[Patch, avr] Provide correct LD offset bound in avr_address_cost

[Patch, avr] Provide correct LD offset bound in avr_address_cost

[Senthil Kumar Selvaraj]

Hi,

  This patch fixes cost computation in avr_address_cost - instead of the
  hardcoded 61, it uses the already existing MAX_LD_OFFSET(mode) macro.

  This showed up when investigating a code size regression in the ivopts
  pass. That pass computes address_cost with and without an offset to
  decide on the right induction variable candidate(s). The legitimate
  address target hook returns false for offsets more than 63, so the
  pass calls the TARGET_ADDRESS_COST hook with 62 as the offset.

  The hook implementation returns 18, and the ivopts pass concludes that
  the cost of address with *any* offset is 18, which is not true - the higher
  cost is incurred only with offsets bigger than MAX_LD_OFFSET. This in
  turn results in a suboptimal choice of induction variables in the
  ivopts pass. The patch changes the hardcoded 61 to use the mode
  specific MAX_LD_OFFSET instead.

  Regression testing with just that fix showed one additional
  compilation timeout. That turned out to be the same as
  https://lists.nongnu.org/archive/html/avr-gcc-list/2014-03/msg00010.html
  - the middle end takes too much time to decide on the best strategy to
  multiply DImode values on a 64 bit host. This already causes timeouts
  for a few builtin-arith-overflow-* tests (see
  https://gcc.gnu.org/ml/gcc-testresults/2016-09/msg02018.html), so it
  isn't really related to this fix. Just providing a cost estimate for
  DImode mul fixes the timeout though, so the patch does that by scaling
  SImode costs by 2 for DImode muls.

  With both changes in, there are no regressions, and the
  builtin-arith-overflow-* tests now PASS and don't timeout.

  Ok to commit to trunk and backport to 5.x?

Regards
Senthil
   

gcc/ChangeLog:

2016-09-21  Senthil Kumar Selvaraj  <senthil_kumar.selvaraj@atmel.com>

        * config/avr/avr.c (avr_rtx_costs_1): Handle DImode MULT.
        (avr_address_cost): Replace 61 with MAX_LD_OFFSET(mode).



diff --git gcc/config/avr/avr.c gcc/config/avr/avr.c
index 148a61d..29f0022 100644
--- gcc/config/avr/avr.c
+++ gcc/config/avr/avr.c
@@ -10257,6 +10257,7 @@ avr_rtx_costs_1 (rtx x, machine_mode mode, int outer_code ATTRIBUTE_UNUSED,
           break;
 
 	case SImode:
+	case DImode:
 	  if (AVR_HAVE_MUL)
             {
               if (!speed)
@@ -10282,7 +10283,10 @@ avr_rtx_costs_1 (rtx x, machine_mode mode, int outer_code ATTRIBUTE_UNUSED,
                 *total = COSTS_N_INSNS (AVR_HAVE_JMP_CALL ? 5 : 4);
             }
 
-          return true;
+	   if (mode == DImode)
+	     *total *= 2;
+
+	   return true;
 
 	default:
 	  return false;
@@ -10863,7 +10867,7 @@ avr_address_cost (rtx x, machine_mode mode ATTRIBUTE_UNUSED,
       && (REG_P (XEXP (x, 0))
           || GET_CODE (XEXP (x, 0)) == SUBREG))
     {
-      if (INTVAL (XEXP (x, 1)) >= 61)
+      if (INTVAL (XEXP (x, 1)) > MAX_LD_OFFSET(mode))
         cost = 18;
     }
   else if (CONSTANT_ADDRESS_P (x))

Re: [Patch, avr] Provide correct LD offset bound in avr_address_cost

[Denis Chertykov]

2016-09-22 7:52 GMT+03:00 Senthil Kumar Selvaraj
<senthil_kumar.selvaraj@atmel.com>:
> Hi,
>
>   This patch fixes cost computation in avr_address_cost - instead of the
>   hardcoded 61, it uses the already existing MAX_LD_OFFSET(mode) macro.
>
>   This showed up when investigating a code size regression in the ivopts
>   pass. That pass computes address_cost with and without an offset to
>   decide on the right induction variable candidate(s). The legitimate
>   address target hook returns false for offsets more than 63, so the
>   pass calls the TARGET_ADDRESS_COST hook with 62 as the offset.
>
>   The hook implementation returns 18, and the ivopts pass concludes that
>   the cost of address with *any* offset is 18, which is not true - the higher
>   cost is incurred only with offsets bigger than MAX_LD_OFFSET. This in
>   turn results in a suboptimal choice of induction variables in the
>   ivopts pass. The patch changes the hardcoded 61 to use the mode
>   specific MAX_LD_OFFSET instead.
>
>   Regression testing with just that fix showed one additional
>   compilation timeout. That turned out to be the same as
>   https://lists.nongnu.org/archive/html/avr-gcc-list/2014-03/msg00010.html
>   - the middle end takes too much time to decide on the best strategy to
>   multiply DImode values on a 64 bit host. This already causes timeouts
>   for a few builtin-arith-overflow-* tests (see
>   https://gcc.gnu.org/ml/gcc-testresults/2016-09/msg02018.html), so it
>   isn't really related to this fix. Just providing a cost estimate for
>   DImode mul fixes the timeout though, so the patch does that by scaling
>   SImode costs by 2 for DImode muls.
>
>   With both changes in, there are no regressions, and the
>   builtin-arith-overflow-* tests now PASS and don't timeout.
>
>   Ok to commit to trunk and backport to 5.x?

A deep analysis...
Approved.


>
> Regards
> Senthil
>
>
> gcc/ChangeLog:
>
> 2016-09-21  Senthil Kumar Selvaraj  <senthil_kumar.selvaraj@atmel.com>
>
>         * config/avr/avr.c (avr_rtx_costs_1): Handle DImode MULT.
>         (avr_address_cost): Replace 61 with MAX_LD_OFFSET(mode).
>
>
>
> diff --git gcc/config/avr/avr.c gcc/config/avr/avr.c
> index 148a61d..29f0022 100644
> --- gcc/config/avr/avr.c
> +++ gcc/config/avr/avr.c
> @@ -10257,6 +10257,7 @@ avr_rtx_costs_1 (rtx x, machine_mode mode, int outer_code ATTRIBUTE_UNUSED,
>            break;
>
>         case SImode:
> +       case DImode:
>           if (AVR_HAVE_MUL)
>              {
>                if (!speed)
> @@ -10282,7 +10283,10 @@ avr_rtx_costs_1 (rtx x, machine_mode mode, int outer_code ATTRIBUTE_UNUSED,
>                  *total = COSTS_N_INSNS (AVR_HAVE_JMP_CALL ? 5 : 4);
>              }
>
> -          return true;
> +          if (mode == DImode)
> +            *total *= 2;
> +
> +          return true;
>
>         default:
>           return false;
> @@ -10863,7 +10867,7 @@ avr_address_cost (rtx x, machine_mode mode ATTRIBUTE_UNUSED,
>        && (REG_P (XEXP (x, 0))
>            || GET_CODE (XEXP (x, 0)) == SUBREG))
>      {
> -      if (INTVAL (XEXP (x, 1)) >= 61)
> +      if (INTVAL (XEXP (x, 1)) > MAX_LD_OFFSET(mode))
>          cost = 18;
>      }
>    else if (CONSTANT_ADDRESS_P (x))