On 15/04/14 11:56, Kyrill Tkachov wrote:
> Hi all,
>
> This patch relates to PR60663 where cse got confused due to asm statements
> being given a cost of zero in the arm backend. Jakub already put in a fix to
> cse for 4.9.0 (http://gcc.gnu.org/ml/gcc-patches/2014-04/msg00512.html) but
> we should still fix the costs in arm.
>
> This patch does that by estimating the number of instructions in the asm
> statement, adding the cost of the input operands and making sure that it's at
> least COSTS_N_INSNS (1).
>
> Tested and bootstrapped on arm-none-linux-gnueabihf.
>
> Ok for trunk?
>
> Thanks,
> Kyrill
>
> 2014-04-15 Kyrylo Tkachov <[email protected]>
>
> PR rtl-optimization/60663
> * config/arm/arm.c (arm_new_rtx_costs): Improve ASM_OPERANDS case,
> avoid 0 cost.
>
I'd be inclined to use 1 + number of operands, rather than MAX.
OK with that change.
R.
>
> pr60663.patch
>
>
> diff --git a/gcc/config/arm/arm.c b/gcc/config/arm/arm.c
> index 91e4cd8..ce7ee82 100644
> --- a/gcc/config/arm/arm.c
> +++ b/gcc/config/arm/arm.c
> @@ -10758,10 +10758,16 @@ arm_new_rtx_costs (rtx x, enum rtx_code code, enum
> rtx_code outer_code,
> return true;
>
> case ASM_OPERANDS:
> - /* Just a guess. Cost one insn per input. */
> - *cost = COSTS_N_INSNS (ASM_OPERANDS_INPUT_LENGTH (x));
> - return true;
> + {
> + /* Just a guess. Guess number of instructions in the asm
> + plus one insn per input. Always a minimum of COSTS_N_INSNS (1)
> + though (see PR60663). */
> + int asm_length = asm_str_count (ASM_OPERANDS_TEMPLATE (x));
> + int num_operands = ASM_OPERANDS_INPUT_LENGTH (x);
>
> + *cost = COSTS_N_INSNS (MAX (1, asm_length + num_operands));
> + return true;
> + }
> default:
> if (mode != VOIDmode)
> *cost = COSTS_N_INSNS (ARM_NUM_REGS (mode));
>
