Re: [PATCH 03/10] target/ppc: Implemented vector divide instructions

[Richard Henderson]

On 3/30/22 14:25, Lucas Mateus Castro(alqotel) wrote:
> +#define TRANS_VDIV_VMOD(FLAGS, NAME, VECE, FNI4_FUNC, FNI8_FUNC)        \
> +static bool trans_##NAME(DisasContext *ctx, arg_VX *a)                  \
> +{                                                                       \
> +    static const GVecGen3 op[2] = {                                     \
> +        {                                                               \
> +            .fni4 = FNI4_FUNC,                                          \
> +            .fni8 = FNI8_FUNC,                                          \
> +            .vece = MO_32                                               \
> +        },                                                              \
> +        {                                                               \
> +            .fni4 = FNI4_FUNC,                                          \
> +            .fni8 = FNI8_FUNC,                                          \
> +            .vece = MO_64                                               \
> +        },                                                              \
> +    };                                                                  \

There is zero point in having a two element array here:
(1) VECE is a constant
(2) The unused array element is actively wrong.

> +#define DIV_VEC(NAME, SZ, DIV)                                          \
> +static void do_vx_##NAME(TCGv_##SZ t, TCGv_##SZ a, TCGv_##SZ b)         \
> +{                                                                       \
> +    TCGv_##SZ zero = tcg_constant_##SZ(0), one = tcg_constant_##SZ(1);  \
> +    /*                                                                  \
> +     *  If N/0 the instruction used by the backend might deliver        \
> +     *  a signal to the process and the hardware returns 0 when         \
> +     *  N/0, so if b = 0 return 0/1                                     \
> +     */                                                                 \
> +    tcg_gen_movcond_##SZ(TCG_COND_EQ, a, b, zero, zero, a);             \
> +    tcg_gen_movcond_##SZ(TCG_COND_EQ, b, b, zero, one, b);              \
> +    DIV(t, a, b);                                                       \
> +}

The manual says N/0 = undefined.  I don't think it's important to require 0.

The signed versions still need to check for int_min / -1, which will fault on x86. 
Compare vs gen_op_arith_div{w,d}.


r~