[PATCH v2 22/25] powerpc32: move xxxxx_dcache_range() functions inline

[Scott Wood]

On Tue, 2015-09-22 at 19:34 +0000, Joakim Tjernlund wrote:
> On Tue, 2015-09-22 at 13:58 -0500, Scott Wood wrote:
> > On Tue, 2015-09-22 at 18:12 +0000, Joakim Tjernlund wrote:
> > > On Tue, 2015-09-22 at 18:51 +0200, Christophe Leroy wrote:
> > > > flush/clean/invalidate _dcache_range() functions are all very
> > > > similar and are quite short. They are mainly used in __dma_sync()
> > > > perf_event locate them in the top 3 consumming functions during
> > > > heavy ethernet activity
> > > > 
> > > > They are good candidate for inlining, as __dma_sync() does
> > > > almost nothing but calling them
> > > > 
> > > > Signed-off-by: Christophe Leroy <christophe.leroy at c-s.fr>
> > > > ---
> > > > New in v2
> > > > 
> > > >  arch/powerpc/include/asm/cacheflush.h | 55 
> > > > +++++++++++++++++++++++++++--
> > > >  arch/powerpc/kernel/misc_32.S         | 65 --------------------------
> > > > ----
> > > > -----
> > > >  arch/powerpc/kernel/ppc_ksyms.c       |  2 ++
> > > >  3 files changed, 54 insertions(+), 68 deletions(-)
> > > > 
> > > > diff --git a/arch/powerpc/include/asm/cacheflush.h 
> > > > b/arch/powerpc/include/asm/cacheflush.h
> > > > index 6229e6b..6169604 100644
> > > > --- a/arch/powerpc/include/asm/cacheflush.h
> > > > +++ b/arch/powerpc/include/asm/cacheflush.h
> > > > @@ -47,12 +47,61 @@ static inline void 
> > > > __flush_dcache_icache_phys(unsigned long physaddr)
> > > >  }
> > > >  #endif
> > > >  
> > > > -extern void flush_dcache_range(unsigned long start, unsigned long 
> > > > stop);
> > > >  #ifdef CONFIG_PPC32
> > > > -extern void clean_dcache_range(unsigned long start, unsigned long 
> > > > stop);
> > > > -extern void invalidate_dcache_range(unsigned long start, unsigned 
> > > > long 
> > > > stop);
> > > > +/*
> > > > + * Write any modified data cache blocks out to memory and invalidate 
> > > > them.
> > > > + * Does not invalidate the corresponding instruction cache blocks.
> > > > + */
> > > > +static inline void flush_dcache_range(unsigned long start, unsigned 
> > > > long 
> > > > stop)
> > > > +{
> > > > +   void *addr = (void *)(start & ~(L1_CACHE_BYTES - 1));
> > > > +   unsigned int size = stop - (unsigned long)addr + (L1_CACHE_BYTES -
> > > > 1);
> > > > +   unsigned int i;
> > > > +
> > > > +   for (i = 0; i < size >> L1_CACHE_SHIFT; i++, addr += 
> > > > L1_CACHE_BYTES)
> > > > +           dcbf(addr);
> > > > +   if (i)
> > > > +           mb();   /* sync */
> > > > +}
> > > 
> > > This feels optimized for the uncommon case when there is no 
> > > invalidation.
> > 
> > If you mean the "if (i)", yes, that looks odd.
> 
> Yes.
> 
> > 
> > > I THINK it would be better to bail early 
> > 
> > Bail under what conditions?
> 
> test for "i = 0" and return. 

Why bother?

> 
> > 
> > > and use do { .. } while(--i); instead.
> > 
> > GCC knows how to optimize loops.  Please don't make them less readable.
> 
> Been a while since I checked but it used to be bad att transforming post 
> inc to pre inc/dec
> I remain unconvinced until I have seen it.

I would expect it to use bdnz for this loop, as the loop variable isn't 
referenced in the loop body.

And generally the one proposing uglification-for-optimization should provide 
the evidence. :-)

-Scott