perf report: Show branch flags/cycles in --branch-history callgraph view
v3 cover perf report: Show branch flags/cycles in --branch-history callgraph view
v3: 1. Display the count for tsx abort, remove the abort percentage.
2. Since the branch history code has a loop detection that removes small loops in util/machine.c:remove_loops(). It would be nice to note how many loops were removed. So it adds the note on some callchain entries.
v2: Just a rebase to Arnaldo's perf/core branch, no functional changes.
Initial post
perf record -g -b ...
perf report --branch-history
Currently it only shows the branches from the LBR in the callgraph view. It would be useful to annotate branch predictions and TSX aborts and also timed LBR cycles also in the callgraph view.
This would allow a quick overview where branch predictions are and how costly basic blocks are.
For example:
# Overhead Source:Line Symbol Shared Object Predicted Abort Cycles
# ........ ........................... ........................... ................. ......... ..... ......
#
38.25% div.c:45 [.] main div 97.6% 0 3
|
---main div.c:42 (cycles:2)
compute_flag div.c:28 (cycles:2)
compute_flag div.c:27 (cycles:1)
rand rand.c:28 (cycles:1)
rand rand.c:28 (cycles:1)
__random random.c:298 (cycles:1)
__random random.c:297 (cycles:1)
__random random.c:295 (cycles:1)
__random random.c:295 (cycles:1)
__random random.c:295 (cycles:1)
__random random.c:295 (cycles:9)
|
|--36.73%--__random_r random_r.c:392 (cycles:9)
| __random_r random_r.c:357 (cycles:1)
| __random random.c:293 (cycles:1)
| __random random.c:293 (cycles:1)
| __random random.c:291 (cycles:1)
| __random random.c:291 (cycles:1)
| __random random.c:291 (cycles:1)
| __random random.c:288 (cycles:1)
| rand rand.c:27 (cycles:1)
| rand rand.c:26 (cycles:1)
| rand@plt +4194304 (cycles:1)
| rand@plt +4194304 (cycles:1)
| compute_flag div.c:25 (cycles:1)
| compute_flag div.c:22 (cycles:1)
| main div.c:40 (cycles:1)
| main div.c:40 (cycles:16)
| main div.c:39 (cycles:16)
| |
| |--29.93%--main div.c:39 (predicted:50.6%, cycles:1, removed loops:12954)
| | main div.c:44 (predicted:50.6%, cycles:1)
| | |
| | --22.69%--main div.c:42 (cycles:2, removed loops:21736)
| | compute_flag div.c:28 (cycles:2)
| | |
| | --10.52%--compute_flag div.c:27 (cycles:1)
| | rand rand.c:28 (cycles:1)
| | rand rand.c:28 (cycles:1)
| | __random random.c:298 (cycles:1)
| | __random random.c:297 (cycles:1)
| | __random random.c:295 (cycles:1)
| | __random random.c:295 (cycles:1)
| | __random random.c:295 (cycles:1)Jin Yao (6):
perf report: Add branch flag to callchain cursor node perf report: Caculate and return the branch flag counting perf report: Create a symbol_conf flag for showing branch flag
counting
perf report: Show branch info in callchain entry for stdio mode perf report: Show branch info in callchain entry for browser mode perf report: Display columns Predicted/Abort/Cycles in
--branch-history
tools/perf/Documentation/perf-report.txt | 12 ++ tools/perf/builtin-report.c | 9 +- tools/perf/ui/browsers/hists.c | 20 +++- tools/perf/ui/stdio/hist.c | 35 +++++- tools/perf/util/callchain.c | 197 ++++++++++++++++++++++++++++++- tools/perf/util/callchain.h | 20 +++- tools/perf/util/hist.c | 3 + tools/perf/util/hist.h | 3 + tools/perf/util/machine.c | 69 ++++++++--- tools/perf/util/sort.c | 113 +++++++++++++++++- tools/perf/util/sort.h | 3 + tools/perf/util/symbol.h | 1 + 12 files changed, 457 insertions(+), 28 deletions(-)
Signature · 1 line
2.7.4
PATCH v3 1/6 perf report: Add branch flag to callchain cursor node
Since the branch ip has been added to call stack for easier browsing, this patch adds more branch information. For example, add a flag to indicate if this ip is a branch, and also add with the branch flag.
Then we can know if the cursor node represents a branch and know what the branch flag it has.
The branch history code has a loop detection pass that removes small loops. It would be nice for knowing how many loops were removed. This patch records the number of removed loops in a callchain cursor node.
--- tools/perf/util/callchain.c | 13 +++++++-- tools/perf/util/callchain.h | 7 ++++- tools/perf/util/machine.c | 69 ++++++++++++++++++++++++++++++++++----------- 3 files changed, 70 insertions(+), 19 deletions(-)
tools/perf/util/callchain.c +11 -2
@@ -730,7 +730,8 @@ merge_chain_branch(struct callchain_cursor *cursor,list_for_each_entry_safe (list, next_list, &src->val, list) {callchain_cursor_append (cursor, list->ip, - list->ms.map, list->ms.sym); + list->ms.map, list->ms.sym, + false,NULL ,0 );list_del (&list->list);free (list); }
@@ -767,7 +768,9 @@ int callchain_merge(struct callchain_cursor *cursor, }int callchain_cursor_append (struct callchain_cursor *cursor, -u64 ip,struct map *map,struct symbol *sym) +u64 ip,struct map *map,struct symbol *sym, +bool branch,struct branch_flags *flags, +int removed_loop) {struct callchain_cursor_node *node = *cursor->last;
@@ -782,6 +785,12 @@ int callchain_cursor_append(struct callchain_cursor *cursor, node->ip = ip; node->map = map; node->sym = sym; + node->branch = branch; + node->removed_loop = removed_loop; + +if (flags) +memcpy (&node->branch_flags, flags, +sizeof (struct branch_flags )); cursor->nr++;
tools/perf/util/callchain.h +6 -1
@@ -129,6 +129,9 @@ struct callchain_cursor_node {u64 ip;struct map *map;struct symbol *sym; +bool branch; +struct branch_flags branch_flags; +int removed_loop;struct callchain_cursor_node *next; };
@@ -183,7 +186,9 @@ static inline void callchain_cursor_reset(struct callchain_cursor *cursor) }int callchain_cursor_append (struct callchain_cursor *cursor,u64 ip, -struct map *map,struct symbol *sym); +struct map *map,struct symbol *sym, +bool branch,struct branch_flags *flags, +int removed_loop);/* Close a cursor writing session. Initialize for the reader */ static inline void callchain_cursor_commit (struct callchain_cursor *cursor)
tools/perf/util/machine.c +53 -16
149 lines not on this page · raw message
Signature · 1 line
2.7.4
PATCH v3 2/6 perf report: Caculate and return the branch flag counting
In reply to Jin YaoCreate some branch counters in per callchain list entry. Each counter is for a branch flag. For example, predicted_count counts all the *predicted* branches. The counters get updated by processing the callchain cursor nodes.
It also provides functions to retrieve or print the values of counters in callchain list.
Besides the counting for branch flags, it also counts and returns the number of removed loops.
--- tools/perf/util/callchain.c | 184 +++++++++++++++++++++++++++++++++++++++++++- tools/perf/util/callchain.h | 13 ++++ 2 files changed, 196 insertions(+), 1 deletion(-)
tools/perf/util/callchain.c +183 -1
@@ -440,6 +440,20 @@ fill_node(struct callchain_node *node, struct callchain_cursor *cursor) call->ip = cursor_node->ip; call->ms.sym = cursor_node->sym; call->ms.map = cursor_node->map; + +if (cursor_node->branch) { + call->branch_count =1 ; + +if (cursor_node->branch_flags.predicted) + call->predicted_count =1 ; + +if (cursor_node->branch_flags.abort) + call->abort_count =1 ; + + call->cycles_count = cursor_node->branch_flags.cycles; + call->removed_loop = cursor_node->removed_loop; + } +list_add_tail (&call->list, &node->val);callchain_cursor_advance (cursor);
@@ -499,8 +513,22 @@ static enum match_result match_chain(struct callchain_cursor_node *node, right = node->ip; } -if (left == right) +if (left == right) { +if (node->branch) { + cnode->branch_count++; + +if (node->branch_flags.predicted) + cnode->predicted_count++; + +if (node->branch_flags.abort) + cnode->abort_count++; + + cnode->cycles_count += node->branch_flags.cycles; + cnode->removed_loop += node->removed_loop; + } +return MATCH_EQ ; + }return left > right ?MATCH_GT :MATCH_LT ; }
@@ -948,6 +976,160 @@ int callchain_node__fprintf_value(struct callchain_node *node,return 0 ; } +static void callchain_counts_value (struct callchain_node *node, +u64 *branch_count,u64 *predicted_count, +u64 *abort_count,u64 *cycles_count) +{ +struct callchain_list *clist; + +list_for_each_entry (clist, &node->val, list) { +if (branch_count) + *branch_count += clist->branch_count; + +if (predicted_count) + *predicted_count += clist->predicted_count; + +if (abort_count) + *abort_count += clist->abort_count; + +if (cycles_count) + *cycles_count += clist->cycles_count; + } +} + +static int callchain_node_branch_counts_cumul (struct callchain_node *node, +u64 *branch_count, +u64 *predicted_count, +u64 *abort_count, +u64 *cycles_count) +{ +struct callchain_node *child; +struct rb_node *n; + + n =rb_first (&node->rb_root_in); +while (n) { + child = rb_entry(n, struct callchain_node, rb_node_in); + n =rb_next (n); + +callchain_node_branch_counts_cumul (child, branch_count, + predicted_count, + abort_count, + cycles_count); + +callchain_counts_value (child, branch_count, + predicted_count, abort_count, + cycles_count); + } + +return 0 ; +} + +int callchain_branch_counts (struct callchain_root *root, +u64 *branch_count,u64 *predicted_count, +u64 *abort_count,u64 *cycles_count) +{ +if (branch_count) + *branch_count =0 ; + +if (predicted_count) + *predicted_count =0 ; + +if (abort_count) + *abort_count =0 ; + +if (cycles_count) + *cycles_count =0 ; + +return callchain_node_branch_counts_cumul (&root->node, + branch_count, + predicted_count, + abort_count, + cycles_count); +} + +static int callchain_counts_printf (FILE *fp,char *bf,int bfsize, +u64 branch_count,u64 predicted_count, +u64 abort_count,u64 cycles_count, +int removed_loop) +{ +double predicted_percent =0.0 ; +const char *null_str ="" ; +char removed_str[32 ]; +char *str; +u64 cycles =0 ; + +if (branch_count ==0 ) { +if (fp) +return fprintf (fp," (calltrace)" ); + +return scnprintf (bf, bfsize," (calltrace)" ); + } + +if (removed_loop) { +scnprintf (removed_str,sizeof (removed_str), +", removed loops:%d" , removed_loop); + str = removed_str; + }else + str = (char *)null_str; + + predicted_percent = predicted_count *100.0 / branch_count; + cycles = cycles_count / branch_count; + +if ((predicted_percent >=100.0 ) && (abort_count ==0 )) { +if (fp) +return fprintf (fp," (cycles:%" PRId64"%s)" , + cycles, str); + +return scnprintf (bf, bfsize," (cycles:%" PRId64"%s)" , + cycles, str); + } + +if ((predicted_percent <100.0 ) && (abort_count ==0 )) { +if (fp) +return fprintf (fp, +" (predicted:%.1f%%, cycles:%" PRId64"%s)" , + predicted_percent, cycles, str); + +return scnprintf (bf, bfsize, +" (predicted:%.1f%%, cycles:%" PRId64"%s)" , + predicted_percent, cycles, str); + } + +if (fp) +return fprintf (fp, +" (predicted:%.1f%%, abort:%" PRId64", cycles:%" PRId64"%s)" , + predicted_percent, abort_count, cycles, str); + +return scnprintf (bf, bfsize, +" (predicted:%.1f%%, abort:%" PRId64", cycles:%" PRId64"%s)" , + predicted_percent, abort_count, cycles, str); +} + +int callchain_list_counts__printf_value (struct callchain_node *node, +struct callchain_list *clist, +FILE *fp,char *bf,int bfsize) +{ +u64 branch_count, predicted_count; +u64 abort_count, cycles_count; +int removed_loop =0 ; + + branch_count = clist->branch_count; + predicted_count = clist->predicted_count; + abort_count = clist->abort_count; + cycles_count = clist->cycles_count; + +if (node) { +struct callchain_list *call; + +list_for_each_entry (call, &node->val, list) + removed_loop += call->removed_loop; + } + +return callchain_counts_printf (fp, bf, bfsize, branch_count, + predicted_count, abort_count, + cycles_count, removed_loop); +} +static void free_callchain_node (struct callchain_node *node) {struct callchain_list *list, *tmp;
tools/perf/util/callchain.h +13 -0
@@ -115,6 +115,11 @@ struct callchain_list {bool unfolded;bool has_children; }; +u64 branch_count; +u64 predicted_count; +u64 abort_count; +u64 cycles_count; +int removed_loop;char *srcline;struct list_head list; };
@@ -266,8 +271,16 @@ char *callchain_node__scnprintf_value(struct callchain_node *node,int callchain_node__fprintf_value (struct callchain_node *node,FILE *fp,u64 total); +int callchain_list_counts__printf_value (struct callchain_node *node, +struct callchain_list *clist, +FILE *fp,char *bf,int bfsize); +void free_callchain (struct callchain_root *root);void decay_callchain (struct callchain_root *root);int callchain_node__make_parent_list (struct callchain_node *node); +int callchain_branch_counts (struct callchain_root *root, +u64 *branch_count,u64 *predicted_count, +u64 *abort_count,u64 *cycles_count); +#endif /* __PERF_CALLCHAIN_H */
Signature · 1 line
2.7.4
PATCH v3 3/6 perf report: Create a symbol_conf flag for showing branch flag counting
In reply to Jin YaoCreate a new flag show_branchflag_count in symbol_conf. The flag is used to control if showing the branch flag counting information. The flag depends on if the perf.data has branch data and if user chooses the "branch-history" option in perf report command line.
--- tools/perf/builtin-report.c | 3 +++ tools/perf/util/symbol.h | 1 + 2 files changed, 4 insertions(+)
tools/perf/builtin-report.c +3 -0
@@ -911,6 +911,9 @@ int cmd_report(int argc, const char **argv, const char *prefix __maybe_unused)if (itrace_synth_opts.last_branch) has_br_stack = true; +if (has_br_stack && branch_call_mode) + symbol_conf.show_branchflag_count = true; +/* * Branch mode is a tristate: * -1 means default, so decide based on the file having branch data.
tools/perf/util/symbol.h +1 -0
@@ -100,6 +100,7 @@ struct symbol_conf { show_total_period, use_callchain, cumulate_callchain, + show_branchflag_count, exclude_other, show_cpu_utilization, initialized,
Signature · 1 line
2.7.4
PATCH v3 4/6 perf report: Show branch info in callchain entry for stdio mode
In reply to Jin YaoIf the branch is 100% predicated then the "predicated" is hide. Similarly, if there is no branch tsx abort, the "abort" is hide. There is only cycles shown (cycle is supported on skylake platform, older platform would be 0).
If no removed loops, the "removed loops" is hide.
One example is:
|--29.93%--main div.c:39 (predicted:50.6%, cycles:1, removed loops:12954) | main div.c:44 (predicted:50.6%, cycles:1) | | | --22.69%--main div.c:42 (cycles:2, removed loops:21736) | compute_flag div.c:28 (cycles:2) | | | --10.52%--compute_flag div.c:27 (cycles:1) | rand rand.c:28 (cycles:1) | rand rand.c:28 (cycles:1) | __random random.c:298 (cycles:1) | __random random.c:297 (cycles:1) | __random random.c:295 (cycles:1) | __random random.c:295 (cycles:1) | __random random.c:295 (cycles:1) | __random random.c:295 (cycles:6)
--- tools/perf/ui/stdio/hist.c | 35 +++++++++++++++++++++++++++++++---- 1 file changed, 31 insertions(+), 4 deletions(-)
tools/perf/ui/stdio/hist.c +31 -4
@@ -41,7 +41,9 @@ static size_t ipchain__fprintf_graph(FILE *fp, struct callchain_node *node, {int i;size_t ret =0 ; -char bf[1024 ]; +char bf[1024 ], *alloc_str =NULL ; +char buf[64 ]; +const char *str; ret +=callchain__fprintf_left_margin (fp, left_margin);for (i =0 ; i < depth; i++) {
@@ -56,8 +58,26 @@ static size_t ipchain__fprintf_graph(FILE *fp, struct callchain_node *node, }else ret +=fprintf (fp,"%s" ," " ); } -fputs (callchain_list__sym_name (chain, bf,sizeof (bf), false), fp); + + str =callchain_list__sym_name (chain, bf,sizeof (bf), false); + +if (symbol_conf.show_branchflag_count) { +if (!period) +callchain_list_counts__printf_value (node, chain,NULL , + buf,sizeof (buf)); +else +callchain_list_counts__printf_value (NULL , chain,NULL , + buf,sizeof (buf)); + +if (asprintf (&alloc_str,"%s%s" , str, buf) <0 ) + str ="Not enough memory!" ; +else + str = alloc_str; + } + +fputs (str, fp);fputc ('\n' , fp); +free (alloc_str);return ret; }
@@ -219,8 +239,15 @@ static size_t callchain__fprintf_graph(FILE *fp, struct rb_root *root, }else ret +=callchain__fprintf_left_margin (fp, left_margin); - ret +=fprintf (fp,"%s\n" ,callchain_list__sym_name (chain, bf,sizeof (bf), - false)); + ret +=fprintf (fp,"%s" , +callchain_list__sym_name (chain, bf, +sizeof (bf), + false)); + +if (symbol_conf.show_branchflag_count) + ret +=callchain_list_counts__printf_value ( +NULL , chain, fp,NULL ,0 ); + ret +=fprintf (fp,"\n" );if (++entries_printed == callchain_param.print_limit)break ;
Signature · 1 line
2.7.4
PATCH v3 5/6 perf report: Show branch info in callchain entry for browser mode
In reply to Jin YaoIf the branch is 100% predicated then the "predicated" is hide. Similarly, if there is no branch tsx abort, the "abort" is hide. There is only cycles shown (cycle is supported on skylake platform, older platform would be 0).
If no removed loops, the "removed loops" is hide.
--- tools/perf/ui/browsers/hists.c | 20 ++++++++++++++++++-- 1 file changed, 18 insertions(+), 2 deletions(-)
tools/perf/ui/browsers/hists.c +18 -2
@@ -738,6 +738,7 @@ static int hist_browser__show_callchain_list(struct hist_browser *browser, struct callchain_print_arg *arg) {char bf[1024 ], *alloc_str; +char buf[64 ], *alloc_str2;const char *str;if (arg->row_offset !=0 ) {
@@ -746,12 +747,26 @@ static int hist_browser__show_callchain_list(struct hist_browser *browser, } alloc_str =NULL ; + alloc_str2 =NULL ; + str =callchain_list__sym_name (chain, bf,sizeof (bf), browser->show_dso); -if (need_percent) { -char buf[64 ]; +if (symbol_conf.show_branchflag_count) { +if (need_percent) +callchain_list_counts__printf_value (node, chain,NULL , + buf,sizeof (buf)); +else +callchain_list_counts__printf_value (NULL , chain,NULL , + buf,sizeof (buf)); + +if (asprintf (&alloc_str2,"%s%s" , str, buf) <0 ) + str ="Not enough memory!" ; +else + str = alloc_str2; + } +if (need_percent) {callchain_node__scnprintf_value (node, buf,sizeof (buf), total);
@@ -764,6 +779,7 @@ static int hist_browser__show_callchain_list(struct hist_browser *browser,free (alloc_str); +free (alloc_str2);return 1 ; }
Signature · 1 line
2.7.4
PATCH v3 6/6 perf report: Display columns Predicted/Abort/Cycles in --branch-history
In reply to Jin YaoUse current sort mechanism but the real .se_cmp() just returns 0 so that new columns "Predicted", "Abort" and "Cycles" are created in display but actually these keys are not the sort keys.
For example:
Overhead Source:Line Symbol Shared Object Predicted Abort Cycles ........ ............ ........ ............. ......... ..... ......
38.25% div.c:45 [.] main div 97.6% 0 3
--- tools/perf/Documentation/perf-report.txt | 12 ++++ tools/perf/builtin-report.c | 6 +- tools/perf/util/hist.c | 3 + tools/perf/util/hist.h | 3 + tools/perf/util/sort.c | 113 ++++++++++++++++++++++++++++++- tools/perf/util/sort.h | 3 + 6 files changed, 138 insertions(+), 2 deletions(-)
tools/perf/Documentation/perf-report.txt +12 -0
@@ -335,6 +335,18 @@ OPTIONS --branch-history:: Add the addresses of sampled taken branches to the callstack. This allows to examine the path the program took to each sample. + + Also show with some branch flags that can be: + - Predicted: display the average percentage of predicated branches. + (predicated number / total number) + - Abort: display the number of tsx aborted branches. + - Cycles: cycles in basic block. + + - removed loops: the number of removed loops. + It also shows the number of removed loops in callchain list. + The branch history code has a loop detection that removes small loops. + It would be nice for knowing how many loops were removed. + The data collection must have used -b (or -j) and -g. --objdump=<path>::
tools/perf/builtin-report.c +5 -1
@@ -670,6 +670,10 @@ const char report_callchain_help[] = "Display call graph (stack chain/backtrace)CALLCHAIN_REPORT_HELP "\n\t\t\t\tDefault: " CALLCHAIN_DEFAULT_OPT ; +#define CALLCHAIN_BRANCH_SORT_ORDER \ + "srcline,symbol,dso,callchain_branch_predicted," \ + "callchain_branch_abort,callchain_branch_cycles" +int cmd_report (int argc,const char **argv,const char *prefix__maybe_unused ) {struct perf_session *session;
@@ -930,7 +934,7 @@ int cmd_report(int argc, const char **argv, const char *prefix __maybe_unused) symbol_conf.use_callchain = true;callchain_register_param (&callchain_param);if (sort_order ==NULL ) - sort_order ="srcline,symbol,dso" ; + sort_order =CALLCHAIN_BRANCH_SORT_ORDER ; }if (report.mem_mode) {
tools/perf/util/hist.c +3 -0
@@ -176,6 +176,9 @@ void hists__calc_col_len(struct hists *hists, struct hist_entry *h)hists__new_col_len (hists,HISTC_MEM_LVL ,21 +3 );hists__new_col_len (hists,HISTC_LOCAL_WEIGHT ,12 );hists__new_col_len (hists,HISTC_GLOBAL_WEIGHT ,12 ); +hists__new_col_len (hists,HISTC_CALLCHAIN_BRANCH_PREDICTED ,9 ); +hists__new_col_len (hists,HISTC_CALLCHAIN_BRANCH_ABORT ,5 ); +hists__new_col_len (hists,HISTC_CALLCHAIN_BRANCH_CYCLES ,6 );if (h->srcline) { len =MAX (strlen (h->srcline),strlen (sort_srcline.se_header));
tools/perf/util/hist.h +3 -0
@@ -56,6 +56,9 @@ enum hist_column {HISTC_SRCLINE_FROM ,HISTC_SRCLINE_TO ,HISTC_TRACE , +HISTC_CALLCHAIN_BRANCH_PREDICTED , +HISTC_CALLCHAIN_BRANCH_ABORT , +HISTC_CALLCHAIN_BRANCH_CYCLES , HISTC_NR_COLS, /* Last entry */ };
tools/perf/util/sort.c +112 -1
131 lines not on this page · raw message
tools/perf/util/sort.h +3 -0
@@ -224,6 +224,9 @@ enum sort_type {SORT_CYCLES ,SORT_SRCLINE_FROM ,SORT_SRCLINE_TO , +SORT_CALLCHAIN_BRANCH_PREDICTED , +SORT_CALLCHAIN_BRANCH_ABORT , +SORT_CALLCHAIN_BRANCH_CYCLES ,/* memory mode specific sort keys */ __SORT_MEMORY_MODE,
Signature · 1 line
2.7.4
Quoted 2 lines
| |
| |--29.93%--main div.c:39 (predicted:50.6%, cycles:1, removed loops:12954)Removed loops should be divided by the number of samples to normalize it.
Also I would call it "iterations"
-Andi
If I change it to, for example: removed iterations: 20%
Is that OK?
Thanks
Jin Yao
Quoted 8 lines · Andi Kleen
Quoted 2 lines
| | | |--29.93%--main div.c:39 (predicted:50.6%, cycles:1, removed loops:12954)Removed loops should be divided by the number of samples to normalize it.
Also I would call it "iterations"
-Andi
Quoted 3 lines · Yao
If I change it to, for example: removed iterations: 20%
Is that OK?
No it should be average iterations, not percentage.
Also "iterations", not "removed iterations"
-Andi