perf report: Implement visual marker for macro fusion in annotate
cover perf report: Implement visual marker for macro fusion in annotate
Macro fusion merges two instructions to a single micro-op. Intel core platform performs this hardware optimization under limited circumstances. For example, CMP + JCC can be "fused" and executed /retired together. While with sampling this can result in the sample sometimes being on the JCC and sometimes on the CMP. So for the fused instruction pair, they could be considered together.
In general, the fused instruction pairs are:
cmp/test/add/sub/and/inc/dec + jcc.
This patch series marks the case clearly by joining the fused instruction pair in the arrow of the jump.
For example:
│ ┌──cmpl $0x0,argp_program_version_hook
81.93 │ │──je 20
│ │ lock cmpxchg %esi,0x38a9a4(%rip)
│ │↓ jne 29
│ │↓ jmp 43
11.47 │20:└─→cmpxch %esi,0x38a999(%rip)Jin Yao (2):
perf report: Check for fused instruction pair
perf report: Implement visual marker for macro fusion in annotate
tools/perf/arch/x86/util/Build | 1 + tools/perf/arch/x86/util/fused.c | 20 ++++++++++++++++++++ tools/perf/ui/browser.c | 27 +++++++++++++++++++++++++++ tools/perf/ui/browser.h | 2 ++ tools/perf/ui/browsers/annotate.c | 30 ++++++++++++++++++++++++++++++ tools/perf/util/Build | 1 + tools/perf/util/annotate.c | 5 +++++ tools/perf/util/annotate.h | 1 + tools/perf/util/fused.c | 11 +++++++++++ tools/perf/util/fused.h | 8 ++++++++ 10 files changed, 106 insertions(+) create mode 100644 tools/perf/arch/x86/util/fused.c create mode 100644 tools/perf/util/fused.c create mode 100644 tools/perf/util/fused.h
Signature · 1 line
2.7.4
PATCH 1/2 perf report: Check for fused instruction pair
Macro fusion merges two instructions to a single micro-op. Intel core platform performs this hardware optimization under limited circumstances. For example, CMP + JCC can be "fused" and executed /retired together. While with sampling this can result in the sample sometimes being on the JCC and sometimes on the CMP. So for the fused instruction pair, they could be considered together.
In general, the fused instruction pairs are:
cmp/test/add/sub/and/inc/dec + jcc.
This patch adds a new function which checks if 2 x86 instructions are in a "fused" pair. For non-x86 arch, the function just returns false.
--- tools/perf/arch/x86/util/Build | 1 + tools/perf/arch/x86/util/fused.c | 20 ++++++++++++++++++++ tools/perf/util/Build | 1 + tools/perf/util/fused.c | 11 +++++++++++ tools/perf/util/fused.h | 8 ++++++++ 5 files changed, 41 insertions(+) create mode 100644 tools/perf/arch/x86/util/fused.c create mode 100644 tools/perf/util/fused.c create mode 100644 tools/perf/util/fused.h
tools/perf/arch/x86/util/Build +1 -0
@@ -4,6 +4,7 @@ libperf-y += pmu.o libperf-y += kvm-stat.o libperf-y += perf_regs.o libperf-y += group.o +libperf-y += fused.o libperf-$(CONFIG_DWARF) += dwarf-regs.o libperf-$(CONFIG_BPF_PROLOGUE) += dwarf-regs.o
tools/perf/arch/x86/util/fused.c new file +20 -0
@@ -0,0 +1,20 @@ +#include <string.h> +#include "../../util/fused.h" + +bool fused_insn_pair (const char *insn1,const char *insn2) +{ +if (strstr (insn2,"jmp" )) +return false; + +if ((strstr (insn1,"cmp" ) && !strstr (insn1,"xchg" )) || +strstr (insn1,"test" ) || +strstr (insn1,"add" ) || +strstr (insn1,"sub" ) || +strstr (insn1,"and" ) || +strstr (insn1,"inc" ) || +strstr (insn1,"dec" )) { +return true; + } + +return false; +}
tools/perf/util/Build +1 -0
@@ -93,6 +93,7 @@ libperf-y += drv_configs.o libperf-y += units.o libperf-y += time-utils.o libperf-y += expr-bison.o +libperf-y += fused.o libperf-$(CONFIG_LIBBPF) += bpf-loader.o libperf-$(CONFIG_BPF_PROLOGUE) += bpf-prologue.o
tools/perf/util/fused.c new file +11 -0
@@ -0,0 +1,11 @@ +#include <linux/compiler.h> +#include <linux/types.h> +#include <string.h> + +#include "fused.h" + +bool __weak fused_insn_pair (const char *insn1__maybe_unused , +const char *insn2__maybe_unused ) +{ +return false; +}
tools/perf/util/fused.h new file +8 -0
@@ -0,0 +1,8 @@ +#ifndef __PERF_FUSED_H +#define __PERF_FUSED_H + +#include <linux/types.h> + +bool fused_insn_pair (const char *insn1,const char *insn2); + +#endif /* __PERF_FUSED_H */
Signature · 1 line
2.7.4
Arnaldo Carvalho de Melo
maintainerRe: [PATCH v1 1/2] perf report: Check for fused instruction pair
Em Wed, Jun 14, 2017 at 10:53:40AM +0800, Jin Yao escreveu:
Quoted 7 lines
Macro fusion merges two instructions to a single micro-op. Intel core platform performs this hardware optimization under limited circumstances. For example, CMP + JCC can be "fused" and executed /retired together. While with sampling this can result in the sample sometimes being on the JCC and sometimes on the CMP. So for the fused instruction pair, they could be considered together.
doing it as a weak function that will be overriden by the host arch doesn't work, as we also support cross-annotation. So you have to take into account perf_evsel__env_arch(evsel), etc.
Please search for perf_evsel__env_arch(evsel) in the annotation source files to see how it is used.
- Arnaldo
Quoted 103 lines
In general, the fused instruction pairs are:
cmp/test/add/sub/and/inc/dec + jcc.
This patch adds a new function which checks if 2 x86 instructions are in a "fused" pair. For non-x86 arch, the function just returns false.
Signed-off-by Jin Yao <[email protected]>--- tools/perf/arch/x86/util/Build | 1 + tools/perf/arch/x86/util/fused.c | 20 ++++++++++++++++++++ tools/perf/util/Build | 1 + tools/perf/util/fused.c | 11 +++++++++++ tools/perf/util/fused.h | 8 ++++++++ 5 files changed, 41 insertions(+) create mode 100644 tools/perf/arch/x86/util/fused.c create mode 100644 tools/perf/util/fused.c create mode 100644 tools/perf/util/fused.htools/perf/arch/x86/util/Build +1
-0@@ -4,6 +4,7 @@ libperf-y += pmu.o libperf-y += kvm-stat.o libperf-y += perf_regs.o libperf-y += group.o +libperf-y += fused.o libperf-$(CONFIG_DWARF) += dwarf-regs.o libperf-$(CONFIG_BPF_PROLOGUE) += dwarf-regs.otools/perf/arch/x86/util/fused.c new file +20
-0@@ -0,0 +1,20 @@ +#include <string.h> +#include "../../util/fused.h" + +bool fused_insn_pair (const char *insn1,const char *insn2) +{ +if (strstr (insn2,"jmp" )) +return false; + +if ((strstr (insn1,"cmp" ) && !strstr (insn1,"xchg" )) || +strstr (insn1,"test" ) || +strstr (insn1,"add" ) || +strstr (insn1,"sub" ) || +strstr (insn1,"and" ) || +strstr (insn1,"inc" ) || +strstr (insn1,"dec" )) { +return true; + } + +return false; +}tools/perf/util/Build +1
-0@@ -93,6 +93,7 @@ libperf-y += drv_configs.o libperf-y += units.o libperf-y += time-utils.o libperf-y += expr-bison.o +libperf-y += fused.o libperf-$(CONFIG_LIBBPF) += bpf-loader.o libperf-$(CONFIG_BPF_PROLOGUE) += bpf-prologue.otools/perf/util/fused.c new file +11
-0@@ -0,0 +1,11 @@ +#include <linux/compiler.h> +#include <linux/types.h> +#include <string.h> + +#include "fused.h" + +bool __weak fused_insn_pair (const char *insn1__maybe_unused , +const char *insn2__maybe_unused ) +{ +return false; +}tools/perf/util/fused.h new file +8
-0@@ -0,0 +1,8 @@ +#ifndef __PERF_FUSED_H +#define __PERF_FUSED_H + +#include <linux/types.h> + +bool fused_insn_pair (const char *insn1,const char *insn2); + +#endif /* __PERF_FUSED_H */ Signature · 1 line
2.7.4
Re: [PATCH v1 1/2] perf report: Check for fused instruction pair
Quoted 8 lines · Arnaldo Carvalho de Melo
Em Wed, Jun 14, 2017 at 10:53:40AM +0800, Jin Yao escreveu:
Quoted 7 lines
Macro fusion merges two instructions to a single micro-op. Intel core platform performs this hardware optimization under limited circumstances. For example, CMP + JCC can be "fused" and executed /retired together. While with sampling this can result in the sample sometimes being on the JCC and sometimes on the CMP. So for the fused instruction pair, they could be considered together.
Quoted 9 lines
doing it as a weak function that will be overriden by the host arch doesn't work, as we also support cross-annotation. So you have to take into account
perf_evsel__env_arch(evsel), etc.Please search for
perf_evsel__env_arch(evsel)in the annotation source files to see how it is used.- Arnaldo
Hi Arnaldo,
Thanks so much for pointing out that the weak function doesn't work.
I have changed it to arch-specific function and just send out v2 series for reviewing.
Thanks
Jin Yao
PATCH 2/2 perf report: Implement visual marker for macro fusion in annotate
In reply to Jin YaoFor marking the fused instructions clearly, This patch adds a line before the first instruction of pair and joins it with the arrow of the jump.
For example, when je is selected in annotate view, the line before cmpl is displayed and joins the arrow of je.
│ ┌──cmpl $0x0,argp_program_version_hook
81.93 │ │──je 20
│ │ lock cmpxchg %esi,0x38a9a4(%rip)
│ │↓ jne 29
│ │↓ jmp 43
11.47 │20:└─→cmpxch %esi,0x38a999(%rip)That means the cmpl+je is fused instruction pair and they should be considered together.
--- tools/perf/ui/browser.c | 27 +++++++++++++++++++++++++++ tools/perf/ui/browser.h | 2 ++ tools/perf/ui/browsers/annotate.c | 30 ++++++++++++++++++++++++++++++ tools/perf/util/annotate.c | 5 +++++ tools/perf/util/annotate.h | 1 + 5 files changed, 65 insertions(+)
tools/perf/ui/browser.c +27 -0
@@ -738,6 +738,33 @@ void __ui_browser__line_arrow(struct ui_browser *browser, unsigned int column,__ui_browser__line_arrow_down (browser, column, start, end); } +void ui_browser__mark_fused (struct ui_browser *browser,unsigned int column, +unsigned int row,bool arrow_down) +{ +unsigned int end_row; + +if (row >= browser->top_idx) + end_row = row - browser->top_idx; +else +return ; + +SLsmg_set_char_set (1 ); + +if (arrow_down) { +ui_browser__gotorc (browser, end_row, column -1 ); +SLsmg_write_char (SLSMG_ULCORN_CHAR ); +ui_browser__gotorc (browser, end_row, column); +SLsmg_draw_hline (2 ); +ui_browser__gotorc (browser, end_row +1 , column -1 ); +SLsmg_draw_vline (1 ); + }else { +ui_browser__gotorc (browser, end_row, column); +SLsmg_draw_hline (2 ); + } + +SLsmg_set_char_set (0 ); +} +void ui_browser__init (void ) {int i =0 ;
tools/perf/ui/browser.h +2 -0
@@ -43,6 +43,8 @@ void ui_browser__printf(struct ui_browser *browser, const char *fmt, ...);void ui_browser__write_graph (struct ui_browser *browser,int graph);void __ui_browser__line_arrow (struct ui_browser *browser,unsigned int column,u64 start,u64 end); +void ui_browser__mark_fused (struct ui_browser *browser,unsigned int column, +unsigned int row,bool arrow_down);void __ui_browser__show_title (struct ui_browser *browser,const char *title);void ui_browser__show_title (struct ui_browser *browser,const char *title); int ui_browser__show(struct ui_browser *browser, const char *title,
tools/perf/ui/browsers/annotate.c +30 -0
@@ -9,6 +9,7 @@#include "../../util/symbol.h" #include "../../util/evsel.h" #include "../../util/config.h" +#include "../../util/fused.h" #include <inttypes.h> #include <pthread.h> #include <linux/kernel.h>
@@ -269,6 +270,28 @@ static bool disasm_line__is_valid_jump(struct disasm_line *dl, struct symbol *syreturn true; } +static bool is_fused (struct disasm_line *cursor) +{ +struct disasm_line *pos =list_prev_entry (cursor, node); +const char *name; + +if (!pos) +return false; + +if (ins__is_lock (&pos->ins)) + name = pos->ops.locked.ins.name; +else + name = pos->ins.name; + +if (!name || !cursor->ins.name) +return false; + +if (fused_insn_pair (name, cursor->ins.name)) +return true; + +return false; +} +static void annotate_browser__draw_current_jump (struct ui_browser *browser) { struct annotate_browser *ab = container_of(browser, struct annotate_browser, b);
@@ -304,6 +327,13 @@ static void annotate_browser__draw_current_jump(struct ui_browser *browser)ui_browser__set_color (browser,HE_COLORSET_JUMP_ARROWS );__ui_browser__line_arrow (browser, pcnt_width +2 + ab->addr_width, from, to); + +if (is_fused (cursor)) { +ui_browser__mark_fused (browser, + pcnt_width +3 + ab->addr_width, + from -1 , + to > from ? true : false); + } }static unsigned int annotate_browser__refresh (struct ui_browser *browser)
tools/perf/util/annotate.c +5 -0
@@ -502,6 +502,11 @@ bool ins__is_ret(const struct ins *ins)return ins->ops == &ret_ops; } +bool ins__is_lock (const struct ins *ins) +{ +return ins->ops == &lock_ops; +} +static int ins__key_cmp (const void *name,const void *insp) {const struct ins *ins = insp;
tools/perf/util/annotate.h +1 -0
@@ -52,6 +52,7 @@ struct ins_ops {bool ins__is_jump (const struct ins *ins);bool ins__is_call (const struct ins *ins);bool ins__is_ret (const struct ins *ins); +bool ins__is_lock (const struct ins *ins);int ins__scnprintf (struct ins *ins,char *bf,size_t size,struct ins_operands *ops);struct annotation ;
Signature · 1 line
2.7.4
Arnaldo Carvalho de Melo
maintainerEm Wed, Jun 14, 2017 at 10:53:39AM +0800, Jin Yao escreveu:
Quoted 15 lines
Macro fusion merges two instructions to a single micro-op. Intel core platform performs this hardware optimization under limited circumstances. For example, CMP + JCC can be "fused" and executed /retired together. While with sampling this can result in the sample sometimes being on the JCC and sometimes on the CMP. So for the fused instruction pair, they could be considered together.
In general, the fused instruction pairs are:
cmp/test/add/sub/and/inc/dec + jcc.
This patch series marks the case clearly by joining the fused instruction pair in the arrow of the jump.
Quoted 8 lines
For example:
│ ┌──cmpl $0x0,argp_program_version_hook 81.93 │ │──je 20 │ │ lock cmpxchg %esi,0x38a9a4(%rip) │ │↓ jne 29 │ │↓ jmp 43 11.47 │20:└─→cmpxch %esi,0x38a999(%rip)
Try to have these example outputs in the changesets, not just in the patch series header.
- Arnaldo
Quoted 21 lines
Jin Yao (2):
perf report: Check for fused instruction pair
perf report: Implement visual marker for macro fusion in annotatetools/perf/arch/x86/util/Build | 1 + tools/perf/arch/x86/util/fused.c | 20 ++++++++++++++++++++ tools/perf/ui/browser.c | 27 +++++++++++++++++++++++++++ tools/perf/ui/browser.h | 2 ++ tools/perf/ui/browsers/annotate.c | 30 ++++++++++++++++++++++++++++++ tools/perf/util/Build | 1 + tools/perf/util/annotate.c | 5 +++++ tools/perf/util/annotate.h | 1 + tools/perf/util/fused.c | 11 +++++++++++ tools/perf/util/fused.h | 8 ++++++++ 10 files changed, 106 insertions(+) create mode 100644 tools/perf/arch/x86/util/fused.c create mode 100644 tools/perf/util/fused.c create mode 100644 tools/perf/util/fused.hSignature · 1 line
2.7.4
Arnaldo Carvalho de Melo
maintainerEm Fri, Jun 16, 2017 at 01:16:55PM -0300, Arnaldo Carvalho de Melo escreveu:
Quoted 25 lines
Em Wed, Jun 14, 2017 at 10:53:39AM +0800, Jin Yao escreveu:
Quoted 15 lines
Macro fusion merges two instructions to a single micro-op. Intel core platform performs this hardware optimization under limited circumstances. For example, CMP + JCC can be "fused" and executed /retired together. While with sampling this can result in the sample sometimes being on the JCC and sometimes on the CMP. So for the fused instruction pair, they could be considered together.
In general, the fused instruction pairs are:
cmp/test/add/sub/and/inc/dec + jcc.
This patch series marks the case clearly by joining the fused instruction pair in the arrow of the jump.
Quoted 8 lines
For example:
│ ┌──cmpl $0x0,argp_program_version_hook 81.93 │ │──je 20 │ │ lock cmpxchg %esi,0x38a9a4(%rip) │ │↓ jne 29 │ │↓ jmp 43 11.47 │20:└─→cmpxch %esi,0x38a999(%rip)
Quoted 2 lines
Try to have these example outputs in the changesets, not just in the patch series header.
Ok, I went trigger happy, sorry, it is in the second patch, I had looked just at the first :-\
- Arnaldo
Quoted 23 lines
- Arnaldo
Quoted 21 lines
Jin Yao (2):
perf report: Check for fused instruction pair
perf report: Implement visual marker for macro fusion in annotatetools/perf/arch/x86/util/Build | 1 + tools/perf/arch/x86/util/fused.c | 20 ++++++++++++++++++++ tools/perf/ui/browser.c | 27 +++++++++++++++++++++++++++ tools/perf/ui/browser.h | 2 ++ tools/perf/ui/browsers/annotate.c | 30 ++++++++++++++++++++++++++++++ tools/perf/util/Build | 1 + tools/perf/util/annotate.c | 5 +++++ tools/perf/util/annotate.h | 1 + tools/perf/util/fused.c | 11 +++++++++++ tools/perf/util/fused.h | 8 ++++++++ 10 files changed, 106 insertions(+) create mode 100644 tools/perf/arch/x86/util/fused.c create mode 100644 tools/perf/util/fused.c create mode 100644 tools/perf/util/fused.hSignature · 1 line
2.7.4