Re: [PATCH 2/3] time: allow gcc to fold constants when using msecs_to_jiffies

[Nicholas Mc Guire <der.herr@hofr.at>]

On Sun, 05 Apr 2015, Joe Perches wrote:

> On Sun, 2015-04-05 at 09:23 +0200, Nicholas Mc Guire wrote:
> > The majority of the msecs_to_jiffies() users in the kernel are passing in
> > constants which would allow gcc to do constant folding by checking with
> > __builtin_constant_p() in msecs_to_jiffies().
> > 
> > The original msecs_to_jiffies is renamed to __msecs_to_jiffies and aside
> > from the removal of the check for negative values being moved out, is
> > unaltered.
> 
> At least for gcc 4.9, this doesn't allow the compiler
> to optimize / precalculation msecs_to_jiffies calls
> with a constant.
> 
> This does: (on top of your patch x86-64 defconfig)
> 
> $ size vmlinux.o.*
>    text	   data	    bss	    dec	    hex	filename
> 11770523	1505971	1018454	14294948	 da1fa4	vmlinux.o.next-b0a12fb5bc8
> 11770530	1505971	1018454	14294955	 da1fab	vmlinux.o.next-b0a12fb5bc8-inline
> 11768734	1505971	1018454	14293159	 da18a7	vmlinux.o.next-b0a12fb5bc8-macro
> 
> I think this should still move the if (m) < 0 back into the
> original __msecs_to_jiffies function.
>

could you check if you can reproduce the results below ?
my assumption was that gcc would always optimize out an 
if(CONST < 0) return CONST; reducing it to the return CONST; 
only and thus this should not make any difference but Im not 
that familiar with gcc.

gcc versions here are:
 for x86 gcc version 4.7.2 (Debian 4.7.2-5) 
 for powerpc it is a gcc version 4.9.2 (crosstool-NG 1.20.0)
 for arm gcc version 4.9.2 20140904 (prerelease) (crosstool-NG linaro-1.13.1-4.9-2014.09 - Linaro GCC 4.9-2014.09)

Procedure used:
root@debian:~/linux-next# make distclean
root@debian:~/linux-next# make defconfig
root@debian:~/linux-next# make drivers/net/wireless/p54/p54usb.lst
root@debian:~/linux-next# make drivers/net/wireless/p54/p54usb.s

same setup in unpatched /usr/src/linux-next/

e.g:
root@debian:/usr/src/linux-next# grep msecs_to_jiffies drivers/net/wireless/p54/p54usb.c
        timeout = jiffies + msecs_to_jiffies(1000);
        timeout = jiffies + msecs_to_jiffies(1000);

So both calls are constants and should be optimized out if it works as
expected.

without the patch applied:

root@debian:/usr/src/linux-next# grep msecs_to_jiffies  drivers/net/wireless/p54/p54usb.s
        call    msecs_to_jiffies        #
        call    msecs_to_jiffies        #
root@debian:/usr/src/linux-next# grep msecs_to_jiffies drivers/net/wireless/p54/p54usb.lst
        timeout = jiffies + msecs_to_jiffies(1000);
                        e19: R_X86_64_PC32      msecs_to_jiffies+0xfffffffffffffffc
        timeout = jiffies + msecs_to_jiffies(1000);
        timeout = jiffies + msecs_to_jiffies(1000);
                        fd8: R_X86_64_PC32      msecs_to_jiffies+0xfffffffffffffffc
        timeout = jiffies + msecs_to_jiffies(1000);


with the patch applied this then gives me:

root@debian:~/linux-next# grep msecs_to_jiffies  drivers/net/wireless/p54/p54usb.s
root@debian:~/linux-next# grep msecs_to_jiffies drivers/net/wireless/p54/p54usb.lst
        timeout = jiffies + msecs_to_jiffies(1000);
        timeout = jiffies + msecs_to_jiffies(1000);
        timeout = jiffies + msecs_to_jiffies(1000);
        timeout = jiffies + msecs_to_jiffies(1000);

Conversely in kernel/sched/core.c the msecs_to_jiffies is not a constant
and the result is that it calls __msecs_to_jiffies

patched:
root@debian:~/linux-next# grep msecs_to_jiffies kernel/sched/core.s
        call    __msecs_to_jiffies      #

unpatched:
root@debian:/usr/src/linux-next# grep msecs_to_jiffies kernel/sched/core.s
        call    msecs_to_jiffies        #


Could you check if you get these results for this test-case ?
If this really were compiler dependant that would be very bad.

I did move the < 0 check - but that did not change the situation here.
but it well may be that there are some cases where this does make a
difference 

thx!
hofrat