[Bug tree-optimization/18179] New: vectorizer: wrong alignment/step/initial-address computed for struct accesses

[Bug tree-optimization/18179] New: vectorizer: wrong alignment/step/initial-address computed for struct accesses

[dorit at il dot ibm dot com]

The following testcase:

===========================================
struct s{
  int m;
  int n[N][N][N];
};

struct test1{
  struct s a; /* array a.n is unaligned */
  int b;
  int c;
  struct s e[N]; /* array e.n is aligned */
};

int main ()
{
  int i,j;
  struct test1 tmp1;

  for (i = 0; i < N; i++)
    for (j = 3; j < N-3; j++)
      {
        tmp1.e[i].n[1][2][j] = 8;
      }

  /* check results:  */
  for (i = 0; i < N; i++)
    for (j = 3; j < N-3; j++)
    {
      if (tmp1.e[i].n[1][2][j] != 8)
          abort ();
    }

  /* not consecutive */
  for (i = 0; i < N; i++)
    for (j = 3; j < N-3; j++)
      {
        tmp1.e[j].n[1][2][j] = 8;
      }

  /* check results:  */
  for (i = 0; i < N; i++)
    for (j = 3; j < N-3; j++)
    {
      if (tmp1.e[j].n[1][2][j] != 8)
          abort ();
    }

  return 0;
}
===========================================

currently produces wrong results. The problem is that only the indexes of the 
last (rightmost) component are analyzed, and the rest are ignored, so we end up 
computing wrong alignment (for the first loop) and wrong step (for the second 
loop, where we conclude that the access is consecutive where it really isn't). 
Both these problems will be addressed soon by the patch that switches to use  
get_inner_reference (http://gcc.gnu.org/ml/gcc-patches/2004-10/msg01140.html).

-- 
           Summary: vectorizer: wrong alignment/step/initial-address
                    computed for struct accesses
           Product: gcc
           Version: 4.0.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P2
         Component: tree-optimization
        AssignedTo: irar at il dot ibm dot com
        ReportedBy: dorit at il dot ibm dot com
                CC: dorit at il dot ibm dot com,gcc-bugs at gcc dot gnu dot
                    org
 GCC build triplet: powerpc-apple-darwin7.0.0
  GCC host triplet: powerpc-apple-darwin7.0.0
GCC target triplet: powerpc-apple-darwin7.0.0


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=18179

[Bug tree-optimization/18179] vectorizer: wrong alignment/step/initial-address computed for struct accesses

[dorit at il dot ibm dot com]

------- Additional Comments From dorit at il dot ibm dot com  2004-10-27 14:41 -------
Created an attachment (id=7417)
 --> (http://gcc.gnu.org/bugzilla/attachment.cgi?id=7417&action=view)
testcase.

Forgot to mention - occurs when compiling with '-O2 -ftree-vectorize
-maltivec'.

Attached is the complete testcase.

-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=18179

[Bug tree-optimization/18179] vectorizer: wrong alignment/step/initial-address computed for struct accesses

[pinskia at gcc dot gnu dot org]

------- Additional Comments From pinskia at gcc dot gnu dot org  2004-10-27 17:44 -------
Confirmed.

-- 
           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
     Ever Confirmed|                            |1
           Keywords|                            |wrong-code
   Last reconfirmed|0000-00-00 00:00:00         |2004-10-27 17:44:19
               date|                            |


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=18179

[Bug tree-optimization/18179] vectorizer: wrong alignment/step/initial-address computed for struct accesses

[dorit at il dot ibm dot com]

------- Additional Comments From dorit at il dot ibm dot com  2004-11-04 14:44 -------
patch: http://gcc.gnu.org/ml/gcc-patches/2004-11/msg00277.html


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=18179

[Bug tree-optimization/18179] vectorizer: wrong alignment/step/initial-address computed for struct accesses

[cvs-commit at gcc dot gnu dot org]

------- Additional Comments From cvs-commit at gcc dot gnu dot org  2004-12-15 13:02 -------
Subject: Bug 18179

CVSROOT:	/cvs/gcc
Module name:	gcc
Branch: 	autovect-branch
Changes by:	dorit@gcc.gnu.org	2004-12-15 13:02:47

Modified files:
	gcc            : ChangeLog.autovect tree-vectorizer.c 
	                 tree-vectorizer.h 
	gcc/testsuite  : ChangeLog.autovect 
Added files:
	gcc/testsuite/gcc.dg/vect: vect-90.c vect-91.c 

Log message:
	2004-12-14  Ira Rosen  <irar@il.ibm.com>
	
	PR tree-optimization/18179
	* tree-vectorizer.h (struct _stmt_vec_info): Add new fields:
	initial_offset, step, misalignment and base_aligned_p and macros
	to access them.
	* tree-vectorizer.c (vect_analyze_offset_expr): New function.
	(new_stmt_vec_info): Initialize the new fields of stmt_vec_info.
	(vect_create_data_ref_ptr): Fix printings.
	(vect_get_base_and_offset): Call get_inner_reference and
	vect_analyze_offset_expr. Renamed (previous name
	vect_get_base_and_bit_offset).
	(vect_create_addr_base_for_vector_ref): Build address_base by
	combining base and initial_offset fields of vect_stmt_info.
	(vect_update_inits_of_dr): Update offset of data-ref instead of its
	access-fn. Remove argument.
	(vect_update_inits_of_drs): Call vect_update_inits_of_dr with new
	arguments.
	(vect_compute_data_ref_alignment): Check misalignment info in
	vect_stmt_info. Remove argument.
	(vect_compute_data_refs_alignment): Call
	vect_compute_data_ref_alignment with correct argument.
	(vect_analyze_data_ref_access): Check access info in vect_stmt_info.
	(vect_analyze_pointer_ref_access): Update step and initial_offset
	fields of vect_stmt_info.
	(vect_get_memtag_and_dr): Call vect_get_base_and_offset and set the
	fields of stmt_vec_info. Reorganize. Add memtag retrieval. Renamed
	(previous name vect_get_symb_and_dr).
	(vect_analyze_data_refs): Find vectype and pass it to
	vect_get_memtag_and_dr. Remove memtag retrieval.
	(vect_get_first_index): Remove.
	(vect_compute_array_base_alignment): Remove.
	(vect_compute_array_ref_alignment): Remove.

Patches:
http://gcc.gnu.org/cgi-bin/cvsweb.cgi/gcc/gcc/ChangeLog.autovect.diff?cvsroot=gcc&only_with_tag=autovect-branch&r1=1.1.2.9&r2=1.1.2.10
http://gcc.gnu.org/cgi-bin/cvsweb.cgi/gcc/gcc/tree-vectorizer.c.diff?cvsroot=gcc&only_with_tag=autovect-branch&r1=2.25.2.5&r2=2.25.2.6
http://gcc.gnu.org/cgi-bin/cvsweb.cgi/gcc/gcc/tree-vectorizer.h.diff?cvsroot=gcc&only_with_tag=autovect-branch&r1=2.7.2.2&r2=2.7.2.3
http://gcc.gnu.org/cgi-bin/cvsweb.cgi/gcc/gcc/testsuite/ChangeLog.autovect.diff?cvsroot=gcc&only_with_tag=autovect-branch&r1=1.1.2.2&r2=1.1.2.3
http://gcc.gnu.org/cgi-bin/cvsweb.cgi/gcc/gcc/testsuite/gcc.dg/vect/vect-90.c.diff?cvsroot=gcc&only_with_tag=autovect-branch&r1=NONE&r2=1.1.2.1
http://gcc.gnu.org/cgi-bin/cvsweb.cgi/gcc/gcc/testsuite/gcc.dg/vect/vect-91.c.diff?cvsroot=gcc&only_with_tag=autovect-branch&r1=NONE&r2=1.1.2.1



-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=18179

[Bug tree-optimization/18179] vectorizer: wrong alignment/step/initial-address computed for struct accesses

[cvs-commit at gcc dot gnu dot org]

------- Additional Comments From cvs-commit at gcc dot gnu dot org  2004-12-29 13:12 -------
Subject: Bug 18179

CVSROOT:	/cvs/gcc
Module name:	gcc
Changes by:	dorit@gcc.gnu.org	2004-12-29 13:11:59

Modified files:
	gcc            : ChangeLog tree-vectorizer.c 

Log message:
	2004-12-29  Ira Rosen  <irar@il.ibm.com>
	
	PR tree-optimization/18179
	* tree-vectorizer.c (vect_get_symbl_and_dr): Reorganize. Add memtag
	retrieval.
	(vect_analyze_data_refs): Remove memtag retrieval.

Patches:
http://gcc.gnu.org/cgi-bin/cvsweb.cgi/gcc/gcc/ChangeLog.diff?cvsroot=gcc&r1=2.6975&r2=2.6976
http://gcc.gnu.org/cgi-bin/cvsweb.cgi/gcc/gcc/tree-vectorizer.c.diff?cvsroot=gcc&r1=2.50&r2=2.51



-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=18179

[Bug tree-optimization/18179] vectorizer: wrong alignment/step/initial-address computed for struct accesses

[cvs-commit at gcc dot gnu dot org]

------- Additional Comments From cvs-commit at gcc dot gnu dot org  2004-12-29 13:16 -------
Subject: Bug 18179

CVSROOT:	/cvs/gcc
Module name:	gcc
Changes by:	dorit@gcc.gnu.org	2004-12-29 13:16:09

Modified files:
	gcc            : ChangeLog tree-vectorizer.h tree-vectorizer.c 

Log message:
	2004-12-29  Ira Rosen  <irar@il.ibm.com>
	
	PR tree-optimization/18179
	* tree-vectorizer.h (struct _stmt_vec_info): Add new fields:
	initial_offset, step, misalignment and base_aligned_p and macros
	to access them.
	* tree-vectorizer.c (vect_analyze_offset_expr): New function.
	(new_stmt_vec_info): Initialize the new fields of stmt_vec_info.
	
	(vect_create_data_ref_ptr): Fix printings.

Patches:
http://gcc.gnu.org/cgi-bin/cvsweb.cgi/gcc/gcc/ChangeLog.diff?cvsroot=gcc&r1=2.6976&r2=2.6977
http://gcc.gnu.org/cgi-bin/cvsweb.cgi/gcc/gcc/tree-vectorizer.h.diff?cvsroot=gcc&r1=2.7&r2=2.8
http://gcc.gnu.org/cgi-bin/cvsweb.cgi/gcc/gcc/tree-vectorizer.c.diff?cvsroot=gcc&r1=2.51&r2=2.52



-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=18179

[Bug tree-optimization/18179] vectorizer: wrong alignment/step/initial-address computed for struct accesses

[cvs-commit at gcc dot gnu dot org]

------- Additional Comments From cvs-commit at gcc dot gnu dot org  2004-12-29 13:20 -------
Subject: Bug 18179

CVSROOT:	/cvs/gcc
Module name:	gcc
Changes by:	dorit@gcc.gnu.org	2004-12-29 13:20:38

Modified files:
	gcc            : ChangeLog tree-vectorizer.c 

Log message:
	2004-12-29  Ira Rosen  <irar@il.ibm.com>
	
	PR tree-optimization/18179
	* tree-vectorizer.c (vect_get_base_and_offset): Rename (previous name
	vect_get_base_and_bit_offset).
	(vect_get_memtag_and_dr): Rename (previous name vect_get_symbl_and_dr).

Patches:
http://gcc.gnu.org/cgi-bin/cvsweb.cgi/gcc/gcc/ChangeLog.diff?cvsroot=gcc&r1=2.6977&r2=2.6978
http://gcc.gnu.org/cgi-bin/cvsweb.cgi/gcc/gcc/tree-vectorizer.c.diff?cvsroot=gcc&r1=2.52&r2=2.53



-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=18179

[Bug tree-optimization/18179] vectorizer: wrong alignment/step/initial-address computed for struct accesses

[cvs-commit at gcc dot gnu dot org]

------- Additional Comments From cvs-commit at gcc dot gnu dot org  2004-12-29 13:39 -------
Subject: Bug 18179

CVSROOT:	/cvs/gcc
Module name:	gcc
Changes by:	dorit@gcc.gnu.org	2004-12-29 13:38:34

Modified files:
	gcc            : ChangeLog tree-vectorizer.c 

Log message:
	2004-12-29  Ira Rosen  <irar@il.ibm.com>
	
	PR tree-optimization/18179
	* tree-vectorizer.c (vect_get_base_and_offset): Call get_inner_reference
	and vect_analyze_offset_expr.
	(vect_create_addr_base_for_vector_ref): Build address_base by combining
	base and initial_offset fields of vect_stmt_info.
	(vect_update_inits_of_dr): Update offset of data-ref instead of its
	access-fn. Remove argument.
	(vect_update_inits_of_drs): Call vect_update_inits_of_dr with new
	arguments.
	(vect_compute_data_ref_alignment): Check misalignment info in
	vect_stmt_info. Remove argument.
	(vect_compute_data_refs_alignment): Call vect_compute_data_ref_alignment
	with correct argument.
	(vect_analyze_data_ref_access): Check access info in vect_stmt_info.
	(vect_analyze_pointer_ref_access): Update step and initial_offset fields
	of vect_stmt_info.
	(vect_get_memtag_and_dr): Call vect_get_base_and_offset and set the
	fields of stmt_vec_info.
	(vect_analyze_data_refs): Find vectype and pass it to
	vect_get_memtag_and_dr.
	(vect_get_first_index): Remove.
	(vect_compute_array_base_alignment): Remove.
	(vect_compute_array_ref_alignment): Remove
	(vect_create_data_ref_ptr): Use TYPE_SIZE_UNIT instead GET_MODE_SIZE.
	(vect_gen_niters_for_prolog_loop): Use TREE_CONSTANT instead
	host_integerp.
	(vectorizable_load): Use size arithmetics.

Patches:
http://gcc.gnu.org/cgi-bin/cvsweb.cgi/gcc/gcc/ChangeLog.diff?cvsroot=gcc&r1=2.6978&r2=2.6979
http://gcc.gnu.org/cgi-bin/cvsweb.cgi/gcc/gcc/tree-vectorizer.c.diff?cvsroot=gcc&r1=2.53&r2=2.54



-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=18179

[Bug tree-optimization/18179] vectorizer: wrong alignment/step/initial-address computed for struct accesses

[aj at gcc dot gnu dot org]

------- Additional Comments From aj at gcc dot gnu dot org  2005-01-01 09:49 -------
I see the following new failures on Linux/x86-64 since Ira's patch went in: 
 
  gcc.dg/tree-ssa/ifc-20040816-1.c (test for excess errors) 
   gcc.dg/vect/pr18400.c (test for excess errors) 
   gcc.dg/vect/pr18400.c scan-tree-dump-times vectorized 1 loops 1 
   gcc.dg/vect/vect-1.c (test for excess errors) 
   gcc.dg/vect/vect-1.c scan-tree-dump-times vectorized 3 loops 1 
   gcc.dg/vect/vect-10.c (test for excess errors) 
   gcc.dg/vect/vect-10.c scan-tree-dump-times vectorized 0 loops 1 
   gcc.dg/vect/vect-11.c (test for excess errors) 
   gcc.dg/vect/vect-12.c (test for excess errors) 
   gcc.dg/vect/vect-13.c (test for excess errors) 
   gcc.dg/vect/vect-14.c (test for excess errors) 
   gcc.dg/vect/vect-15.c (test for excess errors) 
   gcc.dg/vect/vect-16.c (test for excess errors) 
   gcc.dg/vect/vect-17.c (test for excess errors) 
   gcc.dg/vect/vect-18.c (test for excess errors) 
   gcc.dg/vect/vect-19.c (test for excess errors) 
   gcc.dg/vect/vect-2.c (test for excess errors) 
   gcc.dg/vect/vect-2.c scan-tree-dump-times vectorized 1 loops 1 
   gcc.dg/vect/vect-20.c (test for excess errors) 
   gcc.dg/vect/vect-21.c (test for excess errors) 
   gcc.dg/vect/vect-22.c (test for excess errors) 
   gcc.dg/vect/vect-25.c (test for excess errors) 
[...] 
 
The log files shows the following: 
Executing on host: /builds/gcc/misc/gcc/xgcc 
-B/builds/gcc/misc/gcc/ /cvs/gcc/gcc/testsuite/gcc.dg/vect/pr18400.c   -O2 
-ftree-vec 
torize -fdump-tree-vect-stats -msse2  -lm   -o ./pr18400.exe    (timeout = 300) 
/cvs/gcc/gcc/testsuite/gcc.dg/vect/pr18400.c: In function 'main1': 
/cvs/gcc/gcc/testsuite/gcc.dg/vect/pr18400.c:9: internal compiler error: tree 
check: accessed operand 2 of nop_expr with 1 operands in 
vect_analyze_offset_expr, at tree-vectorizer.c:1452 
Please submit a full bug report, 
with preprocessed source if appropriate. 
See <URL:http://gcc.gnu.org/bugs.html> for instructions. 
compiler exited with status 1 
output is: 
/cvs/gcc/gcc/testsuite/gcc.dg/vect/pr18400.c: In function 'main1': 
/cvs/gcc/gcc/testsuite/gcc.dg/vect/pr18400.c:9: internal compiler error: tree 
check: accessed operand 2 of nop_expr with 1 operands in 
vect_analyze_offset_expr, at tree-vectorizer.c:1452 
Please submit a full bug report, 
with preprocessed source if appropriate. 
See <URL:http://gcc.gnu.org/bugs.html> for instructions. 
 
The ICE message is the same in all cases. 

-- 
           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |aj at gcc dot gnu dot org


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=18179

[Bug tree-optimization/18179] vectorizer: wrong alignment/step/initial-address computed for struct accesses

[pinskia at gcc dot gnu dot org]

------- Additional Comments From pinskia at gcc dot gnu dot org  2005-01-06 00:43 -------
Fixed.

-- 
           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|NEW                         |RESOLVED
         Resolution|                            |FIXED
   Target Milestone|---                         |4.0.0


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=18179