RFA: patch for PR22072

RFA: patch for PR22072

[Vladimir Makarov]

[-- Attachment #1: Type: text/plain, Size: 700 bytes --]

The following patch solves one problem reported in PR22072 (additional 
move generation).  The reason for this problem can be found on

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=22072

The old version of the corrected code was actually a typo.  The fix 
should be reliable because it creates additional conflicts (for early 
clobbers as it should be).  In any case I've checked it on a bootstrap 
on x86_64.  SPEC2000 changes are quite minor (the same performance in 
measure error range and quit insignificant code size changes).

Ok to commit?

2009-10-06  Vladimir Makarov  <vmakarov@redhat.com>

    PR middle-end/22072
    * ira-lives.c (check_and_make_def_conflict): Process all operands.




[-- Attachment #2: pr22072.patch --]
[-- Type: text/plain, Size: 629 bytes --]

Index: gcc/ira-lives.c
===================================================================
--- gcc/ira-lives.c	(revision 152403)
+++ gcc/ira-lives.c	(working copy)
@@ -500,7 +500,7 @@
   for (use = 0; use < recog_data.n_operands; use++)
     {
       if (use == def || recog_data.operand_type[use] == OP_OUT)
-	return;
+	continue;
       
       if (recog_op_alt[use][alt].anything_ok)
 	use_cl = ALL_REGS;
@@ -513,7 +513,7 @@
       if ((use_match = recog_op_alt[use][alt].matches) >= 0)
 	{
 	  if (use_match == def)
-	    return;
+	    continue;
 	  
 	  if (recog_op_alt[use_match][alt].anything_ok)
 	    use_cl = ALL_REGS;

Re: RFA: patch for PR22072

[Jeff Law]

On 10/06/09 16:09, Vladimir Makarov wrote:
> The following patch solves one problem reported in PR22072 (additional 
> move generation).  The reason for this problem can be found on
>
> http://gcc.gnu.org/bugzilla/show_bug.cgi?id=22072
>
> The old version of the corrected code was actually a typo.  The fix 
> should be reliable because it creates additional conflicts (for early 
> clobbers as it should be).  In any case I've checked it on a bootstrap 
> on x86_64.  SPEC2000 changes are quite minor (the same performance in 
> measure error range and quit insignificant code size changes).
>
> Ok to commit?
>
> 2009-10-06  Vladimir Makarov <vmakarov@redhat.com>
>
>    PR middle-end/22072
>    * ira-lives.c (check_and_make_def_conflict): Process all operands.
>
>
>
OK
jeff

ps.  I was hoping it'd fix my oddball allocation problem, but no such 
luck :-)

Re: RFA: patch for PR22072

[H.J. Lu]

On Tue, Oct 6, 2009 at 3:09 PM, Vladimir Makarov <vmakarov@redhat.com> wrote:
> The following patch solves one problem reported in PR22072 (additional move
> generation).  The reason for this problem can be found on
>
> http://gcc.gnu.org/bugzilla/show_bug.cgi?id=22072
>
> The old version of the corrected code was actually a typo.  The fix should
> be reliable because it creates additional conflicts (for early clobbers as
> it should be).  In any case I've checked it on a bootstrap on x86_64.
>  SPEC2000 changes are quite minor (the same performance in measure error
> range and quit insignificant code size changes).
>
> Ok to commit?
>
> 2009-10-06  Vladimir Makarov  <vmakarov@redhat.com>
>
>   PR middle-end/22072
>   * ira-lives.c (check_and_make_def_conflict): Process all operands.
>

This patch caused:

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=44031


-- 
H.J.