On 24/03/14 17:15, Kyrill Tkachov wrote:
> Hi all,
>
> This patch adds proper rtx costing logic for floating point fma operations on
> arm. It also handles the fma+neg combinations that can be expressed with
> vfms,vfnma or vfnms instructions.
> Not much else to say here...
>
> Tested and boostrapped on arm-none-linux-gnueabihf.
>
> Ok for next stage1?
>
> Thanks,
> Kyrill
>
> 2014-03-24 Kyrylo Tkachov <kyrylo.tkac...@arm.com>
>
> * config/arm/arm.c (arm_new_rtx_costs): Handle FMA.
>
>
> arm-fma-costs.patch
>
>
> commit 9cebc525c432e4a87ead5b513b70c99dff6628c8
> Author: Kyrylo Tkachov <kyrylo.tkac...@arm.com>
> Date: Thu Mar 20 09:13:06 2014 +0000
>
> [ARM] Cost FMA operations properly.
>
> diff --git a/gcc/config/arm/arm.c b/gcc/config/arm/arm.c
> index 5619ebd..e7a0530 100644
> --- a/gcc/config/arm/arm.c
> +++ b/gcc/config/arm/arm.c
> @@ -10673,6 +10673,36 @@ arm_new_rtx_costs (rtx x, enum rtx_code code, enum
> rtx_code outer_code,
> *cost = LIBCALL_COST (1);
> return false;
>
> + case FMA:
> + if (TARGET_32BIT && TARGET_HARD_FLOAT && TARGET_FMA)
> + {
> + rtx op0 = XEXP (x, 0);
> + rtx op1 = XEXP (x, 1);
> + rtx op2 = XEXP (x, 2);
> +
> + *cost = COSTS_N_INSNS (1);
> +
> + /* vfms or vfnma. */
> + if (GET_CODE (op0) == NEG)
> + op0 = XEXP (op0, 0);
> +
> + /* vfnms or vfnma. */
> + if (GET_CODE (op2) == NEG)
> + op2 = XEXP (op2, 0);
> +
> + *cost += rtx_cost (op0, FMA, 0, speed_p);
> + *cost += rtx_cost (op1, FMA, 1, speed_p);
> + *cost += rtx_cost (op2, FMA, 2, speed_p);
> +
> + if (speed_p)
> + *cost += extra_cost->fp[mode ==DFmode].fma;
> +
> + return true;
> + }
> +
> + *cost = LIBCALL_COST (1);
This should be LIBCALL_COST (3) (there are 3 arguments to the operation).
Otherwise OK.
R.
