Re: [PATCH v3 05/16] tcg/loongarch64: Lower add/sub_vec to vadd/vsub

[Richard Henderson]

On 9/1/23 22:02, Jiajie Chen wrote:
> Lower the following ops:
>
> - add_vec
> - sub_vec
>
> Signed-off-by: Jiajie Chen <c@jia.je>
> ---
>   tcg/loongarch64/tcg-target-con-set.h |  1 +
>   tcg/loongarch64/tcg-target-con-str.h |  1 +
>   tcg/loongarch64/tcg-target.c.inc     | 60 ++++++++++++++++++++++++++++
>   3 files changed, 62 insertions(+)
>
> diff --git a/tcg/loongarch64/tcg-target-con-set.h 
> b/tcg/loongarch64/tcg-target-con-set.h
> index 8c8ea5d919..2d5dce75c3 100644
> --- a/tcg/loongarch64/tcg-target-con-set.h
> +++ b/tcg/loongarch64/tcg-target-con-set.h
> @@ -32,4 +32,5 @@ C_O1_I2(r, rZ, ri)
>   C_O1_I2(r, rZ, rJ)
>   C_O1_I2(r, rZ, rZ)
>   C_O1_I2(w, w, wM)
> +C_O1_I2(w, w, wA)
>   C_O1_I4(r, rZ, rJ, rZ, rZ)
> diff --git a/tcg/loongarch64/tcg-target-con-str.h 
> b/tcg/loongarch64/tcg-target-con-str.h
> index a8a1c44014..2ba9c135ac 100644
> --- a/tcg/loongarch64/tcg-target-con-str.h
> +++ b/tcg/loongarch64/tcg-target-con-str.h
> @@ -27,3 +27,4 @@ CONST('Z', TCG_CT_CONST_ZERO)
>   CONST('C', TCG_CT_CONST_C12)
>   CONST('W', TCG_CT_CONST_WSZ)
>   CONST('M', TCG_CT_CONST_VCMP)
> +CONST('A', TCG_CT_CONST_VADD)
> diff --git a/tcg/loongarch64/tcg-target.c.inc b/tcg/loongarch64/tcg-target.c.inc
> index 129dd92910..0edcf5be35 100644
> --- a/tcg/loongarch64/tcg-target.c.inc
> +++ b/tcg/loongarch64/tcg-target.c.inc
> @@ -177,6 +177,7 @@ static TCGReg tcg_target_call_oarg_reg(TCGCallReturnKind 
> kind, int slot)
>   #define TCG_CT_CONST_C12   0x1000
>   #define TCG_CT_CONST_WSZ   0x2000
>   #define TCG_CT_CONST_VCMP  0x4000
> +#define TCG_CT_CONST_VADD  0x8000
>   
>   #define ALL_GENERAL_REGS   MAKE_64BIT_MASK(0, 32)
>   #define ALL_VECTOR_REGS    MAKE_64BIT_MASK(32, 32)
> @@ -214,6 +215,9 @@ static bool tcg_target_const_match(int64_t val, TCGType 
> type, int ct, int vece)
>       if ((ct & TCG_CT_CONST_VCMP) && -0x10 <= vec_val && vec_val <= 0x1f) {
>           return true;
>       }
> +    if ((ct & TCG_CT_CONST_VADD) && -0x1f <= vec_val && vec_val <= 0x1f) {
> +        return true;
> +    }
>       return false;
>   }
>   
> @@ -1646,6 +1650,18 @@ static void tcg_out_vec_op(TCGContext *s, TCGOpcode opc,
>           [TCG_COND_LTU] = {OPC_VSLTI_BU, OPC_VSLTI_HU, OPC_VSLTI_WU, 
> OPC_VSLTI_DU},
>       };
>       LoongArchInsn insn;
> +    static const LoongArchInsn add_vec_insn[4] = {
> +        OPC_VADD_B, OPC_VADD_H, OPC_VADD_W, OPC_VADD_D
> +    };
> +    static const LoongArchInsn add_vec_imm_insn[4] = {
> +        OPC_VADDI_BU, OPC_VADDI_HU, OPC_VADDI_WU, OPC_VADDI_DU
> +    };
> +    static const LoongArchInsn sub_vec_insn[4] = {
> +        OPC_VSUB_B, OPC_VSUB_H, OPC_VSUB_W, OPC_VSUB_D
> +    };
> +    static const LoongArchInsn sub_vec_imm_insn[4] = {
> +        OPC_VSUBI_BU, OPC_VSUBI_HU, OPC_VSUBI_WU, OPC_VSUBI_DU
> +    };
>   
>       a0 = args[0];
>       a1 = args[1];
> @@ -1712,6 +1728,44 @@ static void tcg_out_vec_op(TCGContext *s, TCGOpcode opc,
>           }
>           tcg_out32(s, encode_vdvjvk_insn(insn, a0, a1, a2));
>           break;
> +    case INDEX_op_add_vec:
> +        if (const_args[2]) {
> +            int64_t value = sextract64(a2, 0, 8 << vece);
> +            /* Try vaddi/vsubi */
> +            if (0 <= value && value <= 0x1f) {
> +                tcg_out32(s, encode_vdvjuk5_insn(add_vec_imm_insn[vece], a0, \
> +                                                 a1, value));
> +                break;
> +            } else if (-0x1f <= value && value < 0) {
> +                tcg_out32(s, encode_vdvjuk5_insn(sub_vec_imm_insn[vece], a0, \
> +                                                 a1, -value));
> +                break;
> +            }
> +
> +            /* constraint TCG_CT_CONST_VADD ensures unreachable */
> +            g_assert_not_reached();
> +        }
> +        tcg_out32(s, encode_vdvjvk_insn(add_vec_insn[vece], a0, a1, a2));
> +        break;
> +    case INDEX_op_sub_vec:
> +        if (const_args[2]) {
> +            int64_t value = sextract64(a2, 0, 8 << vece);
> +            /* Try vaddi/vsubi */
> +            if (0 <= value && value <= 0x1f) {
> +                tcg_out32(s, encode_vdvjuk5_insn(sub_vec_imm_insn[vece], a0, \
> +                                                 a1, value));
> +                break;
> +            } else if (-0x1f <= value && value < 0) {
> +                tcg_out32(s, encode_vdvjuk5_insn(add_vec_imm_insn[vece], a0, \
> +                                                 a1, -value));
> +                break;
> +            }
> +
> +            /* constraint TCG_CT_CONST_VADD ensures unreachable */
> +            g_assert_not_reached();
> +        }
> +        tcg_out32(s, encode_vdvjvk_insn(sub_vec_insn[vece], a0, a1, a2));

It would be nice to share code here.  Perhaps

    case INDEX_op_sub_vec:
       if (!const_args[2]) {
           tcg_out32(s, encode_vdvjvk_insn(sub_vec_insn[vece], a0, a1, a2));
           break;
       }
       a2 = -a2;
       goto do_addi_vec;
   case INDEX_op_add_vec:
       if (!const_args[2]) {
           tcg_out32(s, encode_vdvjvk_insn(add_vec_insn[vece], a0, a1, a2));
           break;
       }
   do_addi_vec:
       ...

or a helper function.

Otherwise,
Reviewed-by: Richard Henderson <richard.henderson@linaro.org>


r~