[PATCH][ARM] Do not lower cost of setting core reg to constant. It doesn't have any effect

[PATCH][ARM] Do not lower cost of setting core reg to constant. It doesn't have any effect

[Kyrill Tkachov]

[-- Attachment #1: Type: text/plain, Size: 655 bytes --]

Hi all,

This hunk that slightly reduces the cost of immediate moves doesn't actually have any effect.
In the whole of SPEC2006 it didn't make a difference. In any case, I'd like to move to a point
where we use COSTS_N_INSNS units for our costs and not increment decrement them by one.

This patch removes that bit of logic and makes it slightly cleaner to look at. As far as I know
its logic has never been confirmed in practice.

Bootstrapped and tested on arm.

Ok for trunk?

Thanks,
Kyrill

2015-04-22  Kyrylo Tkachov  <kyrylo.tkachov@arm.com>

     * config/arm/arm.c (arm_new_rtx_costs): Do not lower cost
     immediate moves.

[-- Warning: decoded text below may be mangled, UTF-8 assumed --]
[-- Attachment #2: arm-costs-reg-core-lower.patch --]
[-- Type: text/x-patch; name=arm-costs-reg-core-lower.patch, Size: 1001 bytes --]

commit e225669ff70f09520007b7898b170fb8fa75281f
Author: Kyrylo Tkachov <kyrylo.tkachov@arm.com>
Date:   Wed Apr 8 10:18:23 2015 +0100

    [ARM] Do not lower cost of setting core reg to constant. It doesn't have any effect

diff --git a/gcc/config/arm/arm.c b/gcc/config/arm/arm.c
index 0ef05c9..03988ac 100644
--- a/gcc/config/arm/arm.c
+++ b/gcc/config/arm/arm.c
@@ -9725,11 +9725,7 @@ arm_new_rtx_costs (rtx x, enum rtx_code code, enum rtx_code outer_code,
 	     and we would otherwise be unable to work out the true cost.  */
 	  *cost = rtx_cost (SET_DEST (x), SET, 0, speed_p);
 	  outer_code = SET;
-	  /* Slightly lower the cost of setting a core reg to a constant.
-	     This helps break up chains and allows for better scheduling.  */
-	  if (REG_P (SET_DEST (x))
-	      && REGNO (SET_DEST (x)) <= LR_REGNUM)
-	    *cost -= 1;
+
 	  x = SET_SRC (x);
 	  /* Immediate moves with an immediate in the range [0, 255] can be
 	     encoded in 16 bits in Thumb mode.  */

Re: [PATCH][ARM] Do not lower cost of setting core reg to constant. It doesn't have any effect

[Kyrill Tkachov]

Ping.
https://gcc.gnu.org/ml/gcc-patches/2015-04/msg01330.html

Thanks,
Kyrill

On 22/04/15 17:18, Kyrill Tkachov wrote:
> Hi all,
>
> This hunk that slightly reduces the cost of immediate moves doesn't actually have any effect.
> In the whole of SPEC2006 it didn't make a difference. In any case, I'd like to move to a point
> where we use COSTS_N_INSNS units for our costs and not increment decrement them by one.
>
> This patch removes that bit of logic and makes it slightly cleaner to look at. As far as I know
> its logic has never been confirmed in practice.
>
> Bootstrapped and tested on arm.
>
> Ok for trunk?
>
> Thanks,
> Kyrill
>
> 2015-04-22  Kyrylo Tkachov  <kyrylo.tkachov@arm.com>
>
>       * config/arm/arm.c (arm_new_rtx_costs): Do not lower cost
>       immediate moves.

Re: [PATCH][ARM] Do not lower cost of setting core reg to constant. It doesn't have any effect

[Marcus Shawcroft]

On 22 April 2015 at 17:18, Kyrill Tkachov <kyrylo.tkachov@arm.com> wrote:

> 2015-04-22  Kyrylo Tkachov  <kyrylo.tkachov@arm.com>
>
>     * config/arm/arm.c (arm_new_rtx_costs): Do not lower cost
>     immediate moves.

OK
/Marcus

Re: [PATCH][ARM] Do not lower cost of setting core reg to constant. It doesn't have any effect

[Marcus Shawcroft]

On 30 April 2015 at 16:22, Marcus Shawcroft <marcus.shawcroft@gmail.com> wrote:
> On 22 April 2015 at 17:18, Kyrill Tkachov <kyrylo.tkachov@arm.com> wrote:
>
>> 2015-04-22  Kyrylo Tkachov  <kyrylo.tkachov@arm.com>
>>
>>     * config/arm/arm.c (arm_new_rtx_costs): Do not lower cost
>>     immediate moves.
>
> OK
> /Marcus

Ignore that, I'm not allowed to make that call. Wait for Ramana.
/Marcus

Re: [PATCH][ARM] Do not lower cost of setting core reg to constant. It doesn't have any effect

[Richard Sandiford]

Kyrill Tkachov <kyrylo.tkachov@arm.com> writes:
> Hi all,
>
> This hunk that slightly reduces the cost of immediate moves doesn't
> actually have any effect.  In the whole of SPEC2006 it didn't make a
> difference. In any case, I'd like to move to a point where we use
> COSTS_N_INSNS units for our costs and not increment decrement them by
> one.

I wonder whether that's always a good idea though?  COSTS_N_INSNS exists
to allow these kinds of fractional costs.  It sounds like they're
not useful in this particular case, but e.g. one case where they
can be useful is if you want to say when optimising for size that two
instructions are the same size but one is slightly more preferable for
speed reasons.  This might help prefer a 2-instruction shift/add
sequence over a load-immediate followed by a general multiplication;
both have the same size, but the shift/add is often faster.  Giving
multiplication a slightly higher cost than COSTS_N_INSNS (1) makes
that clear.

It's not perfect of course.  Add too many fractional costs together
and a carry will give you an extra full instruction at some fairly
arbitrary point.  Maybe size costs should be a (size, speed) pair.

Thanks,
Richard

Re: [PATCH][ARM] Do not lower cost of setting core reg to constant. It doesn't have any effect

[Kyrill Tkachov]

Hi Richard,

On 06/05/15 08:15, Richard Sandiford wrote:
> Kyrill Tkachov <kyrylo.tkachov@arm.com> writes:
>> Hi all,
>>
>> This hunk that slightly reduces the cost of immediate moves doesn't
>> actually have any effect.  In the whole of SPEC2006 it didn't make a
>> difference. In any case, I'd like to move to a point where we use
>> COSTS_N_INSNS units for our costs and not increment decrement them by
>> one.
> I wonder whether that's always a good idea though?  COSTS_N_INSNS exists
> to allow these kinds of fractional costs.  It sounds like they're
> not useful in this particular case, but e.g. one case where they
> can be useful is if you want to say when optimising for size that two
> instructions are the same size but one is slightly more preferable for
> speed reasons.  This might help prefer a 2-instruction shift/add
> sequence over a load-immediate followed by a general multiplication;
> both have the same size, but the shift/add is often faster.  Giving
> multiplication a slightly higher cost than COSTS_N_INSNS (1) makes
> that clear.

Yeah, that's the mult synthesis code in expmed.c
I think the problem here with the costs infrastructure
is that the backend doesn't know for what purpose it's
assigning costs. If it's told to assign a cost to a PLUS
it has no way of knowing that the add is being considered
as part of a replacement for a mult-immediate or whether
to merge it with a shift during combine or as part of an
address calculation during ivpopts or whatever other place
in the midend that calls rtx costs.

>
> It's not perfect of course.  Add too many fractional costs together
> and a carry will give you an extra full instruction at some fairly
> arbitrary point.  Maybe size costs should be a (size, speed) pair.

Yeah, that's a concern as well.
In any case, I'm not going to remove such increments/decrements
of costs from the backend without analysing their impact on
code quality.

Thanks,
Kyrill

>
> Thanks,
> Richard
>