Re: [RFC v2 58/76] target/riscv: rvv-0.9: slide instructions

[Richard Henderson]

On 7/22/20 2:16 AM, frank.ch...@sifive.com wrote:
> -#define GEN_VEXT_VSLIDEDOWN_VX(NAME, ETYPE, H, CLEAR_FN)                  \
> +#define GEN_VEXT_VSLIDEDOWN_VX(NAME, ETYPE, H)                            \
>  void HELPER(NAME)(void *vd, void *v0, target_ulong s1, void *vs2,         \
>                    CPURISCVState *env, uint32_t desc)                      \
>  {                                                                         \
> -    uint32_t vlmax = env_archcpu(env)->cfg.vlen;                          \
> +    uint32_t vlmax = vext_max_elems(desc, sizeof(ETYPE), false);          \
>      uint32_t vm = vext_vm(desc);                                          \
> -    uint32_t vta = vext_vta(desc);                                        \
>      uint32_t vl = env->vl;                                                \
>      target_ulong offset = s1, i;                                          \
>                                                                            \
>      for (i = 0; i < vl; ++i) {                                            \
> +        /* offset may be a large value, which j may overflow */           \
>          target_ulong j = i + offset;                                      \
> +        bool is_valid = (offset >= vlmax || j >= vlmax) ? false : true;   \

This is... silly verbose.
But also, the test is partially loop invariant and entirely predictable,
allowing loop fission.

>          if (!vm && !vext_elem_mask(v0, i)) {                              \
>              continue;                                                     \
>          }                                                                 \
> -        *((ETYPE *)vd + H(i)) = j >= vlmax ? 0 : *((ETYPE *)vs2 + H(j));  \
> +        *((ETYPE *)vd + H(i)) = is_valid ? *((ETYPE *)vs2 + H(j)) : 0;  \
>      }                                                                     \
> -    CLEAR_FN(vd, vta, vl, vl * sizeof(ETYPE), vlmax * sizeof(ETYPE));     \
>  }

E.g.

    i_max = s1 < vlmax ? vlmax - s1 : 0;
    for (i = 0; i < i_max; ++i) {
        if (vext_elem_mask(v0, i)) {
            *((ETYPE *)vd + H(i)) = *((ETYPE *)vs2 + H(i + s1));
        }
    }
    for (i = i_max; i < vl; ++i) {
        if (vext_elem_mask(v0, i)) {
            *((ETYPE *)vd + H(i)) = 0;
        }
    }


r~