PR tree-optimization/58380
authorlaw <law@138bc75d-0d04-0410-961f-82ee72b054a4>
Wed, 11 Sep 2013 02:23:48 +0000 (02:23 +0000)
committerlaw <law@138bc75d-0d04-0410-961f-82ee72b054a4>
Wed, 11 Sep 2013 02:23:48 +0000 (02:23 +0000)
* tree-ssa-threadupdate.c (thread_block): Recognize another case
of threading through a buried loop header.

* tree-ssa-threadedge.c (thread_around_empty_blocks): Correct
return value for single successor case.

* g++.dg/torture/pr58380.C: New test.

2013-09-10  Jeff Law  <law@redhat.com>

git-svn-id: svn+ssh://gcc.gnu.org/svn/gcc/trunk@202489 138bc75d-0d04-0410-961f-82ee72b054a4

gcc/ChangeLog
gcc/testsuite/ChangeLog
gcc/testsuite/g++.dg/torture/pr58380.C [new file with mode: 0644]
gcc/tree-ssa-threadedge.c
gcc/tree-ssa-threadupdate.c

index d4d32e1..21c8538 100644 (file)
@@ -1,3 +1,12 @@
+2013-09-10  Jeff Law  <law@redhat.com>
+
+       PR tree-optimization/58380
+       * tree-ssa-threadupdate.c (thread_block): Recognize another case
+       of threading through a buried loop header.
+
+       * tree-ssa-threadedge.c (thread_around_empty_blocks): Correct
+       return value for single successor case.
+
 2013-09-10  Jan Hubicka  <jh@suse.cz>
 
        * ipa-devirt.c (ipa_devirt): Enable with LTO.
index 8d58e2d..ae9e6e2 100644 (file)
@@ -1,3 +1,7 @@
+2013-09-10  Jeff Law  <law@redhat.com>
+
+       * g++.dg/torture/pr58380.C: New test.
+
 2013-09-10  Jan Hubicka  <jh@suse.cz>
            Paolo Carlini  <paolo.carlini@oracle.com>
 
diff --git a/gcc/testsuite/g++.dg/torture/pr58380.C b/gcc/testsuite/g++.dg/torture/pr58380.C
new file mode 100644 (file)
index 0000000..b2c185d
--- /dev/null
@@ -0,0 +1,171 @@
+// { dg-do compile }
+class iplugin_factory;
+class idocument_plugin_factory {
+  virtual idocument_plugin_factory *create_plugin(iplugin_factory &, int &);
+};
+template <typename _Iterator, typename> class __normal_iterator {
+  _Iterator _M_current;
+
+public:
+  _Iterator iterator_type;
+  __normal_iterator(const _Iterator &p1) : _M_current(p1) {}
+  void operator++();
+  _Iterator &base() { return _M_current; }
+};
+
+template <typename _IteratorL, typename _IteratorR, typename _Container>
+int operator!=(__normal_iterator<_IteratorL, _Container> &p1,
+               __normal_iterator<_IteratorR, _Container> &p2) {
+  return p1.base() != p2.base();
+}
+
+class new_allocator {
+public:
+  typedef int *const_pointer;
+  int *allocate();
+};
+template <typename> class allocator : public new_allocator {};
+
+class basic_string {
+public:
+  basic_string(char *);
+};
+struct __uninitialized_copy {
+  template <typename _InputIterator, typename _ForwardIterator>
+  static _ForwardIterator __uninit_copy(_InputIterator p1, _InputIterator p2,
+                                        _ForwardIterator p3) try {
+    for (; p1 != p2; ++p1, ++p3)
+      ;
+    return p3;
+  }
+  catch (...) {
+  }
+};
+
+template <typename _InputIterator, typename _ForwardIterator>
+_ForwardIterator uninitialized_copy(_InputIterator p1, _InputIterator p2,
+                                    _ForwardIterator p3) {
+  return __uninitialized_copy::__uninit_copy(p1, p2, p3);
+}
+
+template <typename _InputIterator, typename _ForwardIterator, typename _Tp>
+_ForwardIterator __uninitialized_copy_a(_InputIterator p1, _InputIterator p2,
+                                        _ForwardIterator p3, allocator<_Tp> &) {
+  return uninitialized_copy(p1, p2, p3);
+}
+
+struct _Vector_base {
+  struct _Vector_impl : allocator<int> {
+    int *_M_start;
+    int *_M_finish;
+  };
+  allocator<int> &_M_get_Tp_allocator() {}
+  _Vector_base() {}
+  _Vector_base(int p1) { _M_create_storage(p1); }
+  _Vector_impl _M_impl;
+  int *_M_allocate(int p1) { p1 ? _M_impl.allocate() : 0; }
+  void _M_create_storage(int p1) {
+    this->_M_impl._M_start = this->_M_allocate(p1);
+  }
+};
+
+class vector : _Vector_base {
+  _Vector_base _Base;
+
+public:
+  vector() {}
+  vector(const vector &p1) : _Base(p1.size()) {
+    this->_M_impl._M_finish = __uninitialized_copy_a(
+        p1.begin(), p1.end(), this->_M_impl._M_start, _M_get_Tp_allocator());
+  }
+  ~vector();
+  __normal_iterator<typename allocator<int>::const_pointer, int> begin() const {
+    return this->_M_impl._M_start;
+  }
+  __normal_iterator<typename allocator<int>::const_pointer, int> end() const {
+    return this->_M_impl._M_finish;
+  }
+  int size() const { return this->_M_impl._M_finish - this->_M_impl._M_start; }
+};
+class iplugin_factory {
+public:
+  typedef enum {
+    STABLE,
+    EXPERIMENTAL
+  } quality_t;
+};
+class plugin_factory : public iplugin_factory {
+public:
+  plugin_factory(const int &, const basic_string &, const basic_string &,
+                 const basic_string &, quality_t);
+};
+template <typename plugin_t>
+class document_plugin_factory : plugin_factory, idocument_plugin_factory {
+public:
+  document_plugin_factory(const int &p1, const basic_string &,
+                          const basic_string &, const basic_string &, quality_t)
+      : plugin_factory(0, 0, 0, 0, STABLE) {}
+  idocument_plugin_factory *create_plugin(iplugin_factory &p1, int &p2) {
+    plugin_t(p1, p2);
+  }
+};
+
+class container {
+public:
+  template <typename init_t> container(init_t &);
+};
+template <class init_t> class initializer_t : init_t {
+public:
+  initializer_t(const init_t &p1) : init_t(p1) {}
+};
+
+class composition_t {};
+template <typename lhs_t, typename rhs_t>
+const initializer_t<composition_t> operator+(const initializer_t<lhs_t> &,
+                                             const initializer_t<rhs_t> &);
+template <typename value_t> class value_initializer_t {
+public:
+  value_initializer_t(const value_t &p1) : m_value(p1) {}
+  value_t m_value;
+};
+
+template <typename value_t>
+initializer_t<value_initializer_t<value_t> > init_value(const value_t &p1) {
+  initializer_t<value_initializer_t<value_t> >(
+      value_initializer_t<value_t>(p1));
+}
+
+class name_t {};
+class label_t {};
+class description_t {};
+class owner_initializer_t {};
+template <typename owner_t>
+initializer_t<owner_initializer_t> init_owner(owner_t &);
+class set : vector {};
+class node {
+public:
+  node(iplugin_factory &, int &);
+};
+initializer_t<name_t> init_name();
+initializer_t<label_t> init_label();
+initializer_t<description_t> init_description();
+template <typename base_t> class mesh_selection_sink : base_t {
+public:
+  mesh_selection_sink(iplugin_factory &p1, int &p2)
+      : base_t(p1, p2),
+        m_mesh_selection(init_owner(*this) + init_name() + init_label() +
+                         init_description() + init_value(set())) {}
+  container m_mesh_selection;
+};
+
+class selection_to_stdout : mesh_selection_sink<node> {
+public:
+  selection_to_stdout(iplugin_factory &p1, int &p2)
+      : mesh_selection_sink(p1, p2) {}
+  static iplugin_factory &get_factory() {
+    document_plugin_factory<selection_to_stdout>(0, "", 0, "",
+                                                 iplugin_factory::EXPERIMENTAL);
+  }
+};
+
+void selection_to_stdout_factory() { selection_to_stdout::get_factory(); }
index 14bc4e3..839f185 100644 (file)
@@ -770,7 +770,19 @@ thread_around_empty_blocks (edge taken_edge,
   gsi = gsi_start_nondebug_bb (bb);
 
   /* If the block has no statements, but does have a single successor, then
-     it's just a forwarding block and we can thread through it trivially.  */
+     it's just a forwarding block and we can thread through it trivially. 
+
+     However, note that just threading through empty blocks with single
+     successors is not inherently profitable.  For the jump thread to
+     be profitable, we must avoid a runtime conditional.
+
+     By taking the return value from the recursive call, we get the
+     desired effect of returning TRUE when we found a profitable jump
+     threading opportunity and FALSE otherwise. 
+
+     This is particularly important when this routine is called after
+     processing a joiner block.  Returning TRUE too aggressively in
+     that case results in pointless duplication of the joiner block.  */
   if (gsi_end_p (gsi))
     {
       if (single_succ_p (bb))
@@ -781,15 +793,16 @@ thread_around_empty_blocks (edge taken_edge,
            {
              bitmap_set_bit (visited, taken_edge->dest->index);
              path->safe_push (taken_edge);
-             thread_around_empty_blocks (taken_edge,
-                                         dummy_cond,
-                                         handle_dominating_asserts,
-                                         simplify,
-                                         visited,
-                                         path);
-             return true;
+             return thread_around_empty_blocks (taken_edge,
+                                                dummy_cond,
+                                                handle_dominating_asserts,
+                                                simplify,
+                                                visited,
+                                                path);
            }
        }
+
+      /* We have a block with no statements, but multiple successors?  */
       return false;
     }
 
index 3c3d3bc..6d43dc3 100644 (file)
@@ -659,8 +659,10 @@ thread_block (basic_block bb, bool noloop_only)
             threading path.  We do not try and thread this elsewhere, so
             just cancel the jump threading request by clearing the AUX
             field now.  */
-         if (bb->loop_father != e2->src->loop_father
-             && !loop_exit_edge_p (e2->src->loop_father, e2))
+         if ((bb->loop_father != e2->src->loop_father
+              && !loop_exit_edge_p (e2->src->loop_father, e2))
+             || (e2->src->loop_father != e2->dest->loop_father
+                 && !loop_exit_edge_p (e2->src->loop_father, e2)))
            {
              /* Since this case is not handled by our special code
                 to thread through a loop header, we must explicitly