Re: [PATCH 3/3] kmemtrace: Use tracepoints instead of markers.

[Mathieu Desnoyers <mathieu.desnoyers@polymtl.ca>]

* Eduard - Gabriel Munteanu (eduard.munteanu@linux360.ro) wrote:
> On Fri, Jan 02, 2009 at 06:42:54PM -0500, Mathieu Desnoyers wrote:
> > Because whatever slab_buffer_size() does will be done on the fastpath of
> > the instrumented code *even when instrumentation is disabled*, and this
> > is something we need to avoid above all.
> 
> I had doubts about this, so I tried it myself. It seems that when using
> -O2 it generates optimal code, without computing a << b unnecessarily.
> It only precomputes it at -O0. Here's how I tested...
> 
> #include <stdio.h>
> 
> #define unlikely(x)	__builtin_expect(!!(x), 0)
> 
> static int do_something_enabled;
> 
> static void print_that(unsigned long num)
> {
> 	printf("input << 5 == %lu\n", num);
> }
> 
> static inline void do_something(unsigned long num)
> {
> 	if (unlikely(do_something_enabled)) /* Like DEFINE_TRACE does. */
> 		print_that(num);
> }
> 
> static void call_do_something(unsigned long in)
> {
> 	do_something(in << 5);
> }
> 
> int main(int argc, char **argv)
> {
> 	unsigned long in;
> 
> 	if (argc != 3) {
> 		printf("Wrong number of arguments!\n");
> 		return 0;
> 	}
> 
> 	sscanf(argv[1], "%d", &do_something_enabled);
> 	sscanf(argv[2], "%lu", &in);
> 
> 	call_do_something(in);
> 
> 	return 0;
> }
> 
> 
> Snippet of objdump output when using -O2:
> 
> static inline void do_something(unsigned long num)
> {
> 	if (unlikely(do_something_enabled))
>   400635:	85 c0                	test   %eax,%eax
>   400637:	74 be                	je     4005f7 <main+0x17>
> print_that():
> /home/edi/prj/src/inlineargs/inlineargs.c:9
> 
> static int do_something_enabled;
> 
> static void print_that(unsigned long num)
> {
> 	printf("input << 5 == %lu\n", num);
>   400639:	48 c1 e6 05          	shl    $0x5,%rsi
>   40063d:	bf 6e 07 40 00       	mov    $0x40076e,%edi
>   400642:	31 c0                	xor    %eax,%eax
>   400644:	e8 5f fe ff ff       	callq  4004a8 <printf@plt>
> 
> 
> Snippet of objdump output when using -O0:
> 
> static void call_do_something(unsigned long in)
> {
>   4005fd:	55                   	push   %rbp
>   4005fe:	48 89 e5             	mov    %rsp,%rbp
>   400601:	48 83 ec 10          	sub    $0x10,%rsp
>   400605:	48 89 7d f8          	mov    %rdi,-0x8(%rbp)
> /home/edi/prj/src/inlineargs/inlineargs.c:20
> 	do_something(in << 5);
>   400609:	48 8b 45 f8          	mov    -0x8(%rbp),%rax
>   40060d:	48 89 c7             	mov    %rax,%rdi
>   400610:	48 c1 e7 05          	shl    $0x5,%rdi
>   400614:	e8 02 00 00 00       	callq  40061b <do_something>
> 
> 
> Look at that shl, it indicates the left-shift (<< 5). In the first case it's
> deferred as much as possible. However, in the second case, it's done
> before calling that inline. Also confirmed with GCC using breakpoints on
> that shl.
> 
> Can we take this as general behaviour, i.e. fn(a()), where fn() is inlined
> and a() has no side-effects, will only compute a() when needed, at least on
> GCC and when -O2 is in effect?
> 
> It only seems natural to me GCC would do this on a regular basis.
> 

Hopefully it does, especially when there are no side-effects. Can you
also try with -Os ?

Mathieu


> 
> 	Cheers,
> 	Eduard
> 

-- 
Mathieu Desnoyers
OpenPGP key fingerprint: 8CD5 52C3 8E3C 4140 715F  BA06 3F25 A8FE 3BAE 9A68