[PATCH][ARM] Fix costing of sign-extending load in rtx costs

[PATCH][ARM] Fix costing of sign-extending load in rtx costs

[Kyrill Tkachov]

[-- Attachment #1: Type: text/plain, Size: 869 bytes --]

Hi all,

Another costs issue that came out of the investigation for PR 65932 is that
sign-extending loads get a higher cost than they should in the arm backend.
The problem is that when handling a sign-extend of a MEM we add the cost
of the load_sign_extend cost field and then recursively add the cost of the inner MEM
rtx, which is bogus. This will end up adding an extra load cost on it.

The solution in this patch is to just remove that recursive step.
With this patch from various CSE dumps I see much more sane costs assign to these
expressions (such as 12 instead of 32 or higher).

Bootstrapped and tested on arm-none-linux-gnueabihf.

Ok for trunk?

Thanks,
Kyrill

2016-04-27  Kyrylo Tkachov  <kyrylo.tkachov@arm.com>

     * config/arm/arm.c (arm_new_rtx_costs, SIGN_EXTEND case):
     Don't add cost of inner memory when handling sign-extended
     loads.

[-- Attachment #2: arm-ldsh-cost.patch --]
[-- Type: text/x-patch, Size: 520 bytes --]

diff --git a/gcc/config/arm/arm.c b/gcc/config/arm/arm.c
index 7781b4b449ed48a8d902802d8e6a5c8e1ae7793f..7f2babe7339de3586de190bbe2cf8112919dd96f 100644
--- a/gcc/config/arm/arm.c
+++ b/gcc/config/arm/arm.c
@@ -10911,8 +10911,6 @@ arm_new_rtx_costs (rtx x, enum rtx_code code, enum rtx_code outer_code,
       if ((arm_arch4 || GET_MODE (XEXP (x, 0)) == SImode)
 	  && MEM_P (XEXP (x, 0)))
 	{
-	  *cost = rtx_cost (XEXP (x, 0), VOIDmode, code, 0, speed_p);
-
 	  if (mode == DImode)
 	    *cost += COSTS_N_INSNS (1);
 

Re: [PATCH][ARM] Fix costing of sign-extending load in rtx costs

[Kyrill Tkachov]

Ping.
https://gcc.gnu.org/ml/gcc-patches/2016-04/msg01655.html

Thanks,
Kyrill

On 27/04/16 15:13, Kyrill Tkachov wrote:
> Hi all,
>
> Another costs issue that came out of the investigation for PR 65932 is that
> sign-extending loads get a higher cost than they should in the arm backend.
> The problem is that when handling a sign-extend of a MEM we add the cost
> of the load_sign_extend cost field and then recursively add the cost of the inner MEM
> rtx, which is bogus. This will end up adding an extra load cost on it.
>
> The solution in this patch is to just remove that recursive step.
> With this patch from various CSE dumps I see much more sane costs assign to these
> expressions (such as 12 instead of 32 or higher).
>
> Bootstrapped and tested on arm-none-linux-gnueabihf.
>
> Ok for trunk?
>
> Thanks,
> Kyrill
>
> 2016-04-27  Kyrylo Tkachov  <kyrylo.tkachov@arm.com>
>
>     * config/arm/arm.c (arm_new_rtx_costs, SIGN_EXTEND case):
>     Don't add cost of inner memory when handling sign-extended
>     loads.

Re: [PATCH][ARM] Fix costing of sign-extending load in rtx costs

[Kyrill Tkachov]

Ping.
Thanks,
Kyrill

On 11/05/16 10:00, Kyrill Tkachov wrote:
> Ping.
> https://gcc.gnu.org/ml/gcc-patches/2016-04/msg01655.html
>
> Thanks,
> Kyrill
>
> On 27/04/16 15:13, Kyrill Tkachov wrote:
>> Hi all,
>>
>> Another costs issue that came out of the investigation for PR 65932 is that
>> sign-extending loads get a higher cost than they should in the arm backend.
>> The problem is that when handling a sign-extend of a MEM we add the cost
>> of the load_sign_extend cost field and then recursively add the cost of the inner MEM
>> rtx, which is bogus. This will end up adding an extra load cost on it.
>>
>> The solution in this patch is to just remove that recursive step.
>> With this patch from various CSE dumps I see much more sane costs assign to these
>> expressions (such as 12 instead of 32 or higher).
>>
>> Bootstrapped and tested on arm-none-linux-gnueabihf.
>>
>> Ok for trunk?
>>
>> Thanks,
>> Kyrill
>>
>> 2016-04-27  Kyrylo Tkachov  <kyrylo.tkachov@arm.com>
>>
>>     * config/arm/arm.c (arm_new_rtx_costs, SIGN_EXTEND case):
>>     Don't add cost of inner memory when handling sign-extended
>>     loads.
>

Re: [PATCH][ARM] Fix costing of sign-extending load in rtx costs

[Ramana Radhakrishnan]

On 27/04/16 15:13, Kyrill Tkachov wrote:
> Hi all,
> 
> Another costs issue that came out of the investigation for PR 65932 is that
> sign-extending loads get a higher cost than they should in the arm backend.
> The problem is that when handling a sign-extend of a MEM we add the cost
> of the load_sign_extend cost field and then recursively add the cost of the inner MEM
> rtx, which is bogus. This will end up adding an extra load cost on it.
> 
> The solution in this patch is to just remove that recursive step.
> With this patch from various CSE dumps I see much more sane costs assign to these
> expressions (such as 12 instead of 32 or higher).
> 
> Bootstrapped and tested on arm-none-linux-gnueabihf.
> 
> Ok for trunk?
> 
> Thanks,
> Kyrill
> 
> 2016-04-27  Kyrylo Tkachov  <kyrylo.tkachov@arm.com>
> 
>     * config/arm/arm.c (arm_new_rtx_costs, SIGN_EXTEND case):
>     Don't add cost of inner memory when handling sign-extended
>     loads.


Ok ... it took me a while to work out that this was sane.


regards
Ramana