Messages in this thread | | | Date | Mon, 5 Jan 2009 13:09:33 -0500 | From | Mathieu Desnoyers <> | Subject | Re: [PATCH 3/3] kmemtrace: Use tracepoints instead of markers. |
| |
* Eduard - Gabriel Munteanu (eduard.munteanu@linux360.ro) wrote: > On Mon, Jan 05, 2009 at 11:05:34AM -0500, Mathieu Desnoyers wrote: > > Hopefully it does, especially when there are no side-effects. Can you > > also try with -Os ? > > > > Mathieu > > Here's the disassembled code when using -Os. It seems it's optimised, as > with -O2. My GCC's version is 4.3.2 (Gentoo Linux). > > If you want to test yourself, the output was generated with 'objdump -d > -S -a'. > > What do you think? > > static void print_that(unsigned long num) > { > printf("input << 5 == %lu\n", num); > 40062d: 48 c1 e6 05 shl $0x5,%rsi > 400631: bf 5e 07 40 00 mov $0x40075e,%edi > 400636: 31 c0 xor %eax,%eax > 400638: e8 6b fe ff ff callq 4004a8 <printf@plt> > sscanf(argv[2], "%lu", &in); > > call_do_something(in); > > return 0; > } > 40063d: 5a pop %rdx > 40063e: 59 pop %rcx > 40063f: 31 c0 xor %eax,%eax > 400641: 5b pop %rbx > 400642: c3 retq > > > Eduard >
It looks good. Although I wonder if gcc will still optimize this in more complicated functions. Just to be 100% sure, I would recommend testing it in larger functions like schedule() in the kernel tree. But so far it looks like it does not hurt much to leave a small supplementary operation in the unlikely() branch.
Mathieu
-- Mathieu Desnoyers OpenPGP key fingerprint: 8CD5 52C3 8E3C 4140 715F BA06 3F25 A8FE 3BAE 9A68
| |