Re: [PATCH v5 21/23] powerpc: Simplify test in __dma_sync()

[Christophe Leroy]

Le 04/02/2016 12:37, Denis Kirjanov a Ãcrit :
> On 2/4/16, Christophe Leroy <christophe.leroy@xxxxxx> wrote:
> > This simplification helps the compiler. We now have only one test
> > instead of two, so it reduces the number of branches.
> >
> > Signed-off-by: Christophe Leroy <christophe.leroy@xxxxxx>
> > ---
> > v2: new
> > v3: no change
> > v4: no change
> > v5: no change
> >
> >   arch/powerpc/mm/dma-noncoherent.c | 2 +-
> >   1 file changed, 1 insertion(+), 1 deletion(-)
> >
> > diff --git a/arch/powerpc/mm/dma-noncoherent.c
> > b/arch/powerpc/mm/dma-noncoherent.c
> > index 169aba4..2dc74e5 100644
> > --- a/arch/powerpc/mm/dma-noncoherent.c
> > +++ b/arch/powerpc/mm/dma-noncoherent.c
> > @@ -327,7 +327,7 @@ void __dma_sync(void *vaddr, size_t size, int direction)
> >   		 * invalidate only when cache-line aligned otherwise there is
> >   		 * the potential for discarding uncommitted data from the cache
> >   		 */
> > -		if ((start & (L1_CACHE_BYTES - 1)) || (size & (L1_CACHE_BYTES - 1)))
> > +		if ((start | end) & (L1_CACHE_BYTES - 1))
> >   			flush_dcache_range(start, end);
> >   		else
> >   			invalidate_dcache_range(start, end);
> The previous version of address cache-line aligned check reads perfectly fine.
> What's the benefit of this micro optimization?
With this optimisation we avoid one unneccessary test and two associated 
jumps. Taking into account that __dma_sync() is one of the top ten CPU 
consummers, I believe it is worth it:

Without the patch:

c000d894:    70 6a 00 0f     andi.   r10,r3,15
c000d898:    39 29 00 0f     addi    r9,r9,15
c000d89c:    54 63 00 36     rlwinm  r3,r3,0,0,27
c000d8a0:    7d 23 48 50     subf    r9,r3,r9
c000d8a4:    41 82 00 84     beq     c000d928 <__dma_sync+0xb8>
[...]
c000d8c0:    7c 00 04 ac     sync
c000d8c4:    4e 80 00 20     blr
[...]
c000d928:    70 8a 00 0f     andi.   r10,r4,15
c000d92c:    40 a2 ff 7c     bne     c000d8a8 <__dma_sync+0x38>
c000d930:    55 2a e1 3f     rlwinm. r10,r9,28,4,31
c000d934:    41 a2 ff 8c     beq     c000d8c0 <__dma_sync+0x50>

With the patch:

c000d894:    7c 89 1b 78     or      r9,r4,r3
c000d898:    71 2a 00 0f     andi.   r10,r9,15
c000d89c:    54 63 00 36     rlwinm  r3,r3,0,0,27
c000d8a0:    38 84 00 0f     addi    r4,r4,15
c000d8a4:    7c 83 20 50     subf    r4,r3,r4
c000d8a8:    41 82 00 84     beq     c000d92c <__dma_sync+0xbc>
[...]
c000d8c4:    7c 00 04 ac     sync
c000d8c8:    4e 80 00 20     blr
[...]
c000d92c:    54 89 e1 3f     rlwinm. r9,r4,28,4,31
c000d930:    41 a2 ff 94     beq     c000d8c4 <__dma_sync+0x54>


Christophe
> > --
> > 2.1.0
> >
> > _______________________________________________
> > Linuxppc-dev mailing list
> > Linuxppc-dev@xxxxxxxxxxxxxxxx
> > https://lists.ozlabs.org/listinfo/linuxppc-dev