This is the mail archive of the libc-alpha@sourceware.org mailing list for the glibc project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: [PATCH] i386: Assembly Implementation cleanup


On Wed, Aug 12, 2015 at 10:58:16AM -0300, Adhemerval Zanella wrote:
> 
> 
> On 12-08-2015 09:44, H.J. Lu wrote:
> > On Wed, Aug 12, 2015 at 5:38 AM, Adhemerval Zanella
> > <adhemerval.zanella@linaro.org> wrote:
> >>
> >>
> >> On 11-08-2015 21:22, H.J. Lu wrote:
> >>> On Fri, Jul 3, 2015 at 8:41 AM, Adhemerval Zanella
> >>> <adhemerval.zanella@linaro.org> wrote:
> >>>> This patch removes i386 assembly implementation for epoll_wait, mmap,
> >>>> mmap64, semtimeop now that i386 have 6 argument syscall support from
> >>>> C code.
> >>>>
> >>>> The mmap implementation used is the wordsize-32 generic linux one.
> >>>>
> >>>> Tested in i386.
> >>>>
> >>>> ---
> >>>>
> >>>>         * sysdeps/unix/sysv/linux/i386/mmap.c: New file.
> >>>>         * sysdeps/unix/sysv/linux/i386/epoll_pwait.S: Remove file.
> >>>>         * sysdeps/unix/sysv/linux/i386/mmap.S: Likewise.
> >>>>         * sysdeps/unix/sysv/linux/i386/mmap64.S: Likewise.
> >>>>         * sysdeps/unix/sysv/linux/i386/semtimedop.S: Likewise.
> >>>>
> >>>
> >>> What is the motivation?  6 argument syscalls aren't really inlined.
> >>> They put arguments on stack and call __libc_do_syscall.  It is
> >>> slower than assembly code.
> >>>
> >>
> >> Same for ARM: code maintainability on functions that are not really
> >> performance-critical.  This avoids issues like the x86 pthread_cond_*
> >> assembly implementation (which are now aimed to be removed), that
> >> required specific fixes just for x86.
> > 
> > Is there a way to truly inline syscall with 6 arguments, not by calling
> > __libc_do_syscall?
> > 
> 
> I checked around some other projects that implements i386 syscall and
> neither seems to have a workaround for this.  I also recall that Joseph
> has said GCC does not really support to not cobbler the required register,
> so that's why he has implemented through a helper function.  Anyway, as
> Tovarld has asked, I think these are not really heavy hitter (and kernel
> implementation is much more complex than the time to save/restore the
> register in stack).

Correct, only questionable is semtimedop that could be made fast by
using vdso to avoid syscall. 

For mmap* its 2000+ cycles if we map single page and usually
more so it doesnt't matter. Epoll_pwait will likely block so difference
of assembly is again very small.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]