Hi!

On the following testcase, starting with r221675 aka PR65177 fix
we get ICE, because FSM discovery finds a path that includes the same blocks
multiple times, like:
 Registering FSM jump thread: (9, 4) incoming edge;  (4, 5)  (5, 12)  (12, 14)  
(14, 5)  (5, 12) nocopy; (5, 12) 
All these bbs belong to the same loop, with bb14 being the header and bb12
the latch.  And the copy_bbs/duplicate_thread_path don't seem to be really
prepared to duplicate the same basic block more than once.

fsm_find_control_statement_thread_paths has guard against recursion, but it
adds to the hash_set the PHI nodes.  On the testcase, bb5 is added to the
path first through one of the PHIs:
# c_3 = PHI <c_39(4), b_17(14)>
# b_33 = PHI <b_32(4), b_17(14)>
and the second time through the other PHI.

The following patch fixes that by adding to the has_set the basic blocks
containing the PHIs instead of the PHIs.

Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?

2015-04-11  Jakub Jelinek  <ja...@redhat.com>

        PR tree-optimization/65735
        * tree-ssa-threadedge.c (fsm_find_control_statement_thread_paths):
        Remove visited_phis argument, add visited_bbs, avoid recursing into the
        same bb rather than just into the same phi node.
        (thread_through_normal_block): Adjust caller.

        * gcc.c-torture/compile/pr65735.c: New test.

--- gcc/tree-ssa-threadedge.c.jj        2015-02-16 22:18:34.000000000 +0100
+++ gcc/tree-ssa-threadedge.c   2015-04-11 16:13:51.906916300 +0200
@@ -1015,7 +1015,7 @@ static int max_threaded_paths;
 
 static void
 fsm_find_control_statement_thread_paths (tree expr,
-                                        hash_set<gimple> *visited_phis,
+                                        hash_set<basic_block> *visited_bbs,
                                         vec<basic_block, va_gc> *&path,
                                         bool seen_loop_phi)
 {
@@ -1034,7 +1034,7 @@ fsm_find_control_statement_thread_paths
     return;
 
   /* Avoid infinite recursion.  */
-  if (visited_phis->add (def_stmt))
+  if (visited_bbs->add (var_bb))
     return;
 
   gphi *phi = as_a <gphi *> (def_stmt);
@@ -1109,7 +1109,7 @@ fsm_find_control_statement_thread_paths
        {
          vec_safe_push (path, bbi);
          /* Recursively follow SSA_NAMEs looking for a constant definition.  */
-         fsm_find_control_statement_thread_paths (arg, visited_phis, path,
+         fsm_find_control_statement_thread_paths (arg, visited_bbs, path,
                                                   seen_loop_phi);
 
          path->pop ();
@@ -1391,13 +1391,13 @@ thread_through_normal_block (edge e,
       vec<basic_block, va_gc> *bb_path;
       vec_alloc (bb_path, n_basic_blocks_for_fn (cfun));
       vec_safe_push (bb_path, e->dest);
-      hash_set<gimple> *visited_phis = new hash_set<gimple>;
+      hash_set<basic_block> *visited_bbs = new hash_set<basic_block>;
 
       max_threaded_paths = PARAM_VALUE (PARAM_MAX_FSM_THREAD_PATHS);
-      fsm_find_control_statement_thread_paths (cond, visited_phis, bb_path,
+      fsm_find_control_statement_thread_paths (cond, visited_bbs, bb_path,
                                               false);
 
-      delete visited_phis;
+      delete visited_bbs;
       vec_free (bb_path);
     }
   return 0;
--- gcc/testsuite/gcc.c-torture/compile/pr65735.c.jj    2015-04-11 
16:14:33.173263982 +0200
+++ gcc/testsuite/gcc.c-torture/compile/pr65735.c       2015-04-11 
16:14:06.000000000 +0200
@@ -0,0 +1,21 @@
+/* PR tree-optimization/65735 */
+
+int foo (void);
+
+void
+bar (int a, int b, int c)
+{
+  while (!a)
+    {
+      c = foo ();
+      if (c == 7)
+       c = b;
+      switch (c)
+       {
+       case 1:
+         a = b++;
+         if (b)
+           b = 1;
+       }
+    }
+}

        Jakub

Reply via email to