Re: [WPA PATCH 3/5] Benefit analysis.

[Jan Hubicka <hubicka at ucw dot cz>]

Hi,

> +/* Whether the obstack below has been initialized.  */
> +static bool bbbs_obstack_initialized;

Testing for NULL would probably works too.
> +
> +  if (dir)
> +    expr = fold_build2 (gimple_cond_code (stmt), boolean_type_node,
> +			invar, expr);
> +  else
> +    expr = fold_build2 (gimple_cond_code (stmt), boolean_type_node,
> +			expr, invar);
> +
> +  if (!expr || TREE_CODE_CLASS (TREE_CODE (expr)) != tcc_comparison)
> +    return;

Should not be there some logic testing if same conditonal appears
multiple times in thefunction?

Otherwise it seems fine.  DO you have how much it works in practice?
I would like to merge this with the inliner heruistics.  That is make it
use inliner size/time estimates (that do ignore some specific
instructions) and combine it with the param benefits.

I guess all I need to do on this route is to add new size/time vectors
computed by inliner and use the bitmaps after propagation arguments same
way as we do for indirect inlining, right?

Honza