Patchwork [take,2] Fix PR tree-optimization/49960 ,Fix self data dependence

login
register
mail settings
Submitter Razya Ladelsky
Date Nov. 24, 2011, 12:32 p.m.
Message ID <OF40CD4C0B.22310E4B-ONC2257952.0044D957-C2257952.0044EEF9@il.ibm.com>
Download mbox | patch
Permalink /patch/127493/
State New
Headers show

Comments

Razya Ladelsky - Nov. 24, 2011, 12:32 p.m.
Jakub Jelinek <jakub@redhat.com> wrote on 21/11/2011 07:25:10 PM:

> From: Jakub Jelinek <jakub@redhat.com>
> To: Razya Ladelsky/Haifa/IBM@IBMIL
> Cc: GCC Patches <gcc-patches@gcc.gnu.org>, Richard Guenther 
> <richard.guenther@gmail.com>
> Date: 21/11/2011 07:25 PM
> Subject: Re: [PATCH, take 2] Fix PR tree-optimization/49960 ,Fix 
> self data dependence
> 
> On Mon, Nov 21, 2011 at 06:56:55PM +0200, Razya Ladelsky wrote:
> > I have some non-affine cases for which compute_affine_dependence is 
called 
> > (as it is called for 
> > ALL dependences from compte_all_depepndences()), and no harm is done.
> > I looked a little bit closer into the code, and this is what happens 
for 
> > non affine accesses:
> > 
> > initialize_data_dependence_relation() assigns 
> > DDR_ARE_DEPENDENT (res) = chrec_dont_know for the dr.
> 
> It can be chrec_known too (that's actually the only interesting case for 
us
> unless it is a read-read ddr), but you're right that likely
> object_address_invariant_in_loop_p should be false.
> 
> > Then, compute_affine_depepndence()
> > tests if (DDR_ARE_DEPENDENT (ddr) == NULL_TREE), and does nothing 
> > otherwise.
> > Since the dr was initialized with chrec_dont_know, no harm is done.
> > 
> > Anyway, since it is useless for your gather case, I'll just remove it, 

> > along with compute_self_dependence().
> > 
> > OK?
> 
> Yes, patch preapproved.
> 
>    Jakub
> 

Hi,

Going to commit this patch today.
I bootstrapped and regtested on  ppc64-suse-linux.

ChangeLog:
 
        * tree-data-ref.c (initialize_data_dependence_relation): Update 
comment for the 
        self dependence case.
        (compute_self_dependence): Remove.
        * tree-vect-data-refs.c (vect_analyze_data_refs): Remove call to 
        compute_self_dependenc. 


Thanks,
Razya
=

Patch

Index: tree-data-ref.c
===================================================================
--- tree-data-ref.c	(revision 181675)
+++ tree-data-ref.c	(working copy)
@@ -1385,8 +1385,7 @@  initialize_data_dependence_relation (struct data_r
       return res;
     }
 
-  /* When the references are exactly the same, don't spend time doing
-     the data dependence tests, just initialize the ddr and return.  */
+  /* The case where the references are exactly the same.  */
   if (operand_equal_p (DR_REF (a), DR_REF (b), 0))
     {
      if (loop_nest
@@ -4132,19 +4131,6 @@  compute_affine_dependence (struct data_dependence_
     fprintf (dump_file, ")\n");
 }
 
-/* This computes the dependence relation for the same data
-   reference into DDR.  */
-
-void
-compute_self_dependence (struct data_dependence_relation *ddr)
-{
-  if (DDR_ARE_DEPENDENT (ddr) != NULL_TREE)
-    return;
-
-  if (DDR_LOOP_NEST (ddr))
-    compute_affine_dependence (ddr, VEC_index (loop_p, DDR_LOOP_NEST (ddr), 0));
-}
-
 /* Compute in DEPENDENCE_RELATIONS the data dependence graph for all
    the data references in DATAREFS, in the LOOP_NEST.  When
    COMPUTE_SELF_AND_RR is FALSE, don't compute read-read and self
Index: tree-vect-data-refs.c
===================================================================
--- tree-vect-data-refs.c	(revision 181675)
+++ tree-vect-data-refs.c	(working copy)
@@ -3122,7 +3122,6 @@  vect_analyze_data_refs (loop_vec_info loop_vinfo,
 	  ddr = VEC_index (ddr_p, ddrs, k);
 	  gcc_assert (DDR_A (ddr) == olddr && DDR_B (ddr) == olddr);
 	  newddr = initialize_data_dependence_relation (dr, dr, nest);
-	  compute_self_dependence (newddr);
 	  VEC_replace (ddr_p, ddrs, k, newddr);
 	  free_dependence_relation (ddr);
 	  VEC_replace (data_reference_p, datarefs, i, dr);