[Bug tree-optimization/56933] [4.9 Regression] Vectorizer missing read-write dependency for interleaved accesses

[Bug tree-optimization/56933] [4.9 Regression] Vectorizer missing read-write dependency for interleaved accesses

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=56933

Richard Biener <rguenth at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |ASSIGNED
   Last reconfirmed|                            |2013-04-15
         AssignedTo|unassigned at gcc dot       |rguenth at gcc dot gnu.org
                   |gnu.org                     |
   Target Milestone|---                         |4.9.0
     Ever Confirmed|0                           |1

--- Comment #1 from Richard Biener <rguenth at gcc dot gnu.org> 2013-04-15 10:29:19 UTC ---
Mine.

[Bug tree-optimization/56933] [4.9 Regression] Vectorizer missing read-write dependency for interleaved accesses

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=56933

--- Comment #2 from Richard Biener <rguenth at gcc dot gnu.org> 2013-04-15 11:18:43 UTC ---
/* { dg-do run } */

extern void abort (void);
void __attribute__((noinline,noclone))
foo (double *b, double *d, double *f)
{
  int i;
  for (i = 0; i < 1024; i++)
    {
      d[2*i] = 2. * d[2*i];
      d[2*i+1] = 4. * d[2*i+1];
      b[i] = d[2*i] - 1.;
      f[i] = d[2*i+1] + 2.;
    }
}
int main()
{
  double b[1024], d[2*1024], f[1024];
  int i;
  for (i = 0; i < 2*1024; i++)
    d[i] = 1.;
  foo (b, d, f);
  for (i = 0; i < 1024; i+= 2)
    {
      if (d[2*i] != 2.)
        abort ();
      if (d[2*i+1] != 4.)
        abort ();
    }
  for (i = 0; i < 1024; i++)
    {
      if (b[i] != 1.)
        abort ();
      if (f[i] != 6.)
        abort ();
    }
  return 0;
}

/* { dg-final { cleanup-tree-dump "vect" } } */

[Bug tree-optimization/56933] [4.9 Regression] Vectorizer missing read-write dependency for interleaved accesses

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=56933

Richard Biener <rguenth at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|ASSIGNED                    |RESOLVED
         Resolution|                            |FIXED

--- Comment #3 from Richard Biener <rguenth at gcc dot gnu.org> 2013-04-15 14:09:14 UTC ---
Author: rguenth
Date: Mon Apr 15 14:08:41 2013
New Revision: 197972

URL: http://gcc.gnu.org/viewcvs?rev=197972&root=gcc&view=rev
Log:
2013-04-15  Richard Biener  <rguenther@suse.de>

    PR tree-optimization/56933
    * tree-vectorizer.h (struct _stmt_vec_info): Remove read_write_dep
    member.
    (GROUP_READ_WRITE_DEPENDENCE): Remove.
    (STMT_VINFO_GROUP_READ_WRITE_DEPENDENCE): Likewise.
    * tree-vect-data-refs.c (vect_analyze_group_access): Move
    dependence check ...
    vect_analyze_data_ref_dependence (vect_analyze_data_ref_dependence):
    ... here.
    * tree-vect-stmts.c (new_stmt_vec_info): Do not initialize
    GROUP_READ_WRITE_DEPENDENCE.

    * gcc.dg/vect/pr56933.c: New testcase.

Added:
    trunk/gcc/testsuite/gcc.dg/vect/pr56933.c
Modified:
    trunk/gcc/ChangeLog
    trunk/gcc/testsuite/ChangeLog
    trunk/gcc/tree-vect-data-refs.c
    trunk/gcc/tree-vect-stmts.c
    trunk/gcc/tree-vectorizer.h

[Bug tree-optimization/56933] [4.9 Regression] Vectorizer missing read-write dependency for interleaved accesses

[bernd.edlinger at hotmail dot de]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=56933

Bernd Edlinger <bernd.edlinger at hotmail dot de> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |bernd.edlinger at hotmail dot de

--- Comment #4 from Bernd Edlinger <bernd.edlinger at hotmail dot de> ---
The test case gcc.dg/vect/pr56933.c fails on a pentium II,
because of invalid instruction.

A fairly obvious fix (which works for me) would be:

--- gcc/testsuite/gcc.dg/vect/pr56933.c.jj      2013-04-15 16:08:41.000000000
+0200
+++ gcc/testsuite/gcc.dg/vect/pr56933.c 2013-08-22 21:12:19.000000000 +0200
@@ -1,4 +1,5 @@
 /* { dg-do run } */
+/* { dg-require-effective-target sse2_runtime } */

 extern void abort (void);
 void __attribute__((noinline,noclone))

[Bug tree-optimization/56933] [4.9 Regression] Vectorizer missing read-write dependency for interleaved accesses

[jakub at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=56933

Jakub Jelinek <jakub at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |jakub at gcc dot gnu.org

--- Comment #5 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
(In reply to Bernd Edlinger from comment #4)
> The test case gcc.dg/vect/pr56933.c fails on a pentium II,
> because of invalid instruction.
> 
> A fairly obvious fix (which works for me) would be:

There is nothing obvious on it (maybe obviously broken).
The vect.exp tests aren't i?86/x86_64 only, so you definitely can't add
unguarded requirements for specific CPUs, it would cause the test not to be run
on ppcs/arm/etc.

The bug in this test is that it specifies { dg-do run }, either it shouldn't
(then the default will be run if CPU supports the -msse2 flags in this case, or
compile otherwise), or it should use the tree-vect.h + check_vect () framework
like many other tests.

> --- gcc/testsuite/gcc.dg/vect/pr56933.c.jj      2013-04-15
> 16:08:41.000000000 +0200
> +++ gcc/testsuite/gcc.dg/vect/pr56933.c 2013-08-22 21:12:19.000000000 +0200
> @@ -1,4 +1,5 @@
>  /* { dg-do run } */
> +/* { dg-require-effective-target sse2_runtime } */
> 
>  extern void abort (void);
>  void __attribute__((noinline,noclone))

[Bug tree-optimization/56933] [4.9 Regression] Vectorizer missing read-write dependency for interleaved accesses

[bernd.edlinger at hotmail dot de]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=56933

--- Comment #7 from Bernd Edlinger <bernd.edlinger at hotmail dot de> ---
Created attachment 30712
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=30712&action=edit
fixed test case

Looking deeper into the matter it seems like this an example
where vectorisation is not supposed to happen.
But this is still vectorized and causes the crash here:

   for (i = 0; i < 2*1024; i++)
     d[i] = 1.;


So this is my proposed fix for the test case:
- added check_vect() to prevent invalid instruction as Jakub suggested
- addeded scan-tree-dump to detect the READ_WRITE depencency
  to make sure that the vectorization in foo() did not happen
  for the right reason.

[Bug tree-optimization/56933] [4.9 Regression] Vectorizer missing read-write dependency for interleaved accesses

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=56933

--- Comment #8 from Richard Biener <rguenth at gcc dot gnu.org> ---
(In reply to Bernd Edlinger from comment #7)
> Created attachment 30712 [details]
> fixed test case
> 
> Looking deeper into the matter it seems like this an example
> where vectorisation is not supposed to happen.
> But this is still vectorized and causes the crash here:
> 
>    for (i = 0; i < 2*1024; i++)
>      d[i] = 1.;

That loop is initialization only and can be vectorized.  I have committed
a fix (hopefully), so you may want to check again.

> 
> So this is my proposed fix for the test case:
> - added check_vect() to prevent invalid instruction as Jakub suggested
> - addeded scan-tree-dump to detect the READ_WRITE depencency
>   to make sure that the vectorization in foo() did not happen
>   for the right reason.