RE: Patch to handle perfect loop unrolling in tree_transform_and_unroll_loop

["Bingfeng Mei" <bmei at broadcom dot com>]

BTW: the following lines in original tree_transform_and_unroll_loop (I didn't change) are
problematic. Some complex loops won't meet requirements of single_pred(loop->latch) (e.g., 20060905-1.c)
should this function is called to unroll them. 
 
      exit_bb = single_pred (loop->latch);
      new_exit = find_edge (exit_bb, rest);
      new_exit->count = loop_preheader_edge (loop)->count;
      new_exit->probability = REG_BR_PROB_BASE / (new_est_niter + 1);

      rest->count += new_exit->count;
      rest->frequency += EDGE_FREQUENCY (new_exit);

      new_nonexit = single_pred_edge (loop->latch);
      prob = new_nonexit->probability;
      new_nonexit->probability = REG_BR_PROB_BASE - new_exit->probability;
      new_nonexit->count = exit_bb->count - new_exit->count;



> -----Original Message-----
> From: fearyourself@gmail.com [mailto:fearyourself@gmail.com] 
> On Behalf Of Jean Christophe Beyler
> Sent: 28 October 2009 15:24
> To: Bingfeng Mei
> Cc: gcc-patches@gcc.gnu.org; Zdenek Dvorak
> Subject: Re: Patch to handle perfect loop unrolling in 
> tree_transform_and_unroll_loop
> 
> I am only looking at your patch so I haven't tried your code, so it is
> very possible some of these comments are simply stupid:
> 
> - I wonder the usefulness of :
> +  if(perfect_unrolling)
> +    {
> +      new_exit = exit;
> +      rest = exit->dest;
> +      new_est_niter = est_niter / factor;
> +    }
> 
> as far as I can tell, rest is only used in the non perfect case
> 
> - Not important at all:
> +      gcc_assert(EDGE_COUNT(exit_bb->succs) == 2);
> doesn't follow the normal indentation for the '('
> 
> - More to the point, I did something similar except I added a test to
> see if the factor requested by the user was above the limit of
> n_iters. If so, I say that the unrolling is perfect and issue a
> warning.
> 
> Do you want me to look a bit more?
> Jc
> 
> On Wed, Oct 28, 2009 at 10:43 AM, Bingfeng Mei 
> <bmei@broadcom.com> wrote:
> > Hello,
> > This patch is to handle perfect loop unrolling as
> > discussed in http://gcc.gnu.org/ml/gcc/2009-10/msg00333.html.
> >
> > I merged the original code into tree_transform_and_unroll_loop,
> > and added some code to get probability/frequency right.
> >
> > This is related to http://gcc.gnu.org/bugzilla/show_bug.cgi?id=36712
> > A tree-level unrolling pass will just call this modified function.
> > I will submit a separate patch if needed.
> >
> > 2009-10-28 ?Bingfeng Mei <bmei@broadcom.com>
> >
> > ? ? ? ?* tree_transform_and_unroll_loop 
> (tree_transform_and_unroll_loop):
> > ? ? ? ?To handle perfect loop unrolling. 
> (perfect_loop_unrolling_p): New function
> > ? ? ? ?to check whether a loop can perfect unrolling.
> >
> > Cheers,
> > Bingfeng Mei
> >
> >
> 
>