[Bug tree-optimization/50133] New: [4.7 Regression] ICE: SIGSEGV in vect_finish_stmt_generation (gimple.h:4821) with -ftree-vectorize -fno-tree-loop-im

[Bug tree-optimization/50133] New: [4.7 Regression] ICE: SIGSEGV in vect_finish_stmt_generation (gimple.h:4821) with -ftree-vectorize -fno-tree-loop-im

[zsojka at seznam dot cz]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=50133

             Bug #: 50133
           Summary: [4.7 Regression] ICE: SIGSEGV in
                    vect_finish_stmt_generation (gimple.h:4821) with
                    -ftree-vectorize -fno-tree-loop-im
    Classification: Unclassified
           Product: gcc
           Version: 4.7.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: tree-optimization
        AssignedTo: unassigned@gcc.gnu.org
        ReportedBy: zsojka@seznam.cz
              Host: x86_64-pc-linux-gnu
            Target: x86_64-pc-linux-gnu


Created attachment 25057
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=25057
testcase (gcc.dg/pr46184.c)

Compiler output:
$ gcc -O -ftree-vectorize -fno-tree-loop-im testcase.c
==6443== Invalid read of size 8
==6443==    at 0xB053E1: vect_finish_stmt_generation(gimple_statement_d*,
gimple_statement_d*, gimple_stmt_iterator*) (gimple.h:4821)
==6443==    by 0xB05C8E: vect_init_vector(gimple_statement_d*, tree_node*,
tree_node*, gimple_stmt_iterator*) (tree-vect-stmts.c:1082)
==6443==    by 0xB07625: vectorizable_load(gimple_statement_d*,
gimple_stmt_iterator*, gimple_statement_d**, _slp_tree*, _slp_instance*)
(tree-vect-stmts.c:4626)
==6443==    by 0xB1268C: vect_transform_stmt(gimple_statement_d*,
gimple_stmt_iterator*, bool*, _slp_tree*, _slp_instance*)
(tree-vect-stmts.c:5197)
==6443==    by 0xB2270B: vect_transform_loop(_loop_vec_info*)
(tree-vect-loop.c:5332)
==6443==    by 0xB2F678: vectorize_loops() (tree-vectorizer.c:205)
==6443==    by 0x8A6046: execute_one_pass(opt_pass*) (passes.c:2063)
==6443==    by 0x8A63B4: execute_pass_list(opt_pass*) (passes.c:2118)
==6443==    by 0x8A63C6: execute_pass_list(opt_pass*) (passes.c:2119)
==6443==    by 0x8A63C6: execute_pass_list(opt_pass*) (passes.c:2119)
==6443==    by 0x9FE69D: tree_rest_of_compilation(tree_node*)
(tree-optimize.c:420)
==6443==    by 0x653A55: cgraph_expand_function(cgraph_node*)
(cgraphunit.c:1797)
==6443==    by 0x6557BA: cgraph_optimize() (cgraphunit.c:1856)
==6443==    by 0x655E49: cgraph_finalize_compilation_unit() (cgraphunit.c:1310)
==6443==    by 0x537367: c_write_global_declarations() (c-decl.c:9881)
==6443==    by 0x99C191: toplev_main(int, char**) (toplev.c:564)
==6443==    by 0x6734D2C: (below main) (in /lib64/libc-2.12.2.so)
==6443==  Address 0x0 is not stack'd, malloc'd or (recently) free'd
==6443== 
testcase.c: In function 'foo':
testcase.c:4:1: internal compiler error: Segmentation fault
Please submit a full bug report,
with preprocessed source if appropriate.
See <http://gcc.gnu.org/bugs.html> for instructions.

Tested revisions:
r177889 - crash
r177543 - crash
r173150 - OK

[Bug tree-optimization/50133] [4.7 Regression] ICE: SIGSEGV in vect_finish_stmt_generation (gimple.h:4821) with -ftree-vectorize -fno-tree-loop-im

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=50133

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
   Target Milestone|---                         |4.7.0

[Bug tree-optimization/50133] [4.7 Regression] ICE: SIGSEGV in vect_finish_stmt_generation (gimple.h:4821) with -ftree-vectorize -fno-tree-loop-im

[irar at il dot ibm.com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=50133

Ira Rosen <irar at il dot ibm.com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |irar at il dot ibm.com

--- Comment #1 from Ira Rosen <irar at il dot ibm.com> 2011-08-21 06:02:10 UTC ---
When creating a vector from an invariant scalar load, we do 
gsi_next (&gsi2); 
in order to insert the vector after the load. In this testcase the load is the
last stmt in its basic block, which causes the failure when we try to call
gsi_stmt().

This patch fixes the failure:
Index: tree-vect-stmts.c
===================================================================
--- tree-vect-stmts.c   (revision 177929)
+++ tree-vect-stmts.c   (working copy)
@@ -1435,6 +1435,9 @@ vect_finish_stmt_generation (gimple stmt, gimple v
     }

   si = *gsi;
+  if (gsi_end_p (si))
+    si = gsi_for_stmt (stmt);
+
   if (is_gimple_debug (gsi_stmt (si)))
     {
       gsi_next_nondebug (&si);

but it's a hack.

We could add a new parameter insert_after to vect_finish_stmt_generation (), to
indicate that VECT_STMT should be inserted after GSI and not before. For this
case we would have to pass it to vect_init_vector() first.

Ira

[Bug tree-optimization/50133] [4.7 Regression] ICE: SIGSEGV in vect_finish_stmt_generation (gimple.h:4821) with -ftree-vectorize -fno-tree-loop-im

[irar at il dot ibm.com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=50133

Ira Rosen <irar at il dot ibm.com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
   Last reconfirmed|                            |2011-08-21
     Ever Confirmed|0                           |1

[Bug tree-optimization/50133] [4.7 Regression] ICE: SIGSEGV in vect_finish_stmt_generation (gimple.h:4821) with -ftree-vectorize -fno-tree-loop-im

[jakub at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=50133

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

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

--- Comment #2 from Jakub Jelinek <jakub at gcc dot gnu.org> 2011-08-21 17:43:08 UTC ---
Insert before gsi_end_p should be fine, you just can't copy locus from it in
that case.  So you could as well do something like:
  si = *gsi;
  if (gsi_end_p (si))
    {
      gsi_prev (&si);
      gsi_prev_nondebug (&si);
    }
  else if (is_gimple_debug (gsi_stmt (si)))
    {
      gsi_next_nondebug (&si);
      gcc_assert (!gsi_end_p (si));
    }

where the first gsi_prev would move from the end to the newly inserted stmt and
gsi_prev_nondebug to the previous stmt (skipping any debug stmts).
This wouldn't work when inserting into a completely empty bb, but does the
vectorizer ever do something like that?

[Bug tree-optimization/50133] [4.7 Regression] ICE: SIGSEGV in vect_finish_stmt_generation (gimple.h:4821) with -ftree-vectorize -fno-tree-loop-im

[jakub at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=50133

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|NEW                         |ASSIGNED
         AssignedTo|unassigned at gcc dot       |jakub at gcc dot gnu.org
                   |gnu.org                     |

--- Comment #3 from Jakub Jelinek <jakub at gcc dot gnu.org> 2011-08-22 09:31:23 UTC ---
Created attachment 25074
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=25074
gcc47-pr50133.patch

Untested fix.

[Bug tree-optimization/50133] [4.7 Regression] ICE: SIGSEGV in vect_finish_stmt_generation (gimple.h:4821) with -ftree-vectorize -fno-tree-loop-im

[jakub at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=50133

--- Comment #4 from Jakub Jelinek <jakub at gcc dot gnu.org> 2011-08-22 09:59:51 UTC ---
Or, alternatively can't we use:
  gimple_set_location (vec_stmt, gimple_location (stmt));
instead (and remove all the si handling stuff)?
It seems before PR37482 we were using the scalar stmt's location instead of the
gsi, which makes much more sense to me.  Ira, any reason why you've changed it?
I mean the asserts obviously with the patch had to be removed, but I wonder
about the gimple_set_location line...

[Bug tree-optimization/50133] [4.7 Regression] ICE: SIGSEGV in vect_finish_stmt_generation (gimple.h:4821) with -ftree-vectorize -fno-tree-loop-im

[jakub at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=50133

--- Comment #5 from Jakub Jelinek <jakub at gcc dot gnu.org> 2011-08-22 15:01:47 UTC ---
Author: jakub
Date: Mon Aug 22 15:01:42 2011
New Revision: 177958

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=177958
Log:
    PR tree-optimization/50133
    * tree-vect-stmts.c (vect_finish_stmt_generation): Copy location
    from stmt instead of some statement around gsi.

    * gcc.dg/pr50133.c: New test.

Added:
    trunk/gcc/testsuite/gcc.dg/pr50133.c
Modified:
    trunk/gcc/ChangeLog
    trunk/gcc/testsuite/ChangeLog
    trunk/gcc/tree-vect-stmts.c

[Bug tree-optimization/50133] [4.7 Regression] ICE: SIGSEGV in vect_finish_stmt_generation (gimple.h:4821) with -ftree-vectorize -fno-tree-loop-im

[jakub at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=50133

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

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

--- Comment #6 from Jakub Jelinek <jakub at gcc dot gnu.org> 2011-08-23 09:41:14 UTC ---
Fixed.