Andi,

The below is compared result on older machine(cpu info is attached):
That shows new code get better performance up to 1.6x.

Bytes: ORG_TIME: NEW_TIME: ORG vs NEW:
7       0.87    0.76    1.14
16      0.99    0.68    1.45
18      1.07    0.77    1.38
21      1.09    0.78    1.39
25      1.11    0.77    1.44
30      1.12    0.73    1.53
36      1.15    0.75    1.53
38      1.12    0.75    1.49
62      1.18    0.77    1.53
75      1.25    0.79    1.58
85      1.28    0.80    1.60
120     1.33    0.82    1.62
193     1.45    0.88    1.64
245     1.48    0.96    1.54
256     1.45    0.90    1.61
356     1.61    1.02    1.57
601     1.78    1.22    1.45
958     2.04    1.47    1.38
1024    2.07    1.48    1.39
2048    2.80    2.21    1.26

Thanks
Ling

2014-04-08 0:42 GMT+08:00, Andi Kleen <andi@firstfloor.org>:
> ling.ma.program@gmail.com writes:
>
>> From: Ling Ma <ling.ml@alibaba-inc.com>
>>
>> In this patch we manage to reduce miss branch prediction by
>> avoiding using branch instructions and force destination to be aligned
>> with general 64bit instruction.
>> Below compared results shows we improve performance up to 1.8x
>> (We modified test suit from Ondra, send after this patch)
>
> You didn't specify the CPU?
>
> I assume it's some Atom, as nothing else uses these open coded functions
> anymore?
>
> -Andi
>
> --
> ak@linux.intel.com -- Speaking for myself only
>