Re: [RFC][PATCH] accel/tcg: Use lookup_and_goto_ptr() for linux-user in translator_use_goto_tb()

[Richard Henderson <richard.henderson@linaro.org>]

On 7/30/23 10:56, Helge Deller wrote:
> I'm quite unclear about translator_use_goto_tb() for qemu-user
> emulation....(and in general).
> 
> Based on the function name, the function translator_use_goto_tb() shall
> help to decide if a program should use goto_tb() and exit_tb() to jump
> to the next instruction.
> 
> Currently, if the destination is on the same page, it returns true.
> I wonder, if it shouldn't return false in this case instead, because
> arches have code like this: (taken from target/hppa/translate.c):
>      if (... && translator_use_goto_tb(ctx, f)) {
>          tcg_gen_goto_tb(which);
>          tcg_gen_movi_reg(cpu_iaoq_f, f);
>          tcg_gen_movi_reg(cpu_iaoq_b, b);
>          tcg_gen_exit_tb(ctx->base.tb, which);
>      } else {
>          copy_iaoq_entry(cpu_iaoq_f, f, cpu_iaoq_b);
>          copy_iaoq_entry(cpu_iaoq_b, b, ctx->iaoq_n_var);
>          tcg_gen_lookup_and_goto_ptr();
>      }
> 
> Shouldn't, if the destination is on the same page, the (faster?)
> path with tcg_gen_lookup_and_goto_ptr() be taken instead?

No, because tcg_gen_lookup_and_goto_ptr is not the faster path.
That always involves a lookup, then an indirect branch.

The goto_tb path is linked, so only requires a lookup once, and the branch may be direct 
(depending on the host architecture).


r~