[Bug tree-optimization/47271] [4.6 Regression] if-conversion removes a test (if), the function generates invalid outputs

[spop at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=47271

--- Comment #10 from Sebastian Pop <spop at gcc dot gnu.org> 2011-01-24 
18:55:48 UTC ---
The BB of the phi node to be rewritten should post-dominate the BBs in which
the arguments of the phi node are defined.  The following gcc_assert will ICE
when the translation done in predicate_scalar_phi is unsafe:

diff --git a/gcc/tree-if-conv.c b/gcc/tree-if-conv.c
index cc7ad8a..9a3e2b5 100644
--- a/gcc/tree-if-conv.c
+++ b/gcc/tree-if-conv.c
@@ -992,6 +992,7 @@ if_convertible_loop_p_1 (struct loop *loop,
     return false;

   calculate_dominance_info (CDI_DOMINATORS);
+  calculate_dominance_info (CDI_POST_DOMINATORS);

   /* Allow statements that can be handled during if-conversion.  */
   ifc_bbs = get_loop_body_in_if_conv_order (loop);
@@ -1244,6 +1245,15 @@ predicate_scalar_phi (gimple phi, tree cond,
       arg_1 = gimple_phi_arg_def (phi, 1);
     }

+      gcc_assert ((TREE_CODE (arg_0) != SSA_NAME
+           || dominated_by_p (CDI_POST_DOMINATORS,
+                      gimple_bb (SSA_NAME_DEF_STMT (arg_0)),
+                      bb))
+          && (TREE_CODE (arg_1) != SSA_NAME
+              || dominated_by_p (CDI_POST_DOMINATORS,
+                     gimple_bb (SSA_NAME_DEF_STMT (arg_1)),
+                     bb)));
+
       /* Build new RHS using selected condition and arguments.  */
       rhs = build3 (COND_EXPR, TREE_TYPE (res),
             unshare_expr (cond), arg_0, arg_1);