Le 10/02/2022 à 15:26, Steven Rostedt a écrit :
> On Thu, 10 Feb 2022 09:47:52 +0100
> Christophe Leroy <christophe.leroy@xxxxxxxxxx> wrote:
>
>> On a ppc32 build with CONFIG_CC_OPTIMISE_FOR_SIZE,
>> trace_trigger_soft_disabled() appears more than 50 times in vmlinux.
>>
>> That function is rather big for an inlined function, and
>> it doesn't benefit much from inlining as its only parameter
>> is a pointer to a struct in memory:
>
> The number of parameters is not the reason for it being inlined. It's in a
> *very* hot path, and a function call causes a noticeable performance hit.
Fair enough
>
>
>>
>>
>> It doesn't benefit much from inlining as its only parameter is a
>> pointer to a struct in memory so no constant folding is involved.
>>
>> Uninline it and move it into kernel/trace/trace_events_trigger.c
>>
>> It reduces the size of vmlinux by approximately 10 kbytes.
>
> If you have an issue with the size, perhaps the function can be modified to
> condense it. I'm happy to have a size reduction, but I will NACK making it
> into a function call.
>
Well, my first issue is that I get it duplicated 50 times because GCC
find it too big to get inlined. So it is a function call anyway.
For instance, when building arch/powerpc/kernel/irq.o which -Winline, I get:
./include/linux/perf_event.h:1169:20: error: inlining failed in call to
'perf_fetch_caller_regs': call is unlikely and code size would grow
[-Werror=inline]
./include/linux/perf_event.h:1169:20: error: inlining failed in call to
'perf_fetch_caller_regs': call is unlikely and code size would grow
[-Werror=inline]
./include/linux/perf_event.h:1169:20: error: inlining failed in call to
'perf_fetch_caller_regs': call is unlikely and code size would grow
[-Werror=inline]
./include/linux/perf_event.h:1169:20: error: inlining failed in call to
'perf_fetch_caller_regs': call is unlikely and code size would grow
[-Werror=inline]
./include/linux/trace_events.h:712:1: error: inlining failed in call to
'trace_trigger_soft_disabled': call is unlikely and code size would grow
[-Werror=inline]
./include/linux/trace_events.h:712:1: error: inlining failed in call to
'trace_trigger_soft_disabled': call is unlikely and code size would grow
[-Werror=inline]
./include/linux/trace_events.h:712:1: error: inlining failed in call to
'trace_trigger_soft_disabled': call is unlikely and code size would grow
[-Werror=inline]
./include/linux/trace_events.h:712:1: error: inlining failed in call to
'trace_trigger_soft_disabled': call is unlikely and code size would grow
[-Werror=inline]
If having it a function call is really an issue, then it should be
__always_inline
I will see the impact on size when we do so.
What is in the hot path really ? It is the entire function or only the
first test ?
What about:
static inline bool
trace_trigger_soft_disabled(struct trace_event_file *file)
{
unsigned long eflags = file->flags;
if (!(eflags & EVENT_FILE_FL_TRIGGER_COND))
return outlined_trace_trigger_soft_disabled(...);
return false;
}
Or is there more in the hot path ?
Thanks
Christophe
