Re: [PATCH] [media] zl10353: use div_u64 instead of do_div

[Ard Biesheuvel]

On 12 February 2016 at 22:01, Arnd Bergmann <arnd@xxxxxxxx> wrote:
> On Friday 12 February 2016 13:21:33 Nicolas Pitre wrote:
>> On Fri, 12 Feb 2016, Arnd Bergmann wrote:
>>
>> > On Friday 12 February 2016 14:32:20 Mauro Carvalho Chehab wrote:
>> > > Em Fri, 12 Feb 2016 15:27:18 +0100
>> > > Arnd Bergmann <arnd@xxxxxxxx> escreveu:
>> > >
>> > > > I noticed a build error in some randconfig builds in the zl10353 driver:
>> > > >
>> > > > dvb-frontends/zl10353.c:138: undefined reference to `____ilog2_NaN'
>> > > > dvb-frontends/zl10353.c:138: undefined reference to `__aeabi_uldivmod'
>> > > >
>> > > > The problem can be tracked down to the use of -fprofile-arcs (using
>> > > > CONFIG_GCOV_PROFILE_ALL) in combination with CONFIG_PROFILE_ALL_BRANCHES
>> > > > on gcc version 4.9 or higher, when it fails to reliably optimize
>> > > > constant expressions.
>> > > >
>> > > > Using div_u64() instead of do_div() makes the code slightly more
>> > > > readable by both humans and by gcc, which gives the compiler enough
>> > > > of a break to figure it all out.
>> > >
>> > > I'm not against this patch, but we have 94 occurrences of do_div()
>> > > just at the media subsystem. If this is failing here, it would likely
>> > > fail with other drivers. So, I guess we should either fix do_div() or
>> > > convert all such occurrences to do_div64().
>> >
>> > I agree that it's possible that the same problem exists elsewhere, but this is
>> > the only one that I ever saw (in five ranconfig builds out of 8035 last week).
>> >
>> > I also tried changing do_div() to be an inline function with just a small
>> > macro wrapper around it for the odd calling conventions, which also made this
>> > error go away. I would assume that Nico had a good reason for doing do_div()
>> > the way he did.
>>
>> The do_div() calling convention predates my work on it.  I assume it was
>> originally done this way to better map onto the x86 instruction.
>
> Right, this goes back to the dawn of time.
>
>> > In some other files, I saw the object code grow by a few
>> > instructions, but the examples I looked at were otherwise identical.
>> >
>> > I can imagine that there might be cases where the constant-argument optimization
>> > of do_div fails when we go through an inline function in some combination
>> > of Kconfig options and compiler version, though I don't think that was
>> > the case here.
>>
>> What could be tried is to turn __div64_const32() into a static inline
>> and see if that makes a difference with those gcc versions we currently
>> accept.
>>
>> > Nico, any other thoughts on this?
>>
>> This is all related to the gcc bug for which I produced a test case
>> here:
>>
>> http://article.gmane.org/gmane.linux.kernel.cross-arch/29801
>>
>> Do you know if this is fixed in recent gcc?
>
> I have a fairly recent gcc, but I also never got around to submit
> it properly.
>
> However, I did stumble over an older patch I did now, which I could
> not remember what it was good for. It does fix the problem, and
> it seems to be a better solution.
>
>         Arnd
>
> diff --git a/include/linux/compiler.h b/include/linux/compiler.h
> index b5acbb404854..b5ff9881bef8 100644
> --- a/include/linux/compiler.h
> +++ b/include/linux/compiler.h
> @@ -148,7 +148,7 @@ void ftrace_likely_update(struct ftrace_branch_data *f, int val, int expect);
>   */
>  #define if(cond, ...) __trace_if( (cond , ## __VA_ARGS__) )
>  #define __trace_if(cond) \
> -       if (__builtin_constant_p((cond)) ? !!(cond) :                   \
> +       if (__builtin_constant_p(!!(cond)) ? !!(cond) :                 \
>         ({                                                              \
>                 int ______r;                                            \
>                 static struct ftrace_branch_data                        \
>

I remember seeing this patch, but I don't remember the exact context.
But when you think about it, !!cond can be a build time constant even
if cond is not, as long as you can prove statically that cond != 0. So
I think this change is obviously correct, and an improvement since it
will remove the profiling overhead of branches that are not true
branches in the first place.