[Bug middle-end/49536] New: latent bug in FSF gcc with creation of vector of arrays

[Bug middle-end/49536] New: latent bug in FSF gcc with creation of vector of arrays

[howarth at nitro dot med.uc.edu]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49536

           Summary: latent bug in FSF gcc with creation of vector of
                    arrays
           Product: gcc
           Version: 4.7.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: middle-end
        AssignedTo: unassigned@gcc.gnu.org
        ReportedBy: howarth@nitro.med.uc.edu


Duncan Sands has determined that http://llvm.org/bugs/show_bug.cgi?id=10042
exposes a bug in FSF gcc where vectors of arrays are being created. He proposes
that the following patch to FSF gcc exposes the latent issue...


Index: tree-vect-data-refs.c
===================================================================
--- tree-vect-data-refs.c    (revision 174743)
+++ tree-vect-data-refs.c    (working copy)
@@ -2734,6 +2734,10 @@
       scalar_type = TREE_TYPE (DR_REF (dr));
       STMT_VINFO_VECTYPE (stmt_info) =
                 get_vectype_for_scalar_type (scalar_type);
+      gcc_assert(!STMT_VINFO_VECTYPE (stmt_info) ||
+                 TREE_CODE (STMT_VINFO_VECTYPE (stmt_info)) != VECTOR_TYPE ||
+                 TREE_CODE (TREE_TYPE (STMT_VINFO_VECTYPE (stmt_info)))
+                 != ARRAY_TYPE);
       if (!STMT_VINFO_VECTYPE (stmt_info))
         {
           if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS))

by triggering an assertion failure when the tree-vectorizer tries to form a
vector of arrays.

[Bug middle-end/49536] latent bug in FSF gcc with creation of vector of arrays

[baldrick at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49536

--- Comment #1 from Duncan Sands <baldrick at gcc dot gnu.org> 2011-06-27 07:46:26 UTC ---
Created attachment 24605
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=24605
testcase reduced from source/hbuild.f from xplor-nih 2.27

To reproduce, apply the patch to gcc.  Then compile hbuild_reduced.f from the
attached tar file like this:

gfortran -S hbuild_reduced.f -g -fdefault-integer-8 -ffixed-line-length-120
-fno-second-underscore -O3

[Bug middle-end/49536] latent bug with creation of vector of arrays

[ebotcazou at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49536

Eric Botcazou <ebotcazou at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |ebotcazou at gcc dot
                   |                            |gnu.org
            Summary|latent bug in FSF gcc with  |latent bug with creation of
                   |creation of vector of       |vector of arrays
                   |arrays                      |

--- Comment #2 from Eric Botcazou <ebotcazou at gcc dot gnu.org> 2011-06-27 08:42:51 UTC ---
There is no other GCC as what you call "FSF gcc" as far as we are concerned.

[Bug middle-end/49536] latent bug in FSF gcc with creation of vector of arrays

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49536

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |ASSIGNED
   Last reconfirmed|                            |2011.06.27 09:09:35
         AssignedTo|unassigned at gcc dot       |rguenth at gcc dot gnu.org
                   |gnu.org                     |
            Summary|latent bug with creation of |latent bug in FSF gcc with
                   |vector of arrays            |creation of vector of
                   |                            |arrays
     Ever Confirmed|0                           |1

--- Comment #3 from Richard Guenther <rguenth at gcc dot gnu.org> 2011-06-27 09:09:35 UTC ---
Confirmed.  I also get into the issue that get_vectype_for_scalar_type_and_size
is called with a vector typed scalar_type (from basic-block vectorization).

Your case ends up with character(kind=1)[1:4] which has SImode.

The issue is that we really would only care for the component mode
(and we make sure it is either int or float mode), but not really for
what type it does represent (I can imagine we'll happily create a
vector type of RECORD_TYPE components this way).

Nothing on the GIMPLE level is prepared for non-scalar types here
though, so I'm not sure what issues we run into with this right now.
OTOH vectorization will very likely fail later anyway if we are not
only presented with a memory copy loop.  So, for gimple hygiene
we should reject component types we cannot easily handle.

[Bug middle-end/49536] latent bug with creation of vector of arrays

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49536

--- Comment #4 from Richard Guenther <rguenth at gcc dot gnu.org> 2011-06-27 10:31:33 UTC ---
Author: rguenth
Date: Mon Jun 27 10:31:30 2011
New Revision: 175475

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=175475
Log:
2011-06-27  Richard Guenther  <rguenther@suse.de>

    PR tree-optimization/49536
    * tree-vect-stmts.c (get_vectype_for_scalar_type_and_size):
    For non-scalar inner types use a scalar type according to
    the scalar inner mode.

Modified:
    trunk/gcc/ChangeLog
    trunk/gcc/tree-vect-stmts.c

[Bug middle-end/49536] latent bug with creation of vector of arrays

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49536

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|ASSIGNED                    |RESOLVED
         Resolution|                            |FIXED
   Target Milestone|---                         |4.7.0

--- Comment #5 from Richard Guenther <rguenth at gcc dot gnu.org> 2011-06-27 10:32:12 UTC ---
Fixed.

[Bug middle-end/49536] latent bug with creation of vector of arrays

[howarth at nitro dot med.uc.edu]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49536

--- Comment #6 from Jack Howarth <howarth at nitro dot med.uc.edu> 2011-06-27 13:08:38 UTC ---
Can we get this fix backported to gcc-4_5-branch and gcc-4_6-branch as well?
Dragonegg currently only builds against those FSF gcc releases.

[Bug middle-end/49536] latent bug with creation of vector of arrays

[howarth at nitro dot med.uc.edu]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49536

--- Comment #7 from Jack Howarth <howarth at nitro dot med.uc.edu> 2011-06-27 16:46:17 UTC ---
I can confirm that the adjusted backport of...

Index: gcc/tree-vect-stmts.c
===================================================================
--- gcc/tree-vect-stmts.c    (revision 175535)
+++ gcc/tree-vect-stmts.c    (working copy)
@@ -4660,6 +4660,15 @@ get_vectype_for_scalar_type (tree scalar
       && GET_MODE_BITSIZE (inner_mode) != TYPE_PRECISION (scalar_type))
     return NULL_TREE;

+  /* We shouldn't end up building VECTOR_TYPEs of non-scalar components.
+     When the component mode passes the above test simply use a type
+     corresponding to that mode.  The theory is that any use that
+     would cause problems with this will disable vectorization anyway.  */
+  if (!SCALAR_FLOAT_TYPE_P (scalar_type)
+      && !INTEGRAL_TYPE_P (scalar_type)
+      && !POINTER_TYPE_P (scalar_type))
+    scalar_type = lang_hooks.types.type_for_mode (inner_mode, 1);
+
   /* FORNOW: Only a single vector size per mode (UNITS_PER_SIMD_WORD)
      is expected.  */
   nunits = UNITS_PER_SIMD_WORD (inner_mode) / nbytes;

on current gcc-4_5-branch solves the ICE under dragonegg in the original
testcase from  http://llvm.org/bugs/show_bug.cgi?id=10042.

[Bug middle-end/49536] latent bug with creation of vector of arrays

[baldrick at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49536

--- Comment #8 from Duncan Sands <baldrick at gcc dot gnu.org> 2012-02-10 12:34:21 UTC ---
Author: baldrick
Date: Fri Feb 10 12:34:17 2012
New Revision: 184090

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=184090
Log:
Backport the fix for PR tree-optimization/49536 from mainline.
Patch by Jack Howarth, approved by Richard Guenther.

Modified:
    branches/gcc-4_6-branch/gcc/ChangeLog
    branches/gcc-4_6-branch/gcc/tree-vect-stmts.c