Inc, Ingo Molnar <mingo@redhat.com>
27 * 31 *
28 * Improvements and fixes by: 32 * Improvements and fixes by:
29 * 33 *
@@ -46,6 +50,7 @@
46#include "util/evlist.h" 50#include "util/evlist.h"
47#include "util/evsel.h" 51#include "util/evsel.h"
48#include "util/debug.h" 52#include "util/debug.h"
53#include "util/color.h"
49#include "util/header.h" 54#include "util/header.h"
50#include "util/cpumap.h" 55#include "util/cpumap.h"
51#include "util/thread.h" 56#include "util/thread.h"
@@ -65,14 +70,107 @@ static struct perf_event_attr default_attrs[] = {
65 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS }, 70 { .type = PERF_TYPE_SOFTWARE, .config = PERF_COUNT_SW_PAGE_FAULTS },
66 71
67 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES }, 72 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CPU_CYCLES },
73 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND },
74 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_STALLED_CYCLES_BACKEND },
68 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS }, 75 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_INSTRUCTIONS },
69 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS }, 76 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_INSTRUCTIONS },
70 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES }, 77 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_BRANCH_MISSES },
71 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CACHE_REFERENCES },
72 { .type = PERF_TYPE_HARDWARE, .config = PERF_COUNT_HW_CACHE_MISSES },
73 78
74}; 79};
75 80
81/*
82 * Detailed stats (-d), covering the L1 and last level data caches:
83 */
84static struct perf_event_attr detailed_attrs[] = {
85
86 { .type = PERF_TYPE_HW_CACHE,
87 .config =
88 PERF_COUNT_HW_CACHE_L1D << 0 |
89 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
90 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
91
92 { .type = PERF_TYPE_HW_CACHE,
93 .config =
94 PERF_COUNT_HW_CACHE_L1D << 0 |
95 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
96 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
97
98 { .type = PERF_TYPE_HW_CACHE,
99 .config =
100 PERF_COUNT_HW_CACHE_LL << 0 |
101 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
102 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
103
104 { .type = PERF_TYPE_HW_CACHE,
105 .config =
106 PERF_COUNT_HW_CACHE_LL << 0 |
107 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
108 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
109};
110
111/*
112 * Very detailed stats (-d -d), covering the instruction cache and the TLB caches:
113 */
114static struct perf_event_attr very_detailed_attrs[] = {
115
116 { .type = PERF_TYPE_HW_CACHE,
117 .config =
118 PERF_COUNT_HW_CACHE_L1I << 0 |
119 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
120 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
121
122 { .type = PERF_TYPE_HW_CACHE,
123 .config =
124 PERF_COUNT_HW_CACHE_L1I << 0 |
125 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
126 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
127
128 { .type = PERF_TYPE_HW_CACHE,
129 .config =
130 PERF_COUNT_HW_CACHE_DTLB << 0 |
131 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
132 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
133
134 { .type = PERF_TYPE_HW_CACHE,
135 .config =
136 PERF_COUNT_HW_CACHE_DTLB << 0 |
137 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
138 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
139
140 { .type = PERF_TYPE_HW_CACHE,
141 .config =
142 PERF_COUNT_HW_CACHE_ITLB << 0 |
143 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
144 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
145
146 { .type = PERF_TYPE_HW_CACHE,
147 .config =
148 PERF_COUNT_HW_CACHE_ITLB << 0 |
149 (PERF_COUNT_HW_CACHE_OP_READ << 8) |
150 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
151
152};
153
154/*
155 * Very, very detailed stats (-d -d -d), adding prefetch events:
156 */
157static struct perf_event_attr very_very_detailed_attrs[] = {
158
159 { .type = PERF_TYPE_HW_CACHE,
160 .config =
161 PERF_COUNT_HW_CACHE_L1D << 0 |
162 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
163 (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16) },
164
165 { .type = PERF_TYPE_HW_CACHE,
166 .config =
167 PERF_COUNT_HW_CACHE_L1D << 0 |
168 (PERF_COUNT_HW_CACHE_OP_PREFETCH << 8) |
169 (PERF_COUNT_HW_CACHE_RESULT_MISS << 16) },
170};
171
172
173
76struct perf_evlist *evsel_list; 174struct perf_evlist *evsel_list;
77 175
78static bool system_wide = false; 176static bool system_wide = false;
@@ -86,6 +184,8 @@ static pid_t target_pid = -1;
86static pid_t target_tid = -1; 184static pid_t target_tid = -1;
87static pid_t child_pid = -1; 185static pid_t child_pid = -1;
88static bool null_run = false; 186static bool null_run = false;
187static int detailed_run = 0;
188static bool sync_run = false;
89static bool big_num = true; 189static bool big_num = true;
90static int big_num_opt = -1; 190static int big_num_opt = -1;
91static const char *cpu_list; 191static const char *cpu_list;
@@ -156,7 +256,15 @@ static double stddev_stats(struct stats *stats)
156 256
157struct stats runtime_nsecs_stats[MAX_NR_CPUS]; 257struct stats runtime_nsecs_stats[MAX_NR_CPUS];
158struct stats runtime_cycles_stats[MAX_NR_CPUS]; 258struct stats runtime_cycles_stats[MAX_NR_CPUS];
259struct stats runtime_stalled_cycles_front_stats[MAX_NR_CPUS];
260struct stats runtime_stalled_cycles_back_stats[MAX_NR_CPUS];
159struct stats runtime_branches_stats[MAX_NR_CPUS]; 261struct stats runtime_branches_stats[MAX_NR_CPUS];
262struct stats runtime_cacherefs_stats[MAX_NR_CPUS];
263struct stats runtime_l1_dcache_stats[MAX_NR_CPUS];
264struct stats runtime_l1_icache_stats[MAX_NR_CPUS];
265struct stats runtime_ll_cache_stats[MAX_NR_CPUS];
266struct stats runtime_itlb_cache_stats[MAX_NR_CPUS];
267struct stats runtime_dtlb_cache_stats[MAX_NR_CPUS];
160struct stats walltime_nsecs_stats; 268struct stats walltime_nsecs_stats;
161 269
162static int create_perf_stat_counter(struct perf_evsel *evsel) 270static int create_perf_stat_counter(struct perf_evsel *evsel)
@@ -167,16 +275,17 @@ static int create_perf_stat_counter(struct perf_evsel *evsel)
167 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED | 275 attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
168 PERF_FORMAT_TOTAL_TIME_RUNNING; 276 PERF_FORMAT_TOTAL_TIME_RUNNING;
169 277
278 attr->inherit = !no_inherit;
279
170 if (system_wide) 280 if (system_wide)
171 return perf_evsel__open_per_cpu(evsel, evsel_list->cpus, false, false); 281 return perf_evsel__open_per_cpu(evsel, evsel_list->cpus, false);
172 282
173 attr->inherit = !no_inherit;
174 if (target_pid == -1 && target_tid == -1) { 283 if (target_pid == -1 && target_tid == -1) {
175 attr->disabled = 1; 284 attr->disabled = 1;
176 attr->enable_on_exec = 1; 285 attr->enable_on_exec = 1;
177 } 286 }
178 287
179 return perf_evsel__open_per_thread(evsel, evsel_list->threads, false, false); 288 return perf_evsel__open_per_thread(evsel, evsel_list->threads, false);
180} 289}
181 290
182/* 291/*
@@ -192,6 +301,37 @@ static inline int nsec_counter(struct perf_evsel *evsel)
192} 301}
193 302
194/* 303/*
304 * Update various tracking values we maintain to print
305 * more semantic information such as miss/hit ratios,
306 * instruction rates, etc:
307 */
308static void update_shadow_stats(struct perf_evsel *counter, u64 *count)
309{
310 if (perf_evsel__match(counter, SOFTWARE, SW_TASK_CLOCK))
311 update_stats(&runtime_nsecs_stats[0], count[0]);
312 else if (perf_evsel__match(counter, HARDWARE, HW_CPU_CYCLES))
313 update_stats(&runtime_cycles_stats[0], count[0]);
314 else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_FRONTEND))
315 update_stats(&runtime_stalled_cycles_front_stats[0], count[0]);
316 else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_BACKEND))
317 update_stats(&runtime_stalled_cycles_back_stats[0], count[0]);
318 else if (perf_evsel__match(counter, HARDWARE, HW_BRANCH_INSTRUCTIONS))
319 update_stats(&runtime_branches_stats[0], count[0]);
320 else if (perf_evsel__match(counter, HARDWARE, HW_CACHE_REFERENCES))
321 update_stats(&runtime_cacherefs_stats[0], count[0]);
322 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_L1D))
323 update_stats(&runtime_l1_dcache_stats[0], count[0]);
324 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_L1I))
325 update_stats(&runtime_l1_icache_stats[0], count[0]);
326 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_LL))
327 update_stats(&runtime_ll_cache_stats[0], count[0]);
328 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_DTLB))
329 update_stats(&runtime_dtlb_cache_stats[0], count[0]);
330 else if (perf_evsel__match(counter, HW_CACHE, HW_CACHE_ITLB))
331 update_stats(&runtime_itlb_cache_stats[0], count[0]);
332}
333
334/*
195 * Read out the results of a single counter: 335 * Read out the results of a single counter:
196 * aggregate counts across CPUs in system-wide mode 336 * aggregate counts across CPUs in system-wide mode
197 */ 337 */
@@ -216,12 +356,7 @@ static int read_counter_aggr(struct perf_evsel *counter)
216 /* 356 /*
217 * Save the full runtime - to allow normalization during printout: 357 * Save the full runtime - to allow normalization during printout:
218 */ 358 */
219 if (perf_evsel__match(counter, SOFTWARE, SW_TASK_CLOCK)) 359 update_shadow_stats(counter, count);
220 update_stats(&runtime_nsecs_stats[0], count[0]);
221 if (perf_evsel__match(counter, HARDWARE, HW_CPU_CYCLES))
222 update_stats(&runtime_cycles_stats[0], count[0]);
223 if (perf_evsel__match(counter, HARDWARE, HW_BRANCH_INSTRUCTIONS))
224 update_stats(&runtime_branches_stats[0], count[0]);
225 360
226 return 0; 361 return 0;
227} 362}
@@ -241,12 +376,7 @@ static int read_counter(struct perf_evsel *counter)
241 376
242 count = counter->counts->cpu[cpu].values; 377 count = counter->counts->cpu[cpu].values;
243 378
244 if (perf_evsel__match(counter, SOFTWARE, SW_TASK_CLOCK)) 379 update_shadow_stats(counter, count);
245 update_stats(&runtime_nsecs_stats[cpu], count[0]);
246 if (perf_evsel__match(counter, HARDWARE, HW_CPU_CYCLES))
247 update_stats(&runtime_cycles_stats[cpu], count[0]);
248 if (perf_evsel__match(counter, HARDWARE, HW_BRANCH_INSTRUCTIONS))
249 update_stats(&runtime_branches_stats[cpu], count[0]);
250 } 380 }
251 381
252 return 0; 382 return 0;
@@ -314,13 +444,18 @@ static int run_perf_stat(int argc __used, const char **argv)
314 444
315 list_for_each_entry(counter, &evsel_list->entries, node) { 445 list_for_each_entry(counter, &evsel_list->entries, node) {
316 if (create_perf_stat_counter(counter) < 0) { 446 if (create_perf_stat_counter(counter) < 0) {
317 if (errno == -EPERM || errno == -EACCES) { 447 if (errno == EINVAL || errno == ENOSYS || errno == ENOENT) {
448 if (verbose)
449 ui__warning("%s event is not supported by the kernel.\n",
450 event_name(counter));
451 continue;
452 }
453
454 if (errno == EPERM || errno == EACCES) {
318 error("You may not have permission to collect %sstats.\n" 455 error("You may not have permission to collect %sstats.\n"
319 "\t Consider tweaking" 456 "\t Consider tweaking"
320 " /proc/sys/kernel/perf_event_paranoid or running as root.", 457 " /proc/sys/kernel/perf_event_paranoid or running as root.",
321 system_wide ? "system-wide " : ""); 458 system_wide ? "system-wide " : "");
322 } else if (errno == ENOENT) {
323 error("%s event is not supported. ", event_name(counter));
324 } else { 459 } else {
325 error("open_counter returned with %d (%s). " 460 error("open_counter returned with %d (%s). "
326 "/bin/dmesg may provide additional information.\n", 461 "/bin/dmesg may provide additional information.\n",
@@ -371,6 +506,16 @@ static int run_perf_stat(int argc __used, const char **argv)
371 return WEXITSTATUS(status); 506 return WEXITSTATUS(status);
372} 507}
373 508
509static void print_noise_pct(double total, double avg)
510{
511 double pct = 0.0;
512
513 if (avg)
514 pct = 100.0*total/avg;
515
516 fprintf(stderr, " ( +-%6.2f%% )", pct);
517}
518
374static void print_noise(struct perf_evsel *evsel, double avg) 519static void print_noise(struct perf_evsel *evsel, double avg)
375{ 520{
376 struct perf_stat *ps; 521 struct perf_stat *ps;
@@ -379,15 +524,14 @@ static void print_noise(struct perf_evsel *evsel, double avg)
379 return; 524 return;
380 525
381 ps = evsel->priv; 526 ps = evsel->priv;
382 fprintf(stderr, " ( +- %7.3f%% )", 527 print_noise_pct(stddev_stats(&ps->res_stats[0]), avg);
383 100 * stddev_stats(&ps->res_stats[0]) / avg);
384} 528}
385 529
386static void nsec_printout(int cpu, struct perf_evsel *evsel, double avg) 530static void nsec_printout(int cpu, struct perf_evsel *evsel, double avg)
387{ 531{
388 double msecs = avg / 1e6; 532 double msecs = avg / 1e6;
389 char cpustr[16] = { '\0', }; 533 char cpustr[16] = { '\0', };
390 const char *fmt = csv_output ? "%s%.6f%s%s" : "%s%18.6f%s%-24s"; 534 const char *fmt = csv_output ? "%s%.6f%s%s" : "%s%18.6f%s%-25s";
391 535
392 if (no_aggr) 536 if (no_aggr)
393 sprintf(cpustr, "CPU%*d%s", 537 sprintf(cpustr, "CPU%*d%s",
@@ -403,8 +547,191 @@ static void nsec_printout(int cpu, struct perf_evsel *evsel, double avg)
403 return; 547 return;
404 548
405 if (perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK)) 549 if (perf_evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK))
406 fprintf(stderr, " # %10.3f CPUs ", 550 fprintf(stderr, " # %8.3f CPUs utilized ", avg / avg_stats(&walltime_nsecs_stats));
407 avg / avg_stats(&walltime_nsecs_stats)); 551}
552
553static void print_stalled_cycles_frontend(int cpu, struct perf_evsel *evsel __used, double avg)
554{
555 double total, ratio = 0.0;
556 const char *color;
557
558 total = avg_stats(&runtime_cycles_stats[cpu]);
559
560 if (total)
561 ratio = avg / total * 100.0;
562
563 color = PERF_COLOR_NORMAL;
564 if (ratio > 50.0)
565 color = PERF_COLOR_RED;
566 else if (ratio > 30.0)
567 color = PERF_COLOR_MAGENTA;
568 else if (ratio > 10.0)
569 color = PERF_COLOR_YELLOW;
570
571 fprintf(stderr, " # ");
572 color_fprintf(stderr, color, "%6.2f%%", ratio);
573 fprintf(stderr, " frontend cycles idle ");
574}
575
576static void print_stalled_cycles_backend(int cpu, struct perf_evsel *evsel __used, double avg)
577{
578 double total, ratio = 0.0;
579 const char *color;
580
581 total = avg_stats(&runtime_cycles_stats[cpu]);
582
583 if (total)
584 ratio = avg / total * 100.0;
585
586 color = PERF_COLOR_NORMAL;
587 if (ratio > 75.0)
588 color = PERF_COLOR_RED;
589 else if (ratio > 50.0)
590 color = PERF_COLOR_MAGENTA;
591 else if (ratio > 20.0)
592 color = PERF_COLOR_YELLOW;
593
594 fprintf(stderr, " # ");
595 color_fprintf(stderr, color, "%6.2f%%", ratio);
596 fprintf(stderr, " backend cycles idle ");
597}
598
599static void print_branch_misses(int cpu, struct perf_evsel *evsel __used, double avg)
600{
601 double total, ratio = 0.0;
602 const char *color;
603
604 total = avg_stats(&runtime_branches_stats[cpu]);
605
606 if (total)
607 ratio = avg / total * 100.0;
608
609 color = PERF_COLOR_NORMAL;
610 if (ratio > 20.0)
611 color = PERF_COLOR_RED;
612 else if (ratio > 10.0)
613 color = PERF_COLOR_MAGENTA;
614 else if (ratio > 5.0)
615 color = PERF_COLOR_YELLOW;
616
617 fprintf(stderr, " # ");
618 color_fprintf(stderr, color, "%6.2f%%", ratio);
619 fprintf(stderr, " of all branches ");
620}
621
622static void print_l1_dcache_misses(int cpu, struct perf_evsel *evsel __used, double avg)
623{
624 double total, ratio = 0.0;
625 const char *color;
626
627 total = avg_stats(&runtime_l1_dcache_stats[cpu]);
628
629 if (total)
630 ratio = avg / total * 100.0;
631
632 color = PERF_COLOR_NORMAL;
633 if (ratio > 20.0)
634 color = PERF_COLOR_RED;
635 else if (ratio > 10.0)
636 color = PERF_COLOR_MAGENTA;
637 else if (ratio > 5.0)
638 color = PERF_COLOR_YELLOW;
639
640 fprintf(stderr, " # ");
641 color_fprintf(stderr, color, "%6.2f%%", ratio);
642 fprintf(stderr, " of all L1-dcache hits ");
643}
644
645static void print_l1_icache_misses(int cpu, struct perf_evsel *evsel __used, double avg)
646{
647 double total, ratio = 0.0;
648 const char *color;
649
650 total = avg_stats(&runtime_l1_icache_stats[cpu]);
651
652 if (total)
653 ratio = avg / total * 100.0;
654
655 color = PERF_COLOR_NORMAL;
656 if (ratio > 20.0)
657 color = PERF_COLOR_RED;
658 else if (ratio > 10.0)
659 color = PERF_COLOR_MAGENTA;
660 else if (ratio > 5.0)
661 color = PERF_COLOR_YELLOW;
662
663 fprintf(stderr, " # ");
664 color_fprintf(stderr, color, "%6.2f%%", ratio);
665 fprintf(stderr, " of all L1-icache hits ");
666}
667
668static void print_dtlb_cache_misses(int cpu, struct perf_evsel *evsel __used, double avg)
669{
670 double total, ratio = 0.0;
671 const char *color;
672
673 total = avg_stats(&runtime_dtlb_cache_stats[cpu]);
674
675 if (total)
676 ratio = avg / total * 100.0;
677
678 color = PERF_COLOR_NORMAL;
679 if (ratio > 20.0)
680 color = PERF_COLOR_RED;
681 else if (ratio > 10.0)
682 color = PERF_COLOR_MAGENTA;
683 else if (ratio > 5.0)
684 color = PERF_COLOR_YELLOW;
685
686 fprintf(stderr, " # ");
687 color_fprintf(stderr, color, "%6.2f%%", ratio);
688 fprintf(stderr, " of all dTLB cache hits ");
689}
690
691static void print_itlb_cache_misses(int cpu, struct perf_evsel *evsel __used, double avg)
692{
693 double total, ratio = 0.0;
694 const char *color;
695
696 total = avg_stats(&runtime_itlb_cache_stats[cpu]);
697
698 if (total)
699 ratio = avg / total * 100.0;
700
701 color = PERF_COLOR_NORMAL;
702 if (ratio > 20.0)
703 color = PERF_COLOR_RED;
704 else if (ratio > 10.0)
705 color = PERF_COLOR_MAGENTA;
706 else if (ratio > 5.0)
707 color = PERF_COLOR_YELLOW;
708
709 fprintf(stderr, " # ");
710 color_fprintf(stderr, color, "%6.2f%%", ratio);
711 fprintf(stderr, " of all iTLB cache hits ");
712}
713
714static void print_ll_cache_misses(int cpu, struct perf_evsel *evsel __used, double avg)
715{
716 double total, ratio = 0.0;
717 const char *color;
718
719 total = avg_stats(&runtime_ll_cache_stats[cpu]);
720
721 if (total)
722 ratio = avg / total * 100.0;
723
724 color = PERF_COLOR_NORMAL;
725 if (ratio > 20.0)
726 color = PERF_COLOR_RED;
727 else if (ratio > 10.0)
728 color = PERF_COLOR_MAGENTA;
729 else if (ratio > 5.0)
730 color = PERF_COLOR_YELLOW;
731
732 fprintf(stderr, " # ");
733 color_fprintf(stderr, color, "%6.2f%%", ratio);
734 fprintf(stderr, " of all LL-cache hits ");
408} 735}
409 736
410static void abs_printout(int cpu, struct perf_evsel *evsel, double avg) 737static void abs_printout(int cpu, struct perf_evsel *evsel, double avg)
@@ -416,9 +743,9 @@ static void abs_printout(int cpu, struct perf_evsel *evsel, double avg)
416 if (csv_output) 743 if (csv_output)
417 fmt = "%s%.0f%s%s"; 744 fmt = "%s%.0f%s%s";
418 else if (big_num) 745 else if (big_num)
419 fmt = "%s%'18.0f%s%-24s"; 746 fmt = "%s%'18.0f%s%-25s";
420 else 747 else
421 fmt = "%s%18.0f%s%-24s"; 748 fmt = "%s%18.0f%s%-25s";
422 749
423 if (no_aggr) 750 if (no_aggr)
424 sprintf(cpustr, "CPU%*d%s", 751 sprintf(cpustr, "CPU%*d%s",
@@ -441,23 +768,83 @@ static void abs_printout(int cpu, struct perf_evsel *evsel, double avg)
441 if (total) 768 if (total)
442 ratio = avg / total; 769 ratio = avg / total;
443 770
444 fprintf(stderr, " # %10.3f IPC ", ratio); 771 fprintf(stderr, " # %5.2f insns per cycle ", ratio);
772
773 total = avg_stats(&runtime_stalled_cycles_front_stats[cpu]);
774 total = max(total, avg_stats(&runtime_stalled_cycles_back_stats[cpu]));
775
776 if (total && avg) {
777 ratio = total / avg;
778 fprintf(stderr, "\n # %5.2f stalled cycles per insn", ratio);
779 }
780
445 } else if (perf_evsel__match(evsel, HARDWARE, HW_BRANCH_MISSES) && 781 } else if (perf_evsel__match(evsel, HARDWARE, HW_BRANCH_MISSES) &&
446 runtime_branches_stats[cpu].n != 0) { 782 runtime_branches_stats[cpu].n != 0) {
447 total = avg_stats(&runtime_branches_stats[cpu]); 783 print_branch_misses(cpu, evsel, avg);
784 } else if (
785 evsel->attr.type == PERF_TYPE_HW_CACHE &&
786 evsel->attr.config == ( PERF_COUNT_HW_CACHE_L1D |
787 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
788 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
789 runtime_l1_dcache_stats[cpu].n != 0) {
790 print_l1_dcache_misses(cpu, evsel, avg);
791 } else if (
792 evsel->attr.type == PERF_TYPE_HW_CACHE &&
793 evsel->attr.config == ( PERF_COUNT_HW_CACHE_L1I |
794 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
795 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
796 runtime_l1_icache_stats[cpu].n != 0) {
797 print_l1_icache_misses(cpu, evsel, avg);
798 } else if (
799 evsel->attr.type == PERF_TYPE_HW_CACHE &&
800 evsel->attr.config == ( PERF_COUNT_HW_CACHE_DTLB |
801 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
802 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
803 runtime_dtlb_cache_stats[cpu].n != 0) {
804 print_dtlb_cache_misses(cpu, evsel, avg);
805 } else if (
806 evsel->attr.type == PERF_TYPE_HW_CACHE &&
807 evsel->attr.config == ( PERF_COUNT_HW_CACHE_ITLB |
808 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
809 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
810 runtime_itlb_cache_stats[cpu].n != 0) {
811 print_itlb_cache_misses(cpu, evsel, avg);
812 } else if (
813 evsel->attr.type == PERF_TYPE_HW_CACHE &&
814 evsel->attr.config == ( PERF_COUNT_HW_CACHE_LL |
815 ((PERF_COUNT_HW_CACHE_OP_READ) << 8) |
816 ((PERF_COUNT_HW_CACHE_RESULT_MISS) << 16)) &&
817 runtime_ll_cache_stats[cpu].n != 0) {
818 print_ll_cache_misses(cpu, evsel, avg);
819 } else if (perf_evsel__match(evsel, HARDWARE, HW_CACHE_MISSES) &&
820 runtime_cacherefs_stats[cpu].n != 0) {
821 total = avg_stats(&runtime_cacherefs_stats[cpu]);
448 822
449 if (total) 823 if (total)
450 ratio = avg * 100 / total; 824 ratio = avg * 100 / total;
451 825
452 fprintf(stderr, " # %10.3f %% ", ratio); 826 fprintf(stderr, " # %8.3f %% of all cache refs ", ratio);
453 827
828 } else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_FRONTEND)) {
829 print_stalled_cycles_frontend(cpu, evsel, avg);
830 } else if (perf_evsel__match(evsel, HARDWARE, HW_STALLED_CYCLES_BACKEND)) {
831 print_stalled_cycles_backend(cpu, evsel, avg);
832 } else if (perf_evsel__match(evsel, HARDWARE, HW_CPU_CYCLES)) {
833 total = avg_stats(&runtime_nsecs_stats[cpu]);
834
835 if (total)
836 ratio = 1.0 * avg / total;
837
838 fprintf(stderr, " # %8.3f GHz ", ratio);
454 } else if (runtime_nsecs_stats[cpu].n != 0) { 839 } else if (runtime_nsecs_stats[cpu].n != 0) {
455 total = avg_stats(&runtime_nsecs_stats[cpu]); 840 total = avg_stats(&runtime_nsecs_stats[cpu]);
456 841
457 if (total) 842 if (total)
458 ratio = 1000.0 * avg / total; 843 ratio = 1000.0 * avg / total;
459 844
460 fprintf(stderr, " # %10.3f M/sec", ratio); 845 fprintf(stderr, " # %8.3f M/sec ", ratio);
846 } else {
847 fprintf(stderr, " ");
461 } 848 }
462} 849}
463 850
@@ -504,8 +891,7 @@ static void print_counter_aggr(struct perf_evsel *counter)
504 avg_enabled = avg_stats(&ps->res_stats[1]); 891 avg_enabled = avg_stats(&ps->res_stats[1]);
505 avg_running = avg_stats(&ps->res_stats[2]); 892 avg_running = avg_stats(&ps->res_stats[2]);
506 893
507 fprintf(stderr, " (scaled from %.2f%%)", 894 fprintf(stderr, " [%5.2f%%]", 100 * avg_running / avg_enabled);
508 100 * avg_running / avg_enabled);
509 } 895 }
510 fprintf(stderr, "\n"); 896 fprintf(stderr, "\n");
511} 897}
@@ -547,10 +933,8 @@ static void print_counter(struct perf_evsel *counter)
547 if (!csv_output) { 933 if (!csv_output) {
548 print_noise(counter, 1.0); 934 print_noise(counter, 1.0);
549 935
550 if (run != ena) { 936 if (run != ena)
551 fprintf(stderr, " (scaled from %.2f%%)", 937 fprintf(stderr, " (%.2f%%)", 100.0 * run / ena);
552 100.0 * run / ena);
553 }
554 } 938 }
555 fputc('\n', stderr); 939 fputc('\n', stderr);
556 } 940 }
@@ -590,13 +974,14 @@ static void print_stat(int argc, const char **argv)
590 } 974 }
591 975
592 if (!csv_output) { 976 if (!csv_output) {
593 fprintf(stderr, "\n"); 977 if (!null_run)
594 fprintf(stderr, " %18.9f seconds time elapsed", 978 fprintf(stderr, "\n");
979 fprintf(stderr, " %17.9f seconds time elapsed",
595 avg_stats(&walltime_nsecs_stats)/1e9); 980 avg_stats(&walltime_nsecs_stats)/1e9);
596 if (run_count > 1) { 981 if (run_count > 1) {
597 fprintf(stderr, " ( +- %7.3f%% )", 982 fprintf(stderr, " ");
598 100*stddev_stats(&walltime_nsecs_stats) / 983 print_noise_pct(stddev_stats(&walltime_nsecs_stats),
599 avg_stats(&walltime_nsecs_stats)); 984 avg_stats(&walltime_nsecs_stats));
600 } 985 }
601 fprintf(stderr, "\n\n"); 986 fprintf(stderr, "\n\n");
602 } 987 }
@@ -658,6 +1043,10 @@ static const struct option options[] = {
658 "repeat command and print average + stddev (max: 100)"), 1043 "repeat command and print average + stddev (max: 100)"),
659 OPT_BOOLEAN('n', "null", &null_run, 1044 OPT_BOOLEAN('n', "null", &null_run,
660 "null run - dont start any counters"), 1045 "null run - dont start any counters"),
1046 OPT_INCR('d', "detailed", &detailed_run,
1047 "detailed run - start a lot of events"),
1048 OPT_BOOLEAN('S', "sync", &sync_run,
1049 "call sync() before starting a run"),
661 OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL, 1050 OPT_CALLBACK_NOOPT('B', "big-num", NULL, NULL,
662 "print large numbers with thousands\' separators", 1051 "print large numbers with thousands\' separators",
663 stat__set_big_num), 1052 stat__set_big_num),
@@ -673,6 +1062,70 @@ static const struct option options[] = {
673 OPT_END() 1062 OPT_END()
674}; 1063};
675 1064
1065/*
1066 * Add default attributes, if there were no attributes specified or
1067 * if -d/--detailed, -d -d or -d -d -d is used:
1068 */
1069static int add_default_attributes(void)
1070{
1071 struct perf_evsel *pos;
1072 size_t attr_nr = 0;
1073 size_t c;
1074
1075 /* Set attrs if no event is selected and !null_run: */
1076 if (null_run)
1077 return 0;
1078
1079 if (!evsel_list->nr_entries) {
1080 for (c = 0; c < ARRAY_SIZE(default_attrs); c++) {
1081 pos = perf_evsel__new(default_attrs + c, c + attr_nr);
1082 if (pos == NULL)
1083 return -1;
1084 perf_evlist__add(evsel_list, pos);
1085 }
1086 attr_nr += c;
1087 }
1088
1089 /* Detailed events get appended to the event list: */
1090
1091 if (detailed_run < 1)
1092 return 0;
1093
1094 /* Append detailed run extra attributes: */
1095 for (c = 0; c < ARRAY_SIZE(detailed_attrs); c++) {
1096 pos = perf_evsel__new(detailed_attrs + c, c + attr_nr);
1097 if (pos == NULL)
1098 return -1;
1099 perf_evlist__add(evsel_list, pos);
1100 }
1101 attr_nr += c;
1102
1103 if (detailed_run < 2)
1104 return 0;
1105
1106 /* Append very detailed run extra attributes: */
1107 for (c = 0; c < ARRAY_SIZE(very_detailed_attrs); c++) {
1108 pos = perf_evsel__new(very_detailed_attrs + c, c + attr_nr);
1109 if (pos == NULL)
1110 return -1;
1111 perf_evlist__add(evsel_list, pos);
1112 }
1113
1114 if (detailed_run < 3)
1115 return 0;
1116
1117 /* Append very, very detailed run extra attributes: */
1118 for (c = 0; c < ARRAY_SIZE(very_very_detailed_attrs); c++) {
1119 pos = perf_evsel__new(very_very_detailed_attrs + c, c + attr_nr);
1120 if (pos == NULL)
1121 return -1;
1122 perf_evlist__add(evsel_list, pos);
1123 }
1124
1125
1126 return 0;
1127}
1128
676int cmd_stat(int argc, const char **argv, const char *prefix __used) 1129int cmd_stat(int argc, const char **argv, const char *prefix __used)
677{ 1130{
678 struct perf_evsel *pos; 1131 struct perf_evsel *pos;
@@ -718,17 +1171,8 @@ int cmd_stat(int argc, const char **argv, const char *prefix __used)
718 usage_with_options(stat_usage, options); 1171 usage_with_options(stat_usage, options);
719 } 1172 }
720 1173
721 /* Set attrs and nr_counters if no event is selected and !null_run */ 1174 if (add_default_attributes())
722 if (!null_run && !evsel_list->nr_entries) { 1175 goto out;
723 size_t c;
724
725 for (c = 0; c < ARRAY_SIZE(default_attrs); ++c) {
726 pos = perf_evsel__new(&default_attrs[c], c);
727 if (pos == NULL)
728 goto out;
729 perf_evlist__add(evsel_list, pos);
730 }
731 }
732 1176
733 if (target_pid != -1) 1177 if (target_pid != -1)
734 target_tid = target_pid; 1178 target_tid = target_pid;
@@ -772,6 +1216,10 @@ int cmd_stat(int argc, const char **argv, const char *prefix __used)
772 for (run_idx = 0; run_idx < run_count; run_idx++) { 1216 for (run_idx = 0; run_idx < run_count; run_idx++) {
773 if (run_count != 1 && verbose) 1217 if (run_count != 1 && verbose)
774 fprintf(stderr, "[ perf stat: executing run #%d ... ]\n", run_idx + 1); 1218 fprintf(stderr, "[ perf stat: executing run #%d ... ]\n", run_idx + 1);
1219
1220 if (sync_run)
1221 sync();
1222
775 status = run_perf_stat(argc, argv); 1223 status = run_perf_stat(argc, argv);
776 } 1224 }
777 1225
generated by cgit v1.2.2 (git 2.25.0) at 2025-04-25 21:28:27 -0400