Fix for data deps classical distance vector

Fix for data deps classical distance vector

[Sebastian Pop]

Hi,

I've just committed to trunk the attached patch.  It should fix the ICE
that you saw when trying to represent non integer distance vectors.
The patch passed bootstrap and test on amd64-linux.

Sebastian

Re: Fix for data deps classical distance vector

[Sebastian Pop]

[-- Attachment #1: Type: text/plain, Size: 274 bytes --]

On 7/29/07, Sebastian Pop <sebpop@gmail.com> wrote:
> Hi,
>
> I've just committed to trunk the attached patch.  It should fix the ICE
> that you saw when trying to represent non integer distance vectors.
> The patch passed bootstrap and test on amd64-linux.
>
> Sebastian
>

[-- Attachment #2: 785_self_dr.diff --]
[-- Type: text/x-diff, Size: 998 bytes --]

	* tree-data-ref.c (add_multivariate_self_dist): Parametric access
	functions cannot be represented as classical distance vectors.

Index: gcc/tree-data-ref.c
===================================================================
--- gcc/tree-data-ref.c	(revision 126917)
+++ gcc/tree-data-ref.c	(working copy)
@@ -2810,10 +2810,14 @@ add_multivariate_self_dist (struct data_
   lambda_vector dist_v;
   int v1, v2, cd;
 
-  /* Polynomials with more than 2 variables are not handled yet.  */
-  if (TREE_CODE (c_0) != INTEGER_CST)
+  /* Polynomials with more than 2 variables are not handled yet.  When
+     the evolution steps are parameters, it is not possible to
+     represent the dependence using classical distance vectors.  */
+  if (TREE_CODE (c_0) != INTEGER_CST
+      || TREE_CODE (CHREC_RIGHT (c_1)) != INTEGER_CST
+      || TREE_CODE (CHREC_RIGHT (c_2)) != INTEGER_CST)
     {
-      DDR_ARE_DEPENDENT (ddr) = chrec_dont_know;
+      DDR_AFFINE_P (ddr) = false;
       return;
     }