Re: DragonflyBSD fast syscall support and x86 improvements

[David Xu <bsddiy@xxxxxxx>]

I was wrong, it seems sysenter is useful, and change is really
minimal if you only want to support MODERN cpu

> Hi,
>
> Today I have tested sysenter sycall vs int0x80 syscall, I have done
> the work several monthes ago, now I have imported it into perforce,
> the branch name is davidxu_sysenter.
>
> Here are some benchmark results:
>
> 1. pure geteuid() syscall:
>
> x int80_syscall.txt
> + sysenter_syscall.txt
> +--------------------------------------------------------------------------+
> |                                                                       +  |
> |      x                                                                +  |
> |     xxx                                                              ++  |
> |x x xxxx                                                   +      +  ++++ |
> |   |_A_|                                                         |___A_M_||
> +--------------------------------------------------------------------------+
>     N           Min           Max        Median           Avg        Stddev
> x  10       2390403       2562187       2525720     2508262.9     53926.058
> +  10       3855570       4169162       4139662     4103079.4     95933.507
> Difference at 95.0% confidence
> 	1.59482e+06 ± 73117.4
> 	63.5825% ± 2.91506%
> 	(Student's t, pooled s = 77817.9)
>
> 2. mysql super-smack:
>
> x mysql_int80.txt
> + mysql_sysenter.txt
> +--------------------------------------------------------------------------+
> |         x                                                    +           |
> |         x                                                    + +         |
> |         x xx                                              +  +++         |
> |x    x   xxxx x x        x           x                +    +  +++++      +|
> |    |______MA________|                                     |___A___|      |
> +--------------------------------------------------------------------------+
>     N           Min           Max        Median           Avg        Stddev
> x  15      19212.09      19911.76      19417.23     19449.863     165.25167
> +  15      20250.48      20603.26      20408.12     20411.395     76.260797
> Difference at 95.0% confidence
> 	961.531 ± 96.2398
> 	4.94364% ± 0.494809%
> 	(Student's t, pooled s = 128.693)
>
>
> Hardware is dual-core Pentium-D. I thought sysenter is useless as I
> said on dragonflybsd mail list, but the fact is a little different
> here, so I was wrong.
>
> David Xu