Re: [PATCH, sched] Cleanup and improve multipass_dfa_lookahead_guard

[Maxim Kuvyrkov]

On Jun 2, 2014, at 7:54 PM, Andreas Schwab <sch...@linux-m68k.org> wrote:

> Regtested on ia64-suse-linux and installed as obvious.
> 
> Andreas.
> 
>       * config/ia64/ia64.c
>       (ia64_first_cycle_multipass_dfa_lookahead_guard): Check
>       pending_data_specs first.
> 
> diff --git a/gcc/config/ia64/ia64.c b/gcc/config/ia64/ia64.c
> index 118e5bf..4c5390b 100644
> --- a/gcc/config/ia64/ia64.c
> +++ b/gcc/config/ia64/ia64.c
> @@ -7536,7 +7536,7 @@ ia64_first_cycle_multipass_dfa_lookahead_guard (rtx 
> insn, int ready_index)
> 
>   /* Size of ALAT is 32.  As far as we perform conservative
>      data speculation, we keep ALAT half-empty.  */
> -  if ((TODO_SPEC (insn) & BEGIN_DATA) && pending_data_specs >= 16)
> +  if (pending_data_specs >= 16 && (TODO_SPEC (insn) & BEGIN_DATA))
>     return ready_index == 0 ? -1 : 1;
> 
>   if (ready_index == 0)

Thanks for fixing this.  To make it a bit more robust I would suggest using 
same condition as in ia64_variable_issue():
===
  if (sched_deps_info->generate_spec_deps && !sel_sched_p ())
    /* Modulo scheduling does not extend h_i_d when emitting
       new instructions.  Don't use h_i_d, if we don't have to.  */
    {
      if (DONE_SPEC (insn) & BEGIN_DATA)
        pending_data_specs++;
      if (CHECK_SPEC (insn) & BEGIN_DATA)
        pending_data_specs--;
    }
===

The underlaying problem is that TODO_SPEC(insn) should not be accessed for 
sel-sched.  Because pending_data_specs turns out to always be null during 
sel-sched, "pending_data_specs >= 16" can be used as a shorthand for
 "sched_deps_info->generate_spec_deps && !sel_sched_p () && pending_data_specs 
>= 16"

--
Maxim Kuvyrkov
www.linaro.org