src/share/vm/opto/loopnode.hpp

@@ -67,11 +67,13 @@
          HasReductions=128,
          WasSlpAnalyzed=256,
          PassedSlpAnalysis=512,
          DoUnrollOnly=1024,
          VectorizedLoop=2048,
-         HasAtomicPostLoop=4096 };
+         HasAtomicPostLoop=4096,
+         HasRangeChecks=8192,
+         IsMultiversioned=16384};
   char _unswitch_count;
   enum { _unswitch_max=3 };
 
 public:
   // Names for edge indices

@@ -88,10 +90,12 @@
   void mark_was_slp() { _loop_flags |= WasSlpAnalyzed; }
   void mark_passed_slp() { _loop_flags |= PassedSlpAnalysis; }
   void mark_do_unroll_only() { _loop_flags |= DoUnrollOnly; }
   void mark_loop_vectorized() { _loop_flags |= VectorizedLoop; }
   void mark_has_atomic_post_loop() { _loop_flags |= HasAtomicPostLoop; }
+  void mark_has_range_checks() { _loop_flags |=  HasRangeChecks; }
+  void mark_is_multiversioned() { _loop_flags |= IsMultiversioned; }
 
   int unswitch_max() { return _unswitch_max; }
   int unswitch_count() { return _unswitch_count; }
   void set_unswitch_count(int val) {
     assert (val <= unswitch_max(), "too many unswitches");

@@ -222,14 +226,16 @@
   int is_post_loop     () const { return (_loop_flags&PreMainPostFlagsMask) == Post;   }
   int is_reduction_loop() const { return (_loop_flags&HasReductions) == HasReductions; }
   int was_slp_analyzed () const { return (_loop_flags&WasSlpAnalyzed) == WasSlpAnalyzed; }
   int has_passed_slp   () const { return (_loop_flags&PassedSlpAnalysis) == PassedSlpAnalysis; }
   int do_unroll_only      () const { return (_loop_flags&DoUnrollOnly) == DoUnrollOnly; }
-  int is_main_no_pre_loop() const { return _loop_flags & MainHasNoPreLoop; }
+  int is_main_no_pre_loop   () const { return _loop_flags & MainHasNoPreLoop; }
   int is_vectorized_loop    () const { return (_loop_flags & VectorizedLoop) == VectorizedLoop; }
   int has_atomic_post_loop  () const { return (_loop_flags & HasAtomicPostLoop) == HasAtomicPostLoop; }
-  void set_main_no_pre_loop() { _loop_flags |= MainHasNoPreLoop; }
+  int loop_has_range_checks () const { return (_loop_flags & HasRangeChecks) == HasRangeChecks; }
+  int loop_is_multiversioned() const { return (_loop_flags & IsMultiversioned) == IsMultiversioned; }
+  void set_main_no_pre_loop () { _loop_flags |= MainHasNoPreLoop; }
 
   int main_idx() const { return _main_idx; }
 
 
   void set_pre_loop  (CountedLoopNode *main) { assert(is_normal_loop(),""); _loop_flags |= Pre ; _main_idx = main->_idx; }

@@ -540,10 +546,20 @@
   void verify_tree(IdealLoopTree *loop, const IdealLoopTree *parent) const;
 #endif
 
 };
 
+class PostLoopInfo {
+public:
+  Node *new_main_exit;
+  CountedLoopNode *post_head;
+
+  PostLoopInfo() { init(); }
+
+  void init() { new_main_exit = NULL;  post_head = NULL; }
+};
+
 // -----------------------------PhaseIdealLoop---------------------------------
 // Computes the mapping from Nodes to IdealLoopTrees.  Organizes IdealLoopTrees into a
 // loop tree.  Drives the loop-based transformations on the ideal graph.
 class PhaseIdealLoop : public PhaseTransform {
   friend class IdealLoopTree;

@@ -906,12 +922,22 @@
   void do_peeling( IdealLoopTree *loop, Node_List &old_new );
 
   // Add pre and post loops around the given loop.  These loops are used
   // during RCE, unrolling and aligning loops.
   void insert_pre_post_loops( IdealLoopTree *loop, Node_List &old_new, bool peel_only );
+
+  // Add post loop after the given loop.
+  void insert_post_loop( IdealLoopTree *loop, Node_List &old_new,
+                         CountedLoopNode *main_head, CountedLoopEndNode *main_end,
+                         Node *incr, Node *limit, PostLoopInfo &post_loop_info );
+
+  // Add an RCE'd post loop which we will multi-version adapt for run time test path usage
+  void insert_scalar_rced_post_loop( IdealLoopTree *loop, Node_List &old_new );
+
   // Add a vector post loop between a vector main loop and the current post loop
   void insert_vector_post_loop(IdealLoopTree *loop, Node_List &old_new);
+
   // If Node n lives in the back_ctrl block, we clone a private version of n
   // in preheader_ctrl block and return that, otherwise return n.
   Node *clone_up_backedge_goo( Node *back_ctrl, Node *preheader_ctrl, Node *n, VectorSet &visited, Node_Stack &clones );
 
   // Take steps to maximally unroll the loop.  Peel any odd iterations, then

@@ -978,11 +1004,21 @@
   bool is_node_unreachable(Node *n) const {
     return !has_node(n) || n->is_unreachable(_igvn);
   }
 
   // Eliminate range-checks and other trip-counter vs loop-invariant tests.
-  void do_range_check( IdealLoopTree *loop, Node_List &old_new );
+  void do_range_check(IdealLoopTree *loop, Node_List &range_check_list);
+
+  // Check to see if do_range_check(...) cleaned the main loop of range-checks
+  void has_range_checks(IdealLoopTree *loop);
+
+  // Process post loops which have range checks and try to build a multi-version
+  // guard to safely determine if we can execute the post loop which was RCE'd.
+  bool multi_version_post_loops(IdealLoopTree *rce_loop, IdealLoopTree *legacy_loop);
+
+  // Cause the rce'd post loop to optimized away, this happens if we cannot complete multiverioning
+  void poison_rce_post_loop(IdealLoopTree *rce_loop);
 
   // Create a slow version of the loop by cloning the loop
   // and inserting an if to select fast-slow versions.
   ProjNode* create_slow_version_of_loop(IdealLoopTree *loop,
                                         Node_List &old_new,