Re: Tell me why 8 continuos ocbp ins's exist in sh4_flush_purge_region().

[Paul Mundt <lethal@linux-sh.org>]

On Wed, Jun 27, 2012 at 11:52:55AM +0900, MASAO TAKAHASHI wrote:
> Hi everybody.
> 
> Please tell me why 8 continuos "ocbp" instructions exist
>  in sh4_flush_purge_region() as follows.
>   --------------------------------------------------
> 	while (cnt >= 8) {
> 		__ocbp(v); v += L1_CACHE_BYTES;
> 		__ocbp(v); v += L1_CACHE_BYTES;
> 		__ocbp(v); v += L1_CACHE_BYTES;
> 		__ocbp(v); v += L1_CACHE_BYTES;
> 		__ocbp(v); v += L1_CACHE_BYTES;
> 		__ocbp(v); v += L1_CACHE_BYTES;
> 		__ocbp(v); v += L1_CACHE_BYTES;
> 		__ocbp(v); v += L1_CACHE_BYTES;
> 		cnt -= 8;
> 	}
>  -----------------------------------------------------
> 
> The previous version was like this.
>  ------------------------------------------------------- 
> 	for (v = begin; v < end; v+=L1_CACHE_BYTES) {
> 		asm volatile("ocbp	%0"
> 			     : /* no output */
> 			     : "m" (__m(v)));
> 	}
>  --------------------------------------------------------
> 
We unroll the loop for better performance. The same applies for the wback
and invalidate cases, too. If you're passing in a small size then the >8 case is surpassed and you go to line-at-a-time obcp as before.