Re: [PATCH v3 1/7] clk: introduce clk_div_mask() helper

[Stephen Boyd <sboyd@codeaurora.org>]

On 07/07, Andy Shevchenko wrote:
> On Thu, 2015-06-18 at 12:48 -0700, Stephen Boyd wrote:
> > On 03/31, Andy Shevchenko wrote:
> > > diff --git a/include/linux/clk-provider.h b/include/linux/clk
> > > -provider.h
> > > index 5591ea7..20b0b67 100644
> > > --- a/include/linux/clk-provider.h
> > > +++ b/include/linux/clk-provider.h
> > > @@ -353,6 +353,11 @@ struct clk_divider {
> > >  	spinlock_t	*lock;
> > >  };
> > >  
> > > +static inline unsigned long clk_div_mask(u8 width)
> > > +{
> > > +	return (1 << width) - 1;
> > > +}
> > > +
> > 
> > Why not just change drivers to use GENMASK? It's a proven and
> > tested way to generate a bitmask.
> 
> Too many unneeded calculations I suppose.

That's what compiler optimizations are for.

> 
> Compare:
> 	mask = clk_div_mask(mm) << ms;
> 	which is simple ((1 << mm) - 1) << ms
> and
> 	mask = GENMASK(mm + ms - 1, ms);
> 	which is (~0 << ms) & (~0 >> (BITS_PER_LONG - 1 - (mm + ms -
> 1)))

And if mm is 32 then we hit undefined behavior.

> 
> > 
> > So I'd rather see drivers converted to use that macro directly
> > especially because the mask may need to start at some bit that
> > isn't 0.
> 
> If you think the above is not a burden, I can do the conversion to
> GENMASK.
> 
> Though it might make sense when ms = 0 explicitly.
> 

Yes let's use GENMASK. I imagine the extra few instructions are
negligible.

-- 
Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum,
a Linux Foundation Collaborative Project