8024069: replace_in_map() should operate on parent maps

Summary: type information gets lost because replace_in_map() doesn't update parent maps
Reviewed-by: kvn, twisti

src/share/vm/opto/c2_globals.hpp

@@ -638,7 +638,10 @@
           "Find best control for expensive operations")                     \
                                                                             \
   product(bool, UseMathExactIntrinsics, true,                               \
-          "Enables intrinsification of various java.lang.Math funcitons")
+          "Enables intrinsification of various java.lang.Math functions")   \
+                                                                            \
+  experimental(bool, ReplaceInParentMaps, false,                            \
+          "Propagate type improvements in callers of inlinee if possible")
 
 C2_FLAGS(DECLARE_DEVELOPER_FLAG, DECLARE_PD_DEVELOPER_FLAG, DECLARE_PRODUCT_FLAG, DECLARE_PD_PRODUCT_FLAG, DECLARE_DIAGNOSTIC_FLAG, DECLARE_EXPERIMENTAL_FLAG, DECLARE_NOTPRODUCT_FLAG)
 

src/share/vm/opto/callGenerator.cpp

@@ -63,12 +63,12 @@ public:
   }
 
   virtual bool      is_parse() const           { return true; }
-  virtual JVMState* generate(JVMState* jvms);
+  virtual JVMState* generate(JVMState* jvms, Parse* parent_parser);
   int is_osr() { return _is_osr; }
 
 };
 
-JVMState* ParseGenerator::generate(JVMState* jvms) {
+JVMState* ParseGenerator::generate(JVMState* jvms, Parse* parent_parser) {
   Compile* C = Compile::current();
 
   if (is_osr()) {
@@ -80,7 +80,7 @@ JVMState* ParseGenerator::generate(JVMState* jvms) {
     return NULL;  // bailing out of the compile; do not try to parse
   }
 
-  Parse parser(jvms, method(), _expected_uses);
+  Parse parser(jvms, method(), _expected_uses, parent_parser);
   // Grab signature for matching/allocation
 #ifdef ASSERT
   if (parser.tf() != (parser.depth() == 1 ? C->tf() : tf())) {
@@ -119,12 +119,12 @@ class DirectCallGenerator : public CallGenerator {
       _separate_io_proj(separate_io_proj)
   {
   }
-  virtual JVMState* generate(JVMState* jvms);
+  virtual JVMState* generate(JVMState* jvms, Parse* parent_parser);
 
   CallStaticJavaNode* call_node() const { return _call_node; }
 };
 
-JVMState* DirectCallGenerator::generate(JVMState* jvms) {
+JVMState* DirectCallGenerator::generate(JVMState* jvms, Parse* parent_parser) {
   GraphKit kit(jvms);
   bool is_static = method()->is_static();
   address target = is_static ? SharedRuntime::get_resolve_static_call_stub()
@@ -171,10 +171,10 @@ public:
            vtable_index >= 0, "either invalid or usable");
   }
   virtual bool      is_virtual() const          { return true; }
-  virtual JVMState* generate(JVMState* jvms);
+  virtual JVMState* generate(JVMState* jvms, Parse* parent_parser);
 };
 
-JVMState* VirtualCallGenerator::generate(JVMState* jvms) {
+JVMState* VirtualCallGenerator::generate(JVMState* jvms, Parse* parent_parser) {
   GraphKit kit(jvms);
   Node* receiver = kit.argument(0);
 
@@ -276,7 +276,7 @@ class LateInlineCallGenerator : public DirectCallGenerator {
   // Convert the CallStaticJava into an inline
   virtual void do_late_inline();
 
-  virtual JVMState* generate(JVMState* jvms) {
+  virtual JVMState* generate(JVMState* jvms, Parse* parent_parser) {
     Compile *C = Compile::current();
     C->print_inlining_skip(this);
 
@@ -290,7 +290,7 @@ class LateInlineCallGenerator : public DirectCallGenerator {
     // that the late inlining logic can distinguish between fall
     // through and exceptional uses of the memory and io projections
     // as is done for allocations and macro expansion.
-    return DirectCallGenerator::generate(jvms);
+    return DirectCallGenerator::generate(jvms, parent_parser);
   }
 
   virtual void print_inlining_late(const char* msg) {
@@ -389,7 +389,7 @@ void LateInlineCallGenerator::do_late_inline() {
   }
 
   // Now perform the inling using the synthesized JVMState
-  JVMState* new_jvms = _inline_cg->generate(jvms);
+  JVMState* new_jvms = _inline_cg->generate(jvms, NULL);
   if (new_jvms == NULL)  return;  // no change
   if (C->failing())      return;
 
@@ -429,8 +429,8 @@ class LateInlineMHCallGenerator : public LateInlineCallGenerator {
 
   virtual bool is_mh_late_inline() const { return true; }
 
-  virtual JVMState* generate(JVMState* jvms) {
-    JVMState* new_jvms = LateInlineCallGenerator::generate(jvms);
+  virtual JVMState* generate(JVMState* jvms, Parse* parent_parser) {
+    JVMState* new_jvms = LateInlineCallGenerator::generate(jvms, parent_parser);
     if (_input_not_const) {
       // inlining won't be possible so no need to enqueue right now.
       call_node()->set_generator(this);
@@ -477,13 +477,13 @@ class LateInlineStringCallGenerator : public LateInlineCallGenerator {
   LateInlineStringCallGenerator(ciMethod* method, CallGenerator* inline_cg) :
     LateInlineCallGenerator(method, inline_cg) {}
 
-  virtual JVMState* generate(JVMState* jvms) {
+  virtual JVMState* generate(JVMState* jvms, Parse* parent_parser) {
     Compile *C = Compile::current();
     C->print_inlining_skip(this);
 
     C->add_string_late_inline(this);
 
-    JVMState* new_jvms =  DirectCallGenerator::generate(jvms);
+    JVMState* new_jvms =  DirectCallGenerator::generate(jvms, parent_parser);
     return new_jvms;
   }
 };
@@ -498,13 +498,13 @@ class LateInlineBoxingCallGenerator : public LateInlineCallGenerator {
   LateInlineBoxingCallGenerator(ciMethod* method, CallGenerator* inline_cg) :
     LateInlineCallGenerator(method, inline_cg) {}
 
-  virtual JVMState* generate(JVMState* jvms) {
+  virtual JVMState* generate(JVMState* jvms, Parse* parent_parser) {
     Compile *C = Compile::current();
     C->print_inlining_skip(this);
 
     C->add_boxing_late_inline(this);
 
-    JVMState* new_jvms =  DirectCallGenerator::generate(jvms);
+    JVMState* new_jvms =  DirectCallGenerator::generate(jvms, parent_parser);
     return new_jvms;
   }
 };
@@ -540,7 +540,7 @@ public:
   virtual bool      is_virtual() const          { return _is_virtual; }
   virtual bool      is_deferred() const         { return true; }
 
-  virtual JVMState* generate(JVMState* jvms);
+  virtual JVMState* generate(JVMState* jvms, Parse* parent_parser);
 };
 
 
@@ -550,12 +550,12 @@ CallGenerator* CallGenerator::for_warm_call(WarmCallInfo* ci,
   return new WarmCallGenerator(ci, if_cold, if_hot);
 }
 
-JVMState* WarmCallGenerator::generate(JVMState* jvms) {
+JVMState* WarmCallGenerator::generate(JVMState* jvms, Parse* parent_parser) {
   Compile* C = Compile::current();
   if (C->log() != NULL) {
     C->log()->elem("warm_call bci='%d'", jvms->bci());
   }
-  jvms = _if_cold->generate(jvms);
+  jvms = _if_cold->generate(jvms, parent_parser);
   if (jvms != NULL) {
     Node* m = jvms->map()->control();
     if (m->is_CatchProj()) m = m->in(0);  else m = C->top();
@@ -616,7 +616,7 @@ public:
   virtual bool      is_inline()    const    { return _if_hit->is_inline(); }
   virtual bool      is_deferred()  const    { return _if_hit->is_deferred(); }
 
-  virtual JVMState* generate(JVMState* jvms);
+  virtual JVMState* generate(JVMState* jvms, Parse* parent_parser);
 };
 
 
@@ -628,7 +628,7 @@ CallGenerator* CallGenerator::for_predicted_call(ciKlass* predicted_receiver,
 }
 
 
-JVMState* PredictedCallGenerator::generate(JVMState* jvms) {
+JVMState* PredictedCallGenerator::generate(JVMState* jvms, Parse* parent_parser) {
   GraphKit kit(jvms);
   PhaseGVN& gvn = kit.gvn();
   // We need an explicit receiver null_check before checking its type.
@@ -656,7 +656,7 @@ JVMState* PredictedCallGenerator::generate(JVMState* jvms) {
   { PreserveJVMState pjvms(&kit);
     kit.set_control(slow_ctl);
     if (!kit.stopped()) {
-      slow_jvms = _if_missed->generate(kit.sync_jvms());
+      slow_jvms = _if_missed->generate(kit.sync_jvms(), parent_parser);
       if (kit.failing())
         return NULL;  // might happen because of NodeCountInliningCutoff
       assert(slow_jvms != NULL, "must be");
@@ -677,12 +677,12 @@ JVMState* PredictedCallGenerator::generate(JVMState* jvms) {
   kit.replace_in_map(receiver, exact_receiver);
 
   // Make the hot call:
-  JVMState* new_jvms = _if_hit->generate(kit.sync_jvms());
+  JVMState* new_jvms = _if_hit->generate(kit.sync_jvms(), parent_parser);
   if (new_jvms == NULL) {
     // Inline failed, so make a direct call.
     assert(_if_hit->is_inline(), "must have been a failed inline");
     CallGenerator* cg = CallGenerator::for_direct_call(_if_hit->method());
-    new_jvms = cg->generate(kit.sync_jvms());
+    new_jvms = cg->generate(kit.sync_jvms(), parent_parser);
   }
   kit.add_exception_states_from(new_jvms);
   kit.set_jvms(new_jvms);
@@ -874,7 +874,7 @@ public:
   virtual bool      is_inlined()   const    { return true; }
   virtual bool      is_intrinsic() const    { return true; }
 
-  virtual JVMState* generate(JVMState* jvms);
+  virtual JVMState* generate(JVMState* jvms, Parse* parent_parser);
 };
 
 
@@ -884,7 +884,7 @@ CallGenerator* CallGenerator::for_predicted_intrinsic(CallGenerator* intrinsic,
 }
 
 
-JVMState* PredictedIntrinsicGenerator::generate(JVMState* jvms) {
+JVMState* PredictedIntrinsicGenerator::generate(JVMState* jvms, Parse* parent_parser) {
   GraphKit kit(jvms);
   PhaseGVN& gvn = kit.gvn();
 
@@ -904,7 +904,7 @@ JVMState* PredictedIntrinsicGenerator::generate(JVMState* jvms) {
     PreserveJVMState pjvms(&kit);
     kit.set_control(slow_ctl);
     if (!kit.stopped()) {
-      slow_jvms = _cg->generate(kit.sync_jvms());
+      slow_jvms = _cg->generate(kit.sync_jvms(), parent_parser);
       if (kit.failing())
         return NULL;  // might happen because of NodeCountInliningCutoff
       assert(slow_jvms != NULL, "must be");
@@ -922,12 +922,12 @@ JVMState* PredictedIntrinsicGenerator::generate(JVMState* jvms) {
   }
 
   // Generate intrinsic code:
-  JVMState* new_jvms = _intrinsic->generate(kit.sync_jvms());
+  JVMState* new_jvms = _intrinsic->generate(kit.sync_jvms(), parent_parser);
   if (new_jvms == NULL) {
     // Intrinsic failed, so use slow code or make a direct call.
     if (slow_map == NULL) {
       CallGenerator* cg = CallGenerator::for_direct_call(method());
-      new_jvms = cg->generate(kit.sync_jvms());
+      new_jvms = cg->generate(kit.sync_jvms(), parent_parser);
     } else {
       kit.set_jvms(slow_jvms);
       return kit.transfer_exceptions_into_jvms();
@@ -997,7 +997,7 @@ public:
   virtual bool      is_virtual() const          { ShouldNotReachHere(); return false; }
   virtual bool      is_trap() const             { return true; }
 
-  virtual JVMState* generate(JVMState* jvms);
+  virtual JVMState* generate(JVMState* jvms, Parse* parent_parser);
 };
 
 
@@ -1009,7 +1009,7 @@ CallGenerator::for_uncommon_trap(ciMethod* m,
 }
 
 
-JVMState* UncommonTrapCallGenerator::generate(JVMState* jvms) {
+JVMState* UncommonTrapCallGenerator::generate(JVMState* jvms, Parse* parent_parser) {
   GraphKit kit(jvms);
   // Take the trap with arguments pushed on the stack.  (Cf. null_check_receiver).
   int nargs = method()->arg_size();

src/share/vm/opto/callGenerator.hpp

@@ -31,6 +31,8 @@
 #include "opto/type.hpp"
 #include "runtime/deoptimization.hpp"
 
+class Parse;
+
 //---------------------------CallGenerator-------------------------------------
 // The subclasses of this class handle generation of ideal nodes for
 // call sites and method entry points.
@@ -108,7 +110,7 @@ class CallGenerator : public ResourceObj {
   //
   // If the result is NULL, it means that this CallGenerator was unable
   // to handle the given call, and another CallGenerator should be consulted.
-  virtual JVMState* generate(JVMState* jvms) = 0;
+  virtual JVMState* generate(JVMState* jvms, Parse* parent_parser) = 0;
 
   // How to generate a call site that is inlined:
   static CallGenerator* for_inline(ciMethod* m, float expected_uses = -1);

src/share/vm/opto/compile.cpp

@@ -655,7 +655,8 @@ Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr
                   _inlining_progress(false),
                   _inlining_incrementally(false),
                   _print_inlining_list(NULL),
-                  _print_inlining_idx(0) {
+                  _print_inlining_idx(0),
+                  _preserve_jvm_state(0) {
   C = this;
 
   CompileWrapper cw(this);
@@ -763,7 +764,7 @@ Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr
       return;
     }
     JVMState* jvms = build_start_state(start(), tf());
-    if ((jvms = cg->generate(jvms)) == NULL) {
+    if ((jvms = cg->generate(jvms, NULL)) == NULL) {
       record_method_not_compilable("method parse failed");
       return;
     }
@@ -940,7 +941,8 @@ Compile::Compile( ciEnv* ci_env,
     _inlining_progress(false),
     _inlining_incrementally(false),
     _print_inlining_list(NULL),
-    _print_inlining_idx(0) {
+    _print_inlining_idx(0),
+    _preserve_jvm_state(0) {
   C = this;
 
 #ifndef PRODUCT

src/share/vm/opto/compile.hpp

@@ -425,6 +425,9 @@ class Compile : public Phase {
   // Expensive nodes list already sorted?
   bool expensive_nodes_sorted() const;
 
+  // Are we within a PreserveJVMState block?
+  int _preserve_jvm_state;
+
  public:
 
   outputStream* print_inlining_stream() const {
@@ -820,7 +823,9 @@ class Compile : public Phase {
 
   // Decide how to build a call.
   // The profile factor is a discount to apply to this site's interp. profile.
-  CallGenerator*    call_generator(ciMethod* call_method, int vtable_index, bool call_does_dispatch, JVMState* jvms, bool allow_inline, float profile_factor, bool allow_intrinsics = true, bool delayed_forbidden = false);
+  CallGenerator*    call_generator(ciMethod* call_method, int vtable_index, bool call_does_dispatch,
+                                   JVMState* jvms, bool allow_inline, float profile_factor, bool allow_intrinsics = true,
+                                   bool delayed_forbidden = false);
   bool should_delay_inlining(ciMethod* call_method, JVMState* jvms) {
     return should_delay_string_inlining(call_method, jvms) ||
            should_delay_boxing_inlining(call_method, jvms);
@@ -1156,6 +1161,21 @@ class Compile : public Phase {
 
   // Auxiliary method for randomized fuzzing/stressing
   static bool randomized_select(int count);
+
+  // enter a PreserveJVMState block
+  void inc_preserve_jvm_state() {
+    _preserve_jvm_state++;
+  }
+
+  // exit a PreserveJVMState block
+  void dec_preserve_jvm_state() {
+    _preserve_jvm_state--;
+    assert(_preserve_jvm_state >= 0, "_preserve_jvm_state shouldn't be negative");
+  }
+
+  bool has_preserve_jvm_state() const {
+    return _preserve_jvm_state > 0;
+  }
 };
 
 #endif // SHARE_VM_OPTO_COMPILE_HPP

src/share/vm/opto/doCall.cpp

@@ -495,7 +495,7 @@ void Parse::do_call() {
   // because exceptions don't return to the call site.)
   profile_call(receiver);
 
-  JVMState* new_jvms = cg->generate(jvms);
+  JVMState* new_jvms = cg->generate(jvms, this);
   if (new_jvms == NULL) {
     // When inlining attempt fails (e.g., too many arguments),
     // it may contaminate the current compile state, making it
@@ -509,7 +509,7 @@ void Parse::do_call() {
     // intrinsic was expecting to optimize. Should always be possible to
     // get a normal java call that may inline in that case
     cg = C->call_generator(cg->method(), vtable_index, call_does_dispatch, jvms, try_inline, prof_factor(), /* allow_intrinsics= */ false);
-    if ((new_jvms = cg->generate(jvms)) == NULL) {
+    if ((new_jvms = cg->generate(jvms, this)) == NULL) {
       guarantee(failing(), "call failed to generate:  calls should work");
       return;
     }

src/share/vm/opto/graphKit.cpp

@@ -639,6 +639,7 @@ PreserveJVMState::PreserveJVMState(GraphKit* kit, bool clone_map) {
   _map    = kit->map();   // preserve the map
   _sp     = kit->sp();
   kit->set_map(clone_map ? kit->clone_map() : NULL);
+  Compile::current()->inc_preserve_jvm_state();
 #ifdef ASSERT
   _bci    = kit->bci();
   Parse* parser = kit->is_Parse();
@@ -656,6 +657,7 @@ PreserveJVMState::~PreserveJVMState() {
 #endif
   kit->set_map(_map);
   kit->set_sp(_sp);
+  Compile::current()->dec_preserve_jvm_state();
 }
 
 
@@ -1373,17 +1375,70 @@ Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) {
 
 //--------------------------replace_in_map-------------------------------------
 void GraphKit::replace_in_map(Node* old, Node* neww) {
-  this->map()->replace_edge(old, neww);
+  if (old == neww) {
+    return;
+  }
+
+  map()->replace_edge(old, neww);
 
   // Note: This operation potentially replaces any edge
   // on the map.  This includes locals, stack, and monitors
   // of the current (innermost) JVM state.
 
-  // We can consider replacing in caller maps.
-  // The idea would be that an inlined function's null checks
-  // can be shared with the entire inlining tree.
-  // The expense of doing this is that the PreserveJVMState class
-  // would have to preserve caller states too, with a deep copy.
+  if (!ReplaceInParentMaps) {
+    return;
+  }
+
+  // PreserveJVMState doesn't do a deep copy so we can't modify
+  // parents
+  if (Compile::current()->has_preserve_jvm_state()) {
+    return;
+  }
+
+  Parse* parser = is_Parse();
+  bool progress = true;
+  Node* ctrl = map()->in(0);
+  // Follow the chain of parsers and see whether the update can be
+  // done in the map of callers. We can do the replace for a caller if
+  // the current control post dominates the control of a caller.
+  while (parser != NULL && parser->caller() != NULL && progress) {
+    progress = false;
+    Node* parent_map = parser->caller()->map();
+    assert(parser->exits().map()->jvms()->depth() == parser->caller()->depth(), "map mismatch");
+
+    Node* parent_ctrl = parent_map->in(0);
+
+    while (parent_ctrl->is_Region()) {
+      Node* n = parent_ctrl->as_Region()->is_copy();
+      if (n == NULL) {
+        break;
+      }
+      parent_ctrl = n;
+    }
+
+    for (;;) {
+      if (ctrl == parent_ctrl) {
+        // update the map of the exits which is the one that will be
+        // used when compilation resume after inlining
+        parser->exits().map()->replace_edge(old, neww);
+        progress = true;
+        break;
+      }
+      if (ctrl->is_Proj() && ctrl->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none)) {
+        ctrl = ctrl->in(0)->in(0);
+      } else if (ctrl->is_Region()) {
+        Node* n = ctrl->as_Region()->is_copy();
+        if (n == NULL) {
+          break;
+        }
+        ctrl = n;
+      } else {
+        break;
+      }
+    }
+
+    parser = parser->parent_parser();
+  }
 }
 
 

src/share/vm/opto/ifnode.cpp

@@ -1019,7 +1019,7 @@ void IfNode::dominated_by( Node *prev_dom, PhaseIterGVN *igvn ) {
   // be skipped. For example, range check predicate has two checks
   // for lower and upper bounds.
   ProjNode* unc_proj = proj_out(1 - prev_dom->as_Proj()->_con)->as_Proj();
-  if (PhaseIdealLoop::is_uncommon_trap_proj(unc_proj, Deoptimization::Reason_predicate))
+  if (unc_proj->is_uncommon_trap_proj(Deoptimization::Reason_predicate))
     prev_dom = idom;
 
   // Now walk the current IfNode's projections.

src/share/vm/opto/library_call.cpp

@@ -63,7 +63,7 @@ class LibraryIntrinsic : public InlineCallGenerator {
   virtual bool is_virtual()   const { return _is_virtual; }
   virtual bool is_predicted()   const { return _is_predicted; }
   virtual bool does_virtual_dispatch()   const { return _does_virtual_dispatch; }
-  virtual JVMState* generate(JVMState* jvms);
+  virtual JVMState* generate(JVMState* jvms, Parse* parent_parser);
   virtual Node* generate_predicate(JVMState* jvms);
   vmIntrinsics::ID intrinsic_id() const { return _intrinsic_id; }
 };
@@ -556,7 +556,7 @@ void Compile::register_library_intrinsics() {
   // Nothing to do here.
 }
 
-JVMState* LibraryIntrinsic::generate(JVMState* jvms) {
+JVMState* LibraryIntrinsic::generate(JVMState* jvms, Parse* parent_parser) {
   LibraryCallKit kit(jvms, this);
   Compile* C = kit.C;
   int nodes = C->unique();

src/share/vm/opto/loopPredicate.cpp

@@ -41,63 +41,6 @@
  * checks (such as null checks).
 */
 
-//-------------------------------is_uncommon_trap_proj----------------------------
-// Return true if proj is the form of "proj->[region->..]call_uct"
-bool PhaseIdealLoop::is_uncommon_trap_proj(ProjNode* proj, Deoptimization::DeoptReason reason) {
-  int path_limit = 10;
-  assert(proj, "invalid argument");
-  Node* out = proj;
-  for (int ct = 0; ct < path_limit; ct++) {
-    out = out->unique_ctrl_out();
-    if (out == NULL)
-      return false;
-    if (out->is_CallStaticJava()) {
-      int req = out->as_CallStaticJava()->uncommon_trap_request();
-      if (req != 0) {
-        Deoptimization::DeoptReason trap_reason = Deoptimization::trap_request_reason(req);
-        if (trap_reason == reason || reason == Deoptimization::Reason_none) {
-           return true;
-        }
-      }
-      return false; // don't do further after call
-    }
-    if (out->Opcode() != Op_Region)
-      return false;
-  }
-  return false;
-}
-
-//-------------------------------is_uncommon_trap_if_pattern-------------------------
-// Return true  for "if(test)-> proj -> ...
-//                          |
-//                          V
-//                      other_proj->[region->..]call_uct"
-//
-// "must_reason_predicate" means the uct reason must be Reason_predicate
-bool PhaseIdealLoop::is_uncommon_trap_if_pattern(ProjNode *proj, Deoptimization::DeoptReason reason) {
-  Node *in0 = proj->in(0);
-  if (!in0->is_If()) return false;
-  // Variation of a dead If node.
-  if (in0->outcnt() < 2)  return false;
-  IfNode* iff = in0->as_If();
-
-  // we need "If(Conv2B(Opaque1(...)))" pattern for reason_predicate
-  if (reason != Deoptimization::Reason_none) {
-    if (iff->in(1)->Opcode() != Op_Conv2B ||
-       iff->in(1)->in(1)->Opcode() != Op_Opaque1) {
-      return false;
-    }
-  }
-
-  ProjNode* other_proj = iff->proj_out(1-proj->_con)->as_Proj();
-  if (is_uncommon_trap_proj(other_proj, reason)) {
-    assert(reason == Deoptimization::Reason_none ||
-           Compile::current()->is_predicate_opaq(iff->in(1)->in(1)), "should be on the list");
-    return true;
-  }
-  return false;
-}
-
 //-------------------------------register_control-------------------------
 void PhaseIdealLoop::register_control(Node* n, IdealLoopTree *loop, Node* pred) {
   assert(n->is_CFG(), "must be control node");
@@ -147,7 +90,7 @@ void PhaseIdealLoop::register_control(Node* n, IdealLoopTree *loop, Node* pred)
 // This code is also used to clone predicates to clonned loops.
 ProjNode* PhaseIdealLoop::create_new_if_for_predicate(ProjNode* cont_proj, Node* new_entry,
                                                       Deoptimization::DeoptReason reason) {
-  assert(is_uncommon_trap_if_pattern(cont_proj, reason), "must be a uct if pattern!");
+  assert(cont_proj->is_uncommon_trap_if_pattern(reason), "must be a uct if pattern!");
   IfNode* iff = cont_proj->in(0)->as_If();
 
   ProjNode *uncommon_proj = iff->proj_out(1 - cont_proj->_con);
@@ -235,7 +178,7 @@ ProjNode* PhaseIdealLoop::create_new_if_for_predicate(ProjNode* cont_proj, Node*
 ProjNode* PhaseIterGVN::create_new_if_for_predicate(ProjNode* cont_proj, Node* new_entry,
                                                     Deoptimization::DeoptReason reason) {
   assert(new_entry != 0, "only used for clone predicate");
-  assert(PhaseIdealLoop::is_uncommon_trap_if_pattern(cont_proj, reason), "must be a uct if pattern!");
+  assert(cont_proj->is_uncommon_trap_if_pattern(reason), "must be a uct if pattern!");
   IfNode* iff = cont_proj->in(0)->as_If();
 
   ProjNode *uncommon_proj = iff->proj_out(1 - cont_proj->_con);
@@ -422,7 +365,7 @@ Node* PhaseIdealLoop::skip_loop_predicates(Node* entry) {
 ProjNode* PhaseIdealLoop::find_predicate_insertion_point(Node* start_c, Deoptimization::DeoptReason reason) {
   if (start_c == NULL || !start_c->is_Proj())
     return NULL;
-  if (is_uncommon_trap_if_pattern(start_c->as_Proj(), reason)) {
+  if (start_c->as_Proj()->is_uncommon_trap_if_pattern(reason)) {
     return start_c->as_Proj();
   }
   return NULL;
@@ -773,7 +716,7 @@ bool PhaseIdealLoop::loop_predication_impl(IdealLoopTree *loop) {
     ProjNode* proj = if_proj_list.pop()->as_Proj();
     IfNode*   iff  = proj->in(0)->as_If();
 
-    if (!is_uncommon_trap_if_pattern(proj, Deoptimization::Reason_none)) {
+    if (!proj->is_uncommon_trap_if_pattern(Deoptimization::Reason_none)) {
       if (loop->is_loop_exit(iff)) {
         // stop processing the remaining projs in the list because the execution of them
         // depends on the condition of "iff" (iff->in(1)).

src/share/vm/opto/loopnode.cpp

@@ -167,7 +167,7 @@ Node *PhaseIdealLoop::get_early_ctrl_for_expensive(Node *n, Node* earliest) {
     // expensive nodes will notice the loop and skip over it to try to
     // move the node further up.
     if (ctl->is_CountedLoop() && ctl->in(1) != NULL && ctl->in(1)->in(0) != NULL && ctl->in(1)->in(0)->is_If()) {
-      if (!is_uncommon_trap_if_pattern(ctl->in(1)->as_Proj(), Deoptimization::Reason_none)) {
+      if (!ctl->in(1)->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none)) {
         break;
       }
       next = idom(ctl->in(1)->in(0));
@@ -181,7 +181,7 @@ Node *PhaseIdealLoop::get_early_ctrl_for_expensive(Node *n, Node* earliest) {
       } else if (parent_ctl->is_CountedLoopEnd() && parent_ctl->as_CountedLoopEnd()->loopnode() != NULL) {
         next = parent_ctl->as_CountedLoopEnd()->loopnode()->init_control();
       } else if (parent_ctl->is_If()) {
-        if (!is_uncommon_trap_if_pattern(ctl->as_Proj(), Deoptimization::Reason_none)) {
+        if (!ctl->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none)) {
           break;
         }
         assert(idom(ctl) == parent_ctl, "strange");

src/share/vm/opto/loopnode.hpp

@@ -876,13 +876,6 @@ public:
   // Return true if exp is a scaled induction var plus (or minus) constant
   bool is_scaled_iv_plus_offset(Node* exp, Node* iv, int* p_scale, Node** p_offset, int depth = 0);
 
-  // Return true if proj is for "proj->[region->..]call_uct"
-  static bool is_uncommon_trap_proj(ProjNode* proj, Deoptimization::DeoptReason reason);
-  // Return true for    "if(test)-> proj -> ...
-  //                          |
-  //                          V
-  //                      other_proj->[region->..]call_uct"
-  static bool is_uncommon_trap_if_pattern(ProjNode* proj, Deoptimization::DeoptReason reason);
   // Create a new if above the uncommon_trap_if_pattern for the predicate to be promoted
   ProjNode* create_new_if_for_predicate(ProjNode* cont_proj, Node* new_entry,
                                         Deoptimization::DeoptReason reason);

src/share/vm/opto/loopopts.cpp

@@ -238,7 +238,7 @@ void PhaseIdealLoop::dominated_by( Node *prevdom, Node *iff, bool flip, bool exc
   ProjNode* dp_proj  = dp->as_Proj();
   ProjNode* unc_proj = iff->as_If()->proj_out(1 - dp_proj->_con)->as_Proj();
   if (exclude_loop_predicate &&
-      is_uncommon_trap_proj(unc_proj, Deoptimization::Reason_predicate))
+      unc_proj->is_uncommon_trap_proj(Deoptimization::Reason_predicate))
     return; // Let IGVN transformation change control dependence.
 
   IdealLoopTree *old_loop = get_loop(dp);

src/share/vm/opto/multnode.cpp

@@ -24,6 +24,7 @@
 
 #include "precompiled.hpp"
 #include "opto/callnode.hpp"
+#include "opto/cfgnode.hpp"
 #include "opto/matcher.hpp"
 #include "opto/mathexactnode.hpp"
 #include "opto/multnode.hpp"
@@ -150,3 +151,59 @@ const RegMask &ProjNode::out_RegMask() const {
 uint ProjNode::ideal_reg() const {
   return bottom_type()->ideal_reg();
 }
+
+//-------------------------------is_uncommon_trap_proj----------------------------
+// Return true if proj is the form of "proj->[region->..]call_uct"
+bool ProjNode::is_uncommon_trap_proj(Deoptimization::DeoptReason reason) {
+  int path_limit = 10;
+  Node* out = this;
+  for (int ct = 0; ct < path_limit; ct++) {
+    out = out->unique_ctrl_out();
+    if (out == NULL)
+      return false;
+    if (out->is_CallStaticJava()) {
+      int req = out->as_CallStaticJava()->uncommon_trap_request();
+      if (req != 0) {
+        Deoptimization::DeoptReason trap_reason = Deoptimization::trap_request_reason(req);
+        if (trap_reason == reason || reason == Deoptimization::Reason_none) {
+           return true;
+        }
+      }
+      return false; // don't do further after call
+    }
+    if (out->Opcode() != Op_Region)
+      return false;
+  }
+  return false;
+}
+
+//-------------------------------is_uncommon_trap_if_pattern-------------------------
+// Return true  for "if(test)-> proj -> ...
+//                          |
+//                          V
+//                      other_proj->[region->..]call_uct"
+//
+// "must_reason_predicate" means the uct reason must be Reason_predicate
+bool ProjNode::is_uncommon_trap_if_pattern(Deoptimization::DeoptReason reason) {
+  Node *in0 = in(0);
+  if (!in0->is_If()) return false;
+  // Variation of a dead If node.
+  if (in0->outcnt() < 2)  return false;
+  IfNode* iff = in0->as_If();
+
+  // we need "If(Conv2B(Opaque1(...)))" pattern for reason_predicate
+  if (reason != Deoptimization::Reason_none) {
+    if (iff->in(1)->Opcode() != Op_Conv2B ||
+       iff->in(1)->in(1)->Opcode() != Op_Opaque1) {
+      return false;
+    }
+  }
+
+  ProjNode* other_proj = iff->proj_out(1-_con)->as_Proj();
+  if (other_proj->is_uncommon_trap_proj(reason)) {
+    assert(reason == Deoptimization::Reason_none ||
+           Compile::current()->is_predicate_opaq(iff->in(1)->in(1)), "should be on the list");
+    return true;
+  }
+  return false;
+}

src/share/vm/opto/multnode.hpp

@@ -88,6 +88,14 @@ public:
 #ifndef PRODUCT
   virtual void dump_spec(outputStream *st) const;
 #endif
+
+  // Return true if proj is for "proj->[region->..]call_uct"
+  bool is_uncommon_trap_proj(Deoptimization::DeoptReason reason);
+  // Return true for    "if(test)-> proj -> ...
+  //                          |
+  //                          V
+  //                      other_proj->[region->..]call_uct"
+  bool is_uncommon_trap_if_pattern(Deoptimization::DeoptReason reason);
 };
 
 #endif // SHARE_VM_OPTO_MULTNODE_HPP

src/share/vm/opto/parse.hpp

@@ -349,13 +349,15 @@ class Parse : public GraphKit {
   int _est_switch_depth;        // Debugging SwitchRanges.
 #endif
 
+  // parser for the caller of the method of this object
+  Parse* const _parent;
+
  public:
   // Constructor
-  Parse(JVMState* caller, ciMethod* parse_method, float expected_uses);
+  Parse(JVMState* caller, ciMethod* parse_method, float expected_uses, Parse* parent);
 
   virtual Parse* is_Parse() const { return (Parse*)this; }
 
- public:
   // Accessors.
   JVMState*     caller()        const { return _caller; }
   float         expected_uses() const { return _expected_uses; }
@@ -407,6 +409,8 @@ class Parse : public GraphKit {
     return block()->successor_for_bci(bci);
   }
 
+  Parse* parent_parser() const { return _parent; }
+
  private:
   // Create a JVMS & map for the initial state of this method.
   SafePointNode* create_entry_map();

src/share/vm/opto/parse1.cpp

@@ -381,8 +381,8 @@ void Parse::load_interpreter_state(Node* osr_buf) {
 
 //------------------------------Parse------------------------------------------
 // Main parser constructor.
-Parse::Parse(JVMState* caller, ciMethod* parse_method, float expected_uses)
-  : _exits(caller)
+Parse::Parse(JVMState* caller, ciMethod* parse_method, float expected_uses, Parse* parent)
+  : _exits(caller), _parent(parent)
 {
   // Init some variables
   _caller = caller;