RE: [patch] Improve prefetch heuristics

["Fang, Changpeng" <Changpeng dot Fang at amd dot com>]

Good point, Christian:

Yes, this in the alignment issue. We can not say they are on the same
cache line no matter how small delta is. Just as you suggested, we need
a parameter to take into account the alignment issue. The default of 2 is
reasonable for a generic architecture, and we may need some experiments
to verify this.

I will add the debug log entry for the prefetches that are not generated.

Thanks,

Changpeng


-----Original Message-----
From: Christian Borntraeger [mailto:borntraeger@de.ibm.com] 
Sent: Thursday, April 29, 2010 10:08 AM
To: gcc-patches@gcc.gnu.org
Cc: Fang, Changpeng; rguenther@suse.de; sebpop@gmail.com
Subject: Re: [patch] Improve prefetch heuristiWE cs

> Patch1: 0001-Do-not-insert-prefetches-if-they-would-hit-the-same-.patch
> This patch modify the prefetch generation logic. We don't issue a prefetch
> if it would fall on the same cache line with an existing memory reference or
> prefetch.  This patch improves the following benchmarks: 416.gamess (~7%),
> 434.zeusmp (~4%), 454.calculix (~2%) and 445.gobmk (~2%).

>+      /* Don't issue a prefetch if its address falls on the same cache line
>+         with a previous memory reference (prefetch/load/store).  */
>+      if (abs (delta - start_offset) < L1_CACHE_LINE_SIZE)
>+        /* Drop the prefetch.  */

This might need a debug log entry about a dropped prefetch.

>+        continue;
>+      else

I looked a bit further into these patches.
This patch causes a regression with lbm on s390.

This is causes by the fact that ahead is 1 step is 160 and with
cache-line-size=256 this patch drops all prefetches on s390 in 
lbms hotloop.

Thinking more about the whole logic, wouldnt it make sense to check for 
abs (delta - start_offset) < L1_CACHE_LINE_SIZE / 2
                                                ^^^
because we cannot assume that we were starting at the beginning of a cache line?

Or maybe have something like
abs (delta - start_offset) < L1_CACHE_LINE_SIZE / CACHE_LINE_AGAIN_FACTOR

which can be defined by the architecture.

Christian