On Sat, 24 Jun 2006, Arjan van de Ven wrote:
> > >
> > > But doesn't the unlikely help the prediction?
> >
> > nope none at all, at least not on x86/x86-64.
> > (in fact there is no way to help the prediction on those architectures
> > that actually works)
>
> ok that's not entirely accurate, there's some basic heuristics you can
> tweak to; eg often the assumption is "if you jump backwards the branch
> is taken" and stuff like that, but that's generally useful in loops, not
> so much in if () statements.... since for if () both your branches are
> forward.
>
I just tried this code:
--- code ---
extern void marker1(void);
extern void marker2(void);
extern void callme(void *p);
#define unlikely(x) __builtin_expect(!!(x), 0)
void predictme(void *t)
{
marker1();
if (unlikely (t))
callme(t);
marker2();
return;
}
--- end code ---
Compiled it like:
$ gcc -O2 -c predict.c
Then did objdump and got this:
--- objdump ---
00000000 <predictme>:
0: 55 push %ebp
1: 89 e5 mov %esp,%ebp
3: 53 push %ebx
4: 83 ec 04 sub $0x4,%esp
7: 8b 5d 08 mov 0x8(%ebp),%ebx
a: e8 fc ff ff ff call b <predictme+0xb>
b: R_386_PC32 marker1
f: 85 db test %ebx,%ebx
11: 75 08 jne 1b <predictme+0x1b>
13: 58 pop %eax
14: 5b pop %ebx
15: 5d pop %ebp
16: e9 fc ff ff ff jmp 17 <predictme+0x17>
17: R_386_PC32 marker2
1b: 89 1c 24 mov %ebx,(%esp)
1e: e8 fc ff ff ff call 1f <predictme+0x1f>
1f: R_386_PC32 callme
23: eb ee jmp 13 <predictme+0x13>
--- end objdump --
The jne is expected to fail, so we will always continue to 0x13. Now is
this a problem with x86/x86_64?
BTW: taking out the unlikely I get this:
--- objdump ---
00000000 <predictme>:
0: 55 push %ebp
1: 89 e5 mov %esp,%ebp
3: 53 push %ebx
4: 83 ec 04 sub $0x4,%esp
7: 8b 5d 08 mov 0x8(%ebp),%ebx
a: e8 fc ff ff ff call b <predictme+0xb>
b: R_386_PC32 marker1
f: 85 db test %ebx,%ebx
11: 74 08 je 1b <predictme+0x1b>
13: 89 1c 24 mov %ebx,(%esp)
16: e8 fc ff ff ff call 17 <predictme+0x17>
17: R_386_PC32 callme
1b: 58 pop %eax
1c: 5b pop %ebx
1d: 5d pop %ebp
1e: e9 fc ff ff ff jmp 1f <predictme+0x1f>
1f: R_386_PC32 marker2
--- end objdump ---
So it seems that gcc tries to keep the likely's from jumping.
-- Steve
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
[Index of Archives]
[Kernel Newbies]
[Netfilter]
[Bugtraq]
[Photo]
[Stuff]
[Gimp]
[Yosemite News]
[MIPS Linux]
[ARM Linux]
[Linux Security]
[Linux RAID]
[Video 4 Linux]
[Linux for the blind]
[Linux Resources]
