[Bug tree-optimization/51058] [4.7 Regression] ICE: gimple check: expected gimple_assign(error_mark), have gimple_call() in gimple_assign_rhs_code, at gimple.h:1992

[jakub at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=51058

--- Comment #9 from Jakub Jelinek <jakub at gcc dot gnu.org> 2011-11-10 18:40:05 UTC ---
The problem is that we have several SLP instances referring to the same call
stmt, so vectorizable_call is called on the same stmt several times.
But vectorizable_call wants to replace the scalar call with a assignment and
when it does the first time it is called, we get a segfault the second time
because the original scalar call stmt has bb set to NULL by gsi_replace.

--- gcc/tree-vect-stmts.c.jj 12011-11-10 18:09:12.000000000 +0100
+++ gcc/tree-vect-stmts.c 2011-11-10 18:44:07.135949119 +0100
@@ -1886,6 +1886,9 @@ vectorizable_call (gimple stmt, gimple_s
      it defines is mapped to the new definition.  So just replace
      rhs of the statement with something harmless.  */

+  if (slp_node)
+    return true;
+
   type = TREE_TYPE (scalar_dest);
   if (is_pattern_stmt_p (stmt_info))
     lhs = gimple_call_lhs (STMT_VINFO_RELATED_STMT (stmt_info));
@@ -1893,8 +1896,7 @@ vectorizable_call (gimple stmt, gimple_s
     lhs = gimple_call_lhs (stmt);
   new_stmt = gimple_build_assign (lhs, build_zero_cst (type));
   set_vinfo_for_stmt (new_stmt, stmt_info);
-  if (!slp_node)
-    set_vinfo_for_stmt (stmt, NULL);
+  set_vinfo_for_stmt (stmt, NULL);
   STMT_VINFO_STMT (stmt_info) = new_stmt;
   gsi_replace (gsi, new_stmt, false);
   SSA_NAME_DEF_STMT (gimple_assign_lhs (new_stmt)) = new_stmt;

patch fixes the ICE, but then the scalar stmt stays in the tree at the end of
*.vect pass (and at least in this case is DCEd afterwards).
If we wanted to do this replacement, I guess we could do that from
vect_schedule_slp after all vect_schedule_slp_instance calls have returned, but
we'd need to again recursively go through the instance tree.