Some x86 microarchitectures fuse a subset of cmp/test/ALU instructions
with branch instructions, and thus perf annotate highlight such valid
pairs as fused.
When annotated with source, perf uses struct disasm_line to contain
either source or instruction line from objdump output. Usually, a C
statement generates multiple instructions which include such
cmp/test/ALU + branch instruction pairs. But in case of assembly
function, each individual assembly source line generate one
instruction. Perf annotate instruction fusion logic assumes previous
disasm_line as previous instruction line, which is wrong because,
for assembly function, previous disasm_line contains source line.
And thus perf fails to highlight valid fused instruction pairs for
assembly functions.
Fix it by searching backward until we find an instruction line and
consider that disasm_line as fused with current branch instruction.
Before:
│ cmpq %rcx, RIP+8(%rsp)
0.00 │ cmp %rcx,0x88(%rsp)
│ je .Lerror_bad_iret <--- Source line
0.14 │ ┌──je b4 <--- Instruction line
│ │movl %ecx, %eax
After:
│ cmpq %rcx, RIP+8(%rsp)
0.00 │ ┌──cmp %rcx,0x88(%rsp)
│ │je .Lerror_bad_iret
0.14 │ ├──je b4
│ │movl %ecx, %eax
Signed-off-by: Ravi Bangoria <[email protected]>
---
tools/perf/ui/browser.c | 33 ++++++++++++++++++++++---------
tools/perf/ui/browser.h | 2 +-
tools/perf/ui/browsers/annotate.c | 24 +++++++++++++++-------
3 files changed, 42 insertions(+), 17 deletions(-)
diff --git a/tools/perf/ui/browser.c b/tools/perf/ui/browser.c
index 781afe42e90e..fa5bd5c20e96 100644
--- a/tools/perf/ui/browser.c
+++ b/tools/perf/ui/browser.c
@@ -757,25 +757,40 @@ void __ui_browser__line_arrow(struct ui_browser *browser, unsigned int column,
}
void ui_browser__mark_fused(struct ui_browser *browser, unsigned int column,
- unsigned int row, bool arrow_down)
+ unsigned int row, int diff, bool arrow_down)
{
- unsigned int end_row;
+ int end_row;
- if (row >= browser->top_idx)
- end_row = row - browser->top_idx;
- else
+ if (diff <= 0)
return;
SLsmg_set_char_set(1);
if (arrow_down) {
+ if (row + diff <= browser->top_idx)
+ return;
+
+ end_row = row + diff - browser->top_idx;
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_write_char(SLSMG_LTEE_CHAR);
+
+ while (--end_row >= 0 && end_row > (int)(row - browser->top_idx)) {
+ ui_browser__gotorc(browser, end_row, column - 1);
+ SLsmg_draw_vline(1);
+ }
+
+ end_row = (int)(row - browser->top_idx);
+ if (end_row >= 0) {
+ 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);
+ }
} else {
+ if (row < browser->top_idx)
+ return;
+
+ end_row = row - browser->top_idx;
ui_browser__gotorc(browser, end_row, column - 1);
SLsmg_write_char(SLSMG_LTEE_CHAR);
ui_browser__gotorc(browser, end_row, column);
diff --git a/tools/perf/ui/browser.h b/tools/perf/ui/browser.h
index 3678eb88f119..510ce4554050 100644
--- a/tools/perf/ui/browser.h
+++ b/tools/perf/ui/browser.h
@@ -51,7 +51,7 @@ 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);
+ unsigned int row, int diff, 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,
diff --git a/tools/perf/ui/browsers/annotate.c b/tools/perf/ui/browsers/annotate.c
index ef4da4295bf7..e81c2493efdf 100644
--- a/tools/perf/ui/browsers/annotate.c
+++ b/tools/perf/ui/browsers/annotate.c
@@ -125,13 +125,20 @@ static void annotate_browser__write(struct ui_browser *browser, void *entry, int
ab->selection = al;
}
-static bool is_fused(struct annotate_browser *ab, struct disasm_line *cursor)
+static int is_fused(struct annotate_browser *ab, struct disasm_line *cursor)
{
struct disasm_line *pos = list_prev_entry(cursor, al.node);
const char *name;
+ int diff = 1;
+
+ while (pos && pos->al.offset == -1) {
+ pos = list_prev_entry(pos, al.node);
+ if (!ab->opts->hide_src_code)
+ diff++;
+ }
if (!pos)
- return false;
+ return 0;
if (ins__is_lock(&pos->ins))
name = pos->ops.locked.ins.name;
@@ -139,9 +146,11 @@ static bool is_fused(struct annotate_browser *ab, struct disasm_line *cursor)
name = pos->ins.name;
if (!name || !cursor->ins.name)
- return false;
+ return 0;
- return ins__is_fused(ab->arch, name, cursor->ins.name);
+ if (ins__is_fused(ab->arch, name, cursor->ins.name))
+ return diff;
+ return 0;
}
static void annotate_browser__draw_current_jump(struct ui_browser *browser)
@@ -155,6 +164,7 @@ static void annotate_browser__draw_current_jump(struct ui_browser *browser)
struct annotation *notes = symbol__annotation(sym);
u8 pcnt_width = annotation__pcnt_width(notes);
int width;
+ int diff = 0;
/* PLT symbols contain external offsets */
if (strstr(sym->name, "@plt"))
@@ -205,11 +215,11 @@ static void annotate_browser__draw_current_jump(struct ui_browser *browser)
pcnt_width + 2 + notes->widths.addr + width,
from, to);
- if (is_fused(ab, cursor)) {
+ diff = is_fused(ab, cursor);
+ if (diff > 0) {
ui_browser__mark_fused(browser,
pcnt_width + 3 + notes->widths.addr + width,
- from - 1,
- to > from);
+ from - diff, diff, to > from);
}
}
--
2.27.0
AMD family 15h and above microarchs fuse a subset of cmp/test/ALU
instructions with branch instructions[1][2]. Add perf annotate
fused instruction support for these microarchs.
Before:
│ testb $0x80,0x51(%rax)
│ ┌──jne 5b3
0.78 │ │ mov %r13,%rdi
│ │→ callq mark_page_accessed
1.08 │5b3:└─→mov 0x8(%r13),%rax
After:
│ ┌──testb $0x80,0x51(%rax)
│ ├──jne 5b3
0.78 │ │ mov %r13,%rdi
│ │→ callq mark_page_accessed
1.08 │5b3:└─→mov 0x8(%r13),%rax
[1] https://bugzilla.kernel.org/attachment.cgi?id=298553
[2] https://bugzilla.kernel.org/attachment.cgi?id=298555
Reported-by: Kim Phillips <[email protected]>
Signed-off-by: Ravi Bangoria <[email protected]>
---
tools/perf/arch/x86/annotate/instructions.c | 37 ++++++++++++++++++++-
tools/perf/util/annotate.c | 1 +
2 files changed, 37 insertions(+), 1 deletion(-)
diff --git a/tools/perf/arch/x86/annotate/instructions.c b/tools/perf/arch/x86/annotate/instructions.c
index 24ea12ec7e02..46d7124cc4e1 100644
--- a/tools/perf/arch/x86/annotate/instructions.c
+++ b/tools/perf/arch/x86/annotate/instructions.c
@@ -144,8 +144,31 @@ static struct ins x86__instructions[] = {
{ .name = "xorps", .ops = &mov_ops, },
};
-static bool x86__ins_is_fused(struct arch *arch, const char *ins1,
+static bool amd__ins_is_fused(struct arch *arch, const char *ins1,
const char *ins2)
+{
+ if (strstr(ins2, "jmp"))
+ return false;
+
+ /* Family >= 15h supports cmp/test + branch fusion */
+ if (arch->family >= 0x15 && (strstarts(ins1, "test") ||
+ (strstarts(ins1, "cmp") && !strstr(ins1, "xchg")))) {
+ return true;
+ }
+
+ /* Family >= 19h supports some ALU + branch fusion */
+ if (arch->family >= 0x19 && (strstarts(ins1, "add") ||
+ strstarts(ins1, "sub") || strstarts(ins1, "and") ||
+ strstarts(ins1, "inc") || strstarts(ins1, "dec") ||
+ strstarts(ins1, "or") || strstarts(ins1, "xor"))) {
+ return true;
+ }
+
+ return false;
+}
+
+static bool intel__ins_is_fused(struct arch *arch, const char *ins1,
+ const char *ins2)
{
if (arch->family != 6 || arch->model < 0x1e || strstr(ins2, "jmp"))
return false;
@@ -172,6 +195,15 @@ static bool x86__ins_is_fused(struct arch *arch, const char *ins1,
return false;
}
+static bool x86__ins_is_fused(struct arch *arch, const char *ins1,
+ const char *ins2)
+{
+ if (strstarts(arch->vendor, "AuthenticAMD"))
+ return amd__ins_is_fused(arch, ins1, ins2);
+
+ return intel__ins_is_fused(arch, ins1, ins2);
+}
+
static int x86__cpuid_parse(struct arch *arch, char *cpuid)
{
unsigned int family, model, stepping;
@@ -184,6 +216,9 @@ static int x86__cpuid_parse(struct arch *arch, char *cpuid)
if (ret == 3) {
arch->family = family;
arch->model = model;
+ arch->vendor = strndup(cpuid, 12);
+ if (!arch->vendor)
+ return -1;
return 0;
}
diff --git a/tools/perf/util/annotate.c b/tools/perf/util/annotate.c
index 0bae061b2d6d..88326bb990b5 100644
--- a/tools/perf/util/annotate.c
+++ b/tools/perf/util/annotate.c
@@ -77,6 +77,7 @@ struct arch {
bool sorted_instructions;
bool initialized;
void *priv;
+ char *vendor;
unsigned int model;
unsigned int family;
int (*init)(struct arch *arch, char *cpuid);
--
2.27.0
Em Mon, Sep 06, 2021 at 04:26:40PM +0530, Ravi Bangoria escreveu:
> AMD family 15h and above microarchs fuse a subset of cmp/test/ALU
> instructions with branch instructions[1][2]. Add perf annotate
> fused instruction support for these microarchs.
>
> Before:
> │ testb $0x80,0x51(%rax)
> │ ┌──jne 5b3
> 0.78 │ │ mov %r13,%rdi
> │ │→ callq mark_page_accessed
> 1.08 │5b3:└─→mov 0x8(%r13),%rax
>
> After:
> │ ┌──testb $0x80,0x51(%rax)
> │ ├──jne 5b3
> 0.78 │ │ mov %r13,%rdi
> │ │→ callq mark_page_accessed
> 1.08 │5b3:└─→mov 0x8(%r13),%rax
>
> [1] https://bugzilla.kernel.org/attachment.cgi?id=298553
> [2] https://bugzilla.kernel.org/attachment.cgi?id=298555
>
> Reported-by: Kim Phillips <[email protected]>
> Signed-off-by: Ravi Bangoria <[email protected]>
> ---
> tools/perf/arch/x86/annotate/instructions.c | 37 ++++++++++++++++++++-
> tools/perf/util/annotate.c | 1 +
> 2 files changed, 37 insertions(+), 1 deletion(-)
>
> diff --git a/tools/perf/arch/x86/annotate/instructions.c b/tools/perf/arch/x86/annotate/instructions.c
> index 24ea12ec7e02..46d7124cc4e1 100644
> --- a/tools/perf/arch/x86/annotate/instructions.c
> +++ b/tools/perf/arch/x86/annotate/instructions.c
> @@ -144,8 +144,31 @@ static struct ins x86__instructions[] = {
> { .name = "xorps", .ops = &mov_ops, },
> };
>
> -static bool x86__ins_is_fused(struct arch *arch, const char *ins1,
> +static bool amd__ins_is_fused(struct arch *arch, const char *ins1,
> const char *ins2)
> +{
> + if (strstr(ins2, "jmp"))
> + return false;
> +
> + /* Family >= 15h supports cmp/test + branch fusion */
> + if (arch->family >= 0x15 && (strstarts(ins1, "test") ||
> + (strstarts(ins1, "cmp") && !strstr(ins1, "xchg")))) {
> + return true;
> + }
> +
> + /* Family >= 19h supports some ALU + branch fusion */
> + if (arch->family >= 0x19 && (strstarts(ins1, "add") ||
> + strstarts(ins1, "sub") || strstarts(ins1, "and") ||
> + strstarts(ins1, "inc") || strstarts(ins1, "dec") ||
> + strstarts(ins1, "or") || strstarts(ins1, "xor"))) {
> + return true;
> + }
> +
> + return false;
> +}
> +
> +static bool intel__ins_is_fused(struct arch *arch, const char *ins1,
> + const char *ins2)
> {
> if (arch->family != 6 || arch->model < 0x1e || strstr(ins2, "jmp"))
> return false;
> @@ -172,6 +195,15 @@ static bool x86__ins_is_fused(struct arch *arch, const char *ins1,
> return false;
> }
>
> +static bool x86__ins_is_fused(struct arch *arch, const char *ins1,
> + const char *ins2)
> +{
> + if (strstarts(arch->vendor, "AuthenticAMD"))
> + return amd__ins_is_fused(arch, ins1, ins2);
> +
> + return intel__ins_is_fused(arch, ins1, ins2);
> +}
> +
Can we instead make x86__ins_is_fused be a pointer and instead of
storing arch->vendor we set it to one of amd__ins_is_fused() or
intel__ins_is_fused()?
I.e. here:
> static int x86__cpuid_parse(struct arch *arch, char *cpuid)
> {
> unsigned int family, model, stepping;
> @@ -184,6 +216,9 @@ static int x86__cpuid_parse(struct arch *arch, char *cpuid)
> if (ret == 3) {
> arch->family = family;
> arch->model = model;
> + arch->vendor = strndup(cpuid, 12);
x86__ins_is_fused = strstarts(cpuid, "AuthenticAMD") ?
amd__ins_is_fused :
intel__ins_is_fused;
?
> + if (!arch->vendor)
> + return -1;
> return 0;
> }
>
> diff --git a/tools/perf/util/annotate.c b/tools/perf/util/annotate.c
> index 0bae061b2d6d..88326bb990b5 100644
> --- a/tools/perf/util/annotate.c
> +++ b/tools/perf/util/annotate.c
> @@ -77,6 +77,7 @@ struct arch {
> bool sorted_instructions;
> bool initialized;
> void *priv;
> + char *vendor;
> unsigned int model;
> unsigned int family;
> int (*init)(struct arch *arch, char *cpuid);
> --
> 2.27.0
--
- Arnaldo
>> +static bool x86__ins_is_fused(struct arch *arch, const char *ins1,
>> + const char *ins2)
>> +{
>> + if (strstarts(arch->vendor, "AuthenticAMD"))
>> + return amd__ins_is_fused(arch, ins1, ins2);
>> +
>> + return intel__ins_is_fused(arch, ins1, ins2);
>> +}
>> +
>
> Can we instead make x86__ins_is_fused be a pointer and instead of
> storing arch->vendor we set it to one of amd__ins_is_fused() or
> intel__ins_is_fused()?
>
> I.e. here:
>
>> static int x86__cpuid_parse(struct arch *arch, char *cpuid)
>> {
>> unsigned int family, model, stepping;
>> @@ -184,6 +216,9 @@ static int x86__cpuid_parse(struct arch *arch, char *cpuid)
>> if (ret == 3) {
>> arch->family = family;
>> arch->model = model;
>> + arch->vendor = strndup(cpuid, 12);
>
> x86__ins_is_fused = strstarts(cpuid, "AuthenticAMD") ?
> amd__ins_is_fused :
> intel__ins_is_fused;
Sure. Will post v2.
Thanks for the review Arnaldo,
Ravi
Em Fri, Sep 10, 2021 at 04:47:31PM +0530, Ravi Bangoria escreveu:
>
> >> +static bool x86__ins_is_fused(struct arch *arch, const char *ins1,
> >> + const char *ins2)
> >> +{
> >> + if (strstarts(arch->vendor, "AuthenticAMD"))
> >> + return amd__ins_is_fused(arch, ins1, ins2);
> >> +
> >> + return intel__ins_is_fused(arch, ins1, ins2);
> >> +}
> >> +
> >
> > Can we instead make x86__ins_is_fused be a pointer and instead of
> > storing arch->vendor we set it to one of amd__ins_is_fused() or
> > intel__ins_is_fused()?
> >
> > I.e. here:
> >
> >> static int x86__cpuid_parse(struct arch *arch, char *cpuid)
> >> {
> >> unsigned int family, model, stepping;
> >> @@ -184,6 +216,9 @@ static int x86__cpuid_parse(struct arch *arch, char *cpuid)
> >> if (ret == 3) {
> >> arch->family = family;
> >> arch->model = model;
> >> + arch->vendor = strndup(cpuid, 12);
> >
> > x86__ins_is_fused = strstarts(cpuid, "AuthenticAMD") ?
> > amd__ins_is_fused :
> > intel__ins_is_fused;
>
> Sure. Will post v2.
Thanks, if you're quick we may get this into this merge window :-)
> Thanks for the review Arnaldo,
- Arnaldo
On 10-Sep-21 7:49 PM, Arnaldo Carvalho de Melo wrote:
> Em Fri, Sep 10, 2021 at 04:47:31PM +0530, Ravi Bangoria escreveu:
>>
>>>> +static bool x86__ins_is_fused(struct arch *arch, const char *ins1,
>>>> + const char *ins2)
>>>> +{
>>>> + if (strstarts(arch->vendor, "AuthenticAMD"))
>>>> + return amd__ins_is_fused(arch, ins1, ins2);
>>>> +
>>>> + return intel__ins_is_fused(arch, ins1, ins2);
>>>> +}
>>>> +
>>>
>>> Can we instead make x86__ins_is_fused be a pointer and instead of
>>> storing arch->vendor we set it to one of amd__ins_is_fused() or
>>> intel__ins_is_fused()?
>>>
>>> I.e. here:
>>>
>>>> static int x86__cpuid_parse(struct arch *arch, char *cpuid)
>>>> {
>>>> unsigned int family, model, stepping;
>>>> @@ -184,6 +216,9 @@ static int x86__cpuid_parse(struct arch *arch, char *cpuid)
>>>> if (ret == 3) {
>>>> arch->family = family;
>>>> arch->model = model;
>>>> + arch->vendor = strndup(cpuid, 12);
>>>
>>> x86__ins_is_fused = strstarts(cpuid, "AuthenticAMD") ?
>>> amd__ins_is_fused :
>>> intel__ins_is_fused;
>>
>> Sure. Will post v2.
>
> Thanks, if you're quick we may get this into this merge window :-)
v2 Posted: https://lore.kernel.org/lkml/[email protected]
Thanks,
Ravi