Re: [PATCH 05/10] target/ppc: Implemented vector divide extended word

[Richard Henderson]

On 3/30/22 14:25, Lucas Mateus Castro(alqotel) wrote:
> +static void do_vx_dives_i32(TCGv_i32 t, TCGv_i32 a, TCGv_i32 b)
> +{
> +    TCGv_i64 res, val1, val2;
> +    TCGv_i64 zero = tcg_constant_i64(0);
> +    TCGv_i64 one =  tcg_constant_i64(1);
> +
> +    res = tcg_temp_new_i64();
> +    val1 = tcg_temp_new_i64();
> +    val2 = tcg_temp_new_i64();
> +
> +    tcg_gen_ext_i32_i64(val1, a);
> +    tcg_gen_ext_i32_i64(val2, b);
> +
> +    /* return 0 if b = 0, so make b = 1 so the result doesn't fit in 32 bits*/
> +    tcg_gen_movcond_i64(TCG_COND_EQ, val2, val2, zero, one, val2);

Need int_min / -1 check.

> +    /* (a << 32)/b */
> +    tcg_gen_shli_i64(val1, val1, 32);
> +    tcg_gen_div_i64(res, val1, val2);
> +
> +    tcg_gen_ext32s_i64(val1, res);
> +    /* if result is undefined (quotient doesn't fit in 32 bits) return 0 */
> +    tcg_gen_movcond_i64(TCG_COND_EQ, res, res, val1, res, zero);

Again, I don't see the point in producing 0 for undefined.

> +    tcg_gen_ext_i32_i64(val1, a);
> +    tcg_gen_extu_i32_i64(val2, b);

Better with extu for val1, just because a is logically unsigned.


r~