8005071: Incremental inlining for JSR 292

Summary: post parse inlining driven by number of live nodes.
Reviewed-by: twisti, kvn, jrose

src/share/vm/opto/bytecodeInfo.cpp

@@ -46,7 +46,8 @@ InlineTree::InlineTree(Compile* c,
   _method(callee),
   _site_invoke_ratio(site_invoke_ratio),
   _max_inline_level(max_inline_level),
-  _count_inline_bcs(method()->code_size_for_inlining())
+  _count_inline_bcs(method()->code_size_for_inlining()),
+  _subtrees(c->comp_arena(), 2, 0, NULL)
 {
   NOT_PRODUCT(_count_inlines = 0;)
   if (_caller_jvms != NULL) {
@@ -209,17 +210,19 @@ const char* InlineTree::should_not_inline(ciMethod *callee_method, ciMethod* cal
   if ( callee_method->dont_inline())                        return "don't inline by annotation";
   if ( callee_method->has_unloaded_classes_in_signature())  return "unloaded signature classes";
 
-  if (callee_method->force_inline() || callee_method->should_inline()) {
+  if (callee_method->should_inline()) {
     // ignore heuristic controls on inlining
     return NULL;
   }
 
   // Now perform checks which are heuristic
 
+  if (!callee_method->force_inline()) {
     if (callee_method->has_compiled_code() &&
         callee_method->instructions_size() > InlineSmallCode) {
     return "already compiled into a big method";
     }
+  }
 
   // don't inline exception code unless the top method belongs to an
   // exception class
@@ -277,12 +280,15 @@ const char* InlineTree::should_not_inline(ciMethod *callee_method, ciMethod* cal
 //-----------------------------try_to_inline-----------------------------------
 // return NULL if ok, reason for not inlining otherwise
 // Relocated from "InliningClosure::try_to_inline"
-const char* InlineTree::try_to_inline(ciMethod* callee_method, ciMethod* caller_method, int caller_bci, ciCallProfile& profile, WarmCallInfo* wci_result) {
-
+const char* InlineTree::try_to_inline(ciMethod* callee_method, ciMethod* caller_method, int caller_bci, ciCallProfile& profile, WarmCallInfo* wci_result, bool& should_delay) {
   // Old algorithm had funny accumulating BC-size counters
   if (UseOldInlining && ClipInlining
       && (int)count_inline_bcs() >= DesiredMethodLimit) {
+    if (!callee_method->force_inline() || !IncrementalInline) {
       return "size > DesiredMethodLimit";
+    } else if (!C->inlining_incrementally()) {
+      should_delay = true;
+    }
   }
 
   const char *msg = NULL;
@@ -303,8 +309,13 @@ const char* InlineTree::try_to_inline(ciMethod* callee_method, ciMethod* caller_
   if (callee_method->code_size() > MaxTrivialSize) {
 
     // don't inline into giant methods
-    if (C->unique() > (uint)NodeCountInliningCutoff) {
+    if (C->over_inlining_cutoff()) {
+      if ((!callee_method->force_inline() && !caller_method->is_compiled_lambda_form())
+          || !IncrementalInline) {
         return "NodeCountInliningCutoff";
+      } else {
+        should_delay = true;
+      }
     }
 
     if ((!UseInterpreter || CompileTheWorld) &&
@@ -323,7 +334,11 @@ const char* InlineTree::try_to_inline(ciMethod* callee_method, ciMethod* caller_
     return "not an accessor";
   }
   if (inline_level() > _max_inline_level) {
+    if (!callee_method->force_inline() || !IncrementalInline) {
       return "inlining too deep";
+    } else if (!C->inlining_incrementally()) {
+      should_delay = true;
+    }
   }
 
   // detect direct and indirect recursive inlining
@@ -348,7 +363,11 @@ const char* InlineTree::try_to_inline(ciMethod* callee_method, ciMethod* caller_
 
   if (UseOldInlining && ClipInlining
       && (int)count_inline_bcs() + size >= DesiredMethodLimit) {
+    if (!callee_method->force_inline() || !IncrementalInline) {
       return "size > DesiredMethodLimit";
+    } else if (!C->inlining_incrementally()) {
+      should_delay = true;
+    }
   }
 
   // ok, inline this method
@@ -413,8 +432,9 @@ void InlineTree::print_inlining(ciMethod* callee_method, int caller_bci, const c
 }
 
 //------------------------------ok_to_inline-----------------------------------
-WarmCallInfo* InlineTree::ok_to_inline(ciMethod* callee_method, JVMState* jvms, ciCallProfile& profile, WarmCallInfo* initial_wci) {
+WarmCallInfo* InlineTree::ok_to_inline(ciMethod* callee_method, JVMState* jvms, ciCallProfile& profile, WarmCallInfo* initial_wci, bool& should_delay) {
   assert(callee_method != NULL, "caller checks for optimized virtual!");
+  assert(!should_delay, "should be initialized to false");
 #ifdef ASSERT
   // Make sure the incoming jvms has the same information content as me.
   // This means that we can eventually make this whole class AllStatic.
@@ -444,7 +464,7 @@ WarmCallInfo* InlineTree::ok_to_inline(ciMethod* callee_method, JVMState* jvms,
 
   // Check if inlining policy says no.
   WarmCallInfo wci = *(initial_wci);
-  failure_msg = try_to_inline(callee_method, caller_method, caller_bci, profile, &wci);
+  failure_msg = try_to_inline(callee_method, caller_method, caller_bci, profile, &wci, should_delay);
   if (failure_msg != NULL && C->log() != NULL) {
     C->log()->inline_fail(failure_msg);
   }

src/share/vm/opto/c2_globals.hpp

@@ -606,6 +606,16 @@
                                                                             \
   develop(bool, VerifyAliases, false,                                       \
           "perform extra checks on the results of alias analysis")          \
+                                                                            \
+  product(bool, IncrementalInline, true,                                    \
+          "do post parse inlining")                                         \
+                                                                            \
+  develop(bool, AlwaysIncrementalInline, false,                             \
+          "do all inlining incrementally")                                  \
+                                                                            \
+  product(intx, LiveNodeCountInliningCutoff, 20000,                         \
+          "max number of live nodes in a method")                           \
+
 
 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

@@ -262,8 +262,11 @@ CallGenerator* CallGenerator::for_virtual_call(ciMethod* m, int vtable_index) {
 
 // Allow inlining decisions to be delayed
 class LateInlineCallGenerator : public DirectCallGenerator {
+ protected:
   CallGenerator* _inline_cg;
 
+  virtual bool do_late_inline_check(JVMState* jvms) { return true; }
+
  public:
   LateInlineCallGenerator(ciMethod* method, CallGenerator* inline_cg) :
     DirectCallGenerator(method, true), _inline_cg(inline_cg) {}
@@ -279,7 +282,9 @@ class LateInlineCallGenerator : public DirectCallGenerator {
 
     // Record that this call site should be revisited once the main
     // parse is finished.
-    Compile::current()->add_late_inline(this);
+    if (!is_mh_late_inline()) {
+      C->add_late_inline(this);
+    }
 
     // Emit the CallStaticJava and request separate projections so
     // that the late inlining logic can distinguish between fall
@@ -287,8 +292,15 @@ class LateInlineCallGenerator : public DirectCallGenerator {
     // as is done for allocations and macro expansion.
     return DirectCallGenerator::generate(jvms);
   }
-};
 
+  virtual void print_inlining_late(const char* msg) {
+    CallNode* call = call_node();
+    Compile* C = Compile::current();
+    C->print_inlining_insert(this);
+    C->print_inlining(method(), call->jvms()->depth()-1, call->jvms()->bci(), msg);
+  }
+
+};
 
 void LateInlineCallGenerator::do_late_inline() {
   // Can't inline it
@@ -296,6 +308,18 @@ void LateInlineCallGenerator::do_late_inline() {
       call_node()->in(0) == NULL || call_node()->in(0)->is_top())
     return;
 
+  for (int i1 = 0; i1 < method()->arg_size(); i1++) {
+    if (call_node()->in(TypeFunc::Parms + i1)->is_top()) {
+      assert(Compile::current()->inlining_incrementally(), "shouldn't happen during parsing");
+      return;
+    }
+  }
+
+  if (call_node()->in(TypeFunc::Memory)->is_top()) {
+    assert(Compile::current()->inlining_incrementally(), "shouldn't happen during parsing");
+    return;
+  }
+
   CallStaticJavaNode* call = call_node();
 
   // Make a clone of the JVMState that appropriate to use for driving a parse
@@ -324,6 +348,11 @@ void LateInlineCallGenerator::do_late_inline() {
     }
   }
 
+  if (!do_late_inline_check(jvms)) {
+    map->disconnect_inputs(NULL, C);
+    return;
+  }
+
   C->print_inlining_insert(this);
 
   CompileLog* log = C->log();
@@ -360,6 +389,10 @@ void LateInlineCallGenerator::do_late_inline() {
     result = (result_size == 1) ? kit.pop() : kit.pop_pair();
   }
 
+  C->set_has_loops(C->has_loops() || _inline_cg->method()->has_loops());
+  C->env()->notice_inlined_method(_inline_cg->method());
+  C->set_inlining_progress(true);
+
   kit.replace_call(call, result);
 }
 
@@ -368,6 +401,83 @@ CallGenerator* CallGenerator::for_late_inline(ciMethod* method, CallGenerator* i
   return new LateInlineCallGenerator(method, inline_cg);
 }
 
+class LateInlineMHCallGenerator : public LateInlineCallGenerator {
+  ciMethod* _caller;
+  int _attempt;
+  bool _input_not_const;
+
+  virtual bool do_late_inline_check(JVMState* jvms);
+  virtual bool already_attempted() const { return _attempt > 0; }
+
+ public:
+  LateInlineMHCallGenerator(ciMethod* caller, ciMethod* callee, bool input_not_const) :
+    LateInlineCallGenerator(callee, NULL), _caller(caller), _attempt(0), _input_not_const(input_not_const) {}
+
+  virtual bool is_mh_late_inline() const { return true; }
+
+  virtual JVMState* generate(JVMState* jvms) {
+    JVMState* new_jvms = LateInlineCallGenerator::generate(jvms);
+    if (_input_not_const) {
+      // inlining won't be possible so no need to enqueue right now.
+      call_node()->set_generator(this);
+    } else {
+      Compile::current()->add_late_inline(this);
+    }
+    return new_jvms;
+  }
+
+  virtual void print_inlining_late(const char* msg) {
+    if (!_input_not_const) return;
+    LateInlineCallGenerator::print_inlining_late(msg);
+  }
+};
+
+bool LateInlineMHCallGenerator::do_late_inline_check(JVMState* jvms) {
+
+  CallGenerator* cg = for_method_handle_inline(jvms, _caller, method(), _input_not_const);
+
+  if (!_input_not_const) {
+    _attempt++;
+  }
+
+  if (cg != NULL) {
+    assert(!cg->is_late_inline() && cg->is_inline(), "we're doing late inlining");
+    _inline_cg = cg;
+    Compile::current()->dec_number_of_mh_late_inlines();
+    return true;
+  }
+
+  call_node()->set_generator(this);
+  return false;
+}
+
+CallGenerator* CallGenerator::for_mh_late_inline(ciMethod* caller, ciMethod* callee, bool input_not_const) {
+  Compile::current()->inc_number_of_mh_late_inlines();
+  CallGenerator* cg = new LateInlineMHCallGenerator(caller, callee, input_not_const);
+  return cg;
+}
+
+class LateInlineStringCallGenerator : public LateInlineCallGenerator {
+
+ public:
+  LateInlineStringCallGenerator(ciMethod* method, CallGenerator* inline_cg) :
+    LateInlineCallGenerator(method, inline_cg) {}
+
+  virtual JVMState* generate(JVMState* jvms) {
+    Compile *C = Compile::current();
+    C->print_inlining_skip(this);
+
+    C->add_string_late_inline(this);
+
+    JVMState* new_jvms =  DirectCallGenerator::generate(jvms);
+    return new_jvms;
+  }
+};
+
+CallGenerator* CallGenerator::for_string_late_inline(ciMethod* method, CallGenerator* inline_cg) {
+  return new LateInlineStringCallGenerator(method, inline_cg);
+}
+
 
 //---------------------------WarmCallGenerator--------------------------------
 // Internal class which handles initial deferral of inlining decisions.
@@ -586,35 +696,52 @@ JVMState* PredictedCallGenerator::generate(JVMState* jvms) {
 }
 
 
-CallGenerator* CallGenerator::for_method_handle_call(JVMState* jvms, ciMethod* caller, ciMethod* callee) {
+CallGenerator* CallGenerator::for_method_handle_call(JVMState* jvms, ciMethod* caller, ciMethod* callee, bool delayed_forbidden) {
   assert(callee->is_method_handle_intrinsic() ||
          callee->is_compiled_lambda_form(), "for_method_handle_call mismatch");
-  CallGenerator* cg = CallGenerator::for_method_handle_inline(jvms, caller, callee);
-  if (cg != NULL)
+  bool input_not_const;
+  CallGenerator* cg = CallGenerator::for_method_handle_inline(jvms, caller, callee, input_not_const);
+  Compile* C = Compile::current();
+  if (cg != NULL) {
+    if (!delayed_forbidden && AlwaysIncrementalInline) {
+      return CallGenerator::for_late_inline(callee, cg);
+    } else {
       return cg;
+    }
+  }
+  int bci = jvms->bci();
+  ciCallProfile profile = caller->call_profile_at_bci(bci);
+  int call_site_count = caller->scale_count(profile.count());
+
+  if (IncrementalInline && call_site_count > 0 &&
+      (input_not_const || !C->inlining_incrementally() || C->over_inlining_cutoff())) {
+    return CallGenerator::for_mh_late_inline(caller, callee, input_not_const);
+  } else {
     return CallGenerator::for_direct_call(callee);
+  }
 }
 
-CallGenerator* CallGenerator::for_method_handle_inline(JVMState* jvms, ciMethod* caller, ciMethod* callee) {
+CallGenerator* CallGenerator::for_method_handle_inline(JVMState* jvms, ciMethod* caller, ciMethod* callee, bool& input_not_const) {
   GraphKit kit(jvms);
   PhaseGVN& gvn = kit.gvn();
   Compile* C = kit.C;
   vmIntrinsics::ID iid = callee->intrinsic_id();
+  input_not_const = true;
   switch (iid) {
   case vmIntrinsics::_invokeBasic:
     {
       // Get MethodHandle receiver:
       Node* receiver = kit.argument(0);
       if (receiver->Opcode() == Op_ConP) {
+        input_not_const = false;
         const TypeOopPtr* oop_ptr = receiver->bottom_type()->is_oopptr();
         ciMethod* target = oop_ptr->const_oop()->as_method_handle()->get_vmtarget();
         guarantee(!target->is_method_handle_intrinsic(), "should not happen");  // XXX remove
         const int vtable_index = Method::invalid_vtable_index;
-        CallGenerator* cg = C->call_generator(target, vtable_index, false, jvms, true, PROB_ALWAYS);
+        CallGenerator* cg = C->call_generator(target, vtable_index, false, jvms, true, PROB_ALWAYS, true, true);
+        assert (!cg->is_late_inline() || cg->is_mh_late_inline(), "no late inline here");
         if (cg != NULL && cg->is_inline())
           return cg;
-      } else {
-        if (PrintInlining)  C->print_inlining(callee, jvms->depth() - 1, jvms->bci(), "receiver not constant");
       }
     }
     break;
@@ -627,6 +754,7 @@ CallGenerator* CallGenerator::for_method_handle_inline(JVMState* jvms, ciMethod*
       // Get MemberName argument:
       Node* member_name = kit.argument(callee->arg_size() - 1);
       if (member_name->Opcode() == Op_ConP) {
+        input_not_const = false;
         const TypeOopPtr* oop_ptr = member_name->bottom_type()->is_oopptr();
         ciMethod* target = oop_ptr->const_oop()->as_member_name()->get_vmtarget();
 
@@ -661,7 +789,8 @@ CallGenerator* CallGenerator::for_method_handle_inline(JVMState* jvms, ciMethod*
         }
         const int vtable_index = Method::invalid_vtable_index;
         const bool call_is_virtual = target->is_abstract();  // FIXME workaround
-        CallGenerator* cg = C->call_generator(target, vtable_index, call_is_virtual, jvms, true, PROB_ALWAYS);
+        CallGenerator* cg = C->call_generator(target, vtable_index, call_is_virtual, jvms, true, PROB_ALWAYS, true, true);
+        assert (!cg->is_late_inline() || cg->is_mh_late_inline(), "no late inline here");
         if (cg != NULL && cg->is_inline())
           return cg;
       }

src/share/vm/opto/callGenerator.hpp

@@ -68,6 +68,12 @@ class CallGenerator : public ResourceObj {
 
   // is_late_inline: supports conversion of call into an inline
   virtual bool      is_late_inline() const      { return false; }
+  // same but for method handle calls
+  virtual bool      is_mh_late_inline() const   { return false; }
+
+  // for method handle calls: have we tried inlinining the call already?
+  virtual bool      already_attempted() const   { ShouldNotReachHere(); return false; }
+
   // Replace the call with an inline version of the code
   virtual void do_late_inline() { ShouldNotReachHere(); }
 
@@ -112,11 +118,13 @@ class CallGenerator : public ResourceObj {
   static CallGenerator* for_virtual_call(ciMethod* m, int vtable_index);  // virtual, interface
   static CallGenerator* for_dynamic_call(ciMethod* m);   // invokedynamic
 
-  static CallGenerator* for_method_handle_call(  JVMState* jvms, ciMethod* caller, ciMethod* callee);
-  static CallGenerator* for_method_handle_inline(JVMState* jvms, ciMethod* caller, ciMethod* callee);
+  static CallGenerator* for_method_handle_call(  JVMState* jvms, ciMethod* caller, ciMethod* callee, bool delayed_forbidden);
+  static CallGenerator* for_method_handle_inline(JVMState* jvms, ciMethod* caller, ciMethod* callee, bool& input_not_const);
 
   // How to generate a replace a direct call with an inline version
   static CallGenerator* for_late_inline(ciMethod* m, CallGenerator* inline_cg);
+  static CallGenerator* for_mh_late_inline(ciMethod* caller, ciMethod* callee, bool input_not_const);
+  static CallGenerator* for_string_late_inline(ciMethod* m, CallGenerator* inline_cg);
 
   // How to make a call but defer the decision whether to inline or not.
   static CallGenerator* for_warm_call(WarmCallInfo* ci,
@@ -147,6 +155,8 @@ class CallGenerator : public ResourceObj {
                                                 CallGenerator* cg);
   virtual Node* generate_predicate(JVMState* jvms) { return NULL; };
 
+  virtual void print_inlining_late(const char* msg) { ShouldNotReachHere(); }
+
   static void print_inlining(Compile* C, ciMethod* callee, int inline_level, int bci, const char* msg) {
     if (PrintInlining)
       C->print_inlining(callee, inline_level, bci, msg);

src/share/vm/opto/callnode.cpp

@@ -25,6 +25,7 @@
 #include "precompiled.hpp"
 #include "ci/bcEscapeAnalyzer.hpp"
 #include "compiler/oopMap.hpp"
+#include "opto/callGenerator.hpp"
 #include "opto/callnode.hpp"
 #include "opto/escape.hpp"
 #include "opto/locknode.hpp"
@@ -775,16 +776,38 @@ void CallNode::extract_projections(CallProjections* projs, bool separate_io_proj
   // and the exception object may not exist if an exception handler
   // swallows the exception but all the other must exist and be found.
   assert(projs->fallthrough_proj      != NULL, "must be found");
-  assert(projs->fallthrough_catchproj != NULL, "must be found");
-  assert(projs->fallthrough_memproj   != NULL, "must be found");
-  assert(projs->fallthrough_ioproj    != NULL, "must be found");
-  assert(projs->catchall_catchproj    != NULL, "must be found");
+  assert(Compile::current()->inlining_incrementally() || projs->fallthrough_catchproj != NULL, "must be found");
+  assert(Compile::current()->inlining_incrementally() || projs->fallthrough_memproj   != NULL, "must be found");
+  assert(Compile::current()->inlining_incrementally() || projs->fallthrough_ioproj    != NULL, "must be found");
+  assert(Compile::current()->inlining_incrementally() || projs->catchall_catchproj    != NULL, "must be found");
   if (separate_io_proj) {
-    assert(projs->catchall_memproj      != NULL, "must be found");
-    assert(projs->catchall_ioproj       != NULL, "must be found");
+    assert(Compile::current()->inlining_incrementally() || projs->catchall_memproj    != NULL, "must be found");
+    assert(Compile::current()->inlining_incrementally() || projs->catchall_ioproj     != NULL, "must be found");
   }
 }
 
+Node *CallNode::Ideal(PhaseGVN *phase, bool can_reshape) {
+  CallGenerator* cg = generator();
+  if (can_reshape && cg != NULL && cg->is_mh_late_inline() && !cg->already_attempted()) {
+    // Check whether this MH handle call becomes a candidate for inlining
+    ciMethod* callee = cg->method();
+    vmIntrinsics::ID iid = callee->intrinsic_id();
+    if (iid == vmIntrinsics::_invokeBasic) {
+      if (in(TypeFunc::Parms)->Opcode() == Op_ConP) {
+        phase->C->prepend_late_inline(cg);
+        set_generator(NULL);
+      }
+    } else {
+      assert(callee->has_member_arg(), "wrong type of call?");
+      if (in(TypeFunc::Parms + callee->arg_size() - 1)->Opcode() == Op_ConP) {
+        phase->C->prepend_late_inline(cg);
+        set_generator(NULL);
+      }
+    }
+  }
+  return SafePointNode::Ideal(phase, can_reshape);
+}
+
 
 //=============================================================================
 uint CallJavaNode::size_of() const { return sizeof(*this); }

src/share/vm/opto/callnode.hpp

@@ -507,6 +507,7 @@ public:
   Node* exobj;
 };
 
+class CallGenerator;
 
 //------------------------------CallNode---------------------------------------
 // Call nodes now subsume the function of debug nodes at callsites, so they
@@ -517,12 +518,14 @@ public:
   const TypeFunc *_tf;        // Function type
   address      _entry_point;  // Address of method being called
   float        _cnt;          // Estimate of number of times called
+  CallGenerator* _generator;  // corresponding CallGenerator for some late inline calls
 
   CallNode(const TypeFunc* tf, address addr, const TypePtr* adr_type)
     : SafePointNode(tf->domain()->cnt(), NULL, adr_type),
       _tf(tf),
       _entry_point(addr),
-      _cnt(COUNT_UNKNOWN)
+      _cnt(COUNT_UNKNOWN),
+      _generator(NULL)
   {
     init_class_id(Class_Call);
   }
@@ -530,13 +533,16 @@ public:
   const TypeFunc* tf()         const { return _tf; }
   const address  entry_point() const { return _entry_point; }
   const float    cnt()         const { return _cnt; }
+  CallGenerator* generator()   const { return _generator; }
 
   void set_tf(const TypeFunc* tf)       { _tf = tf; }
   void set_entry_point(address p)       { _entry_point = p; }
   void set_cnt(float c)                 { _cnt = c; }
+  void set_generator(CallGenerator* cg) { _generator = cg; }
 
   virtual const Type *bottom_type() const;
   virtual const Type *Value( PhaseTransform *phase ) const;
+  virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
   virtual Node *Identity( PhaseTransform *phase ) { return this; }
   virtual uint        cmp( const Node &n ) const;
   virtual uint        size_of() const = 0;

src/share/vm/opto/cfgnode.cpp

@@ -363,6 +363,49 @@ bool RegionNode::is_unreachable_region(PhaseGVN *phase) const {
   return true; // The Region node is unreachable - it is dead.
 }
 
+bool RegionNode::try_clean_mem_phi(PhaseGVN *phase) {
+  // Incremental inlining + PhaseStringOpts sometimes produce:
+  //
+  // cmpP with 1 top input
+  //           |
+  //          If
+  //         /  \
+  //   IfFalse  IfTrue  /- Some Node
+  //         \  /      /    /
+  //        Region    / /-MergeMem
+  //             \---Phi
+  //
+  //
+  // It's expected by PhaseStringOpts that the Region goes away and is
+  // replaced by If's control input but because there's still a Phi,
+  // the Region stays in the graph. The top input from the cmpP is
+  // propagated forward and a subgraph that is useful goes away. The
+  // code below replaces the Phi with the MergeMem so that the Region
+  // is simplified.
+
+  PhiNode* phi = has_unique_phi();
+  if (phi && phi->type() == Type::MEMORY && req() == 3 && phi->is_diamond_phi(true)) {
+    MergeMemNode* m = NULL;
+    assert(phi->req() == 3, "same as region");
+    for (uint i = 1; i < 3; ++i) {
+      Node *mem = phi->in(i);
+      if (mem && mem->is_MergeMem() && in(i)->outcnt() == 1) {
+        // Nothing is control-dependent on path #i except the region itself.
+        m = mem->as_MergeMem();
+        uint j = 3 - i;
+        Node* other = phi->in(j);
+        if (other && other == m->base_memory()) {
+          // m is a successor memory to other, and is not pinned inside the diamond, so push it out.
+          // This will allow the diamond to collapse completely.
+          phase->is_IterGVN()->replace_node(phi, m);
+          return true;
+        }
+      }
+    }
+  }
+  return false;
+}
+
 //------------------------------Ideal------------------------------------------
 // Return a node which is more "ideal" than the current node.  Must preserve
 // the CFG, but we can still strip out dead paths.
@@ -375,6 +418,10 @@ Node *RegionNode::Ideal(PhaseGVN *phase, bool can_reshape) {
   bool has_phis = false;
   if (can_reshape) {            // Need DU info to check for Phi users
     has_phis = (has_phi() != NULL);       // Cache result
+    if (has_phis && try_clean_mem_phi(phase)) {
+      has_phis = false;
+    }
+
     if (!has_phis) {            // No Phi users?  Nothing merging?
       for (uint i = 1; i < req()-1; i++) {
         Node *if1 = in(i);
@@ -1005,7 +1052,9 @@ const Type *PhiNode::Value( PhaseTransform *phase ) const {
 //------------------------------is_diamond_phi---------------------------------
 // Does this Phi represent a simple well-shaped diamond merge?  Return the
 // index of the true path or 0 otherwise.
-int PhiNode::is_diamond_phi() const {
+// If check_control_only is true, do not inspect the If node at the
+// top, and return -1 (not an edge number) on success.
+int PhiNode::is_diamond_phi(bool check_control_only) const {
   // Check for a 2-path merge
   Node *region = in(0);
   if( !region ) return 0;
@@ -1018,6 +1067,7 @@ int PhiNode::is_diamond_phi() const {
   Node *iff = ifp1->in(0);
   if( !iff || !iff->is_If() ) return 0;
   if( iff != ifp2->in(0) ) return 0;
+  if (check_control_only)  return -1;
   // Check for a proper bool/cmp
   const Node *b = iff->in(1);
   if( !b->is_Bool() ) return 0;

src/share/vm/opto/cfgnode.hpp

@@ -95,6 +95,7 @@ public:
   virtual Node *Identity( PhaseTransform *phase );
   virtual Node *Ideal(PhaseGVN *phase, bool can_reshape);
   virtual const RegMask &out_RegMask() const;
+  bool try_clean_mem_phi(PhaseGVN *phase);
 };
 
 //------------------------------JProjNode--------------------------------------
@@ -181,7 +182,7 @@ public:
   LoopSafety simple_data_loop_check(Node *in) const;
   // Is it unsafe data loop? It becomes a dead loop if this phi node removed.
   bool is_unsafe_data_reference(Node *in) const;
-  int  is_diamond_phi() const;
+  int  is_diamond_phi(bool check_control_only = false) const;
   virtual int Opcode() const;
   virtual bool pinned() const { return in(0) != 0; }
   virtual const TypePtr *adr_type() const { verify_adr_type(true); return _adr_type; }

src/share/vm/opto/compile.cpp

@@ -136,7 +136,7 @@ int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual) {
 
 void Compile::register_intrinsic(CallGenerator* cg) {
   if (_intrinsics == NULL) {
-    _intrinsics = new GrowableArray<CallGenerator*>(60);
+    _intrinsics = new (comp_arena())GrowableArray<CallGenerator*>(comp_arena(), 60, 0, NULL);
   }
   // This code is stolen from ciObjectFactory::insert.
   // Really, GrowableArray should have methods for
@@ -365,6 +365,21 @@ void Compile::update_dead_node_list(Unique_Node_List &useful) {
   }
 }
 
+void Compile::remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful) {
+  int shift = 0;
+  for (int i = 0; i < inlines->length(); i++) {
+    CallGenerator* cg = inlines->at(i);
+    CallNode* call = cg->call_node();
+    if (shift > 0) {
+      inlines->at_put(i-shift, cg);
+    }
+    if (!useful.member(call)) {
+      shift++;
+    }
+  }
+  inlines->trunc_to(inlines->length()-shift);
+}
+
 // Disconnect all useless nodes by disconnecting those at the boundary.
 void Compile::remove_useless_nodes(Unique_Node_List &useful) {
   uint next = 0;
@@ -394,6 +409,9 @@ void Compile::remove_useless_nodes(Unique_Node_List &useful) {
       remove_macro_node(n);
     }
   }
+  // clean up the late inline lists
+  remove_useless_late_inlines(&_string_late_inlines, useful);
+  remove_useless_late_inlines(&_late_inlines, useful);
   debug_only(verify_graph_edges(true/*check for no_dead_code*/);)
 }
 
@@ -611,6 +629,12 @@ Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr
                   _printer(IdealGraphPrinter::printer()),
 #endif
                   _congraph(NULL),
+                  _late_inlines(comp_arena(), 2, 0, NULL),
+                  _string_late_inlines(comp_arena(), 2, 0, NULL),
+                  _late_inlines_pos(0),
+                  _number_of_mh_late_inlines(0),
+                  _inlining_progress(false),
+                  _inlining_incrementally(false),
                   _print_inlining_list(NULL),
                   _print_inlining(0) {
   C = this;
@@ -737,29 +761,13 @@ Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr
       rethrow_exceptions(kit.transfer_exceptions_into_jvms());
     }
 
-    if (!failing() && has_stringbuilder()) {
-      {
-        // remove useless nodes to make the usage analysis simpler
-        ResourceMark rm;
-        PhaseRemoveUseless pru(initial_gvn(), &for_igvn);
-      }
+    assert(IncrementalInline || (_late_inlines.length() == 0 && !has_mh_late_inlines()), "incremental inlining is off");
 
-      {
-        ResourceMark rm;
-        print_method("Before StringOpts", 3);
-        PhaseStringOpts pso(initial_gvn(), &for_igvn);
-        print_method("After StringOpts", 3);
+    if (_late_inlines.length() == 0 && !has_mh_late_inlines() && !failing() && has_stringbuilder()) {
+      inline_string_calls(true);
     }
 
-      // now inline anything that we skipped the first time around
-      while (_late_inlines.length() > 0) {
-        CallGenerator* cg = _late_inlines.pop();
-        cg->do_late_inline();
     if (failing())  return;
-      }
-    }
-    assert(_late_inlines.length() == 0, "should have been processed");
-    dump_inlining();
 
     print_method("Before RemoveUseless", 3);
 
@@ -906,6 +914,9 @@ Compile::Compile( ciEnv* ci_env,
     _dead_node_list(comp_arena()),
     _dead_node_count(0),
     _congraph(NULL),
+    _number_of_mh_late_inlines(0),
+    _inlining_progress(false),
+    _inlining_incrementally(false),
     _print_inlining_list(NULL),
     _print_inlining(0) {
   C = this;
@@ -1760,6 +1771,124 @@ void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) {
   assert(predicate_count()==0, "should be clean!");
 }
 
+// StringOpts and late inlining of string methods
+void Compile::inline_string_calls(bool parse_time) {
+  {
+    // remove useless nodes to make the usage analysis simpler
+    ResourceMark rm;
+    PhaseRemoveUseless pru(initial_gvn(), for_igvn());
+  }
+
+  {
+    ResourceMark rm;
+    print_method("Before StringOpts", 3);
+    PhaseStringOpts pso(initial_gvn(), for_igvn());
+    print_method("After StringOpts", 3);
+  }
+
+  // now inline anything that we skipped the first time around
+  if (!parse_time) {
+    _late_inlines_pos = _late_inlines.length();
+  }
+
+  while (_string_late_inlines.length() > 0) {
+    CallGenerator* cg = _string_late_inlines.pop();
+    cg->do_late_inline();
+    if (failing())  return;
+  }
+  _string_late_inlines.trunc_to(0);
+}
+
+void Compile::inline_incrementally_one(PhaseIterGVN& igvn) {
+  assert(IncrementalInline, "incremental inlining should be on");
+  PhaseGVN* gvn = initial_gvn();
+
+  set_inlining_progress(false);
+  for_igvn()->clear();
+  gvn->replace_with(&igvn);
+
+  int i = 0;
+
+  for (; i <_late_inlines.length() && !inlining_progress(); i++) {
+    CallGenerator* cg = _late_inlines.at(i);
+    _late_inlines_pos = i+1;
+    cg->do_late_inline();
+    if (failing())  return;
+  }
+  int j = 0;
+  for (; i < _late_inlines.length(); i++, j++) {
+    _late_inlines.at_put(j, _late_inlines.at(i));
+  }
+  _late_inlines.trunc_to(j);
+
+  {
+    ResourceMark rm;
+    PhaseRemoveUseless pru(C->initial_gvn(), C->for_igvn());
+  }
+
+  igvn = PhaseIterGVN(gvn);
+}
+
+// Perform incremental inlining until bound on number of live nodes is reached
+void Compile::inline_incrementally(PhaseIterGVN& igvn) {
+  PhaseGVN* gvn = initial_gvn();
+
+  set_inlining_incrementally(true);
+  set_inlining_progress(true);
+  uint low_live_nodes = 0;
+
+  while(inlining_progress() && _late_inlines.length() > 0) {
+
+    if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
+      if (low_live_nodes < (uint)LiveNodeCountInliningCutoff * 8 / 10) {
+        // PhaseIdealLoop is expensive so we only try it once we are
+        // out of loop and we only try it again if the previous helped
+        // got the number of nodes down significantly
+        PhaseIdealLoop ideal_loop( igvn, false, true );
+        if (failing())  return;
+        low_live_nodes = live_nodes();
+        _major_progress = true;
+      }
+
+      if (live_nodes() > (uint)LiveNodeCountInliningCutoff) {
+        break;
+      }
+    }
+
+    inline_incrementally_one(igvn);
+
+    if (failing())  return;
+
+    igvn.optimize();
+
+    if (failing())  return;
+  }
+
+  assert( igvn._worklist.size() == 0, "should be done with igvn" );
+
+  if (_string_late_inlines.length() > 0) {
+    assert(has_stringbuilder(), "inconsistent");
+    for_igvn()->clear();
+    initial_gvn()->replace_with(&igvn);
+
+    inline_string_calls(false);
+
+    if (failing())  return;
+
+    {
+      ResourceMark rm;
+      PhaseRemoveUseless pru(initial_gvn(), for_igvn());
+    }
+
+    igvn = PhaseIterGVN(gvn);
+
+    igvn.optimize();
+  }
+
+  set_inlining_incrementally(false);
+}
+
+
 //------------------------------Optimize---------------------------------------
 // Given a graph, optimize it.
 void Compile::Optimize() {
@@ -1792,6 +1921,12 @@ void Compile::Optimize() {
 
   if (failing())  return;
 
+  inline_incrementally(igvn);
+
+  print_method("Incremental Inline", 2);
+
+  if (failing())  return;
+
   // Perform escape analysis
   if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) {
     if (has_loops()) {
@@ -1914,6 +2049,7 @@ void Compile::Optimize() {
 
  } // (End scope of igvn; run destructor if necessary for asserts.)
 
+ dump_inlining();
   // A method with only infinite loops has no edges entering loops from root
   {
     NOT_PRODUCT( TracePhase t2("graphReshape", &_t_graphReshaping, TimeCompiler); )
@@ -3362,6 +3498,28 @@ void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n
 
 void Compile::dump_inlining() {
   if (PrintInlining) {
+    // Print inlining message for candidates that we couldn't inline
+    // for lack of space or non constant receiver
+    for (int i = 0; i < _late_inlines.length(); i++) {
+      CallGenerator* cg = _late_inlines.at(i);
+      cg->print_inlining_late("live nodes > LiveNodeCountInliningCutoff");
+    }
+    Unique_Node_List useful;
+    useful.push(root());
+    for (uint next = 0; next < useful.size(); ++next) {
+      Node* n  = useful.at(next);
+      if (n->is_Call() && n->as_Call()->generator() != NULL && n->as_Call()->generator()->call_node() == n) {
+        CallNode* call = n->as_Call();
+        CallGenerator* cg = call->generator();
+        cg->print_inlining_late("receiver not constant");
+      }
+      uint max = n->len();
+      for ( uint i = 0; i < max; ++i ) {
+        Node *m = n->in(i);
+        if ( m == NULL ) continue;
+        useful.push(m);
+      }
+    }
     for (int i = 0; i < _print_inlining_list->length(); i++) {
       tty->print(_print_inlining_list->at(i).ss()->as_string());
     }

src/share/vm/opto/compile.hpp

@@ -280,6 +280,8 @@ class Compile : public Phase {
   int                   _orig_pc_slot_offset_in_bytes;
 
   int                   _major_progress;        // Count of something big happening
+  bool                  _inlining_progress;     // progress doing incremental inlining?
+  bool                  _inlining_incrementally;// Are we doing incremental inlining (post parse)
   bool                  _has_loops;             // True if the method _may_ have some loops
   bool                  _has_split_ifs;         // True if the method _may_ have some split-if
   bool                  _has_unsafe_access;     // True if the method _may_ produce faults in unsafe loads or stores.
@@ -369,6 +371,11 @@ class Compile : public Phase {
 
   GrowableArray<CallGenerator*> _late_inlines;        // List of CallGenerators to be revisited after
                                                       // main parsing has finished.
+  GrowableArray<CallGenerator*> _string_late_inlines; // same but for string operations
+
+  int                           _late_inlines_pos;    // Where in the queue should the next late inlining candidate go (emulate depth first inlining)
+  uint                          _number_of_mh_late_inlines; // number of method handle late inlining still pending
+
 
   // Inlining may not happen in parse order which would make
   // PrintInlining output confusing. Keep track of PrintInlining
@@ -491,6 +498,10 @@ class Compile : public Phase {
   int               fixed_slots() const         { assert(_fixed_slots >= 0, "");         return _fixed_slots; }
   void          set_fixed_slots(int n)          { _fixed_slots = n; }
   int               major_progress() const      { return _major_progress; }
+  void          set_inlining_progress(bool z)   { _inlining_progress = z; }
+  int               inlining_progress() const   { return _inlining_progress; }
+  void          set_inlining_incrementally(bool z) { _inlining_incrementally = z; }
+  int               inlining_incrementally() const { return _inlining_incrementally; }
   void          set_major_progress()            { _major_progress++; }
   void        clear_major_progress()            { _major_progress = 0; }
   int               num_loop_opts() const       { return _num_loop_opts; }
@@ -729,7 +740,7 @@ 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_is_virtual, JVMState* jvms, bool allow_inline, float profile_factor, bool allow_intrinsics = true);
+  CallGenerator*    call_generator(ciMethod* call_method, int vtable_index, bool call_is_virtual, 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);
 
   // Report if there were too many traps at a current method and bci.
@@ -765,10 +776,39 @@ class Compile : public Phase {
   WarmCallInfo* pop_warm_call();
 
   // Record this CallGenerator for inlining at the end of parsing.
-  void              add_late_inline(CallGenerator* cg) { _late_inlines.push(cg); }
+  void              add_late_inline(CallGenerator* cg)        {
+    _late_inlines.insert_before(_late_inlines_pos, cg);
+    _late_inlines_pos++;
+  }
+
+  void              prepend_late_inline(CallGenerator* cg)    {
+    _late_inlines.insert_before(0, cg);
+  }
+
+  void              add_string_late_inline(CallGenerator* cg) {
+    _string_late_inlines.push(cg);
+  }
+
+  void remove_useless_late_inlines(GrowableArray<CallGenerator*>* inlines, Unique_Node_List &useful);
 
   void dump_inlining();
 
+  bool over_inlining_cutoff() const {
+    if (!inlining_incrementally()) {
+      return unique() > (uint)NodeCountInliningCutoff;
+    } else {
+      return live_nodes() > (uint)LiveNodeCountInliningCutoff;
+    }
+  }
+
+  void inc_number_of_mh_late_inlines() { _number_of_mh_late_inlines++; }
+  void dec_number_of_mh_late_inlines() { assert(_number_of_mh_late_inlines > 0, "_number_of_mh_late_inlines < 0 !"); _number_of_mh_late_inlines--; }
+  bool has_mh_late_inlines() const     { return _number_of_mh_late_inlines > 0; }
+
+  void inline_incrementally_one(PhaseIterGVN& igvn);
+  void inline_incrementally(PhaseIterGVN& igvn);
+  void inline_string_calls(bool parse_time);
+
   // Matching, CFG layout, allocation, code generation
   PhaseCFG*         cfg()                       { return _cfg; }
   bool              select_24_bit_instr() const { return _select_24_bit_instr; }

src/share/vm/opto/doCall.cpp

@@ -63,7 +63,7 @@ void trace_type_profile(Compile* C, ciMethod *method, int depth, int bci, ciMeth
 
 CallGenerator* Compile::call_generator(ciMethod* callee, int vtable_index, bool call_is_virtual,
                                        JVMState* jvms, bool allow_inline,
-                                       float prof_factor, bool allow_intrinsics) {
+                                       float prof_factor, bool allow_intrinsics, bool delayed_forbidden) {
   ciMethod*       caller   = jvms->method();
   int             bci      = jvms->bci();
   Bytecodes::Code bytecode = caller->java_code_at_bci(bci);
@@ -130,7 +130,9 @@ CallGenerator* Compile::call_generator(ciMethod* callee, int vtable_index, bool
   // MethodHandle.invoke* are native methods which obviously don't
   // have bytecodes and so normal inlining fails.
   if (callee->is_method_handle_intrinsic()) {
-    return CallGenerator::for_method_handle_call(jvms, caller, callee);
+    CallGenerator* cg = CallGenerator::for_method_handle_call(jvms, caller, callee, delayed_forbidden);
+    assert (cg == NULL || !delayed_forbidden || !cg->is_late_inline() || cg->is_mh_late_inline(), "unexpected CallGenerator");
+    return cg;
   }
 
   // Do not inline strict fp into non-strict code, or the reverse
@@ -161,20 +163,27 @@ CallGenerator* Compile::call_generator(ciMethod* callee, int vtable_index, bool
       WarmCallInfo scratch_ci;
       if (!UseOldInlining)
         scratch_ci.init(jvms, callee, profile, prof_factor);
-      WarmCallInfo* ci = ilt->ok_to_inline(callee, jvms, profile, &scratch_ci);
+      bool should_delay = false;
+      WarmCallInfo* ci = ilt->ok_to_inline(callee, jvms, profile, &scratch_ci, should_delay);
       assert(ci != &scratch_ci, "do not let this pointer escape");
       bool allow_inline   = (ci != NULL && !ci->is_cold());
       bool require_inline = (allow_inline && ci->is_hot());
 
       if (allow_inline) {
         CallGenerator* cg = CallGenerator::for_inline(callee, expected_uses);
-        if (require_inline && cg != NULL && should_delay_inlining(callee, jvms)) {
+
+        if (require_inline && cg != NULL) {
           // Delay the inlining of this method to give us the
           // opportunity to perform some high level optimizations
           // first.
+          if (should_delay_inlining(callee, jvms)) {
+            assert(!delayed_forbidden, "strange");
+            return CallGenerator::for_string_late_inline(callee, cg);
+          } else if ((should_delay || AlwaysIncrementalInline) && !delayed_forbidden) {
             return CallGenerator::for_late_inline(callee, cg);
           }
-        if (cg == NULL) {
+        }
+        if (cg == NULL || should_delay) {
           // Fall through.
         } else if (require_inline || !InlineWarmCalls) {
           return cg;

src/share/vm/opto/graphKit.cpp

@@ -1794,10 +1794,15 @@ void GraphKit::replace_call(CallNode* call, Node* result) {
 
   if (ejvms == NULL) {
     // No exception edges to simply kill off those paths
+    if (callprojs.catchall_catchproj != NULL) {
       C->gvn_replace_by(callprojs.catchall_catchproj, C->top());
+    }
+    if (callprojs.catchall_memproj != NULL) {
       C->gvn_replace_by(callprojs.catchall_memproj,   C->top());
+    }
+    if (callprojs.catchall_ioproj != NULL) {
       C->gvn_replace_by(callprojs.catchall_ioproj,    C->top());
-
+    }
     // Replace the old exception object with top
     if (callprojs.exobj != NULL) {
       C->gvn_replace_by(callprojs.exobj, C->top());
@@ -1809,10 +1814,15 @@ void GraphKit::replace_call(CallNode* call, Node* result) {
     SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states();
 
     Node* ex_oop = ekit.use_exception_state(ex_map);
-
+    if (callprojs.catchall_catchproj != NULL) {
       C->gvn_replace_by(callprojs.catchall_catchproj, ekit.control());
+    }
+    if (callprojs.catchall_memproj != NULL) {
       C->gvn_replace_by(callprojs.catchall_memproj,   ekit.reset_memory());
+    }
+    if (callprojs.catchall_ioproj != NULL) {
       C->gvn_replace_by(callprojs.catchall_ioproj,    ekit.i_o());
+    }
 
     // Replace the old exception object with the newly created one
     if (callprojs.exobj != NULL) {

src/share/vm/opto/parse.hpp

@@ -70,7 +70,7 @@ protected:
   InlineTree *build_inline_tree_for_callee(ciMethod* callee_method,
                                            JVMState* caller_jvms,
                                            int caller_bci);
-  const char* try_to_inline(ciMethod* callee_method, ciMethod* caller_method, int caller_bci, ciCallProfile& profile, WarmCallInfo* wci_result);
+  const char* try_to_inline(ciMethod* callee_method, ciMethod* caller_method, int caller_bci, ciCallProfile& profile, WarmCallInfo* wci_result, bool& should_delay);
   const char* should_inline(ciMethod* callee_method, ciMethod* caller_method, int caller_bci, ciCallProfile& profile, WarmCallInfo* wci_result) const;
   const char* should_not_inline(ciMethod* callee_method, ciMethod* caller_method, WarmCallInfo* wci_result) const;
   void        print_inlining(ciMethod *callee_method, int caller_bci, const char *failure_msg) const;
@@ -107,7 +107,7 @@ public:
   // and may be accessed by find_subtree_from_root.
   // The call_method is the dest_method for a special or static invocation.
   // The call_method is an optimized virtual method candidate otherwise.
-  WarmCallInfo* ok_to_inline(ciMethod *call_method, JVMState* caller_jvms, ciCallProfile& profile, WarmCallInfo* wci);
+  WarmCallInfo* ok_to_inline(ciMethod *call_method, JVMState* caller_jvms, ciCallProfile& profile, WarmCallInfo* wci, bool& should_delay);
 
   // Information about inlined method
   JVMState*   caller_jvms()       const { return _caller_jvms; }

src/share/vm/opto/phaseX.cpp

@@ -75,6 +75,13 @@ NodeHash::NodeHash(NodeHash *nh) {
   // nh->_sentinel must be in the current node space
 }
 
+void NodeHash::replace_with(NodeHash *nh) {
+  debug_only(_table = (Node**)badAddress);   // interact correctly w/ operator=
+  // just copy in all the fields
+  *this = *nh;
+  // nh->_sentinel must be in the current node space
+}
+
 //------------------------------hash_find--------------------------------------
 // Find in hash table
 Node *NodeHash::hash_find( const Node *n ) {

src/share/vm/opto/phaseX.hpp

@@ -92,6 +92,7 @@ public:
   }
 
   void   remove_useless_nodes(VectorSet &useful); // replace with sentinel
+  void replace_with(NodeHash* nh);
 
   Node  *sentinel() { return _sentinel; }
 
@@ -386,6 +387,11 @@ public:
   Node  *transform( Node *n );
   Node  *transform_no_reclaim( Node *n );
 
+  void replace_with(PhaseGVN* gvn) {
+    _table.replace_with(&gvn->_table);
+    _types = gvn->_types;
+  }
+
   // Check for a simple dead loop when a data node references itself.
   DEBUG_ONLY(void dead_loop_check(Node *n);)
 };

src/share/vm/opto/stringopts.cpp

@@ -265,7 +265,8 @@ void StringConcat::eliminate_unneeded_control() {
     } else if (n->is_IfTrue()) {
       Compile* C = _stringopts->C;
       C->gvn_replace_by(n, n->in(0)->in(0));
-      C->gvn_replace_by(n->in(0), C->top());
+      // get rid of the other projection
+      C->gvn_replace_by(n->in(0)->as_If()->proj_out(false), C->top());
     }
   }
 }
@@ -439,7 +440,7 @@ StringConcat* PhaseStringOpts::build_candidate(CallStaticJavaNode* call) {
       }
       // Find the constructor call
       Node* result = alloc->result_cast();
-      if (result == NULL || !result->is_CheckCastPP()) {
+      if (result == NULL || !result->is_CheckCastPP() || alloc->in(TypeFunc::Memory)->is_top()) {
         // strange looking allocation
 #ifndef PRODUCT
         if (PrintOptimizeStringConcat) {
@@ -834,6 +835,9 @@ bool StringConcat::validate_control_flow() {
           ptr->in(1)->in(0) != NULL && ptr->in(1)->in(0)->is_If()) {
         // Simple diamond.
         // XXX should check for possibly merging stores.  simple data merges are ok.
+        // The IGVN will make this simple diamond go away when it
+        // transforms the Region. Make sure it sees it.
+        Compile::current()->record_for_igvn(ptr);
         ptr = ptr->in(1)->in(0)->in(0);
         continue;
       }

src/share/vm/runtime/arguments.cpp

@@ -3291,6 +3291,18 @@ jint Arguments::parse(const JavaVMInitArgs* args) {
   if (!EliminateLocks) {
     EliminateNestedLocks = false;
   }
+  if (!Inline) {
+    IncrementalInline = false;
+  }
+#ifndef PRODUCT
+  if (!IncrementalInline) {
+    AlwaysIncrementalInline = false;
+  }
+#endif
+  if (IncrementalInline && FLAG_IS_DEFAULT(MaxNodeLimit)) {
+    // incremental inlining: bump MaxNodeLimit
+    FLAG_SET_DEFAULT(MaxNodeLimit, (intx)75000);
+  }
 #endif
 
   if (PrintAssembly && FLAG_IS_DEFAULT(DebugNonSafepoints)) {