[patch] Avoid division by zero in data-refs dependence analysis

[patch] Avoid division by zero in data-refs dependence analysis

[Ira Rosen]

[-- Attachment #1: Type: text/plain, Size: 379 bytes --]


An attempt to divide by zero was made (causing ICE on the attached test
case) for evolution functions with zero step.

Thanks to Revital for the test case.

Bootstrapped and tested on ppc-linux.
O.K. for mainline?

Thanks,
Ira

ChangeLog entry:

      * tree-data-ref.c (chrec_steps_divide_constant_p): Avoid division
      by 0.

(See attached file: patch.txt)


:ADDPATCH SSA:

[-- Attachment #2: patch.txt --]
[-- Type: text/plain, Size: 2538 bytes --]

Index: testsuite/gcc.dg/vect/vect-117.c
===================================================================
*** testsuite/gcc.dg/vect/vect-117.c    (revision 0)
--- testsuite/gcc.dg/vect/vect-117.c    (revision 0)
***************
*** 0 ****
--- 1,40 ----
+ /* { dg-do compile } */
+ /* { dg-require-effective-target vect_int } */
+ 
+ #include <stdarg.h>
+ #include "tree-vect.h"
+
+ #define N 16
+ 
+ struct
+ {
+   unsigned int x;
+   unsigned int y;
+ } pS [100];
+ 
+ void
+ main1 ()
+ {
+   int i, j;
+   unsigned int ub[N] =
+     { 1, 3, 6, 9, 12, 15, 18, 21, 24, 27, 30, 33, 36, 39, 42, 45 };
+   unsigned int uc[N] =
+     { 1, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15 };
+   unsigned int udiffx, udiffy;
+ 
+   for (i = 0; i < N; i++)
+     {
+       pS[i].x = 0;
+       pS[i].y = 0;
+       for (j = 0; j < N; j++)
+       {
+         udiffx = (ub[j] - uc[j]);
+         udiffy = (ub[j] - uc[j]);
+         pS[i].x = udiffx;
+         pS[i].y = udiffy;
+       }
+     }
+ }
+ 
+ /* { dg-final { cleanup-tree-dump "vect" } } */
+
Index: testsuite/ChangeLog
===================================================================
*** testsuite/ChangeLog (revision 123126)
--- testsuite/ChangeLog (working copy)
***************
*** 1,3 ****
--- 1,7 ----
+ 2007-03-21  Ira Rosen  <irar@il.ibm.com>
+
+       * gcc.dg/vect/vect-117.c: New test.
+
  2007-03-21  Richard Sandiford  <richard@codesourcery.com>

        * lib/target-supports.exp (check_effective_target_nonpic): New
Index: tree-data-ref.c
===================================================================
*** tree-data-ref.c     (revision 123126)
--- tree-data-ref.c     (working copy)
*************** chrec_steps_divide_constant_p (tree chre
*** 3467,3473 ****
      case POLYNOMIAL_CHREC:
        if (evolution_function_is_constant_p (CHREC_RIGHT (chrec)))
        {
!         if (tree_fold_divides_p (CHREC_RIGHT (chrec), cst))
            /* Keep RES to true, and iterate on other dimensions.  */
            return chrec_steps_divide_constant_p (CHREC_LEFT (chrec), cst, res);

--- 3467,3474 ----
      case POLYNOMIAL_CHREC:
        if (evolution_function_is_constant_p (CHREC_RIGHT (chrec)))
        {
!         if (!integer_zerop (CHREC_RIGHT (chrec))
!             && tree_fold_divides_p (CHREC_RIGHT (chrec), cst))
            /* Keep RES to true, and iterate on other dimensions.  */
            return chrec_steps_divide_constant_p (CHREC_LEFT (chrec), cst, res);

Re: [patch] Avoid division by zero in data-refs dependence analysis

[Sebastian Pop]

Hello,

On 3/22/07, Ira Rosen <IRAR@il.ibm.com> wrote:
>
> An attempt to divide by zero was made (causing ICE on the attached test
> case) for evolution functions with zero step.
>

I think that this is a problem in tree-scalar-evolution.c as I fail to
see what subscript is having an evolution of 0 in that testcase.

Anyway, I just think that we should never have a chrec with a
step set to zero.  So probably it is good to have a check for that
gcc_assert (!integer_zerop (CHREC_RIGHT (chrec)))
in one of the functions building chains of recurrences.

Let me see how I can fix the problem.  Can you also send the testcase
as a bug report.

Thank you,
Sebastian

:REVIEWMAIL:

Re: [patch] Avoid division by zero in data-refs dependence analysis

[Ira Rosen]

sebpop@gmail.com wrote on 22/03/2007 17:58:27:

> Hello,
>
> On 3/22/07, Ira Rosen <IRAR@il.ibm.com> wrote:
> >
> > An attempt to divide by zero was made (causing ICE on the attached test
> > case) for evolution functions with zero step.
> >
>
> I think that this is a problem in tree-scalar-evolution.c as I fail to
> see what subscript is having an evolution of 0 in that testcase.
>
> Anyway, I just think that we should never have a chrec with a
> step set to zero.  So probably it is good to have a check for that
> gcc_assert (!integer_zerop (CHREC_RIGHT (chrec)))
> in one of the functions building chains of recurrences.
>
> Let me see how I can fix the problem.  Can you also send the testcase
> as a bug report.

I opened a PR (http://gcc.gnu.org/bugzilla/show_bug.cgi?id=31343).

Thanks,
Ira

>
> Thank you,
> Sebastian
>
> :REVIEWMAIL: