perf diff: diff cycles at basic block level
v2 cover perf diff: diff cycles at basic block level
In some cases small changes in hot loops can show big differences. But it's difficult to identify these differences.
perf diff currently can only diff symbols (functions). We can also expand it to diff cycles of individual programs blocks as reported by timed LBR. This would allow to identify changes in specific code accurately.
With this patch set, for example,
# perf record -b ./div # perf record -b ./div # perf diff -s cycles
# Event 'cycles'
#
# Baseline Block cycles diff [start:end] Shared Object Symbol
# ........ .................................... ................ ....................................
#
49.03% -9 [ 4ef: 520] div [.] main
49.03% 0 [ 4e8: 4ea] div [.] main
49.03% 0 [ 4ef: 500] div [.] main
49.03% 0 [ 4ef: 51c] div [.] main
49.03% 0 [ 4ef: 535] div [.] main
18.82% 0 [ 3ac40: 3ac4d] libc-2.23.so [.] __random_r
18.82% 0 [ 3ac40: 3ac5c] libc-2.23.so [.] __random_r
18.82% 0 [ 3ac40: 3ac76] libc-2.23.so [.] __random_r
18.82% 0 [ 3ac40: 3ac88] libc-2.23.so [.] __random_r
18.82% 0 [ 3ac90: 3ac9c] libc-2.23.so [.] __random_r
16.29% -8 [ 3aac0: 3aac0] libc-2.23.so [.] __random
16.29% 0 [ 3aac0: 3aad2] libc-2.23.so [.] __random
16.29% 0 [ 3aae0: 3aae7] libc-2.23.so [.] __random
16.29% 0 [ 3ab03: 3ab0f] libc-2.23.so [.] __random
16.29% 0 [ 3ab14: 3ab1b] libc-2.23.so [.] __random
16.29% 0 [ 3ab28: 3ab2e] libc-2.23.so [.] __random
16.29% 0 [ 3ab4a: 3ab53] libc-2.23.so [.] __random
8.11% 0 [ 640: 644] div [.] compute_flag
8.11% 0 [ 649: 659] div [.] compute_flag
5.46% 0 [ 3af60: 3af60] libc-2.23.so [.] rand
5.46% 0 [ 3af60: 3af64] libc-2.23.so [.] rand
2.25% 0 [ 490: 490] div [.] rand@plt
0.01% 26 [ c00a27: c00a27] [kernel.vmlinux] [k] native_irq_return_iret
0.00% -157 [ 2bf9f2: 2bfa63] [kernel.vmlinux] [k] update_blocked_averages
0.00% -56 [ 2bf980: 2bf9d3] [kernel.vmlinux] [k] update_blocked_averages
0.00% 48 [ 2bf934: 2bf942] [kernel.vmlinux] [k] update_blocked_averages
0.00% 3 [ 2bfb38: 2bfb67] [kernel.vmlinux] [k] update_blocked_averagesThe 'cycles' is a new perf-diff computation selection, which enables the displaying of cycles difference of same program basic block amongst two perf.data. The program basic block is the code block between two branches in a function.
v2:
---
Keep standard perf diff format.
Following is the v1 output.
# perf diff --basic-block
# Cycles diff Basic block (start:end)
# ........... .......................
#
-208 hrtimer_interrupt (30b9e0:30ba42)
-157 update_blocked_averages (2bf9f2:2bfa63)
-126 interrupt_entry (c00880:c0093a)
-86 hrtimer_interrupt (30bb29:30bb32)
-74 hrtimer_interrupt (30ba65:30bac4)
-56 update_blocked_averages (2bf980:2bf9d3)
48 update_blocked_averages (2bf934:2bf942)
-35 native_write_msr (267900:26790b)
26 native_irq_return_iret (c00a27:c00a27)
22 rcu_check_callbacks (2febb6:2febdc)
-21 __hrtimer_run_queues (30b220:30b2a3)
19 pvclock_gtod_notify (14ba0:14c1b)
-18 task_tick_fair (2c5d29:2c5d41)Jin Yao (7):
perf util: Create block_info structure
perf util: Add block_info in hist_entry
perf diff: Check if all data files with branch stacks perf diff: Use hists to manage basic blocks per symbol perf diff: Link same basic blocks among different data files perf diff: Print the basic block cycles diff
perf diff: Documentation -c cycles option
tools/perf/Documentation/perf-diff.txt | 14 +- tools/perf/builtin-diff.c | 373 ++++++++++++++++++++++++++++++++- tools/perf/ui/stdio/hist.c | 26 +++ tools/perf/util/hist.c | 42 +++- tools/perf/util/hist.h | 9 + tools/perf/util/sort.h | 8 + tools/perf/util/symbol.c | 22 ++ tools/perf/util/symbol.h | 23 ++ 8 files changed, 509 insertions(+), 8 deletions(-)
Signature · 1 line
2.7.4
PATCH v2 1/7 perf util: Create block_info structure
perf diff currently can only diff symbols(functions). We should expand it to diff cycles of individual programs blocks as reported by timed LBR. This would allow to identify changes in specific code accurately.
We need a new structure to maintain the basic block information, such as, symbol(function), start/end address of this block, cycles. This patch creates this structure and with some ops.
--- tools/perf/util/symbol.c | 22 ++++++++++++++++++++++ tools/perf/util/symbol.h | 23 +++++++++++++++++++++++ 2 files changed, 45 insertions(+)
tools/perf/util/symbol.c +22 -0
@@ -2351,3 +2351,25 @@ struct mem_info *mem_info__new(void)refcount_set (&mi->refcnt,1 );return mi; } + +struct block_info *block_info__get (struct block_info *bi) +{ +if (bi) +refcount_inc (&bi->refcnt); +return bi; +} + +void block_info__put (struct block_info *bi) +{ +if (bi &&refcount_dec_and_test (&bi->refcnt)) +free (bi); +} + +struct block_info *block_info__new (void ) +{ +struct block_info *bi =zalloc (sizeof (*bi)); + +if (bi) +refcount_set (&bi->refcnt,1 ); +return bi; +}
tools/perf/util/symbol.h +23 -0
@@ -131,6 +131,17 @@ struct mem_info {refcount_t refcnt; }; +struct block_info { +struct symbol *sym; +u64 start; +u64 end; +u64 cycles; +u64 cycles_aggr; +int num; +int num_aggr; +refcount_t refcnt; +}; +struct addr_location {struct machine *machine;struct thread *thread;
@@ -332,4 +343,16 @@ static inline void __mem_info__zput(struct mem_info **mi)#define mem_info__zput (mi) __mem_info__zput(&mi) +struct block_info *block_info__new (void ); +struct block_info *block_info__get (struct block_info *bi); +void block_info__put (struct block_info *bi); + +static inline void __block_info__zput (struct block_info **bi) +{ +block_info__put (*bi); + *bi =NULL ; +} + +#define block_info__zput (bi) __block_info__zput(&bi) +#endif /* __PERF_SYMBOL */
Signature · 1 line
2.7.4
PATCH v2 2/7 perf util: Add block_info in hist_entry
In reply to Jin YaoThe block_info contains the program basic block information, i.e, contains the start address and the end address of this basic block and how much cycles it takes. We need to compare, sort and even print out the basic block by some orders, i.e. sort by cycles.
For this purpose, we add block_info field to hist_entry. In order not to impact current interface, we creates a new function hists__add_entry_block.
--- tools/perf/util/hist.c | 22 ++++++++++++++++++++-- tools/perf/util/hist.h | 6 ++++++ tools/perf/util/sort.h | 1 + 3 files changed, 27 insertions(+), 2 deletions(-)
tools/perf/util/hist.c +20 -2
@@ -574,6 +574,8 @@ static struct hist_entry *hists__findnew_entry(struct hists *hists,*/ mem_info__zput (entry->mem_info); +block_info__zput (entry->block_info); +/* If the map of an existing hist_entry has * become out-of-date due to an exec() or * similar, update it. Otherwise we will
@@ -645,6 +647,7 @@ __hists__add_entry(struct hists *hists,struct symbol *sym_parent, struct branch_info *bi, struct mem_info *mi, + struct block_info *block_info, struct perf_sample *sample, bool sample_self, struct hist_entry_ops *ops)
@@ -677,6 +680,7 @@ __hists__add_entry(struct hists *hists, .hists = hists, .branch_info = bi, .mem_info = mi, + .block_info = block_info, .transaction = sample->transaction, .raw_data = sample->raw_data, .raw_size = sample->raw_size,
@@ -699,7 +703,7 @@ struct hist_entry *hists__add_entry(struct hists *hists,struct perf_sample *sample, bool sample_self) { - return __hists__add_entry(hists, al, sym_parent, bi, mi, + return __hists__add_entry(hists, al, sym_parent, bi, mi, NULL,sample ,sample_self ,NULL ); }
@@ -712,10 +716,24 @@ struct hist_entry *hists__add_entry_ops(struct hists *hists,struct perf_sample *sample, bool sample_self) { - return __hists__add_entry(hists, al, sym_parent, bi, mi, + return __hists__add_entry(hists, al, sym_parent, bi, mi, NULL,sample ,sample_self ,ops ); } +struct hist_entry *hists__add_entry_block (struct hists *hists, +struct hist_entry_ops *ops, +struct addr_location *al, +struct block_info *block_info) +{ +struct hist_entry entry = { + .ops = ops, + .block_info = block_info, + .hists = hists, + }, *he =hists__findnew_entry (hists, &entry, al, false); + +return he; +} +static int iter_next_nop_entry (struct hist_entry_iter *iter__maybe_unused ,struct addr_location *al__maybe_unused )
tools/perf/util/hist.h +6 -0
@@ -16,6 +16,7 @@ struct addr_location;struct map_symbol ;struct mem_info ;struct branch_info ; +struct block_info ;struct symbol ;enum hist_filter {
@@ -149,6 +150,11 @@ struct hist_entry *hists__add_entry_ops(struct hists *hists,struct perf_sample *sample, bool sample_self); +struct hist_entry *hists__add_entry_block (struct hists *hists, +struct hist_entry_ops *ops, +struct addr_location *al, +struct block_info *bi); +int hist_entry_iter__add (struct hist_entry_iter *iter,struct addr_location *al,int max_stack_depth,void *arg);
tools/perf/util/sort.h +1 -0
@@ -144,6 +144,7 @@ struct hist_entry {long time;struct hists *hists;struct mem_info *mem_info; +struct block_info *block_info;void *raw_data;u32 raw_size;int num_res;
Signature · 1 line
2.7.4
PATCH v2 3/7 perf diff: Check if all data files with branch stacks
In reply to Jin YaoWe will expand perf diff to support diff cycles of individual programs blocks, so it requires all data files having branch stacks.
This patch checks HEADER_BRANCH_STACK in header, and only set the flag has_br_stack when HEADER_BRANCH_STACK are set in all data files.
v2:
---
Move check_file_brstack() from __cmd_diff() to cmd_diff(). Because later patch will check flag 'has_br_stack' before ui_init().
--- tools/perf/builtin-diff.c | 29 +++++++++++++++++++++++++++++ 1 file changed, 29 insertions(+)
tools/perf/builtin-diff.c +29 -0
@@ -32,6 +32,7 @@ struct perf_diff {struct perf_time_interval *ptime_range;int range_size;int range_num; +bool has_br_stack; };/* Diff command specific HPP columns. */
@@ -873,6 +874,31 @@ static int parse_time_str(struct data__file *d, char *abstime_ostr,return ret; } +static int check_file_brstack (void ) +{ +struct data__file *d; +bool has_br_stack; +int i; + +data__for_each_file (i, d) { + d->session =perf_session__new (&d->data, false, &pdiff.tool); +if (!d->session) { +pr_err ("Failed to open %s\n" , d->data.path); +return -1 ; + } + + has_br_stack =perf_header__has_feat (&d->session->header, +HEADER_BRANCH_STACK ); +perf_session__delete (d->session); +if (!has_br_stack) +return 0 ; + } + +/* Set only all files having branch stacks */ + pdiff.has_br_stack = true; +return 0 ; +} +static int __cmd_diff (void ) {struct data__file *d;
@@ -1487,6 +1513,9 @@ int cmd_diff(int argc, const char **argv)if (data_init (argc, argv) <0 )return -1 ; +if (check_file_brstack () <0 ) +return -1 ; +if (ui_init () <0 )return -1 ;
Signature · 1 line
2.7.4
PATCH v2 4/7 perf diff: Use hists to manage basic blocks per symbol
In reply to Jin YaoThe hist__account_cycles() can account cycles per basic block. The basic block information is saved in cycles_hist structure.
This patch processes each symbol, get basic blocks from cycles_hist and add the basic block entries to a new hists (block_hists in hist_entry). Using a hists is because we need to compare, sort and print the basic blocks.
v2:
---
v1 adds the basic block entries to per data-file hists but v2 adds the basic block entries to per symbol hists. That is to keep current perf-diff format. Will show the result in next patches.
--- tools/perf/builtin-diff.c | 194 ++++++++++++++++++++++++++++++++++++++++++++++ tools/perf/util/sort.h | 6 ++ 2 files changed, 200 insertions(+)
tools/perf/builtin-diff.c +194 -0
@@ -20,6 +20,7 @@#include "util/data.h" #include "util/config.h" #include "util/time-utils.h" +#include "util/annotate.h" #include <errno.h> #include <inttypes.h>
@@ -64,6 +65,17 @@ struct data__file {struct diff_hpp_fmt fmt[PERF_HPP_DIFF__MAX_INDEX ]; }; +struct block_hpp_fmt { +struct perf_hpp_fmt fmt; +struct data__file *file; +}; + +struct block_hists { +struct hists hists; +struct perf_hpp_list list; +struct block_hpp_fmt block_fmt; +}; +static struct data__file *data__files;static int data__files_cnt;
@@ -87,11 +99,14 @@ static s64 compute_wdiff_w2;static const char *cpu_list; static DECLARE_BITMAP(cpu_bitmap, MAX_NR_CPUS); +static struct addr_location dummy_al; +enum {COMPUTE_DELTA ,COMPUTE_RATIO ,COMPUTE_WEIGHTED_DIFF ,COMPUTE_DELTA_ABS , +COMPUTE_CYCLES ,COMPUTE_MAX , };
@@ -100,6 +115,7 @@ const char *compute_names[COMPUTE_MAX] = { [COMPUTE_DELTA_ABS] = "delta-abs", [COMPUTE_RATIO] = "ratio", [COMPUTE_WEIGHTED_DIFF] = "wdiff", + [COMPUTE_CYCLES] = "cycles", };static int compute =COMPUTE_DELTA_ABS ;
@@ -234,6 +250,8 @@ static int setup_compute(const struct option *opt, const char *str,for (i =0 ; i <COMPUTE_MAX ; i++)if (!strcmp (cstr, compute_names[i])) { *cp = i; +if (i ==COMPUTE_CYCLES ) +break ;return setup_compute_opt (option); }
@@ -368,6 +386,9 @@ static int diff__process_sample_event(struct perf_tool *tool, goto out_put; } +if (compute ==COMPUTE_CYCLES ) +hist__account_cycles (sample->branch_stack, &al, sample, false); +/* * The total_period is updated here before going to the output * tree since normally only the baseline hists will call
@@ -475,6 +496,155 @@ static void hists__baseline_only(struct hists *hists) } } +static int64_t block_cmp (struct perf_hpp_fmt *fmt__maybe_unused , +struct hist_entry *left,struct hist_entry *right) +{ +struct block_info *bi_l = left->block_info; +struct block_info *bi_r = right->block_info; +int cmp; + +if (!bi_l->sym || !bi_r->sym) { +if (!bi_l->sym && !bi_r->sym) +return 0 ; +else if (!bi_l->sym) +return -1 ; +else +return 1 ; + } + +if (bi_l->sym == bi_r->sym) { +if (bi_l->start == bi_r->start) { +if (bi_l->end == bi_r->end) +return 0 ; +else +return (int64_t )(bi_r->end - bi_l->end); + }else +return (int64_t )(bi_r->start - bi_l->start); + }else { + cmp =strcmp (bi_l->sym->name, bi_r->sym->name); +return cmp; + } + +if (bi_l->sym->start != bi_r->sym->start) +return (int64_t )(bi_r->sym->start - bi_l->sym->start); + +return (int64_t )(bi_r->sym->end - bi_l->sym->end); +} + +static int64_t block_cycles_diff_cmp (struct hist_entry *left, +struct hist_entry *right) +{ +bool pairs_left =hist_entry__has_pairs (left); +bool pairs_right =hist_entry__has_pairs (right); +s64 l, r; + +if (!pairs_left && !pairs_right) +return 0 ; + + l =abs (left->diff.cycles); + r =abs (right->diff.cycles); +return r - l; +} + +static int64_t block_sort (struct perf_hpp_fmt *fmt__maybe_unused , +struct hist_entry *left,struct hist_entry *right) +{ +return block_cycles_diff_cmp (right, left); +} + +static struct block_hists *alloc_block_hists (struct data__file *d) +{ +struct block_hists *block_hists =zalloc (sizeof (*block_hists)); +struct block_hpp_fmt *block_fmt; + +if (block_hists) { +__hists__init (&block_hists->hists, &block_hists->list); +perf_hpp_list__init (&block_hists->list); + + block_fmt = &block_hists->block_fmt; +INIT_LIST_HEAD (&block_fmt->fmt.list); +INIT_LIST_HEAD (&block_fmt->fmt.sort_list); + block_fmt->fmt.cmp = block_cmp; + block_fmt->fmt.sort = block_sort; + block_fmt->file = d; +perf_hpp_list__register_sort_field (&block_hists->list, + &block_fmt->fmt); + } + +return block_hists; +} + +static void init_block_info (struct block_info *bi,struct symbol *sym, +struct cyc_hist *ch,int offset) +{ + bi->sym = sym; + bi->start = ch->start; + bi->end = offset; + bi->cycles = ch->cycles; + bi->cycles_aggr = ch->cycles_aggr; + bi->num = ch->num; + bi->num_aggr = ch->num_aggr; +} + +static void *block_he_zalloc (size_t size) +{ +return zalloc (size +sizeof (struct hist_entry )); +} + +static void block_he_free (void *he) +{ +struct block_info *bi = ((struct hist_entry *)he)->block_info; + +block_info__put (bi); +free (he); +} + +struct hist_entry_ops block_he_ops = { + .new = block_he_zalloc, + .free = block_he_free, +}; + +static int process_block_per_sym (struct hist_entry *he,struct data__file *d) +{ +struct annotation *notes; +struct cyc_hist *ch; + +if (!he->ms.map || !he->ms.sym) +return 0 ; + + notes =symbol__annotation (he->ms.sym); +if (!notes || !notes->src || !notes->src->cycles_hist) +return 0 ; + + he->block_hists =alloc_block_hists (d); +if (!he->block_hists) +return -1 ; + + ch = notes->src->cycles_hist; +for (unsigned int i =0 ; i <symbol__size (he->ms.sym); i++) { +if (ch[i].num_aggr) { +struct block_info *bi; +struct hist_entry *he_block; +struct block_hists *block_hists = he->block_hists; + + bi =block_info__new (); +if (!bi) +return -1 ; + +init_block_info (bi, he->ms.sym, &ch[i], i); + he_block =hists__add_entry_block (&block_hists->hists, + &block_he_ops, + &dummy_al, bi); +if (!he_block) { +block_info__put (bi); +return -1 ; + } + } + } + +return 0 ; +} +static void hists__precompute (struct hists *hists) {struct rb_root_cached *root;
@@ -494,6 +664,9 @@ static void hists__precompute(struct hists *hists) he = rb_entry(next, struct hist_entry, rb_node_in); next =rb_next (&he->rb_node_in); +if (compute ==COMPUTE_CYCLES ) +process_block_per_sym (he, &data__files[0 ]); +data__for_each_file_new (i, d) { pair =get_pair_data (he, d);if (!pair)
@@ -510,6 +683,9 @@ static void hists__precompute(struct hists *hists)case COMPUTE_WEIGHTED_DIFF :compute_wdiff (he, pair);break ; +case COMPUTE_CYCLES : +process_block_per_sym (pair, d); +break ;default :BUG_ON (1 ); }
@@ -713,6 +889,14 @@ hist_entry__cmp_wdiff_idx(struct perf_hpp_fmt *fmt __maybe_unused, sort_compute); } +static int64_t +hist_entry__cmp_cycles_idx (struct perf_hpp_fmt *fmt__maybe_unused , +struct hist_entry *left__maybe_unused , +struct hist_entry *right__maybe_unused ) +{ +return 0 ; +} +static void hists__process (struct hists *hists) {if (show_baseline_only)
@@ -746,6 +930,8 @@ static void data_process(void)struct perf_evsel *evsel_base;bool first = true; +memset (&dummy_al,0 ,sizeof (dummy_al)); +evlist__for_each_entry (evlist_base, evsel_base) {struct hists *hists_base =evsel__hists (evsel_base);struct data__file *d;
@@ -1411,6 +1597,9 @@ static int ui_init(void)case COMPUTE_DELTA_ABS : fmt->sort = hist_entry__cmp_delta_abs_idx;break ; +case COMPUTE_CYCLES : + fmt->sort = hist_entry__cmp_cycles_idx; +break ;default :BUG_ON (1 ); }
@@ -1507,6 +1696,8 @@ int cmd_diff(int argc, const char **argv)if (quiet)perf_quiet_option (); +symbol__annotation_init (); +if (symbol__init (NULL ) <0 )return -1 ;
@@ -1516,6 +1707,9 @@ int cmd_diff(int argc, const char **argv)if (check_file_brstack () <0 )return -1 ; +if (compute ==COMPUTE_CYCLES && !pdiff.has_br_stack) +return -1 ; +if (ui_init () <0 )return -1 ;
tools/perf/util/sort.h +6 -0
@@ -79,6 +79,9 @@ struct hist_entry_diff {/* HISTC_WEIGHTED_DIFF */ s64 wdiff; + +/* PERF_HPP_DIFF__CYCLES */ +s64 cycles; }; };
@@ -143,6 +146,9 @@ struct hist_entry {struct branch_info *branch_info;long time;struct hists *hists; +void *block_hists; +int block_idx; +int block_num;struct mem_info *mem_info;struct block_info *block_info;void *raw_data;
Signature · 1 line
2.7.4
Re: [PATCH v2 4/7] perf diff: Use hists to manage basic blocks per symbol
On Mon, Jun 03, 2019 at 10:36:14PM +0800, Jin Yao wrote:
SNIP
Quoted 15 lines
tools/perf/util/sort.h +3
-0@@ -79,6 +79,9 @@ struct hist_entry_diff {/* HISTC_WEIGHTED_DIFF */ s64 wdiff; + +/* PERF_HPP_DIFF__CYCLES */ +s64 cycles; }; };
tools/perf/util/sort.h struct hist_entry { · line 153
Quoted 8 lines
tools/perf/util/sort.h +3
-0@@ -143,6 +146,9 @@ struct hist_entry {struct branch_info *branch_info;long time;struct hists *hists; +void *block_hists; +int block_idx; +int block_num;struct mem_info *mem_info;struct block_info *block_info;
could you please not add the new block* stuff in here, and instead use the "c2c model" and use yourr own struct on top of hist_entry? we are trying to librarize this stuff and keep only necessary things in here..
you're already using hist_entry_ops, so should be easy
something like:
struct block_hist_entry {
void *block_hists;
int block_idx;
int block_num;
struct block_info *block_info;
struct hist_entry he;
};jirka
Re: [PATCH v2 4/7] perf diff: Use hists to manage basic blocks per symbol
Quoted 48 lines · Jiri Olsa
On Mon, Jun 03, 2019 at 10:36:14PM +0800, Jin Yao wrote:
SNIP
Quoted 15 lines
tools/perf/util/sort.h +3
-0@@ -79,6 +79,9 @@ struct hist_entry_diff {/* HISTC_WEIGHTED_DIFF */ s64 wdiff; + +/* PERF_HPP_DIFF__CYCLES */ +s64 cycles; }; };tools/perf/util/sort.h struct hist_entry { · line 153
Quoted 8 lines
tools/perf/util/sort.h +3
-0@@ -143,6 +146,9 @@ struct hist_entry {struct branch_info *branch_info;long time;struct hists *hists; +void *block_hists; +int block_idx; +int block_num;struct mem_info *mem_info;struct block_info *block_info;could you please not add the new block* stuff in here, and instead use the "c2c model" and use yourr own struct on top of
hist_entry? we are trying to librarize this stuff and keep only necessary things in here..you're already using
hist_entry_ops, so should be easysomething like:
struct block_hist_entry { void *block_hists; int block_idx; int block_num; struct block_info *block_info; struct hist_entry he; };
Quoted 2 lines
jirka
Oh, yes, I should refer to 'c2c_hist_entry'. That avoids to add more stuffs in hist_entry. That's a good idea!
Thanks
Jin Yao
Re: [PATCH v2 4/7] perf diff: Use hists to manage basic blocks per symbol
In reply to Jiri OlsaQuoted 48 lines · Jiri Olsa
On Mon, Jun 03, 2019 at 10:36:14PM +0800, Jin Yao wrote:
SNIP
Quoted 15 lines
tools/perf/util/sort.h +3
-0@@ -79,6 +79,9 @@ struct hist_entry_diff {/* HISTC_WEIGHTED_DIFF */ s64 wdiff; + +/* PERF_HPP_DIFF__CYCLES */ +s64 cycles; }; };tools/perf/util/sort.h struct hist_entry { · line 153
Quoted 8 lines
tools/perf/util/sort.h +3
-0@@ -143,6 +146,9 @@ struct hist_entry {struct branch_info *branch_info;long time;struct hists *hists; +void *block_hists; +int block_idx; +int block_num;struct mem_info *mem_info;struct block_info *block_info;could you please not add the new block* stuff in here, and instead use the "c2c model" and use yourr own struct on top of
hist_entry? we are trying to librarize this stuff and keep only necessary things in here..you're already using
hist_entry_ops, so should be easysomething like:
struct block_hist_entry { void *block_hists; int block_idx; int block_num; struct block_info *block_info; struct hist_entry he; };
Quoted 2 lines
jirka
Hi Jiri,
After more considerations, maybe I can't move these stuffs from hist_entry to block_hist_entry.
Actually we use 2 kinds of hist_entry in this patch series. On kind of hist_entry is for symbol/function. The other kind of hist_entry is for basic block.
@@ -143,6 +146,9 @@ struct hist_entry {
struct branch_info *branch_info;
long time;
struct hists *hists;
+ void *block_hists;
+ int block_idx;
+ int block_num;
struct mem_info *mem_info;
struct block_info *block_info;
The above hist_entry is actually for symbol/function. This patch series collects all basic blocks in a symbol/function, so it needs a hists in struct hist_entry (block_hists) to point to the hists of basic blocks.
Correct me if I'm wrong.
Thanks
Jin Yao
Re: [PATCH v2 4/7] perf diff: Use hists to manage basic blocks per symbol
Quoted 70 lines · Yao
Quoted 41 lines · Jiri Olsa
On Mon, Jun 03, 2019 at 10:36:14PM +0800, Jin Yao wrote:
SNIP
Quoted 13 lines
tools/perf/util/sort.h +3
-0@@ -79,6 +79,9 @@ struct hist_entry_diff { /* HISTC_WEIGHTED_DIFF */ s64 wdiff; + + /* PERF_HPP_DIFF__CYCLES */ + s64 cycles; }; };tools/perf/util/sort.h struct hist_entry { · line 153
Quoted 8 lines
tools/perf/util/sort.h +3
-0@@ -143,6 +146,9 @@ struct hist_entry { struct branch_info *branch_info; long time; struct hists *hists; + void *block_hists; + int block_idx; + int block_num; struct mem_info *mem_info; struct block_info *block_info;could you please not add the new block* stuff in here, and instead use the "c2c model" and use yourr own struct on top of
hist_entry? we are trying to librarize this stuff and keep only necessary things in here..you're already using
hist_entry_ops, so should be easysomething like:
struct
block_hist_entry{
void*block_hists;
intblock_idx;
intblock_num;
struct block_info*block_info;Quoted 7 lines
struct hist_entry he;
};jirka
Hi Jiri,
After more considerations, maybe I can't move these stuffs from
hist_entrytoblock_hist_entry.Actually we use 2 kinds of
hist_entryin this patch series. On kind ofhist_entryis for symbol/function. The other kind ofhist_entryis for basic block.tools/perf/util/sort.h +3
-0@@ -143,6 +146,9 @@ struct hist_entry { struct branch_info *branch_info; long time; struct hists *hists; + void *block_hists; + int block_idx; + int block_num; struct mem_info *mem_info; struct block_info *block_info;The above
hist_entryis actually for symbol/function. This patch series collects all basic blocks in a symbol/function, so it needs a hists in structhist_entry(block_hists) to point to the hists of basic blocks.
Quoted 5 lines
Correct me if I'm wrong.
Thanks
Jin Yao
Hi Jiri,
Either adding a new pointer 'priv' in 'struct map_symbol'?
struct map_symbol {
struct map *map;
struct symbol *sym;+ void *priv;
};
We create a struct outside and assign the pointer to priv. Logically it should make sense since the symbol/function may have private data for processing.
Any idea?
Thanks
Jin Yao
Re: [PATCH v2 4/7] perf diff: Use hists to manage basic blocks per symbol
In reply to Jin, YaoQuoted 49 lines · Yao
Quoted 46 lines · Jiri Olsa
On Mon, Jun 03, 2019 at 10:36:14PM +0800, Jin Yao wrote:
SNIP
Quoted 13 lines
tools/perf/util/sort.h +3
-0@@ -79,6 +79,9 @@ struct hist_entry_diff {/* HISTC_WEIGHTED_DIFF */ s64 wdiff; + +/* PERF_HPP_DIFF__CYCLES */ +s64 cycles; }; };tools/perf/util/sort.h struct hist_entry { · line 153
Quoted 8 lines
tools/perf/util/sort.h +3
-0@@ -143,6 +146,9 @@ struct hist_entry {struct branch_info *branch_info;long time;struct hists *hists; +void *block_hists; +int block_idx; +int block_num;struct mem_info *mem_info;struct block_info *block_info;could you please not add the new block* stuff in here, and instead use the "c2c model" and use yourr own struct on top of
hist_entry? we are trying to librarize this stuff and keep only necessary things in here..you're already using
hist_entry_ops, so should be easysomething like:
struct block_hist_entry { void *block_hists; int block_idx; int block_num; struct block_info *block_info; struct hist_entry he; };
Quoted 7 lines
Quoted 2 lines
jirka
Hi Jiri,
After more considerations, maybe I can't move these stuffs from
hist_entrytoblock_hist_entry.
why?
Quoted 4 lines
Actually we use 2 kinds of
hist_entryin this patch series. On kind ofhist_entryis for symbol/function. The other kind ofhist_entryis for basic block.
correct
so the way I see it the processing goes like this:
1) there's standard hist_entry processing ending up with evsel->hists->rb_root full of hist entries
2) then you process every hist_entry and create
new 'struct hists' for each and fill it with
symbol counts data
you could add 'struct hist_entry_ops' for the 1) processing that adds the 'struct hists' object for each hist_entry
and add another 'struct hist_entry_ops' for 2) processing to carry the block data for each hist_entry
jirka
Re: [PATCH v2 4/7] perf diff: Use hists to manage basic blocks per symbol
Quoted 79 lines · Jiri Olsa
Quoted 49 lines · Yao
Quoted 46 lines · Jiri Olsa
On Mon, Jun 03, 2019 at 10:36:14PM +0800, Jin Yao wrote:
SNIP
Quoted 13 lines
tools/perf/util/sort.h +3
-0@@ -79,6 +79,9 @@ struct hist_entry_diff {/* HISTC_WEIGHTED_DIFF */ s64 wdiff; + +/* PERF_HPP_DIFF__CYCLES */ +s64 cycles; }; };tools/perf/util/sort.h struct hist_entry { · line 153
Quoted 8 lines
tools/perf/util/sort.h +3
-0@@ -143,6 +146,9 @@ struct hist_entry {struct branch_info *branch_info;long time;struct hists *hists; +void *block_hists; +int block_idx; +int block_num;struct mem_info *mem_info;struct block_info *block_info;could you please not add the new block* stuff in here, and instead use the "c2c model" and use yourr own struct on top of
hist_entry? we are trying to librarize this stuff and keep only necessary things in here..you're already using
hist_entry_ops, so should be easysomething like:
struct block_hist_entry { void *block_hists; int block_idx; int block_num; struct block_info *block_info; struct hist_entry he; };Quoted 7 lines
Quoted 2 lines
jirka
Hi Jiri,
After more considerations, maybe I can't move these stuffs from
hist_entrytoblock_hist_entry.why?
Quoted 4 lines
Actually we use 2 kinds of
hist_entryin this patch series. On kind ofhist_entryis for symbol/function. The other kind ofhist_entryis for basic block.correct
so the way I see it the processing goes like this:
1) there's standard
hist_entryprocessing ending upwith evsel->hists->rb_root full of hist entries2) then you process every
hist_entryand create
new 'struct hists' for each and fill it with
symbol counts data
Quoted 8 lines
you could add 'struct
hist_entry_ops' for the 1) processing that adds the 'struct hists' object for eachhist_entryand add another 'struct
hist_entry_ops' for 2) processing to carry the block data for eachhist_entryjirka
Hi Jiri,
Yes, I can use two hist_entry_ops but one thing is still difficult to handle that is the printing of blocks.
One function may contain multiple blocks so I add 'block_num' in 'struct hist_entry' to record the number of blocks.
In patch "perf diff: Print the basic block cycles diff", I reuse most of current code to print the blocks. The major change is:
static int hist_entry__fprintf(struct hist_entry *he, size_t size,
char *bf, size_t bfsz, FILE *fp, bool ignore_callchains) {
+ if (he->block_hists)
+ return hist_entry__block_fprintf(he, bf, size, fp);
+
hist_entry__snprintf(he, &hpp); }
+static int hist_entry__block_fprintf(struct hist_entry *he,
+ char *bf, size_t size,
+ FILE *fp)
+{
+ int ret = 0;
+
+ for (int i = 0; i < he->block_num; i++) {
+ struct perf_hpp hpp = {
+ .buf = bf, + .size = size, + .skip = false,
+ };
+
+ he->block_idx = i;
+ hist_entry__snprintf(he, &hpp);
+
+ if (!hpp.skip)
+ ret += fprintf(fp, "%s\n", bf);
+ }
+
+ return ret;
+}
+
So it looks at least I need to add 'block_num' to 'struct hist_entry', otherwise I can't reuse most of codes.
Any idea for 'block_num'?
Thanks
Jin Yao
Re: [PATCH v2 4/7] perf diff: Use hists to manage basic blocks per symbol
Quoted 106 lines · Yao
Quoted 79 lines · Jiri Olsa
Quoted 49 lines · Yao
Quoted 46 lines · Jiri Olsa
On Mon, Jun 03, 2019 at 10:36:14PM +0800, Jin Yao wrote:
SNIP
Quoted 13 lines
tools/perf/util/sort.h +3
-0@@ -79,6 +79,9 @@ struct hist_entry_diff {/* HISTC_WEIGHTED_DIFF */ s64 wdiff; + +/* PERF_HPP_DIFF__CYCLES */ +s64 cycles; }; };tools/perf/util/sort.h struct hist_entry { · line 153
Quoted 8 lines
tools/perf/util/sort.h +3
-0@@ -143,6 +146,9 @@ struct hist_entry {struct branch_info *branch_info;long time;struct hists *hists; +void *block_hists; +int block_idx; +int block_num;struct mem_info *mem_info;struct block_info *block_info;could you please not add the new block* stuff in here, and instead use the "c2c model" and use yourr own struct on top of
hist_entry? we are trying to librarize this stuff and keep only necessary things in here..you're already using
hist_entry_ops, so should be easysomething like:
struct block_hist_entry { void *block_hists; int block_idx; int block_num; struct block_info *block_info; struct hist_entry he; };Quoted 7 lines
Quoted 2 lines
jirka
Hi Jiri,
After more considerations, maybe I can't move these stuffs from
hist_entrytoblock_hist_entry.why?
Quoted 4 lines
Actually we use 2 kinds of
hist_entryin this patch series. On kind ofhist_entryis for symbol/function. The other kind ofhist_entryis for basic block.correct
so the way I see it the processing goes like this:
1) there's standard
hist_entryprocessing ending upwith evsel->hists->rb_root full of hist entries2) then you process every
hist_entryand create
new 'struct hists' for each and fill it with
symbol counts dataQuoted 8 lines
you could add 'struct
hist_entry_ops' for the 1) processing that adds the 'struct hists' object for eachhist_entryand add another 'struct
hist_entry_ops' for 2) processing to carry the block data for eachhist_entryjirka
Hi Jiri,
Yes, I can use two
hist_entry_opsbut one thing is still difficult to handle that is the printing of blocks.One function may contain multiple blocks so I add '
block_num' in 'structhist_entry' to record the number of blocks.In patch "perf diff: Print the basic block cycles diff", I reuse most of current code to print the blocks. The major change is:
static int hist_entry__fprintf(struct
hist_entry*he,size_tsize,
char *bf,size_tbfsz, FILE *fp, boolignore_callchains) {
Review
Quoted 3 lines
(quoted diff) +3
-0+ if (he->block_hists) + return hist_entry__block_fprintf(he, bf, size, fp); +
you could do it the way we do hierarchy and have
something like 'symbol_conf.report_block'
if (symbol_conf.report_hierarchy)
return hist_entry__hierarchy_fprintf(he, &hpp, hists, fp);and in hist_entry__block_fprintf you cast the hist_entry to your struct.. so you'll have all the data
jirka
Quoted 34 lines
hist_entry__snprintf(he, &hpp); }(quoted diff) +23
-0+static int hist_entry__block_fprintf(struct hist_entry *he, + char *bf, size_t size, + FILE *fp) +{ + int ret = 0; + + for (int i = 0; i < he->block_num; i++) { + struct perf_hpp hpp = { + .buf = bf, + .size = size, + .skip = false, + }; + + he->block_idx = i; + hist_entry__snprintf(he, &hpp); + + if (!hpp.skip) + ret += fprintf(fp, "%s\n", bf); + } + + return ret; +} +So it looks at least I need to add '
block_num' to 'structhist_entry', otherwise I can't reuse most of codes.Any idea for '
block_num'?Thanks
Jin Yao
Re: [PATCH v2 4/7] perf diff: Use hists to manage basic blocks per symbol
Quoted 115 lines · Jiri Olsa
Quoted 106 lines · Yao
Quoted 79 lines · Jiri Olsa
Quoted 49 lines · Yao
Quoted 46 lines · Jiri Olsa
On Mon, Jun 03, 2019 at 10:36:14PM +0800, Jin Yao wrote:
SNIP
Quoted 13 lines
tools/perf/util/sort.h +3
-0@@ -79,6 +79,9 @@ struct hist_entry_diff {/* HISTC_WEIGHTED_DIFF */ s64 wdiff; + +/* PERF_HPP_DIFF__CYCLES */ +s64 cycles; }; };tools/perf/util/sort.h struct hist_entry { · line 153
Quoted 8 lines
tools/perf/util/sort.h +3
-0@@ -143,6 +146,9 @@ struct hist_entry {struct branch_info *branch_info;long time;struct hists *hists; +void *block_hists; +int block_idx; +int block_num;struct mem_info *mem_info;struct block_info *block_info;could you please not add the new block* stuff in here, and instead use the "c2c model" and use yourr own struct on top of
hist_entry? we are trying to librarize this stuff and keep only necessary things in here..you're already using
hist_entry_ops, so should be easysomething like:
struct block_hist_entry { void *block_hists; int block_idx; int block_num; struct block_info *block_info; struct hist_entry he; };Quoted 7 lines
Quoted 2 lines
jirka
Hi Jiri,
After more considerations, maybe I can't move these stuffs from
hist_entrytoblock_hist_entry.why?
Quoted 4 lines
Actually we use 2 kinds of
hist_entryin this patch series. On kind ofhist_entryis for symbol/function. The other kind ofhist_entryis for basic block.correct
so the way I see it the processing goes like this:
1) there's standard
hist_entryprocessing ending upwith evsel->hists->rb_root full of hist entries2) then you process every
hist_entryand create
new 'struct hists' for each and fill it with
symbol counts dataQuoted 8 lines
you could add 'struct
hist_entry_ops' for the 1) processing that adds the 'struct hists' object for eachhist_entryand add another 'struct
hist_entry_ops' for 2) processing to carry the block data for eachhist_entryjirka
Hi Jiri,
Yes, I can use two
hist_entry_opsbut one thing is still difficult to handle that is the printing of blocks.One function may contain multiple blocks so I add '
block_num' in 'structhist_entry' to record the number of blocks.In patch "perf diff: Print the basic block cycles diff", I reuse most of current code to print the blocks. The major change is:
static int hist_entry__fprintf(struct
hist_entry*he,size_tsize,
char *bf,size_tbfsz, FILE *fp, boolignore_callchains) {Review
Quoted 3 lines
(quoted diff) +3
-0+ if (he->block_hists) + return hist_entry__block_fprintf(he, bf, size, fp); +you could do it the way we do hierarchy and have
something like 'symbol_conf.report_block'
Quoted 8 lines
if (symbol_conf.report_hierarchy) return hist_entry__hierarchy_fprintf(he, &hpp, hists, fp);and in
hist_entry__block_fprintfyou cast thehist_entryto your struct.. so you'll have all the datajirka
Thanks Jiri. So it looks I need to define 'struct block_hist_entry' in util/sort.h, then hist_entry__block_fprintf can know this struct. Previously I just defined this struct in builtin-diff.c.
Thanks
Jin Yao
Re: [PATCH v2 4/7] perf diff: Use hists to manage basic blocks per symbol
In reply to Jin YaoOn Mon, Jun 03, 2019 at 10:36:14PM +0800, Jin Yao wrote:
SNIP
Quoted 15 lines
tools/perf/builtin-diff.c +8
-0sort_compute); } +static int64_t +hist_entry__cmp_cycles_idx (struct perf_hpp_fmt *fmt__maybe_unused , +struct hist_entry *left__maybe_unused , +struct hist_entry *right__maybe_unused ) +{ +return 0 ; +} +static void hists__process (struct hists *hists) {if (show_baseline_only)
tools/perf/builtin-diff.c static void data_process(void) · line 933
Quoted 4 lines
tools/perf/builtin-diff.c +1
-0@@ -746,6 +930,8 @@ static void data_process(void)struct perf_evsel *evsel_base;bool first = true; +memset (&dummy_al,0 ,sizeof (dummy_al));
why is this needed? it's zero by static declaration, and it's never set, right?
jirka
Re: [PATCH v2 4/7] perf diff: Use hists to manage basic blocks per symbol
Quoted 22 lines · Jiri Olsa
On Mon, Jun 03, 2019 at 10:36:14PM +0800, Jin Yao wrote:
SNIP
Quoted 18 lines
sort_compute); }tools/perf/builtin-diff.c +8
-0+static int64_t +hist_entry__cmp_cycles_idx (struct perf_hpp_fmt *fmt__maybe_unused , +struct hist_entry *left__maybe_unused , +struct hist_entry *right__maybe_unused ) +{ +return 0 ; +} +static void hists__process (struct hists *hists) {if (show_baseline_only)@@ -746,6 +930,8 @@ static void data_process(void)struct perf_evsel *evsel_base;bool first = true;
Quoted 6 lines
tools/perf/builtin-diff.c static void data_process(void) · line 933
Quoted 1 line
tools/perf/builtin-diff.c +1
-0+memset (&dummy_al,0 ,sizeof (dummy_al));why is this needed? it's zero by static declaration, and it's never set, right?
jirka
C standard says yes for initializing static variables to 0, but I just want to avoid any potential for unexpected behavior. Maybe I was over thinking it. I will remove this line.
Thanks
Jin Yao
Re: [PATCH v2 4/7] perf diff: Use hists to manage basic blocks per symbol
In reply to Jin YaoOn Mon, Jun 03, 2019 at 10:36:14PM +0800, Jin Yao wrote:
SNIP
Quoted 14 lines
tools/perf/builtin-diff.c +3
-0data__for_each_file_new (i, d) { pair =get_pair_data (he, d);if (!pair)@@ -510,6 +683,9 @@ static void hists__precompute(struct hists *hists)case COMPUTE_WEIGHTED_DIFF :compute_wdiff (he, pair);break ; +case COMPUTE_CYCLES : +process_block_per_sym (pair, d); +break ;default :BUG_ON (1 ); }
tools/perf/builtin-diff.c hist_entry__cmp_wdiff_idx(struct perf_hpp_fmt *fmt __maybe_unused, · line 898
Quoted 10 lines
tools/perf/builtin-diff.c +7
-0@@ -713,6 +889,14 @@ hist_entry__cmp_wdiff_idx(struct perf_hpp_fmt *fmt __maybe_unused, sort_compute); } +static int64_t +hist_entry__cmp_cycles_idx (struct perf_hpp_fmt *fmt__maybe_unused , +struct hist_entry *left__maybe_unused , +struct hist_entry *right__maybe_unused ) +{ +return 0 ; +}
this is hist_entry__cmp_nop.. please use it instead and explain in comment why for COMPUTE_CYCLES we need the default sort
jirka
Re: [PATCH v2 4/7] perf diff: Use hists to manage basic blocks per symbol
Quoted 33 lines · Jiri Olsa
On Mon, Jun 03, 2019 at 10:36:14PM +0800, Jin Yao wrote:
SNIP
Quoted 14 lines
data__for_each_file_new(i, d) { pair = get_pair_data(he, d); if (!pair)tools/perf/builtin-diff.c +3
-0@@ -510,6 +683,9 @@ static void hists__precompute(struct hists *hists)case COMPUTE_WEIGHTED_DIFF :compute_wdiff (he, pair);break ; +case COMPUTE_CYCLES : +process_block_per_sym (pair, d); +break ;default :BUG_ON (1 ); }tools/perf/builtin-diff.c hist_entry__cmp_wdiff_idx(struct perf_hpp_fmt *fmt __maybe_unused, · line 898
Quoted 10 lines
tools/perf/builtin-diff.c +7
-0@@ -713,6 +889,14 @@ hist_entry__cmp_wdiff_idx(struct perf_hpp_fmt *fmt __maybe_unused, sort_compute); } +static int64_t +hist_entry__cmp_cycles_idx (struct perf_hpp_fmt *fmt__maybe_unused , +struct hist_entry *left__maybe_unused , +struct hist_entry *right__maybe_unused ) +{ +return 0 ; +}this is
hist_entry__cmp_nop.. please use it instead and explain in comment why forCOMPUTE_CYCLESwe need the default sort
Quoted 2 lines
jirka
fmt->sort should be set otherwise since fmt->sort will be called without checking valid, the crash happens. Yes, I should use hist_entry__cmp_nop instead, and will add some comments for COMPUTE_CYCLES.
Thanks
Jin Yao
PATCH v2 5/7 perf diff: Link same basic blocks among different data files
In reply to Jin YaoThe target is to compare the performance difference (cycles diff) for the same basic blocks in different data files.
The same basic block means same function, same start address and same end address. This patch finds the same basic blocks from different data files and link them together and resort by the cycles diff.
v2:
---
Since now the basic block hists is changed to per symbol, the patch only links the basic block hists for the same symbol in different data files.
--- tools/perf/builtin-diff.c | 93 ++++++++++++++++++++++++++++++++++++++++++++++- tools/perf/util/hist.c | 2 +- tools/perf/util/sort.h | 1 + 3 files changed, 94 insertions(+), 2 deletions(-)
tools/perf/builtin-diff.c +92 -1
@@ -645,6 +645,84 @@ static int process_block_per_sym(struct hist_entry *he, struct data__file *d)return 0 ; } +static int block_pair_cmp (struct hist_entry *a,struct hist_entry *b) +{ +struct block_info *bi_a = a->block_info; +struct block_info *bi_b = b->block_info; +int cmp; + +if (!bi_a->sym || !bi_b->sym) +return -1 ; + +if (bi_a->sym->name && bi_b->sym->name) { + cmp =strcmp (bi_a->sym->name, bi_b->sym->name); +if ((!cmp) && (bi_a->start == bi_b->start) && + (bi_a->end == bi_b->end)) { +return 0 ; + } + } + +return -1 ; +} + +static struct hist_entry *get_block_pair (struct hist_entry *he, +struct hists *hists_pair) +{ +struct rb_root_cached *root = hists_pair->entries_in; +struct rb_node *next =rb_first_cached (root); +int cmp; + +while (next !=NULL ) { + struct hist_entry *he_pair = rb_entry(next, struct hist_entry, + rb_node_in); + + next =rb_next (&he_pair->rb_node_in); + + cmp =block_pair_cmp (he_pair, he); +if (!cmp) +return he_pair; + } + +return NULL ; +} + +static void compute_cycles_diff (struct hist_entry *he, +struct hist_entry *pair) +{ + pair->diff.computed = true; +if (pair->block_info->num && he->block_info->num) { + pair->diff.cycles = + pair->block_info->cycles_aggr / pair->block_info->num_aggr - + he->block_info->cycles_aggr / he->block_info->num_aggr; + } +} + +static void block_hists_match (struct hists *hists_base, +struct hists *hists_pair) +{ +struct rb_root_cached *root = hists_base->entries_in; +struct rb_node *next =rb_first_cached (root); + +while (next !=NULL ) { + struct hist_entry *he = rb_entry(next, struct hist_entry, + rb_node_in); +struct hist_entry *pair =get_block_pair (he, hists_pair); + + next =rb_next (&he->rb_node_in); + +if (pair) { +hist_entry__add_pair (pair, he); +compute_cycles_diff (he, pair); + } + } +} + +static int filter_cb (struct hist_entry *he,void *arg__maybe_unused ) +{ + he->not_collen = true; +return 0 ; +} +static void hists__precompute (struct hists *hists) {struct rb_root_cached *root;
@@ -659,13 +737,19 @@ static void hists__precompute(struct hists *hists)while (next !=NULL ) {struct hist_entry *he, *pair;struct data__file *d; +struct block_hists *b; +struct block_hists *p;int i; he = rb_entry(next, struct hist_entry, rb_node_in); next =rb_next (&he->rb_node_in); -if (compute ==COMPUTE_CYCLES ) +if (compute ==COMPUTE_CYCLES ) {process_block_per_sym (he, &data__files[0 ]); + b = he->block_hists; +if (b) + he->block_num = b->hists.nr_entries; + }data__for_each_file_new (i, d) { pair =get_pair_data (he, d);
@@ -685,6 +769,13 @@ static void hists__precompute(struct hists *hists)break ;case COMPUTE_CYCLES :process_block_per_sym (pair, d); + b = he->block_hists; + p = pair->block_hists; +if (b && p) { +block_hists_match (&b->hists, &p->hists); +hists__output_resort_cb (&p->hists,NULL , + filter_cb); + }break ;default :BUG_ON (1 );
tools/perf/util/hist.c +1 -1
@@ -1846,7 +1846,7 @@ static void output_resort(struct hists *hists, struct ui_progress *prog,__hists__insert_output_entry (&hists->entries, n, min_callchain_hits, use_callchain);hists__inc_stats (hists, n); -if (!n->filtered) +if ((!n->filtered) && (!n->not_collen))hists__calc_col_len (hists, n);if (prog)
tools/perf/util/sort.h +1 -0
@@ -122,6 +122,7 @@ struct hist_entry {char level;u8 filtered; +bool not_collen;u16 callchain_size;union {
Signature · 1 line
2.7.4
PATCH v2 6/7 perf diff: Print the basic block cycles diff
In reply to Jin Yaoperf record -b ./div
perf record -b ./div
Following is the default perf diff output
# perf diff
# Event 'cycles'
#
# Baseline Delta Abs Shared Object Symbol
# ........ ......... ................ ....................................
#
49.03% +0.30% div [.] main
16.29% -0.20% libc-2.23.so [.] __random
18.82% -0.07% libc-2.23.so [.] __random_r
8.11% -0.04% div [.] compute_flag
2.25% +0.01% div [.] rand@plt
0.00% +0.01% [kernel.vmlinux] [k] task_tick_fair
5.46% +0.01% libc-2.23.so [.] rand
0.01% -0.01% [kernel.vmlinux] [k] native_irq_return_iret
0.00% -0.00% [kernel.vmlinux] [k] interrupt_entryThis patch creates a new computation selection 'cycles'.
# perf diff -c cycles
# Event 'cycles'
#
# Baseline Block cycles diff [start:end] Shared Object Symbol
# ........ .................................... ................ ....................................
#
49.03% -9 [ 4ef: 520] div [.] main
49.03% 0 [ 4e8: 4ea] div [.] main
49.03% 0 [ 4ef: 500] div [.] main
49.03% 0 [ 4ef: 51c] div [.] main
49.03% 0 [ 4ef: 535] div [.] main
18.82% 0 [ 3ac40: 3ac4d] libc-2.23.so [.] __random_r
18.82% 0 [ 3ac40: 3ac5c] libc-2.23.so [.] __random_r
18.82% 0 [ 3ac40: 3ac76] libc-2.23.so [.] __random_r
18.82% 0 [ 3ac40: 3ac88] libc-2.23.so [.] __random_r
18.82% 0 [ 3ac90: 3ac9c] libc-2.23.so [.] __random_r
16.29% -8 [ 3aac0: 3aac0] libc-2.23.so [.] __random
16.29% 0 [ 3aac0: 3aad2] libc-2.23.so [.] __random
16.29% 0 [ 3aae0: 3aae7] libc-2.23.so [.] __random
16.29% 0 [ 3ab03: 3ab0f] libc-2.23.so [.] __random
16.29% 0 [ 3ab14: 3ab1b] libc-2.23.so [.] __random
16.29% 0 [ 3ab28: 3ab2e] libc-2.23.so [.] __random
16.29% 0 [ 3ab4a: 3ab53] libc-2.23.so [.] __random
8.11% 0 [ 640: 644] div [.] compute_flag
8.11% 0 [ 649: 659] div [.] compute_flag
5.46% 0 [ 3af60: 3af60] libc-2.23.so [.] rand
5.46% 0 [ 3af60: 3af64] libc-2.23.so [.] rand
2.25% 0 [ 490: 490] div [.] rand@plt
0.01% 26 [ c00a27: c00a27] [kernel.vmlinux] [k] native_irq_return_iret
0.00% -157 [ 2bf9f2: 2bfa63] [kernel.vmlinux] [k] update_blocked_averages
0.00% -56 [ 2bf980: 2bf9d3] [kernel.vmlinux] [k] update_blocked_averages
0.00% 48 [ 2bf934: 2bf942] [kernel.vmlinux] [k] update_blocked_averages
0.00% 3 [ 2bfb38: 2bfb67] [kernel.vmlinux] [k] update_blocked_averages
0.00% 0 [ 2bf968: 2bf97b] [kernel.vmlinux] [k] update_blocked_averages"[start:end]" indicates the basic block range. The output is sorted by "Baseline" and the basic blocks in the same function are sorted by cycles diff.
v2:
---
Keep standard perf diff format and display the 'Baseline' and 'Shared Object'.
--- tools/perf/builtin-diff.c | 63 +++++++++++++++++++++++++++++++++++++++++++--- tools/perf/ui/stdio/hist.c | 26 +++++++++++++++++++ tools/perf/util/hist.c | 18 +++++++++++++ tools/perf/util/hist.h | 3 +++ 4 files changed, 106 insertions(+), 4 deletions(-)
tools/perf/builtin-diff.c +59 -4
@@ -46,6 +46,7 @@ enum {PERF_HPP_DIFF__WEIGHTED_DIFF ,PERF_HPP_DIFF__FORMULA ,PERF_HPP_DIFF__DELTA_ABS , +PERF_HPP_DIFF__CYCLES , PERF_HPP_DIFF__MAX_INDEX };
@@ -125,6 +126,7 @@ static int compute_2_hpp[COMPUTE_MAX] = { [COMPUTE_DELTA_ABS] = PERF_HPP_DIFF__DELTA_ABS, [COMPUTE_RATIO] = PERF_HPP_DIFF__RATIO, [COMPUTE_WEIGHTED_DIFF] = PERF_HPP_DIFF__WEIGHTED_DIFF, + [COMPUTE_CYCLES] = PERF_HPP_DIFF__CYCLES, };#define MAX_COL_WIDTH 70
@@ -163,6 +165,10 @@ static struct header_column { [PERF_HPP_DIFF__FORMULA ] = { .name ="Formula" , .width =MAX_COL_WIDTH , + }, + [PERF_HPP_DIFF__CYCLES ] = { + .name ="Block cycles diff [start:end]" , + .width =36 , } };
@@ -250,8 +256,6 @@ static int setup_compute(const struct option *opt, const char *str,for (i =0 ; i <COMPUTE_MAX ; i++)if (!strcmp (cstr, compute_names[i])) { *cp = i; -if (i ==COMPUTE_CYCLES ) -break ;return setup_compute_opt (option); }
@@ -949,6 +953,14 @@ hist_entry__cmp_wdiff(struct perf_hpp_fmt *fmt, }static int64_t +hist_entry__cmp_cycles (struct perf_hpp_fmt *fmt__maybe_unused , +struct hist_entry *left__maybe_unused , +struct hist_entry *right__maybe_unused ) +{ +return 0 ; +} + +static int64_t hist_entry__cmp_delta_idx (struct perf_hpp_fmt *fmt__maybe_unused ,struct hist_entry *left,struct hist_entry *right) {
@@ -1253,7 +1265,7 @@ static const struct option options[] = {OPT_BOOLEAN ('b' ,"baseline-only" , &show_baseline_only,"Show only items with match in baseline" ),OPT_CALLBACK ('c' ,"compute" , &compute, -"delta,delta-abs,ratio,wdiff:w1,w2 (default delta-abs)" , +"delta,delta-abs,ratio,wdiff:w1,w2 (default delta-abs),cycles" ,"Entries differential computation selection" , setup_compute), OPT_BOOLEAN('p', "period", &show_period,
@@ -1331,6 +1343,33 @@ static int hpp__entry_baseline(struct hist_entry *he, char *buf, size_t size)return ret; } +static int cycles_printf (struct hist_entry *he,struct hist_entry *pair, +struct perf_hpp *hpp,int width) +{ +struct block_hists *p = pair->block_hists; +struct hist_entry *block_he; +struct block_info *bi; +char buf[128 ]; + +if (!p) { + hpp->skip = true; +return 0 ; + } + + block_he =hists__get_entry (&p->hists, he->block_idx); +if (!block_he) { + hpp->skip = true; +return 0 ; + } + + bi = block_he->block_info; +scnprintf (buf,sizeof (buf),"%ld [%12lx:%12lx]" , + block_he->diff.cycles, bi->sym->start + bi->start, + bi->sym->start + bi->end); + +return scnprintf (hpp->buf, hpp->size,"%*s" , width, buf); +} +static int __hpp__color_compare (struct perf_hpp_fmt *fmt,struct perf_hpp *hpp,struct hist_entry *he,int comparison_method)
@@ -1342,8 +1381,12 @@ static int __hpp__color_compare(struct perf_hpp_fmt *fmt,s64 wdiff;char pfmt[20 ] =" " ; -if (!pair) +if (!pair) { +if (comparison_method ==COMPUTE_CYCLES && he->block_idx) + hpp->skip = true; + goto no_print; + }switch (comparison_method) {case COMPUTE_DELTA :
@@ -1378,6 +1421,8 @@ static int __hpp__color_compare(struct perf_hpp_fmt *fmt,return color_snprintf (hpp->buf, hpp->size,get_percent_color (wdiff), pfmt, wdiff); +case COMPUTE_CYCLES : +return cycles_printf (he, pair, hpp, dfmt->header_width);default :BUG_ON (1 ); }
@@ -1407,6 +1452,12 @@ static int hpp__color_wdiff(struct perf_hpp_fmt *fmt,return __hpp__color_compare (fmt, hpp, he,COMPUTE_WEIGHTED_DIFF ); } +static int hpp__color_cycles (struct perf_hpp_fmt *fmt, +struct perf_hpp *hpp,struct hist_entry *he) +{ +return __hpp__color_compare (fmt, hpp, he,COMPUTE_CYCLES ); +} +static void hpp__entry_unpair (struct hist_entry *he,int idx,char *buf,size_t size) {
@@ -1608,6 +1659,10 @@ static void data__hpp_register(struct data__file *d, int idx) fmt->color = hpp__color_delta; fmt->sort = hist_entry__cmp_delta_abs;break ; +case PERF_HPP_DIFF__CYCLES : + fmt->color = hpp__color_cycles; + fmt->sort = hist_entry__cmp_cycles; +break ;default : fmt->sort = hist_entry__cmp_nop;break ;
tools/perf/ui/stdio/hist.c +26 -0
@@ -531,6 +531,29 @@ static int hist_entry__hierarchy_fprintf(struct hist_entry *he,return printed; } +static int hist_entry__block_fprintf (struct hist_entry *he, +char *bf,size_t size, +FILE *fp) +{ +int ret =0 ; + +for (int i =0 ; i < he->block_num; i++) { +struct perf_hpp hpp = { + .buf = bf, + .size = size, + .skip = false, + }; + + he->block_idx = i; +hist_entry__snprintf (he, &hpp); + +if (!hpp.skip) + ret +=fprintf (fp,"%s\n" , bf); + } + +return ret; +} +static int hist_entry__fprintf (struct hist_entry *he,size_t size,char *bf,size_t bfsz,FILE *fp,bool ignore_callchains)
@@ -550,6 +573,9 @@ static int hist_entry__fprintf(struct hist_entry *he, size_t size,if (symbol_conf.report_hierarchy)return hist_entry__hierarchy_fprintf (he, &hpp, hists, fp); +if (he->block_hists) +return hist_entry__block_fprintf (he, bf, size, fp); +hist_entry__snprintf (he, &hpp); ret =fprintf (fp,"%s\n" , bf);
tools/perf/util/hist.c +18 -0
@@ -376,6 +376,24 @@ void hists__delete_entries(struct hists *hists) } } +struct hist_entry *hists__get_entry (struct hists *hists,int idx) +{ +struct rb_node *next =rb_first_cached (&hists->entries); +struct hist_entry *n; +int i =0 ; + +while (next) { + n = rb_entry(next, struct hist_entry, rb_node); +if (i == idx) +return n; + + next =rb_next (&n->rb_node); + i++; + } + +return NULL ; +} +/* * histogram, sorted on item, collects periods */
tools/perf/util/hist.h +3 -0
@@ -184,6 +184,8 @@ void hists__decay_entries(struct hists *hists, bool zap_user, bool zap_kernel);void hists__delete_entries (struct hists *hists);void hists__output_recalc_col_len (struct hists *hists,int max_rows); +struct hist_entry *hists__get_entry (struct hists *hists,int idx); +u64 hists__total_period (struct hists *hists);void hists__reset_stats (struct hists *hists);void hists__inc_stats (struct hists *hists,struct hist_entry *h);
@@ -249,6 +251,7 @@ struct perf_hpp {size_t size;const char *sep;void *ptr; +bool skip; };struct perf_hpp_fmt {
Signature · 1 line
2.7.4
Re: [PATCH v2 6/7] perf diff: Print the basic block cycles diff
Quoted 51 lines · Jin Yao
perf record -b ./div
perf record -b ./divFollowing is the default perf diff output
# perf diff# Event 'cycles'
## Baseline Delta Abs Shared Object Symbol # ........ ......... ................ .................................... # 49.03% +0.30% div [.] main 16.29% -0.20% libc-2.23.so [.] __random 18.82% -0.07% libc-2.23.so [.] __random_r 8.11% -0.04% div [.] compute_flag 2.25% +0.01% div [.] rand@plt 0.00% +0.01% [kernel.vmlinux] [k] task_tick_fair 5.46% +0.01% libc-2.23.so [.] rand 0.01% -0.01% [kernel.vmlinux] [k] native_irq_return_iret 0.00% -0.00% [kernel.vmlinux] [k] interrupt_entryThis patch creates a new computation selection 'cycles'.
# perf diff -c cycles# Event 'cycles'
## Baseline Block cycles diff [start:end] Shared Object Symbol # ........ .................................... ................ .................................... # 49.03% -9 [ 4ef: 520] div [.] main 49.03% 0 [ 4e8: 4ea] div [.] main 49.03% 0 [ 4ef: 500] div [.] main 49.03% 0 [ 4ef: 51c] div [.] main 49.03% 0 [ 4ef: 535] div [.] main 18.82% 0 [ 3ac40: 3ac4d] libc-2.23.so [.] __random_r 18.82% 0 [ 3ac40: 3ac5c] libc-2.23.so [.] __random_r 18.82% 0 [ 3ac40: 3ac76] libc-2.23.so [.] __random_r 18.82% 0 [ 3ac40: 3ac88] libc-2.23.so [.] __random_r 18.82% 0 [ 3ac90: 3ac9c] libc-2.23.so [.] __random_r 16.29% -8 [ 3aac0: 3aac0] libc-2.23.so [.] __random 16.29% 0 [ 3aac0: 3aad2] libc-2.23.so [.] __random 16.29% 0 [ 3aae0: 3aae7] libc-2.23.so [.] __random 16.29% 0 [ 3ab03: 3ab0f] libc-2.23.so [.] __random 16.29% 0 [ 3ab14: 3ab1b] libc-2.23.so [.] __random 16.29% 0 [ 3ab28: 3ab2e] libc-2.23.so [.] __random 16.29% 0 [ 3ab4a: 3ab53] libc-2.23.so [.] __random 8.11% 0 [ 640: 644] div [.] compute_flag 8.11% 0 [ 649: 659] div [.] compute_flag 5.46% 0 [ 3af60: 3af60] libc-2.23.so [.] rand
Quoted 9 lines
5.46% 0 [ 3af60: 3af64] libc-2.23.so [.] rand
2.25% 0 [ 490: 490] div [.] rand@plt
0.01% 26 [ c00a27: c00a27] [kernel.vmlinux] [k] native_irq_return_iret
0.00% -157 [ 2bf9f2: 2bfa63] [kernel.vmlinux] [k] update_blocked_averages
0.00% -56 [ 2bf980: 2bf9d3] [kernel.vmlinux] [k] update_blocked_averages
0.00% 48 [ 2bf934: 2bf942] [kernel.vmlinux] [k] update_blocked_averages
0.00% 3 [ 2bfb38: 2bfb67] [kernel.vmlinux] [k] update_blocked_averages
0.00% 0 [ 2bf968: 2bf97b] [kernel.vmlinux] [k] update_blocked_averagesso what I'd expect would be Baseline column with cycles and another column showing the differrence (in cycles) for given symbol
Quoted 3 lines
"[start:end]" indicates the basic block range. The output is sorted by "Baseline" and the basic blocks in the same function are sorted by cycles diff.
hum, why is there multiple basic blocks [start:end] for a symbol?
thanks,
jirka
Re: [PATCH v2 6/7] perf diff: Print the basic block cycles diff
Quoted 65 lines · Jiri Olsa
Quoted 51 lines · Jin Yao
perf record -b ./div
perf record -b ./divFollowing is the default perf diff output
# perf diff# Event 'cycles'
## Baseline Delta Abs Shared Object Symbol # ........ ......... ................ .................................... # 49.03% +0.30% div [.] main 16.29% -0.20% libc-2.23.so [.] __random 18.82% -0.07% libc-2.23.so [.] __random_r 8.11% -0.04% div [.] compute_flag 2.25% +0.01% div [.] rand@plt 0.00% +0.01% [kernel.vmlinux] [k] task_tick_fair 5.46% +0.01% libc-2.23.so [.] rand 0.01% -0.01% [kernel.vmlinux] [k] native_irq_return_iret 0.00% -0.00% [kernel.vmlinux] [k] interrupt_entryThis patch creates a new computation selection 'cycles'.
# perf diff -c cycles# Event 'cycles'
## Baseline Block cycles diff [start:end] Shared Object Symbol # ........ .................................... ................ .................................... # 49.03% -9 [ 4ef: 520] div [.] main 49.03% 0 [ 4e8: 4ea] div [.] main 49.03% 0 [ 4ef: 500] div [.] main 49.03% 0 [ 4ef: 51c] div [.] main 49.03% 0 [ 4ef: 535] div [.] main 18.82% 0 [ 3ac40: 3ac4d] libc-2.23.so [.] __random_r 18.82% 0 [ 3ac40: 3ac5c] libc-2.23.so [.] __random_r 18.82% 0 [ 3ac40: 3ac76] libc-2.23.so [.] __random_r 18.82% 0 [ 3ac40: 3ac88] libc-2.23.so [.] __random_r 18.82% 0 [ 3ac90: 3ac9c] libc-2.23.so [.] __random_r 16.29% -8 [ 3aac0: 3aac0] libc-2.23.so [.] __random 16.29% 0 [ 3aac0: 3aad2] libc-2.23.so [.] __random 16.29% 0 [ 3aae0: 3aae7] libc-2.23.so [.] __random 16.29% 0 [ 3ab03: 3ab0f] libc-2.23.so [.] __random 16.29% 0 [ 3ab14: 3ab1b] libc-2.23.so [.] __random 16.29% 0 [ 3ab28: 3ab2e] libc-2.23.so [.] __random 16.29% 0 [ 3ab4a: 3ab53] libc-2.23.so [.] __random 8.11% 0 [ 640: 644] div [.] compute_flag 8.11% 0 [ 649: 659] div [.] compute_flag 5.46% 0 [ 3af60: 3af60] libc-2.23.so [.] randQuoted 9 lines
5.46% 0 [ 3af60: 3af64] libc-2.23.so [.] rand 2.25% 0 [ 490: 490] div [.] rand@plt 0.01% 26 [ c00a27: c00a27] [kernel.vmlinux] [k] native_irq_return_iret 0.00% -157 [ 2bf9f2: 2bfa63] [kernel.vmlinux] [k] update_blocked_averages 0.00% -56 [ 2bf980: 2bf9d3] [kernel.vmlinux] [k] update_blocked_averages 0.00% 48 [ 2bf934: 2bf942] [kernel.vmlinux] [k] update_blocked_averages 0.00% 3 [ 2bfb38: 2bfb67] [kernel.vmlinux] [k] update_blocked_averages 0.00% 0 [ 2bf968: 2bf97b] [kernel.vmlinux] [k] update_blocked_averagesso what I'd expect would be Baseline column with cycles and another column showing the differrence (in cycles) for given symbol
Quoted 9 lines
Quoted 3 lines
"[start:end]" indicates the basic block range. The output is sorted by "Baseline" and the basic blocks in the same function are sorted by cycles diff.
hum, why is there multiple basic blocks [start:end] for a symbol?
thanks,
jirka
The basic block is the code between 2 branches (for one branch, for example, jmp, call, ret, interrupt, ...). So it's expected that one function (symbol is function) may contain multiple basic blocks.
The idea is, sorting by baseline to display the hottest functions and the second column shows the cycles diff of blocks in this function (comparing between different perf data files). This would allow to identify performance changes in specific code accurately and effectively.
Thanks
Jin Yao
Re: [PATCH v2 6/7] perf diff: Print the basic block cycles diff
In reply to Jin YaoOn Mon, Jun 03, 2019 at 10:36:16PM +0800, Jin Yao wrote:
SNIP
Quoted 5 lines
tools/perf/builtin-diff.c +0
-1-break ;return setup_compute_opt (option); }
tools/perf/builtin-diff.c hist_entry__cmp_wdiff(struct perf_hpp_fmt *fmt, · line 961
Quoted 9 lines
tools/perf/builtin-diff.c +6
-0@@ -949,6 +953,14 @@ hist_entry__cmp_wdiff(struct perf_hpp_fmt *fmt, }static int64_t +hist_entry__cmp_cycles (struct perf_hpp_fmt *fmt__maybe_unused , +struct hist_entry *left__maybe_unused , +struct hist_entry *right__maybe_unused ) +{ +return 0 ; +}
we have hist_entry__cmp_nop for that
SNIP
Quoted 17 lines
tools/perf/builtin-diff.c +6
-0default :BUG_ON (1 ); }@@ -1407,6 +1452,12 @@ static int hpp__color_wdiff(struct perf_hpp_fmt *fmt,return __hpp__color_compare (fmt, hpp, he,COMPUTE_WEIGHTED_DIFF ); } +static int hpp__color_cycles (struct perf_hpp_fmt *fmt, +struct perf_hpp *hpp,struct hist_entry *he) +{ +return __hpp__color_compare (fmt, hpp, he,COMPUTE_CYCLES ); +} +static void hpp__entry_unpair (struct hist_entry *he,int idx,char *buf,size_t size) {
tools/perf/builtin-diff.c static void data__hpp_register(struct data__file *d, int idx) · line 1664
Quoted 6 lines
tools/perf/builtin-diff.c +3
-0@@ -1608,6 +1659,10 @@ static void data__hpp_register(struct data__file *d, int idx) fmt->color = hpp__color_delta; fmt->sort = hist_entry__cmp_delta_abs;break ; +case PERF_HPP_DIFF__CYCLES : + fmt->color = hpp__color_cycles; + fmt->sort = hist_entry__cmp_cycles;
also please explain in comment why it's nop
jirka
Re: [PATCH v2 6/7] perf diff: Print the basic block cycles diff
Quoted 49 lines · Jiri Olsa
On Mon, Jun 03, 2019 at 10:36:16PM +0800, Jin Yao wrote:
SNIP
Quoted 5 lines
tools/perf/builtin-diff.c +0
-1-break ;return setup_compute_opt (option); }tools/perf/builtin-diff.c hist_entry__cmp_wdiff(struct perf_hpp_fmt *fmt, · line 961
Quoted 9 lines
tools/perf/builtin-diff.c +6
-0@@ -949,6 +953,14 @@ hist_entry__cmp_wdiff(struct perf_hpp_fmt *fmt, }static int64_t +hist_entry__cmp_cycles (struct perf_hpp_fmt *fmt__maybe_unused , +struct hist_entry *left__maybe_unused , +struct hist_entry *right__maybe_unused ) +{ +return 0 ; +}we have
hist_entry__cmp_nopfor thatSNIP
Quoted 17 lines
tools/perf/builtin-diff.c +6
-0default :BUG_ON (1 ); }@@ -1407,6 +1452,12 @@ static int hpp__color_wdiff(struct perf_hpp_fmt *fmt,return __hpp__color_compare (fmt, hpp, he,COMPUTE_WEIGHTED_DIFF ); } +static int hpp__color_cycles (struct perf_hpp_fmt *fmt, +struct perf_hpp *hpp,struct hist_entry *he) +{ +return __hpp__color_compare (fmt, hpp, he,COMPUTE_CYCLES ); +} +static void hpp__entry_unpair (struct hist_entry *he,int idx,char *buf,size_t size) {tools/perf/builtin-diff.c static void data__hpp_register(struct data__file *d, int idx) · line 1664
Quoted 6 lines
tools/perf/builtin-diff.c +3
-0@@ -1608,6 +1659,10 @@ static void data__hpp_register(struct data__file *d, int idx) fmt->color = hpp__color_delta; fmt->sort = hist_entry__cmp_delta_abs;break ; +case PERF_HPP_DIFF__CYCLES : + fmt->color = hpp__color_cycles; + fmt->sort = hist_entry__cmp_cycles;also please explain in comment why it's nop
Quoted 2 lines
jirka
Got it, I will update the patch.
Thanks
Jin Yao
PATCH v2 7/7 perf diff: Documentation -c cycles option
In reply to Jin YaoDocumentation the new computation selection 'cycles'.
--- tools/perf/Documentation/perf-diff.txt | 14 +++++++++++--- 1 file changed, 11 insertions(+), 3 deletions(-)
tools/perf/Documentation/perf-diff.txt +11 -3
@@ -90,9 +90,10 @@ OPTIONS -c:: --compute:: - Differential computation selection - delta, ratio, wdiff, delta-abs - (default is delta-abs). Default can be changed using diff.compute - config option. See COMPARISON METHODS section for more info. + Differential computation selection - delta, ratio, wdiff, cycles, + delta-abs (default is delta-abs). Default can be changed using + diff.compute config option. See COMPARISON METHODS section for + more info. -p:: --period::
@@ -278,6 +279,13 @@ If specified the 'Weighted diff' column is displayed with value 'd' computed as: - WEIGHT-A being the weight of the data file - WEIGHT-B being the weight of the baseline data file +cycles +~~~~~~ +If specified the 'Block cycles diff [start:end]' column is displayed. +It displays the cycles difference of same program basic block amongst +two perf.data. The program basic block is the code block between +two branches in a function (indicated by [start:end]). + SEE ALSO -------- linkperf:perf-record[1], linkperf:perf-report[1]
Signature · 1 line
2.7.4
Quoted 7 lines · Jin Yao
In some cases small changes in hot loops can show big differences. But it's difficult to identify these differences.
perf diff currently can only diff symbols (functions). We can also expand it to diff cycles of individual programs blocks as reported by timed LBR. This would allow to identify changes in specific code accurately.
Quoted 6 lines
With this patch set, for example,
# perf record -b ./div # perf record -b ./div # perf diff -s cycles
Sorry, the command line should be perf diff -c cycles.
Thanks
Jin Yao
Quoted 82 lines
# Event 'cycles'
## Baseline Block cycles diff [start:end] Shared Object Symbol # ........ .................................... ................ .................................... # 49.03% -9 [ 4ef: 520] div [.] main 49.03% 0 [ 4e8: 4ea] div [.] main 49.03% 0 [ 4ef: 500] div [.] main 49.03% 0 [ 4ef: 51c] div [.] main 49.03% 0 [ 4ef: 535] div [.] main 18.82% 0 [ 3ac40: 3ac4d] libc-2.23.so [.] __random_r 18.82% 0 [ 3ac40: 3ac5c] libc-2.23.so [.] __random_r 18.82% 0 [ 3ac40: 3ac76] libc-2.23.so [.] __random_r 18.82% 0 [ 3ac40: 3ac88] libc-2.23.so [.] __random_r 18.82% 0 [ 3ac90: 3ac9c] libc-2.23.so [.] __random_r 16.29% -8 [ 3aac0: 3aac0] libc-2.23.so [.] __random 16.29% 0 [ 3aac0: 3aad2] libc-2.23.so [.] __random 16.29% 0 [ 3aae0: 3aae7] libc-2.23.so [.] __random 16.29% 0 [ 3ab03: 3ab0f] libc-2.23.so [.] __random 16.29% 0 [ 3ab14: 3ab1b] libc-2.23.so [.] __random 16.29% 0 [ 3ab28: 3ab2e] libc-2.23.so [.] __random 16.29% 0 [ 3ab4a: 3ab53] libc-2.23.so [.] __random 8.11% 0 [ 640: 644] div [.] compute_flag 8.11% 0 [ 649: 659] div [.] compute_flag 5.46% 0 [ 3af60: 3af60] libc-2.23.so [.] rand 5.46% 0 [ 3af60: 3af64] libc-2.23.so [.] rand 2.25% 0 [ 490: 490] div [.] rand@plt 0.01% 26 [ c00a27: c00a27] [kernel.vmlinux] [k] native_irq_return_iret 0.00% -157 [ 2bf9f2: 2bfa63] [kernel.vmlinux] [k] update_blocked_averages 0.00% -56 [ 2bf980: 2bf9d3] [kernel.vmlinux] [k] update_blocked_averages 0.00% 48 [ 2bf934: 2bf942] [kernel.vmlinux] [k] update_blocked_averages 0.00% 3 [ 2bfb38: 2bfb67] [kernel.vmlinux] [k] update_blocked_averagesThe 'cycles' is a new perf-diff computation selection, which enables the displaying of cycles difference of same program basic block amongst two perf.data. The program basic block is the code block between two branches in a function.
v2:
---
Keep standard perf diff format.Following is the v1 output.
# perf diff --basic-block# Cycles diff Basic block (start:end)
# ........... ....................... # -208 hrtimer_interrupt (30b9e0:30ba42) -157 update_blocked_averages (2bf9f2:2bfa63) -126 interrupt_entry (c00880:c0093a) -86 hrtimer_interrupt (30bb29:30bb32) -74 hrtimer_interrupt (30ba65:30bac4) -56 update_blocked_averages (2bf980:2bf9d3) 48 update_blocked_averages (2bf934:2bf942) -35 native_write_msr (267900:26790b) 26 native_irq_return_iret (c00a27:c00a27) 22 rcu_check_callbacks (2febb6:2febdc) -21 __hrtimer_run_queues (30b220:30b2a3) 19 pvclock_gtod_notify (14ba0:14c1b) -18 task_tick_fair (2c5d29:2c5d41)Jin Yao (7):
perf util: Createblock_infostructure
perf util: Addblock_infoinhist_entry
perf diff: Check if all data files with branch stacks perf diff: Use hists to manage basic blocks per symbol perf diff: Link same basic blocks among different data files perf diff: Print the basic block cycles diff
perf diff: Documentation -c cycles optiontools/perf/Documentation/perf-diff.txt | 14 +- tools/perf/builtin-diff.c | 373 ++++++++++++++++++++++++++++++++- tools/perf/ui/stdio/hist.c | 26 +++ tools/perf/util/hist.c | 42 +++- tools/perf/util/hist.h | 9 + tools/perf/util/sort.h | 8 + tools/perf/util/symbol.c | 22 ++ tools/perf/util/symbol.h | 23 ++ 8 files changed, 509 insertions(+), 8 deletions(-)
Quoted 6 lines · Jin Yao
In some cases small changes in hot loops can show big differences. But it's difficult to identify these differences.
perf diff currently can only diff symbols (functions). We can also expand it to diff cycles of individual programs blocks as reported by timed LBR. This would allow to identify changes in specific code accurately.
can't compile on Fedora 30
builtin-diff.c: In function ‘block_cycles_diff_cmp’:
builtin-diff.c:544:6: error: absolute value function ‘abs’ given an argument of type ‘s64’ {aka ‘long int’} but has parameter of type ‘int’ which may cause truncation of value [-Werror=absolute-value]
544 | l = abs(left->diff.cycles);
| ^~~
builtin-diff.c:545:6: error: absolute value function ‘abs’ given an argument of type ‘s64’ {aka ‘long int’} but has parameter of type ‘int’ which may cause truncation of value [-Werror=absolute-value]
545 | r = abs(right->diff.cycles);
| ^~~[jolsa@krava perf]$ gcc --version
gcc (GCC) 9.1.1 20190503 (Red Hat 9.1.1-1)
jirka
Quoted 92 lines
With this patch set, for example,
# perf record -b ./div # perf record -b ./div # perf diff -s cycles# Event 'cycles'
## Baseline Block cycles diff [start:end] Shared Object Symbol # ........ .................................... ................ .................................... # 49.03% -9 [ 4ef: 520] div [.] main 49.03% 0 [ 4e8: 4ea] div [.] main 49.03% 0 [ 4ef: 500] div [.] main 49.03% 0 [ 4ef: 51c] div [.] main 49.03% 0 [ 4ef: 535] div [.] main 18.82% 0 [ 3ac40: 3ac4d] libc-2.23.so [.] __random_r 18.82% 0 [ 3ac40: 3ac5c] libc-2.23.so [.] __random_r 18.82% 0 [ 3ac40: 3ac76] libc-2.23.so [.] __random_r 18.82% 0 [ 3ac40: 3ac88] libc-2.23.so [.] __random_r 18.82% 0 [ 3ac90: 3ac9c] libc-2.23.so [.] __random_r 16.29% -8 [ 3aac0: 3aac0] libc-2.23.so [.] __random 16.29% 0 [ 3aac0: 3aad2] libc-2.23.so [.] __random 16.29% 0 [ 3aae0: 3aae7] libc-2.23.so [.] __random 16.29% 0 [ 3ab03: 3ab0f] libc-2.23.so [.] __random 16.29% 0 [ 3ab14: 3ab1b] libc-2.23.so [.] __random 16.29% 0 [ 3ab28: 3ab2e] libc-2.23.so [.] __random 16.29% 0 [ 3ab4a: 3ab53] libc-2.23.so [.] __random 8.11% 0 [ 640: 644] div [.] compute_flag 8.11% 0 [ 649: 659] div [.] compute_flag 5.46% 0 [ 3af60: 3af60] libc-2.23.so [.] rand 5.46% 0 [ 3af60: 3af64] libc-2.23.so [.] rand 2.25% 0 [ 490: 490] div [.] rand@plt 0.01% 26 [ c00a27: c00a27] [kernel.vmlinux] [k] native_irq_return_iret 0.00% -157 [ 2bf9f2: 2bfa63] [kernel.vmlinux] [k] update_blocked_averages 0.00% -56 [ 2bf980: 2bf9d3] [kernel.vmlinux] [k] update_blocked_averages 0.00% 48 [ 2bf934: 2bf942] [kernel.vmlinux] [k] update_blocked_averages 0.00% 3 [ 2bfb38: 2bfb67] [kernel.vmlinux] [k] update_blocked_averagesThe 'cycles' is a new perf-diff computation selection, which enables the displaying of cycles difference of same program basic block amongst two perf.data. The program basic block is the code block between two branches in a function.
v2:
---
Keep standard perf diff format.Following is the v1 output.
# perf diff --basic-block# Cycles diff Basic block (start:end)
# ........... ....................... # -208 hrtimer_interrupt (30b9e0:30ba42) -157 update_blocked_averages (2bf9f2:2bfa63) -126 interrupt_entry (c00880:c0093a) -86 hrtimer_interrupt (30bb29:30bb32) -74 hrtimer_interrupt (30ba65:30bac4) -56 update_blocked_averages (2bf980:2bf9d3) 48 update_blocked_averages (2bf934:2bf942) -35 native_write_msr (267900:26790b) 26 native_irq_return_iret (c00a27:c00a27) 22 rcu_check_callbacks (2febb6:2febdc) -21 __hrtimer_run_queues (30b220:30b2a3) 19 pvclock_gtod_notify (14ba0:14c1b) -18 task_tick_fair (2c5d29:2c5d41)Jin Yao (7):
perf util: Createblock_infostructure
perf util: Addblock_infoinhist_entry
perf diff: Check if all data files with branch stacks perf diff: Use hists to manage basic blocks per symbol perf diff: Link same basic blocks among different data files perf diff: Print the basic block cycles diff
perf diff: Documentation -c cycles optiontools/perf/Documentation/perf-diff.txt | 14 +- tools/perf/builtin-diff.c | 373 ++++++++++++++++++++++++++++++++- tools/perf/ui/stdio/hist.c | 26 +++ tools/perf/util/hist.c | 42 +++- tools/perf/util/hist.h | 9 + tools/perf/util/sort.h | 8 + tools/perf/util/symbol.c | 22 ++ tools/perf/util/symbol.h | 23 ++ 8 files changed, 509 insertions(+), 8 deletions(-)Signature · 1 line
2.7.4
Quoted 18 lines · Jiri Olsa
Quoted 6 lines · Jin Yao
In some cases small changes in hot loops can show big differences. But it's difficult to identify these differences.
perf diff currently can only diff symbols (functions). We can also expand it to diff cycles of individual programs blocks as reported by timed LBR. This would allow to identify changes in specific code accurately.
can't compile on Fedora 30
builtin-diff.c: In function ‘block_cycles_diff_cmp’:
builtin-diff.c:544:6: error: absolute value function ‘abs’ given an argument of type ‘s64’ {aka ‘long int’} but has parameter of type ‘int’ which may cause truncation of value [-Werror=absolute-value]544 | l = abs(left->diff.cycles); | ^~~ builtin-diff.c:545:6: error: absolute value function ‘abs’ given an argument of type ‘s64’ {aka ‘long int’} but has parameter of type ‘int’ which may cause truncation of value [-Werror=absolute-value] 545 | r = abs(right->diff.cycles); | ^~~
Quoted 5 lines
[jolsa@krava perf]$ gcc --version
gcc (GCC) 9.1.1 20190503 (Red Hat 9.1.1-1)jirka
Thanks for pointing out this issue.
I should use labs()
Thanks
Jin Yao
Quoted 92 lines
Quoted 92 lines
With this patch set, for example,
# perf record -b ./div # perf record -b ./div # perf diff -s cycles# Event 'cycles'
## Baseline Block cycles diff [start:end] Shared Object Symbol # ........ .................................... ................ .................................... # 49.03% -9 [ 4ef: 520] div [.] main 49.03% 0 [ 4e8: 4ea] div [.] main 49.03% 0 [ 4ef: 500] div [.] main 49.03% 0 [ 4ef: 51c] div [.] main 49.03% 0 [ 4ef: 535] div [.] main 18.82% 0 [ 3ac40: 3ac4d] libc-2.23.so [.] __random_r 18.82% 0 [ 3ac40: 3ac5c] libc-2.23.so [.] __random_r 18.82% 0 [ 3ac40: 3ac76] libc-2.23.so [.] __random_r 18.82% 0 [ 3ac40: 3ac88] libc-2.23.so [.] __random_r 18.82% 0 [ 3ac90: 3ac9c] libc-2.23.so [.] __random_r 16.29% -8 [ 3aac0: 3aac0] libc-2.23.so [.] __random 16.29% 0 [ 3aac0: 3aad2] libc-2.23.so [.] __random 16.29% 0 [ 3aae0: 3aae7] libc-2.23.so [.] __random 16.29% 0 [ 3ab03: 3ab0f] libc-2.23.so [.] __random 16.29% 0 [ 3ab14: 3ab1b] libc-2.23.so [.] __random 16.29% 0 [ 3ab28: 3ab2e] libc-2.23.so [.] __random 16.29% 0 [ 3ab4a: 3ab53] libc-2.23.so [.] __random 8.11% 0 [ 640: 644] div [.] compute_flag 8.11% 0 [ 649: 659] div [.] compute_flag 5.46% 0 [ 3af60: 3af60] libc-2.23.so [.] rand 5.46% 0 [ 3af60: 3af64] libc-2.23.so [.] rand 2.25% 0 [ 490: 490] div [.] rand@plt 0.01% 26 [ c00a27: c00a27] [kernel.vmlinux] [k] native_irq_return_iret 0.00% -157 [ 2bf9f2: 2bfa63] [kernel.vmlinux] [k] update_blocked_averages 0.00% -56 [ 2bf980: 2bf9d3] [kernel.vmlinux] [k] update_blocked_averages 0.00% 48 [ 2bf934: 2bf942] [kernel.vmlinux] [k] update_blocked_averages 0.00% 3 [ 2bfb38: 2bfb67] [kernel.vmlinux] [k] update_blocked_averagesThe 'cycles' is a new perf-diff computation selection, which enables the displaying of cycles difference of same program basic block amongst two perf.data. The program basic block is the code block between two branches in a function.
v2:
---
Keep standard perf diff format.Following is the v1 output.
# perf diff --basic-block# Cycles diff Basic block (start:end)
# ........... ....................... # -208 hrtimer_interrupt (30b9e0:30ba42) -157 update_blocked_averages (2bf9f2:2bfa63) -126 interrupt_entry (c00880:c0093a) -86 hrtimer_interrupt (30bb29:30bb32) -74 hrtimer_interrupt (30ba65:30bac4) -56 update_blocked_averages (2bf980:2bf9d3) 48 update_blocked_averages (2bf934:2bf942) -35 native_write_msr (267900:26790b) 26 native_irq_return_iret (c00a27:c00a27) 22 rcu_check_callbacks (2febb6:2febdc) -21 __hrtimer_run_queues (30b220:30b2a3) 19 pvclock_gtod_notify (14ba0:14c1b) -18 task_tick_fair (2c5d29:2c5d41)Jin Yao (7):
perf util: Createblock_infostructure
perf util: Addblock_infoinhist_entry
perf diff: Check if all data files with branch stacks perf diff: Use hists to manage basic blocks per symbol perf diff: Link same basic blocks among different data files perf diff: Print the basic block cycles diff
perf diff: Documentation -c cycles optiontools/perf/Documentation/perf-diff.txt | 14 +- tools/perf/builtin-diff.c | 373 ++++++++++++++++++++++++++++++++- tools/perf/ui/stdio/hist.c | 26 +++ tools/perf/util/hist.c | 42 +++- tools/perf/util/hist.h | 9 + tools/perf/util/sort.h | 8 + tools/perf/util/symbol.c | 22 ++ tools/perf/util/symbol.h | 23 ++ 8 files changed, 509 insertions(+), 8 deletions(-)Signature · 1 line
2.7.4