Re: [PATCH 07/20] target/loongarch: Add fixed point arithmetic instruction translation

[Richard Henderson]

On 7/2/21 1:51 AM, Philippe Mathieu-Daudé wrote:
> static bool trans_mul_d(DisasContext *ctx, int rd, int rj, int rk)
> {
>      TCGv t0, t1;
>
>      check_loongarch_64(ctx);
>
>      if (a->rd == 0) {
>          /* Treat as NOP. */
>          return true;
>      }
>
>      t0 = tcg_temp_new();
>      t1 = tcg_temp_new();
>
>      gen_load_gpr(t0, a->rj);
>      gen_load_gpr(t1, a->rk);

Another improvement over mips is possible, while we're at it:

TCGv get_gpr(int regno)
{
    if (regno == 0) {
        return tcg_constant_tl(0);
    } else {
        return cpu_gpr[regno];
    }
}

    t0 = get_gpr(a->rj);
    t1 = get_gpr(a->rk);

>      tcg_gen_mul_i64(cpu_gpr[a->rd], t0, t1);
>
>      tcg_temp_free(t0);
>      tcg_temp_free(t1);

and now the frees are not necessary.

You do have to be careful that you consume the input before you write back to 
cpu_gpr[a->rd].  Previously you had a copy, but now t0 and t1 reference the live register. 
 And of course you cannot write to either t0 or t1, because one of them might be the 
constant 0.


r~