[PATCH] PR85964

[Jan Hubicka]

> 
> This makes tracer not explode with -fno-guess-branch-probabilities.
> I've settled with find_best_successor/predecessor not returning
> anything if _any_ edge in the interesting direction doesn't have
> ->count () initialized (rather than ignoring such edges).
> 
> Honza - I suppose it is on purpose that functions like
> .to_frequency () do not ICE for uninitialized counters?
> It at least looks like "previous" behavior was more sane
> for tracer in the counts/frequencies that were exposed.
> 
> Bootstrapped and tested on x86_64-unknown-linux-gnu.
> 
> Honza, does this look OK to you?
> 
> tracer going wild on this testcase exposes the CFG cleanup
> scalability issue I've posted the following RFC for:
> https://gcc.gnu.org/ml/gcc-patches/2018-05/msg01644.html
> 
> Thanks,
> Richard.
> 
> 2018-05-30  Richard Biener  <rguenther@suse.de>
> 
> 	PR tree-optimization/85964
> 	* tracer.c (better_p): Drop initialized count check, we only
> 	call the function with initialized counts now.
> 	(find_best_successor): Do find a best edge if one
> 	has uninitialized count.
> 	(find_best_predecessor): Likewise.  Do BB frequency check only
> 	if count is initialized.

> 
> Index: gcc/tracer.c
> ===================================================================
> --- gcc/tracer.c	(revision 260896)
> +++ gcc/tracer.c	(working copy)
> @@ -132,8 +132,7 @@ count_insns (basic_block bb)
>  static bool
>  better_p (const_edge e1, const_edge e2)
>  {
> -  if (e1->count ().initialized_p () && e2->count ().initialized_p ()
> -      && ((e1->count () > e2->count ()) || (e1->count () < e2->count  ())))
> +  if ((e1->count () > e2->count ()) || (e1->count () < e2->count ()))
>      return e1->count () > e2->count ();
>    /* This is needed to avoid changes in the decision after
>       CFG is modified.  */
> @@ -152,12 +151,15 @@ find_best_successor (basic_block bb)
>    edge_iterator ei;
>  
>    FOR_EACH_EDGE (e, ei, bb->succs)
> -    if (!best || better_p (e, best))
> -      best = e;
> +    {
> +      if (!e->count ().initialized_p ())
> +	return NULL;
> +      if (!best || better_p (e, best))
> +	best = e;
> +    }
>    if (!best || ignore_bb_p (best->dest))
>      return NULL;
> -  if (best->probability.initialized_p ()
> -      && best->probability.to_reg_br_prob_base () <= probability_cutoff)
> +  if (best->probability.to_reg_br_prob_base () <= probability_cutoff)

Technically we could accept when one edge has large known probability and other unknown,
but in practice it won't matter because w/o profile guessing tracer is useless anyway.

So OK.
Honza

>      return NULL;
>    return best;
>  }
> @@ -172,12 +174,17 @@ find_best_predecessor (basic_block bb)
>    edge_iterator ei;
>  
>    FOR_EACH_EDGE (e, ei, bb->preds)
> -    if (!best || better_p (e, best))
> -      best = e;
> +    {
> +      if (!e->count ().initialized_p ())
> +	return NULL;
> +      if (!best || better_p (e, best))
> +	best = e;
> +    }
>    if (!best || ignore_bb_p (best->src))
>      return NULL;
> -  if (EDGE_FREQUENCY (best) * REG_BR_PROB_BASE
> -      < bb->count.to_frequency (cfun) * branch_ratio_cutoff)
> +  if (bb->count.initialized_p ()
> +      && (best->count ().to_frequency (cfun) * REG_BR_PROB_BASE
> +	  < bb->count.to_frequency (cfun) * branch_ratio_cutoff))
>      return NULL;
>    return best;
>  }