This is the mail archive of the
libc-alpha@sourceware.org
mailing list for the glibc project.
Re: [PATCH] s390: optimize syscall function
- From: Adhemerval Zanella <adhemerval dot zanella at linaro dot org>
- To: Martin Schwidefsky <schwidefsky at de dot ibm dot com>
- Cc: Christian Borntraeger <borntraeger at de dot ibm dot com>, Stefan Liebler <stli at linux dot vnet dot ibm dot com>, Andreas Krebbel <krebbel at linux dot vnet dot ibm dot com>, libc-alpha at sourceware dot org
- Date: Tue, 13 Jun 2017 13:16:34 -0300
- Subject: Re: [PATCH] s390: optimize syscall function
- Authentication-results: sourceware.org; auth=none
- References: <1497007747-191015-1-git-send-email-borntraeger@de.ibm.com> <20170612074958.71090d0f@mschwideX1> <fe60b8c0-7105-0c9a-f62e-0e5540308d31@linaro.org> <20170612155505.07fcc961@mschwideX1> <d3fde8f4-4077-571a-51f1-7f43081dc1ad@linaro.org> <20170613180110.69c14974@mschwideX1>
> On 13 Jun 2017, at 13:01, Martin Schwidefsky <schwidefsky@de.ibm.com> wrote:
>
> On Tue, 13 Jun 2017 11:31:09 -0300
> Adhemerval Zanella <adhemerval.zanella@linaro.org> wrote:
>
>>> On 12/06/2017 10:55, Martin Schwidefsky wrote:
>>> On Mon, 12 Jun 2017 09:04:36 -0300
>>> Adhemerval Zanella <adhemerval.zanella@linaro.org> wrote:
>>>
>>>>> On 12/06/2017 02:49, Martin Schwidefsky wrote:
>>>>> On Fri, 9 Jun 2017 13:29:07 +0200
>>>>> Christian Borntraeger <borntraeger@de.ibm.com> wrote:
>>>>>
>>>>>> Since kernel 2.6.0 all Linux version accept the system call number
>>>>>> in register 1 for svc 0. There is no need to have special handling
>>>>>> that uses EX for system calls < 256. This will simplify and speed
>>>>>> up that code.
>>>>>>
>>>>>> A microbenchmark doing "syscall(__NR_getpid);" in a loops gets faster
>>>>>> by ~12%.
>>>>>>
>>>>>> * sysdeps/unix/sysv/linux/s390/s390-64/syscall.S: Simplify
>>>>>> code by always using SVC 0 instead of EX.
>>>>>> * sysdeps/unix/sysv/linux/s390/s390-64/syscall.S: Likewise.
>>>>>>
>>>>>> Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
>>>>>> ---
>>>>>> sysdeps/unix/sysv/linux/s390/s390-32/syscall.S | 12 +++---------
>>>>>> sysdeps/unix/sysv/linux/s390/s390-64/syscall.S | 12 +++---------
>>>>>> 2 files changed, 6 insertions(+), 18 deletions(-)
>>>>>
>>>>> NAK. E.g. this from glibc:
>>>>>
>>>>> sysdeps/unix/sysv/linux/s390/s390-64:
>>>>>
>>>>> ENTRY (syscall)
>>>>> ...
>>>>> basr %r8,0
>>>>> 0: clg %r1,4f-0b(%r8) /* svc number < 256? */
>>>>> jl 2f
>>>>> 1: svc 0
>>>>> j 3f
>>>>> 2: ex %r1,1b-0b(%r8) /* lsb of R1 is subsituted as SVC number */
>>>>> 3: lg %r15,0(%r15) /* load back chain */
>>>>> cfi_adjust_cfa_offset (-160)
>>>>> lmg %r6,15,48(%r15) /* Load registers. */
>>>>> ...
>>>>>
>>>>> And there are old version of glibc where NR_syscalls was < 256 that
>>>>> exclusively used the SVC instruction with the system call number
>>>>> encoded in the instruction itself.
>>>>>
>>>>
>>>> I am not following the NAK, is the 'svc' way not really allowed for current
>>>> minimum kernel supported for s390 (3.2) or is the 'svc' method not really the
>>>> fastest one for such syscalls?
>>>
>>> I claim lack of coffee after the long weekend ..
>>>
>>> Yes, with version 3.2 as the minimum kernel level for using a glibc with
>>> the prosed patch everything should be fine. The support for system calls
>>> larger than 256 has been added with 2.5.68.
>>
>> Thanks, since it should be ok we can also cleanup the {INLINE,INTERNAL}_SYSCALL
>> macros in s390x sysdep.h as well.
>
> How so? If you have a fixed system call number smaller than 256 the best option
> is to encode the system call number in the instruction. Loading the number to
> %r1 and then doing SVC-0 requires an additional instruction, no?
Indeed, I had the wrong assumption the syscall.S method was used in sysdep.h.