[PATCH] New and semi-improved PHI-OPT

[Andrew Pinski]

On Mar 6, 2005, at 3:11 PM, Kazu Hirata wrote:

> Hi Andrew and Richard,
>
>>> This is port of the PHI-OPT from the tree cleanup branch.  The 
>>> original
>>> patch
>>> was posted here:
>>> http://gcc.gnu.org/ml/gcc-patches/2004-09/msg02959.html> which
>>> describes what the patch does.
>>>
>>> OK? Bootstrapped and tested on powerpc-darwin with no regressions.
>>
>> Do you have some numbers on the compile-time and memory-usage
>> impact vs. performance benefit of this patch?
>
> In particular, it would be interesting to see how many more
> opportunities we catch at tree level and how much work we take away
> from ifcvt at the RTL level.

Also note both the old algorithm and this one is still O(n), the n 
changed
from the number of BBs with one PHIs to the number of BBs that end in a
conditional.  Also one more thing the following code we get better code
than before (this shows up in PR8361 for an example):
_Bool f1(_Bool a, _Bool b)
{
   if (a)
    {
      if (b)
       return 1;
      else
       return 0;
    }
   return 0;
}

Before (PPC):
_f1:
         cmpwi cr7,r3,0
         cmpwi cr6,r4,0
         li r3,1
         beq- cr7,L2
         bnelr+ cr6
L2:
         li r3,0
         blr

After:
_f1:
         cmpwi cr7,r3,0
         li r3,0
         beqlr- cr7
         addic r0,r4,-1
         subfe r3,r0,r4
         blr

Notice how it is smaller code and one less branch so it will be faster.


> +	   for (i = 0;i < PHI_NUM_ARGS (phi); i++)
> +	     {
> +	       if (PHI_ARG_EDGE (phi, i) == e1)
> +	         arg0 = PHI_ARG_DEF_TREE (phi, i);
> +	       else if (PHI_ARG_EDGE (phi, i) == e2)
> +	         arg1 = PHI_ARG_DEF_TREE (phi, i);
> +	     }
>
> This loop can collapse to
>
>   arg0 = PHI_ARG_DEF (phi, e1->dest_idx);
>   arg1 = PHI_ARG_DEF (phi, e2->dest_idx);

Woops, I had forgot about that, well that is because this changed
after I wrote the code and it is not really written up anywhere.