This is the mail archive of the mailing list for the GCC project.

Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: [new-ra] good register classes for webs

Hi Denis,

On Wed, 3 Mar 2004, Gazovik wrote:

> This is a patch for choosing better register classes for webs. This
> patch based on pre-rload not on regclass.
> (You must be more happy with this ;)

Yay, cool.  I think I begin to like this ;-)

> After that I have a four passes of different web_class_... routines.

Yes, this is similar to how I think it should work, very nice.

Btw. I had thought in that past about reducing the memory overhead of the 
cost collection.  With restructuring the order this might be possible.

For instance your web_class_costs only works on one web each time, and 
web_class_insn_alt only works on all references in one insn (i.e. on just 
a limited set of the web_costs entries).  So maybe it would be possible to 
do the loops in select_regclass() like so: ?

  foreach bb
    foreach insn i in bb
      int index = 0
      foreach web w in i (without duplicates)
        web_class_costs (w, NULL, web_costs[index++])
      web_class_insn_alt (i, insn_alternative, web_costs)

  foreach web w
    web_cost = 0
    web_class_costs (w, insn_alternative, web_cost)

web_class_insn_alt would then need to be changed to lookup in the 
web_costs array by number (per insn) of reference instead of by web id.

This saves memory (web_costs needs only be as long as the most webs used 
in one single insn), but at the expense of possibly recalculating 
web_costs for the same web as often as it is referenced in different 
insns.  I don't know which is better.

Anyway, I think you should check it in after fixing some typos:

>         (web_class_insn_alt): New functin. Select insn alternatives


>         according to costs of register classes.
>         (web_class_spill): New functin. Add spill code for refs with


>         * ra-colorize.c (colorize_one_web): Add code which disable

"code _to_ disable"

>         * ra.c (one_pass): Remeber last max uid of insn and use it


>         * pre-reload.h 

Empty change?

> +   /* Firts pass of web_class. Choose the best reg_class for each web.  */

"First", two spaces between sentences.

> +   /* Second pass of web_clsss. Select a better alternative for each insn

"web_class", two spaces.

> +      depend from reg_class of each web.  */

"depending on"?

> !                   for (j = 0; j < N_REG_CLASSES; ++j)

A pity, that there is no better way than iterating over all register 
classes in the inner loops per references (yes, regclass is doing the 

> !   /* FIXME:  */
> !   if (ra_pass != 1)
> !     return;

What doesn't work yet in the later passes, or was it that we don't expect 
to add any code in them?


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]