PING^2: resubmitted IRA improvement patches

[Vladimir Makarov]

On 12/06/2010 11:12 AM, Kenneth Zadeck wrote:
> sorry, for taking so long.   i have been temporarily put on another 
> project.
>
> i applied all of your patches to the trunk, and then reread all of the 
> ira files looking for parts that i did not understand.    many of them 
> relate to the patches at hand but some are things that should have 
> been documented in earlier rounds that were just missed.  When these 
> comments are properly addressed, you can check in your patches.
>
Kenny, thanks for pointing this out.  I'll address all your proposals.  
Many of them are for the code which is on the trunk.  That makes them 
even more valuable.
> ira.c:
>
>   setup_pressure_classes:
>
>     you should describe how you find/compute the set of pressure
>     classes here.
>
>   setup_allocno_and_important_classes:
>
>     there are no globals defined above.
>
>   setup_class_translate_array and setup_class_translate
>
>     what do these translators translate?
>
>   reorder_important_classes:
>
>     what makes a class "important"?  is it just that it is non empty?
>     what purpose do empty classes serve? (i.e. why were they just not
>     deleted to start with? and are some more important
>     than others?  is this why you reorder them?
>
>   setup_reg_class_relations
>
>     what "above" relations?  A line or two describing each relation as 
> it is
>     setup would be useful here.
>
>
> ira-color.c
>
>   assign_hard_reg
>
>     what criteria do you use to select a hard register to assign?
>
>   improve_allocation
>
>     what criteria do you use to improve the allocation?
I already added a lot of comments for this function.  They are now on 
the branch.
>
> ira-costs.c
>
>   setup_cost_classes and setup_regno_cost_classes_by_aclass
>
>     what are the costs based on?
>
>   ira-emit
>
>     a few extra comments would be appreciated here to describe what is
>     going on.
>