From mboxrd@z Thu Jan  1 00:00:00 1970
Return-Path: <gcc-patches-return-292098-listarch-gcc-patches=gcc.gnu.org@gcc.gnu.org>
Received: (qmail 15160 invoked by alias); 18 May 2011 09:39:36 -0000
Received: (qmail 15146 invoked by uid 22791); 18 May 2011 09:39:34 -0000
X-SWARE-Spam-Status: No, hits=-1.4 required=5.0	tests=AWL,BAYES_00,TW_EG,T_RP_MATCHES_RCVD
X-Spam-Check-By: sourceware.org
Received: from mail.codesourcery.com (HELO mail.codesourcery.com) (38.113.113.100)    by sourceware.org (qpsmtpd/0.43rc1) with ESMTP; Wed, 18 May 2011 09:39:17 +0000
Received: (qmail 953 invoked from network); 18 May 2011 09:39:16 -0000
Received: from unknown (HELO ?192.168.1.68?) (vries@127.0.0.2)  by mail.codesourcery.com with ESMTPA; 18 May 2011 09:39:16 -0000
Message-ID: <4DD39392.7000400@codesourcery.com>
Date: Wed, 18 May 2011 13:00:00 -0000
From: Tom de Vries <vries@codesourcery.com>
User-Agent: Mozilla/5.0 (X11; U; Linux x86_64; en-US; rv:1.9.2.17) Gecko/20110424 Lightning/1.0b2 Thunderbird/3.1.10
MIME-Version: 1.0
To: Zdenek Dvorak <rakdver@kam.mff.cuni.cz>
CC: gcc-patches@gcc.gnu.org
Subject: Re: [PATCH, PR45098, 3/10]
References: <4DD21F6E.4050308@codesourcery.com> <4DD220C1.7000405@codesourcery.com> <20110518072620.GA18720@kam.mff.cuni.cz>
In-Reply-To: <20110518072620.GA18720@kam.mff.cuni.cz>
Content-Type: text/plain; charset=ISO-8859-1
Content-Transfer-Encoding: 7bit
Mailing-List: contact gcc-patches-help@gcc.gnu.org; run by ezmlm
Precedence: bulk
List-Id: <gcc-patches.gcc.gnu.org>
List-Archive: <http://gcc.gnu.org/ml/gcc-patches/>
List-Post: <mailto:gcc-patches@gcc.gnu.org>
List-Help: <mailto:gcc-patches-help@gcc.gnu.org>
Sender: gcc-patches-owner@gcc.gnu.org
X-SW-Source: 2011-05/txt/msg01269.txt.bz2

Hi Zdenek,

thanks for the review.

On 05/18/2011 09:26 AM, Zdenek Dvorak wrote:
>> 2011-05-05  Tom de Vries  <tom@codesourcery.com>
>>
>> 	PR target/45098
>> 	* tree-ssa-loop-ivopts.c (computation_cost): Prevent cost of 0.
> 
> this looks strange.  Something like
> 
>   cost = seq_cost (seq, speed);
>   if (MEM_P (rslt))
>     the current code;
>   else
>     cost += rtx_cost (rslt, SET, speed));
> 
> would make more sense to me (if I understand correctly what you are
> trying to achieve).
> 

Sorry for not putting an explanation in the first place.

I'm trying to achieve that the cost of forcing an int into a reg is not
0. Currently, if the expansion results in an rtx expression rather than
an insn, seq is NULL and seq_cost returns 0, after which
computation_cost returns 0:
...
(gdb) call debug_generic_expr (expr)
2000
(gdb) call debug_rtx (rslt)
(const_int 2000 [0x7d0])
(gdb) p seq
$5 = (rtx) 0x0
...

Using an assert on an x86_64 build, I found this case where seq != NULL_RTX:
...
gdb) call debug_generic_expr (expr)
(int) ((unsigned int) ivtmp.11 * 8)
(gdb) call debug_rtx (rslt)
(reg:SI 62)
(gdb) call debug_rtx (seq)
(insn 4 0 0 (parallel [
            (set (reg:SI 62)
                (ashift:SI (subreg:SI (reg:DI 59 [ ivtmp.11 ]) 0)
                    (const_int 3 [0x3])))
            (clobber (reg:CC 17 flags))
        ]) -1
     (nil))
...
we don't need to additionally count a regcopy of r62 on top of seq_cost
in this case.

How about:
...
@@ -2866,6 +2878,8 @@ computation_cost (tree expr, bool speed)
   if (MEM_P (rslt))
     cost += address_cost (XEXP (rslt, 0), TYPE_MODE (type),
 			  TYPE_ADDR_SPACE (type), speed);
+  else if (!REG_P (rslt))
+    cost += (unsigned)rtx_cost (rslt, SET, speed);

   return cost;
 }
...
?

Thanks
- Tom