Re: [RFC PATCH v2 15/44] target/loongarch: Implement vmul/vmuh/vmulw{ev/od}

[gaosong]

HI, Richard

在 2023/3/29 上午4:46, Richard Henderson 写道:

> > +static void
> >         do_vmuh_s(unsigned vece, uint32_t vd_ofs, uint32_t vj_ofs,
> >         
> >         +                      uint32_t vk_ofs, uint32_t oprsz, uint32_t
> >         maxsz)
> >         
> >         +{
> >         
> >         +    static const GVecGen3 op[4] = {
> >         
> >         +        {
> >         
> >         +            .fno = gen_helper_vmuh_b,
> >         
> >         +            .vece = MO_8
> >         
> >         +        },
> >         
> >         +        {
> >         
> >         +            .fno = gen_helper_vmuh_h,
> >         
> >         +            .vece = MO_16
> >         
> >         +        },
> >         
> >         +        {
> >         
> >         +            .fno = gen_helper_vmuh_w,
> >         
> >         +            .vece = MO_32
> >         
> >         +        },
> >         
> >         +        {
> >         
> >         +            .fno = gen_helper_vmuh_d,
> >         
> >         +            .vece = MO_64
> >         
> >         +        },
> >         
> >         +    };
> >         
> >       
>       
>       Could be worth integer expansion, especially for MO_32/MO_64?
>       
>       Should be trivial... 
>     
    For integer expansion.  How about the following code?
    
    static void gen_vmuh_b(TCGv_i64 t, TCGv_i64 a, TCGv_i64 b)
    {
        int i;
        TCGv_i64 t1, t2;
    
        t1 = tcg_temp_new_i64();
        t2 = tcg_temp_new_i64();
    
        tcg_gen_mov_i64(t, tcg_constant_i64(0));
    
        for (i = 0; i < 8; i++) {
            tcg_gen_shri_i64(t1, a, 8 *i);
            tcg_gen_shri_i64(t2, b, 8 *i);
            tcg_gen_ext8s_i64(t1, t1);
            tcg_gen_ext8s_i64(t2, t2);
            tcg_gen_mul_i64(t1, t1, t2);
            tcg_gen_andi_i64(t1, t1, 0xffff);
            tcg_gen_shri_i64(t1, t1, 8); 
            tcg_gen_shli_i64(t1, t1, 8 * i);
            tcg_gen_or_i64(t, t, t1);
        }
    }
    
    static void gen_vmuh_h(TCGv_i64 t, TCGv_i64 a, TCGv_i64 b)
    {
        int i;
        TCGv_i64 t1, t2;
    
        t1 = tcg_temp_new_i64();
        t2 = tcg_temp_new_i64();
    
        tcg_gen_mov_i64(t, tcg_constant_i64(0));
    
        for ( i = 0; i < 4; i++) {
            tcg_gen_shri_i64(t1, a, 16 *i);
            tcg_gen_shri_i64(t2, b, 16*i);
            tcg_gen_ext16s_i64(t1, t1);
            tcg_gen_ext16s_i64(t2, t2);
            tcg_gen_mul_i64(t1, t1, t2);
            tcg_gen_andi_i64(t1, t1, 0xffffffff);
            tcg_gen_shri_i64(t1, t1, 16);
            tcg_gen_shli_i64(t1, t1, 16 * i);
            tcg_gen_or_i64(t, t, t1);
        }
    }
    
    static void gen_vmuh_w(TCGv_i32 t, TCGv_i32 a, TCGv_i32 b)
    {
        TCGv_i64 t1, t2;
    
        t1 = tcg_temp_new_i64();
        t2 = tcg_temp_new_i64();
        tcg_gen_ext_i32_i64(t1, a);
        tcg_gen_ext_i32_i64(t2, b);
        tcg_gen_mul_i64(t2, t1, t2);
        tcg_gen_extrh_i64_i32(t, t2);
    }
    
    static void gen_vmuh_d(TCGv_i64 t, TCGv_i64 a, TCGv_i64 b)
    {
        TCGv_i64 t1;
    
        t1 = tcg_temp_new_i64();
        tcg_gen_muls2_i64(t1, t, a, b);
    }
    
    static void gen_vmuh_bu(TCGv_i64 t, TCGv_i64 a, TCGv_i64 b)
    {
        int i;
        TCGv_i64 t1, t2;
    
        t1 = tcg_temp_new_i64();
        t2 = tcg_temp_new_i64();
    
        tcg_gen_mov_i64(t, tcg_constant_i64(0));
    
        for (i = 0; i < 8; i++) {
            tcg_gen_shri_i64(t1, a, 8 * i);
            tcg_gen_shri_i64(t2, b, 8 * i);
            tcg_gen_ext8u_i64(t1, t1);
            tcg_gen_ext8u_i64(t2, t2);
            tcg_gen_mul_i64(t1, t1, t2);
            tcg_gen_shri_i64(t1, t1, 8);
            tcg_gen_shli_i64(t1, t1, 8 * i);
            tcg_gen_or_i64(t, t, t1);
        }
    }
    
    static void gen_vmuh_hu(TCGv_i64 t, TCGv_i64 a, TCGv_i64 b)
    {
        int i;
        TCGv_i64 t1, t2;
    
        t1 = tcg_temp_new_i64();
        t2 = tcg_temp_new_i64();
    
        tcg_gen_mov_i64(t, tcg_constant_i64(0));
    
        for (i = 0; i < 4; i++) {
            tcg_gen_shri_i64(t1, a, 16 *i);
            tcg_gen_shri_i64(t2, b, 16*i);
            tcg_gen_ext16u_i64(t1, t1);
            tcg_gen_ext16u_i64(t2, t2);
            tcg_gen_mul_i64(t1, t1, t2);
            tcg_gen_shri_i64(t1, t1, 16);
            tcg_gen_shli_i64(t1, t1, 16 * i);
            tcg_gen_or_i64(t, t, t1);
        }
    }
    
    static void gen_vmuh_wu(TCGv_i32 t, TCGv_i32 a, TCGv_i32 b)
    {
        TCGv_i64 t1, t2;
    
        t1 = tcg_temp_new_i64();
        t2 = tcg_temp_new_i64();
        tcg_gen_extu_i32_i64(t1, a);
        tcg_gen_extu_i32_i64(t2, b);
        tcg_gen_mul_i64(t2, t1, t2);
        tcg_gen_extrh_i64_i32(t, t2);
    }
    
    static void gen_vmuh_du(TCGv_i64 t, TCGv_i64 a, TCGv_i64 b)
    {
        TCGv_i64 t1;
    
        t1 = tcg_temp_new_i64();
        tcg_gen_mulu2_i64(t1, t, a, b);
    }
    
    Thanks.
    Song Gao