< prev index next >

src/hotspot/share/opto/graphKit.cpp

Print this page
@@ -21,10 +21,12 @@
   * questions.
   *
   */
  
  #include "precompiled.hpp"
+ #include "ci/ciFlatArrayKlass.hpp"
+ #include "ci/ciInlineKlass.hpp"
  #include "ci/ciUtilities.hpp"
  #include "classfile/javaClasses.hpp"
  #include "ci/ciObjArray.hpp"
  #include "asm/register.hpp"
  #include "compiler/compileLog.hpp"

@@ -35,13 +37,15 @@
  #include "opto/addnode.hpp"
  #include "opto/castnode.hpp"
  #include "opto/convertnode.hpp"
  #include "opto/graphKit.hpp"
  #include "opto/idealKit.hpp"
+ #include "opto/inlinetypenode.hpp"
  #include "opto/intrinsicnode.hpp"
  #include "opto/locknode.hpp"
  #include "opto/machnode.hpp"
+ #include "opto/narrowptrnode.hpp"
  #include "opto/opaquenode.hpp"
  #include "opto/parse.hpp"
  #include "opto/rootnode.hpp"
  #include "opto/runtime.hpp"
  #include "opto/subtypenode.hpp"

@@ -51,19 +55,27 @@
  #include "utilities/powerOfTwo.hpp"
  #include "utilities/growableArray.hpp"
  
  //----------------------------GraphKit-----------------------------------------
  // Main utility constructor.
- GraphKit::GraphKit(JVMState* jvms)
+ GraphKit::GraphKit(JVMState* jvms, PhaseGVN* gvn)
    : Phase(Phase::Parser),
      _env(C->env()),
-     _gvn(*C->initial_gvn()),
+     _gvn((gvn != nullptr) ? *gvn : *C->initial_gvn()),
      _barrier_set(BarrierSet::barrier_set()->barrier_set_c2())
  {
+   assert(gvn == nullptr || !gvn->is_IterGVN() || gvn->is_IterGVN()->delay_transform(), "delay transform should be enabled");
    _exceptions = jvms->map()->next_exception();
    if (_exceptions != nullptr)  jvms->map()->set_next_exception(nullptr);
    set_jvms(jvms);
+ #ifdef ASSERT
+   if (_gvn.is_IterGVN() != nullptr) {
+     assert(_gvn.is_IterGVN()->delay_transform(), "Transformation must be delayed if IterGVN is used");
+     // Save the initial size of _for_igvn worklist for verification (see ~GraphKit)
+     _worklist_size = _gvn.C->igvn_worklist()->size();
+   }
+ #endif
  }
  
  // Private constructor for parser.
  GraphKit::GraphKit()
    : Phase(Phase::Parser),

@@ -855,11 +867,11 @@
    ciMethod* cur_method = jvms->method();
    int       cur_bci   = jvms->bci();
    if (cur_method != nullptr && cur_bci != InvocationEntryBci) {
      Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci);
      return Interpreter::bytecode_should_reexecute(code) ||
-            (is_anewarray && code == Bytecodes::_multianewarray);
+            (is_anewarray && (code == Bytecodes::_multianewarray));
      // Reexecute _multianewarray bytecode which was replaced with
      // sequence of [a]newarray. See Parse::do_multianewarray().
      //
      // Note: interpreter should not have it set since this optimization
      // is limited by dimensions and guarded by flag so in some cases

@@ -971,23 +983,46 @@
      // Add the Locals
      k = in_jvms->locoff();
      l = in_jvms->loc_size();
      out_jvms->set_locoff(p);
      if (!can_prune_locals) {
-       for (j = 0; j < l; j++)
+       for (j = 0; j < l; j++) {
          call->set_req(p++, in_map->in(k+j));
+         Node* local = in_map->in(k+j);
+         // TODO 8325106
+         /*
+         if (false && local->is_InlineType() && local->isa_InlineType()->is_larval()) {
+           tty->print_cr("LARVAL FOUND in LOCAL");
+           in_map->dump(0);
+           local->dump(0);
+         }
+         */
+       }
      } else {
        p += l;  // already set to top above by add_req_batch
      }
  
      // Add the Expression Stack
      k = in_jvms->stkoff();
      l = in_jvms->sp();
      out_jvms->set_stkoff(p);
      if (!can_prune_locals) {
-       for (j = 0; j < l; j++)
+       for (j = 0; j < l; j++) {
          call->set_req(p++, in_map->in(k+j));
+         Node* local = in_map->in(k+j);
+         // TODO 8325106 check if there's a larval on stack in the caller state that has been written in the callee state and update it accordingly
+         /*
+         if (false && local->is_InlineType() && local->isa_InlineType()->is_larval()) {
+           tty->print_cr("LARVAL FOUND on STACK");
+           in_map->dump(0);
+           local->dump(0);
+           map()->replaced_nodes().dump(tty);
+           map()->replaced_nodes().apply(call, 0);
+           tty->print_cr("");
+         }
+         */
+       }
      } else if (can_prune_locals && stack_slots_not_pruned != 0) {
        // Divide stack into {S0,...,S1}, where S0 is set to top.
        uint s1 = stack_slots_not_pruned;
        stack_slots_not_pruned = 0;  // for next iteration
        if (s1 > l)  s1 = l;

@@ -1197,11 +1232,11 @@
  Node* GraphKit::load_object_klass(Node* obj) {
    // Special-case a fresh allocation to avoid building nodes:
    Node* akls = AllocateNode::Ideal_klass(obj, &_gvn);
    if (akls != nullptr)  return akls;
    Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
-   return _gvn.transform(LoadKlassNode::make(_gvn, nullptr, immutable_memory(), k_adr, TypeInstPtr::KLASS));
+   return _gvn.transform(LoadKlassNode::make(_gvn, nullptr, immutable_memory(), k_adr, TypeInstPtr::KLASS, TypeInstKlassPtr::OBJECT));
  }
  
  //-------------------------load_array_length-----------------------------------
  Node* GraphKit::load_array_length(Node* array) {
    // Special-case a fresh allocation to avoid building nodes:

@@ -1246,15 +1281,38 @@
  #endif
  Node* GraphKit::null_check_common(Node* value, BasicType type,
                                    // optional arguments for variations:
                                    bool assert_null,
                                    Node* *null_control,
-                                   bool speculative) {
+                                   bool speculative,
+                                   bool is_init_check) {
    assert(!assert_null || null_control == nullptr, "not both at once");
    if (stopped())  return top();
    NOT_PRODUCT(explicit_null_checks_inserted++);
  
+   if (value->is_InlineType()) {
+     // Null checking a scalarized but nullable inline type. Check the IsInit
+     // input instead of the oop input to avoid keeping buffer allocations alive.
+     InlineTypeNode* vtptr = value->as_InlineType();
+     while (vtptr->get_oop()->is_InlineType()) {
+       vtptr = vtptr->get_oop()->as_InlineType();
+     }
+     null_check_common(vtptr->get_is_init(), T_INT, assert_null, null_control, speculative, true);
+     if (stopped()) {
+       return top();
+     }
+     if (assert_null) {
+       // TODO 8284443 Scalarize here (this currently leads to compilation bailouts)
+       // vtptr = InlineTypeNode::make_null(_gvn, vtptr->type()->inline_klass());
+       // replace_in_map(value, vtptr);
+       // return vtptr;
+       return null();
+     }
+     bool do_replace_in_map = (null_control == nullptr || (*null_control) == top());
+     return cast_not_null(value, do_replace_in_map);
+   }
+ 
    // Construct null check
    Node *chk = nullptr;
    switch(type) {
      case T_LONG   : chk = new CmpLNode(value, _gvn.zerocon(T_LONG)); break;
      case T_INT    : chk = new CmpINode(value, _gvn.intcon(0)); break;

@@ -1354,11 +1412,11 @@
    // Branch to failure if null
    float ok_prob = PROB_MAX;  // a priori estimate:  nulls never happen
    Deoptimization::DeoptReason reason;
    if (assert_null) {
      reason = Deoptimization::reason_null_assert(speculative);
-   } else if (type == T_OBJECT) {
+   } else if (type == T_OBJECT || is_init_check) {
      reason = Deoptimization::reason_null_check(speculative);
    } else {
      reason = Deoptimization::Reason_div0_check;
    }
    // %%% Since Reason_unhandled is not recorded on a per-bytecode basis,

@@ -1428,14 +1486,23 @@
    }
  
    return value;
  }
  
- 
  //------------------------------cast_not_null----------------------------------
  // Cast obj to not-null on this path
  Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) {
+   if (obj->is_InlineType()) {
+     // TODO 8325106 Can we avoid cloning?
+     Node* vt = obj->clone();
+     vt->as_InlineType()->set_is_init(_gvn);
+     vt = _gvn.transform(vt);
+     if (do_replace_in_map) {
+       replace_in_map(obj, vt);
+     }
+     return vt;
+   }
    const Type *t = _gvn.type(obj);
    const Type *t_not_null = t->join_speculative(TypePtr::NOTNULL);
    // Object is already not-null?
    if( t == t_not_null ) return obj;
  

@@ -1558,10 +1625,11 @@
    const TypePtr* adr_type = nullptr; // debug-mode-only argument
    debug_only(adr_type = C->get_adr_type(adr_idx));
    Node* mem = memory(adr_idx);
    Node* ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt, mo, control_dependency, require_atomic_access, unaligned, mismatched, unsafe, barrier_data);
    ld = _gvn.transform(ld);
+ 
    if (((bt == T_OBJECT) && C->do_escape_analysis()) || C->eliminate_boxing()) {
      // Improve graph before escape analysis and boxing elimination.
      record_for_igvn(ld);
    }
    return ld;

@@ -1604,11 +1672,12 @@
                                  Node* adr,
                                  const TypePtr* adr_type,
                                  Node* val,
                                  const Type* val_type,
                                  BasicType bt,
-                                 DecoratorSet decorators) {
+                                 DecoratorSet decorators,
+                                 bool safe_for_replace) {
    // Transformation of a value which could be null pointer (CastPP #null)
    // could be delayed during Parse (for example, in adjust_map_after_if()).
    // Execute transformation here to avoid barrier generation in such case.
    if (_gvn.type(val) == TypePtr::NULL_PTR) {
      val = _gvn.makecon(TypePtr::NULL_PTR);

@@ -1617,10 +1686,17 @@
    if (stopped()) {
      return top(); // Dead path ?
    }
  
    assert(val != nullptr, "not dead path");
+   if (val->is_InlineType()) {
+     // Store to non-flat field. Buffer the inline type and make sure
+     // the store is re-executed if the allocation triggers deoptimization.
+     PreserveReexecuteState preexecs(this);
+     jvms()->set_should_reexecute(true);
+     val = val->as_InlineType()->buffer(this, safe_for_replace);
+   }
  
    C2AccessValuePtr addr(adr, adr_type);
    C2AccessValue value(val, val_type);
    C2ParseAccess access(this, decorators | C2_WRITE_ACCESS, bt, obj, addr);
    if (access.is_raw()) {

@@ -1633,17 +1709,18 @@
  Node* GraphKit::access_load_at(Node* obj,   // containing obj
                                 Node* adr,   // actual address to store val at
                                 const TypePtr* adr_type,
                                 const Type* val_type,
                                 BasicType bt,
-                                DecoratorSet decorators) {
+                                DecoratorSet decorators,
+                                Node* ctl) {
    if (stopped()) {
      return top(); // Dead path ?
    }
  
    C2AccessValuePtr addr(adr, adr_type);
-   C2ParseAccess access(this, decorators | C2_READ_ACCESS, bt, obj, addr);
+   C2ParseAccess access(this, decorators | C2_READ_ACCESS, bt, obj, addr, ctl);
    if (access.is_raw()) {
      return _barrier_set->BarrierSetC2::load_at(access, val_type);
    } else {
      return _barrier_set->load_at(access, val_type);
    }

@@ -1744,11 +1821,12 @@
  }
  
  //-------------------------array_element_address-------------------------
  Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt,
                                        const TypeInt* sizetype, Node* ctrl) {
-   uint shift  = exact_log2(type2aelembytes(elembt));
+   const TypeAryPtr* arytype = _gvn.type(ary)->is_aryptr();
+   uint shift = arytype->is_flat() ? arytype->flat_log_elem_size() : exact_log2(type2aelembytes(elembt));
    uint header = arrayOopDesc::base_offset_in_bytes(elembt);
  
    // short-circuit a common case (saves lots of confusing waste motion)
    jint idx_con = find_int_con(idx, -1);
    if (idx_con >= 0) {

@@ -1776,16 +1854,52 @@
    return ld;
  }
  
  //-------------------------set_arguments_for_java_call-------------------------
  // Arguments (pre-popped from the stack) are taken from the JVMS.
- void GraphKit::set_arguments_for_java_call(CallJavaNode* call) {
-   // Add the call arguments:
-   uint nargs = call->method()->arg_size();
-   for (uint i = 0; i < nargs; i++) {
-     Node* arg = argument(i);
-     call->init_req(i + TypeFunc::Parms, arg);
+ void GraphKit::set_arguments_for_java_call(CallJavaNode* call, bool is_late_inline) {
+   PreserveReexecuteState preexecs(this);
+   if (EnableValhalla) {
+     // Make sure the call is "re-executed", if buffering of inline type arguments triggers deoptimization.
+     // At this point, the call hasn't been executed yet, so we will only ever execute the call once.
+     jvms()->set_should_reexecute(true);
+     int arg_size = method()->get_declared_signature_at_bci(bci())->arg_size_for_bc(java_bc());
+     inc_sp(arg_size);
+   }
+   // Add the call arguments
+   const TypeTuple* domain = call->tf()->domain_sig();
+   uint nargs = domain->cnt();
+   int arg_num = 0;
+   for (uint i = TypeFunc::Parms, idx = TypeFunc::Parms; i < nargs; i++) {
+     Node* arg = argument(i-TypeFunc::Parms);
+     const Type* t = domain->field_at(i);
+     // TODO 8284443 A static call to a mismatched method should still be scalarized
+     if (t->is_inlinetypeptr() && !call->method()->get_Method()->mismatch() && call->method()->is_scalarized_arg(arg_num)) {
+       // We don't pass inline type arguments by reference but instead pass each field of the inline type
+       if (!arg->is_InlineType()) {
+         assert(_gvn.type(arg)->is_zero_type() && !t->inline_klass()->is_null_free(), "Unexpected argument type");
+         arg = InlineTypeNode::make_from_oop(this, arg, t->inline_klass(), t->inline_klass()->is_null_free());
+       }
+       InlineTypeNode* vt = arg->as_InlineType();
+       vt->pass_fields(this, call, idx, true, !t->maybe_null());
+       // If an inline type argument is passed as fields, attach the Method* to the call site
+       // to be able to access the extended signature later via attached_method_before_pc().
+       // For example, see CompiledMethod::preserve_callee_argument_oops().
+       call->set_override_symbolic_info(true);
+       // Register an evol dependency on the callee method to make sure that this method is deoptimized and
+       // re-compiled with a non-scalarized calling convention if the callee method is later marked as mismatched.
+       C->dependencies()->assert_evol_method(call->method());
+       arg_num++;
+       continue;
+     } else if (arg->is_InlineType()) {
+       // Pass inline type argument via oop to callee
+       arg = arg->as_InlineType()->buffer(this);
+     }
+     if (t != Type::HALF) {
+       arg_num++;
+     }
+     call->init_req(idx++, arg);
    }
  }
  
  //---------------------------set_edges_for_java_call---------------------------
  // Connect a newly created call into the current JVMS.

@@ -1819,17 +1933,10 @@
  }
  
  Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj, bool deoptimize) {
    if (stopped())  return top();  // maybe the call folded up?
  
-   // Capture the return value, if any.
-   Node* ret;
-   if (call->method() == nullptr ||
-       call->method()->return_type()->basic_type() == T_VOID)
-         ret = top();
-   else  ret = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
- 
    // Note:  Since any out-of-line call can produce an exception,
    // we always insert an I_O projection from the call into the result.
  
    make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj, deoptimize);
  

@@ -1838,10 +1945,42 @@
      // through and exceptional paths, so replace the projections for
      // the fall through path.
      set_i_o(_gvn.transform( new ProjNode(call, TypeFunc::I_O) ));
      set_all_memory(_gvn.transform( new ProjNode(call, TypeFunc::Memory) ));
    }
+ 
+   // Capture the return value, if any.
+   Node* ret;
+   if (call->method() == nullptr || call->method()->return_type()->basic_type() == T_VOID) {
+     ret = top();
+   } else if (call->tf()->returns_inline_type_as_fields()) {
+     // Return of multiple values (inline type fields): we create a
+     // InlineType node, each field is a projection from the call.
+     ciInlineKlass* vk = call->method()->return_type()->as_inline_klass();
+     uint base_input = TypeFunc::Parms;
+     ret = InlineTypeNode::make_from_multi(this, call, vk, base_input, false, false);
+   } else {
+     ret = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
+     ciType* t = call->method()->return_type();
+     if (t->is_klass()) {
+       const Type* type = TypeOopPtr::make_from_klass(t->as_klass());
+       if (type->is_inlinetypeptr()) {
+         ret = InlineTypeNode::make_from_oop(this, ret, type->inline_klass(), type->inline_klass()->is_null_free());
+       }
+     }
+   }
+ 
+   // We just called the constructor on a value type receiver. Reload it from the buffer
+   if (call->method()->is_object_constructor() && call->method()->holder()->is_inlinetype()) {
+     InlineTypeNode* receiver = call->in(TypeFunc::Parms)->as_InlineType();
+     assert(receiver->is_larval(), "must be larval");
+     assert(receiver->is_allocated(&gvn()), "larval must be buffered");
+     InlineTypeNode* reloaded = InlineTypeNode::make_from_oop(this, receiver->get_oop(), receiver->bottom_type()->inline_klass(), true);
+     assert(!reloaded->is_larval(), "should not be larval anymore");
+     replace_in_map(receiver, reloaded);
+   }
+ 
    return ret;
  }
  
  //--------------------set_predefined_input_for_runtime_call--------------------
  // Reading and setting the memory state is way conservative here.

@@ -1928,80 +2067,83 @@
    Node* ex_ctl = top();
  
    SafePointNode* final_state = stop();
  
    // Find all the needed outputs of this call
-   CallProjections callprojs;
-   call->extract_projections(&callprojs, true);
+   CallProjections* callprojs = call->extract_projections(true);
  
    Unique_Node_List wl;
    Node* init_mem = call->in(TypeFunc::Memory);
    Node* final_mem = final_state->in(TypeFunc::Memory);
    Node* final_ctl = final_state->in(TypeFunc::Control);
    Node* final_io = final_state->in(TypeFunc::I_O);
  
    // Replace all the old call edges with the edges from the inlining result
-   if (callprojs.fallthrough_catchproj != nullptr) {
-     C->gvn_replace_by(callprojs.fallthrough_catchproj, final_ctl);
+   if (callprojs->fallthrough_catchproj != nullptr) {
+     C->gvn_replace_by(callprojs->fallthrough_catchproj, final_ctl);
    }
-   if (callprojs.fallthrough_memproj != nullptr) {
+   if (callprojs->fallthrough_memproj != nullptr) {
      if (final_mem->is_MergeMem()) {
        // Parser's exits MergeMem was not transformed but may be optimized
        final_mem = _gvn.transform(final_mem);
      }
-     C->gvn_replace_by(callprojs.fallthrough_memproj,   final_mem);
+     C->gvn_replace_by(callprojs->fallthrough_memproj,   final_mem);
      add_mergemem_users_to_worklist(wl, final_mem);
    }
-   if (callprojs.fallthrough_ioproj != nullptr) {
-     C->gvn_replace_by(callprojs.fallthrough_ioproj,    final_io);
+   if (callprojs->fallthrough_ioproj != nullptr) {
+     C->gvn_replace_by(callprojs->fallthrough_ioproj,    final_io);
    }
  
    // Replace the result with the new result if it exists and is used
-   if (callprojs.resproj != nullptr && result != nullptr) {
-     C->gvn_replace_by(callprojs.resproj, result);
+   if (callprojs->resproj[0] != nullptr && result != nullptr) {
+     // If the inlined code is dead, the result projections for an inline type returned as
+     // fields have not been replaced. They will go away once the call is replaced by TOP below.
+     assert(callprojs->nb_resproj == 1 || (call->tf()->returns_inline_type_as_fields() && stopped()),
+            "unexpected number of results");
+     C->gvn_replace_by(callprojs->resproj[0], result);
    }
  
    if (ejvms == nullptr) {
      // No exception edges to simply kill off those paths
-     if (callprojs.catchall_catchproj != nullptr) {
-       C->gvn_replace_by(callprojs.catchall_catchproj, C->top());
+     if (callprojs->catchall_catchproj != nullptr) {
+       C->gvn_replace_by(callprojs->catchall_catchproj, C->top());
      }
-     if (callprojs.catchall_memproj != nullptr) {
-       C->gvn_replace_by(callprojs.catchall_memproj,   C->top());
+     if (callprojs->catchall_memproj != nullptr) {
+       C->gvn_replace_by(callprojs->catchall_memproj,   C->top());
      }
-     if (callprojs.catchall_ioproj != nullptr) {
-       C->gvn_replace_by(callprojs.catchall_ioproj,    C->top());
+     if (callprojs->catchall_ioproj != nullptr) {
+       C->gvn_replace_by(callprojs->catchall_ioproj,    C->top());
      }
      // Replace the old exception object with top
-     if (callprojs.exobj != nullptr) {
-       C->gvn_replace_by(callprojs.exobj, C->top());
+     if (callprojs->exobj != nullptr) {
+       C->gvn_replace_by(callprojs->exobj, C->top());
      }
    } else {
      GraphKit ekit(ejvms);
  
      // Load my combined exception state into the kit, with all phis transformed:
      SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states();
      replaced_nodes_exception = ex_map->replaced_nodes();
  
      Node* ex_oop = ekit.use_exception_state(ex_map);
  
-     if (callprojs.catchall_catchproj != nullptr) {
-       C->gvn_replace_by(callprojs.catchall_catchproj, ekit.control());
+     if (callprojs->catchall_catchproj != nullptr) {
+       C->gvn_replace_by(callprojs->catchall_catchproj, ekit.control());
        ex_ctl = ekit.control();
      }
-     if (callprojs.catchall_memproj != nullptr) {
+     if (callprojs->catchall_memproj != nullptr) {
        Node* ex_mem = ekit.reset_memory();
-       C->gvn_replace_by(callprojs.catchall_memproj,   ex_mem);
+       C->gvn_replace_by(callprojs->catchall_memproj,   ex_mem);
        add_mergemem_users_to_worklist(wl, ex_mem);
      }
-     if (callprojs.catchall_ioproj != nullptr) {
-       C->gvn_replace_by(callprojs.catchall_ioproj,    ekit.i_o());
+     if (callprojs->catchall_ioproj != nullptr) {
+       C->gvn_replace_by(callprojs->catchall_ioproj,    ekit.i_o());
      }
  
      // Replace the old exception object with the newly created one
-     if (callprojs.exobj != nullptr) {
-       C->gvn_replace_by(callprojs.exobj, ex_oop);
+     if (callprojs->exobj != nullptr) {
+       C->gvn_replace_by(callprojs->exobj, ex_oop);
      }
    }
  
    // Disconnect the call from the graph
    call->disconnect_inputs(C);

@@ -2011,11 +2153,11 @@
    // optimizer doesn't like that.
    while (wl.size() > 0) {
      _gvn.transform(wl.pop());
    }
  
-   if (callprojs.fallthrough_catchproj != nullptr && !final_ctl->is_top() && do_replaced_nodes) {
+   if (callprojs->fallthrough_catchproj != nullptr && !final_ctl->is_top() && do_replaced_nodes) {
      replaced_nodes.apply(C, final_ctl);
    }
    if (!ex_ctl->is_top() && do_replaced_nodes) {
      replaced_nodes_exception.apply(C, ex_ctl);
    }

@@ -2210,11 +2352,11 @@
      const TypeKlassPtr* tklass = TypeKlassPtr::make(exact_kls, Type::trust_interfaces);
      const TypeOopPtr* xtype = tklass->as_instance_type();
      assert(xtype->klass_is_exact(), "Should be exact");
      // Any reason to believe n is not null (from this profiling or a previous one)?
      assert(ptr_kind != ProfileAlwaysNull, "impossible here");
-     const TypePtr* ptr = (ptr_kind == ProfileMaybeNull && current_type->speculative_maybe_null()) ? TypePtr::BOTTOM : TypePtr::NOTNULL;
+     const TypePtr* ptr = (ptr_kind != ProfileNeverNull && current_type->speculative_maybe_null()) ? TypePtr::BOTTOM : TypePtr::NOTNULL;
      // record the new speculative type's depth
      speculative = xtype->cast_to_ptr_type(ptr->ptr())->is_ptr();
      speculative = speculative->with_inline_depth(jvms()->depth());
    } else if (current_type->would_improve_ptr(ptr_kind)) {
      // Profiling report that null was never seen so we can change the

@@ -2233,11 +2375,11 @@
    }
  
    if (speculative != current_type->speculative()) {
      // Build a type with a speculative type (what we think we know
      // about the type but will need a guard when we use it)
-     const TypeOopPtr* spec_type = TypeOopPtr::make(TypePtr::BotPTR, Type::OffsetBot, TypeOopPtr::InstanceBot, speculative);
+     const TypeOopPtr* spec_type = TypeOopPtr::make(TypePtr::BotPTR, Type::Offset::bottom, TypeOopPtr::InstanceBot, speculative);
      // We're changing the type, we need a new CheckCast node to carry
      // the new type. The new type depends on the control: what
      // profiling tells us is only valid from here as far as we can
      // tell.
      Node* cast = new CheckCastPPNode(control(), n, current_type->remove_speculative()->join_speculative(spec_type));

@@ -2267,23 +2409,34 @@
         java_bc() == Bytecodes::_instanceof ||
         java_bc() == Bytecodes::_aastore) &&
        method()->method_data()->is_mature()) {
      ciProfileData* data = method()->method_data()->bci_to_data(bci());
      if (data != nullptr) {
-       if (!data->as_BitData()->null_seen()) {
-         ptr_kind = ProfileNeverNull;
+       if (java_bc() == Bytecodes::_aastore) {
+         ciKlass* array_type = nullptr;
+         ciKlass* element_type = nullptr;
+         ProfilePtrKind element_ptr = ProfileMaybeNull;
+         bool flat_array = true;
+         bool null_free_array = true;
+         method()->array_access_profiled_type(bci(), array_type, element_type, element_ptr, flat_array, null_free_array);
+         exact_kls = element_type;
+         ptr_kind = element_ptr;
        } else {
-         assert(data->is_ReceiverTypeData(), "bad profile data type");
-         ciReceiverTypeData* call = (ciReceiverTypeData*)data->as_ReceiverTypeData();
-         uint i = 0;
-         for (; i < call->row_limit(); i++) {
-           ciKlass* receiver = call->receiver(i);
-           if (receiver != nullptr) {
-             break;
+         if (!data->as_BitData()->null_seen()) {
+           ptr_kind = ProfileNeverNull;
+         } else {
+           assert(data->is_ReceiverTypeData(), "bad profile data type");
+           ciReceiverTypeData* call = (ciReceiverTypeData*)data->as_ReceiverTypeData();
+           uint i = 0;
+           for (; i < call->row_limit(); i++) {
+             ciKlass* receiver = call->receiver(i);
+             if (receiver != nullptr) {
+               break;
+             }
            }
+           ptr_kind = (i == call->row_limit()) ? ProfileAlwaysNull : ProfileMaybeNull;
          }
-         ptr_kind = (i == call->row_limit()) ? ProfileAlwaysNull : ProfileMaybeNull;
        }
      }
    }
    return record_profile_for_speculation(n, exact_kls, ptr_kind);
  }

@@ -2298,14 +2451,14 @@
  void GraphKit::record_profiled_arguments_for_speculation(ciMethod* dest_method, Bytecodes::Code bc) {
    if (!UseTypeSpeculation) {
      return;
    }
    const TypeFunc* tf    = TypeFunc::make(dest_method);
-   int             nargs = tf->domain()->cnt() - TypeFunc::Parms;
+   int             nargs = tf->domain_sig()->cnt() - TypeFunc::Parms;
    int skip = Bytecodes::has_receiver(bc) ? 1 : 0;
    for (int j = skip, i = 0; j < nargs && i < TypeProfileArgsLimit; j++) {
-     const Type *targ = tf->domain()->field_at(j + TypeFunc::Parms);
+     const Type *targ = tf->domain_sig()->field_at(j + TypeFunc::Parms);
      if (is_reference_type(targ->basic_type())) {
        ProfilePtrKind ptr_kind = ProfileMaybeNull;
        ciKlass* better_type = nullptr;
        if (method()->argument_profiled_type(bci(), i, better_type, ptr_kind)) {
          record_profile_for_speculation(argument(j), better_type, ptr_kind);

@@ -2355,13 +2508,13 @@
  
  void GraphKit::round_double_arguments(ciMethod* dest_method) {
    if (Matcher::strict_fp_requires_explicit_rounding) {
      // (Note:  TypeFunc::make has a cache that makes this fast.)
      const TypeFunc* tf    = TypeFunc::make(dest_method);
-     int             nargs = tf->domain()->cnt() - TypeFunc::Parms;
+     int             nargs = tf->domain_sig()->cnt() - TypeFunc::Parms;
      for (int j = 0; j < nargs; j++) {
-       const Type *targ = tf->domain()->field_at(j + TypeFunc::Parms);
+       const Type *targ = tf->domain_sig()->field_at(j + TypeFunc::Parms);
        if (targ->basic_type() == T_DOUBLE) {
          // If any parameters are doubles, they must be rounded before
          // the call, dprecision_rounding does gvn.transform
          Node *arg = argument(j);
          arg = dprecision_rounding(arg);

@@ -2506,11 +2659,11 @@
    if (!is_leaf) {
      call = new CallStaticJavaNode(call_type, call_addr, call_name, adr_type);
    } else if (flags & RC_NO_FP) {
      call = new CallLeafNoFPNode(call_type, call_addr, call_name, adr_type);
    } else  if (flags & RC_VECTOR){
-     uint num_bits = call_type->range()->field_at(TypeFunc::Parms)->is_vect()->length_in_bytes() * BitsPerByte;
+     uint num_bits = call_type->range_sig()->field_at(TypeFunc::Parms)->is_vect()->length_in_bytes() * BitsPerByte;
      call = new CallLeafVectorNode(call_type, call_addr, call_name, adr_type, num_bits);
    } else {
      call = new CallLeafNode(call_type, call_addr, call_name, adr_type);
    }
  

@@ -2586,10 +2739,11 @@
  Node* GraphKit::sign_extend_short(Node* in) {
    Node* tmp = _gvn.transform(new LShiftINode(in, _gvn.intcon(16)));
    return _gvn.transform(new RShiftINode(tmp, _gvn.intcon(16)));
  }
  
+ 
  //------------------------------merge_memory-----------------------------------
  // Merge memory from one path into the current memory state.
  void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) {
    for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) {
      Node* old_slice = mms.force_memory();

@@ -2882,17 +3036,22 @@
    *ctrl = gvn.transform(r_ok_subtype);
    return gvn.transform(r_not_subtype);
  }
  
  Node* GraphKit::gen_subtype_check(Node* obj_or_subklass, Node* superklass) {
+   const Type* sub_t = _gvn.type(obj_or_subklass);
+   if (sub_t->make_oopptr() != nullptr && sub_t->make_oopptr()->is_inlinetypeptr()) {
+     sub_t = TypeKlassPtr::make(sub_t->inline_klass());
+     obj_or_subklass = makecon(sub_t);
+   }
    bool expand_subtype_check = C->post_loop_opts_phase() ||   // macro node expansion is over
                                ExpandSubTypeCheckAtParseTime; // forced expansion
    if (expand_subtype_check) {
      MergeMemNode* mem = merged_memory();
      Node* ctrl = control();
      Node* subklass = obj_or_subklass;
-     if (!_gvn.type(obj_or_subklass)->isa_klassptr()) {
+     if (!sub_t->isa_klassptr()) {
        subklass = load_object_klass(obj_or_subklass);
      }
  
      Node* n = Phase::gen_subtype_check(subklass, superklass, &ctrl, mem, _gvn, method(), bci());
      set_control(ctrl);

@@ -2906,41 +3065,62 @@
    return _gvn.transform(new IfFalseNode(iff));
  }
  
  // Profile-driven exact type check:
  Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass,
-                                     float prob,
-                                     Node* *casted_receiver) {
+                                     float prob, Node* *casted_receiver) {
    assert(!klass->is_interface(), "no exact type check on interfaces");
- 
+   Node* fail = top();
+   const Type* rec_t = _gvn.type(receiver);
+   if (rec_t->is_inlinetypeptr()) {
+     if (klass->equals(rec_t->inline_klass())) {
+       (*casted_receiver) = receiver; // Always passes
+     } else {
+       (*casted_receiver) = top();    // Always fails
+       fail = control();
+       set_control(top());
+     }
+     return fail;
+   }
    const TypeKlassPtr* tklass = TypeKlassPtr::make(klass, Type::trust_interfaces);
    Node* recv_klass = load_object_klass(receiver);
-   Node* want_klass = makecon(tklass);
-   Node* cmp = _gvn.transform(new CmpPNode(recv_klass, want_klass));
-   Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
-   IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN);
-   set_control( _gvn.transform(new IfTrueNode (iff)));
-   Node* fail = _gvn.transform(new IfFalseNode(iff));
+   fail = type_check(recv_klass, tklass, prob);
  
    if (!stopped()) {
      const TypeOopPtr* receiver_type = _gvn.type(receiver)->isa_oopptr();
-     const TypeOopPtr* recvx_type = tklass->as_instance_type();
-     assert(recvx_type->klass_is_exact(), "");
+     const TypeOopPtr* recv_xtype = tklass->as_instance_type();
+     assert(recv_xtype->klass_is_exact(), "");
  
-     if (!receiver_type->higher_equal(recvx_type)) { // ignore redundant casts
+     if (!receiver_type->higher_equal(recv_xtype)) { // ignore redundant casts
        // Subsume downstream occurrences of receiver with a cast to
        // recv_xtype, since now we know what the type will be.
-       Node* cast = new CheckCastPPNode(control(), receiver, recvx_type);
-       (*casted_receiver) = _gvn.transform(cast);
+       Node* cast = new CheckCastPPNode(control(), receiver, recv_xtype);
+       Node* res = _gvn.transform(cast);
+       if (recv_xtype->is_inlinetypeptr()) {
+         assert(!gvn().type(res)->maybe_null(), "receiver should never be null");
+         res = InlineTypeNode::make_from_oop(this, res, recv_xtype->inline_klass());
+       }
+       (*casted_receiver) = res;
        assert(!(*casted_receiver)->is_top(), "that path should be unreachable");
        // (User must make the replace_in_map call.)
      }
    }
  
    return fail;
  }
  
+ Node* GraphKit::type_check(Node* recv_klass, const TypeKlassPtr* tklass,
+                            float prob) {
+   Node* want_klass = makecon(tklass);
+   Node* cmp = _gvn.transform(new CmpPNode(recv_klass, want_klass));
+   Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
+   IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN);
+   set_control(_gvn.transform(new IfTrueNode (iff)));
+   Node* fail = _gvn.transform(new IfFalseNode(iff));
+   return fail;
+ }
+ 
  //------------------------------subtype_check_receiver-------------------------
  Node* GraphKit::subtype_check_receiver(Node* receiver, ciKlass* klass,
                                         Node** casted_receiver) {
    const TypeKlassPtr* tklass = TypeKlassPtr::make(klass, Type::trust_interfaces)->try_improve();
    Node* want_klass = makecon(tklass);

@@ -2949,13 +3129,16 @@
  
    // Ignore interface type information until interface types are properly tracked.
    if (!stopped() && !klass->is_interface()) {
      const TypeOopPtr* receiver_type = _gvn.type(receiver)->isa_oopptr();
      const TypeOopPtr* recv_type = tklass->cast_to_exactness(false)->is_klassptr()->as_instance_type();
-     if (!receiver_type->higher_equal(recv_type)) { // ignore redundant casts
-       Node* cast = new CheckCastPPNode(control(), receiver, recv_type);
-       (*casted_receiver) = _gvn.transform(cast);
+     if (receiver_type != nullptr && !receiver_type->higher_equal(recv_type)) { // ignore redundant casts
+       Node* cast = _gvn.transform(new CheckCastPPNode(control(), receiver, recv_type));
+       if (recv_type->is_inlinetypeptr()) {
+         cast = InlineTypeNode::make_from_oop(this, cast, recv_type->inline_klass());
+       }
+       (*casted_receiver) = cast;
      }
    }
  
    return slow_ctl;
  }

@@ -3060,11 +3243,24 @@
  
    // (No, this isn't a call, but it's enough like a virtual call
    // to use the same ciMethod accessor to get the profile info...)
    // If we have a speculative type use it instead of profiling (which
    // may not help us)
-   ciKlass* exact_kls = spec_klass == nullptr ? profile_has_unique_klass() : spec_klass;
+   ciKlass* exact_kls = spec_klass;
+   if (exact_kls == nullptr) {
+     if (java_bc() == Bytecodes::_aastore) {
+       ciKlass* array_type = nullptr;
+       ciKlass* element_type = nullptr;
+       ProfilePtrKind element_ptr = ProfileMaybeNull;
+       bool flat_array = true;
+       bool null_free_array = true;
+       method()->array_access_profiled_type(bci(), array_type, element_type, element_ptr, flat_array, null_free_array);
+       exact_kls = element_type;
+     } else {
+       exact_kls = profile_has_unique_klass();
+     }
+   }
    if (exact_kls != nullptr) {// no cast failures here
      if (require_klass == nullptr ||
          C->static_subtype_check(require_klass, TypeKlassPtr::make(exact_kls, Type::trust_interfaces)) == Compile::SSC_always_true) {
        // If we narrow the type to match what the type profile sees or
        // the speculative type, we can then remove the rest of the

@@ -3190,11 +3386,11 @@
    // Do we know the type check always succeed?
    bool known_statically = false;
    if (_gvn.type(superklass)->singleton()) {
      const TypeKlassPtr* superk = _gvn.type(superklass)->is_klassptr();
      const TypeKlassPtr* subk = _gvn.type(obj)->is_oopptr()->as_klass_type();
-     if (subk->is_loaded()) {
+     if (subk != nullptr && subk->is_loaded()) {
        int static_res = C->static_subtype_check(superk, subk);
        known_statically = (static_res == Compile::SSC_always_true || static_res == Compile::SSC_always_false);
      }
    }
  

@@ -3246,35 +3442,52 @@
  // array store bytecode.  Stack must be as-if BEFORE doing the bytecode so the
  // uncommon-trap paths work.  Adjust stack after this call.
  // If failure_control is supplied and not null, it is filled in with
  // the control edge for the cast failure.  Otherwise, an appropriate
  // uncommon trap or exception is thrown.
- Node* GraphKit::gen_checkcast(Node *obj, Node* superklass,
-                               Node* *failure_control) {
+ Node* GraphKit::gen_checkcast(Node *obj, Node* superklass, Node* *failure_control, bool null_free) {
    kill_dead_locals();           // Benefit all the uncommon traps
    const TypeKlassPtr *tk = _gvn.type(superklass)->is_klassptr()->try_improve();
    const TypeOopPtr *toop = tk->cast_to_exactness(false)->as_instance_type();
+   bool safe_for_replace = (failure_control == nullptr);
+   assert(!null_free || toop->is_inlinetypeptr(), "must be an inline type pointer");
  
    // Fast cutout:  Check the case that the cast is vacuously true.
    // This detects the common cases where the test will short-circuit
    // away completely.  We do this before we perform the null check,
    // because if the test is going to turn into zero code, we don't
    // want a residual null check left around.  (Causes a slowdown,
    // for example, in some objArray manipulations, such as a[i]=a[j].)
    if (tk->singleton()) {
-     const TypeOopPtr* objtp = _gvn.type(obj)->isa_oopptr();
-     if (objtp != nullptr) {
-       switch (C->static_subtype_check(tk, objtp->as_klass_type())) {
+     const TypeKlassPtr* kptr = nullptr;
+     const Type* t = _gvn.type(obj);
+     if (t->isa_oop_ptr()) {
+       kptr = t->is_oopptr()->as_klass_type();
+     } else if (obj->is_InlineType()) {
+       ciInlineKlass* vk = t->inline_klass();
+       kptr = TypeInstKlassPtr::make(TypePtr::NotNull, vk, Type::Offset(0));
+     }
+     if (kptr != nullptr) {
+       switch (C->static_subtype_check(tk, kptr)) {
        case Compile::SSC_always_true:
          // If we know the type check always succeed then we don't use
          // the profiling data at this bytecode. Don't lose it, feed it
          // to the type system as a speculative type.
-         return record_profiled_receiver_for_speculation(obj);
+         obj = record_profiled_receiver_for_speculation(obj);
+         if (null_free) {
+           assert(safe_for_replace, "must be");
+           obj = null_check(obj);
+         }
+         assert(stopped() || !toop->is_inlinetypeptr() || obj->is_InlineType(), "should have been scalarized");
+         return obj;
        case Compile::SSC_always_false:
+         if (null_free) {
+           assert(safe_for_replace, "must be");
+           obj = null_check(obj);
+         }
          // It needs a null check because a null will *pass* the cast check.
-         // A non-null value will always produce an exception.
-         if (!objtp->maybe_null()) {
+         if (t->isa_oopptr() != nullptr && !t->is_oopptr()->maybe_null()) {
            bool is_aastore = (java_bc() == Bytecodes::_aastore);
            Deoptimization::DeoptReason reason = is_aastore ?
              Deoptimization::Reason_array_check : Deoptimization::Reason_class_check;
            builtin_throw(reason);
            return top();

@@ -3287,37 +3500,56 @@
        }
      }
    }
  
    ciProfileData* data = nullptr;
-   bool safe_for_replace = false;
    if (failure_control == nullptr) {        // use MDO in regular case only
      assert(java_bc() == Bytecodes::_aastore ||
             java_bc() == Bytecodes::_checkcast,
             "interpreter profiles type checks only for these BCs");
-     data = method()->method_data()->bci_to_data(bci());
-     safe_for_replace = true;
+     if (method()->method_data()->is_mature()) {
+       data = method()->method_data()->bci_to_data(bci());
+     }
    }
  
    // Make the merge point
    enum { _obj_path = 1, _null_path, PATH_LIMIT };
    RegionNode* region = new RegionNode(PATH_LIMIT);
    Node*       phi    = new PhiNode(region, toop);
+   _gvn.set_type(region, Type::CONTROL);
+   _gvn.set_type(phi, toop);
+ 
    C->set_has_split_ifs(true); // Has chance for split-if optimization
  
    // Use null-cast information if it is available
    bool speculative_not_null = false;
    bool never_see_null = ((failure_control == nullptr)  // regular case only
                           && seems_never_null(obj, data, speculative_not_null));
  
+   if (obj->is_InlineType()) {
+     // Re-execute if buffering during triggers deoptimization
+     PreserveReexecuteState preexecs(this);
+     jvms()->set_should_reexecute(true);
+     obj = obj->as_InlineType()->buffer(this, safe_for_replace);
+   }
+ 
    // Null check; get casted pointer; set region slot 3
    Node* null_ctl = top();
-   Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null);
+   Node* not_null_obj = nullptr;
+   if (null_free) {
+     assert(safe_for_replace, "must be");
+     not_null_obj = null_check(obj);
+   } else {
+     not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null);
+   }
  
    // If not_null_obj is dead, only null-path is taken
    if (stopped()) {              // Doing instance-of on a null?
      set_control(null_ctl);
+     if (toop->is_inlinetypeptr()) {
+       return InlineTypeNode::make_null(_gvn, toop->inline_klass());
+     }
      return null();
    }
    region->init_req(_null_path, null_ctl);
    phi   ->init_req(_null_path, null());  // Set null path value
    if (null_ctl == top()) {

@@ -3349,19 +3581,25 @@
      }
    }
  
    if (cast_obj == nullptr) {
      // Generate the subtype check
-     Node* not_subtype_ctrl = gen_subtype_check(not_null_obj, superklass );
+     Node* not_subtype_ctrl = gen_subtype_check(not_null_obj, superklass);
  
      // Plug in success path into the merge
      cast_obj = _gvn.transform(new CheckCastPPNode(control(), not_null_obj, toop));
      // Failure path ends in uncommon trap (or may be dead - failure impossible)
      if (failure_control == nullptr) {
        if (not_subtype_ctrl != top()) { // If failure is possible
          PreserveJVMState pjvms(this);
          set_control(not_subtype_ctrl);
+         Node* obj_klass = nullptr;
+         if (not_null_obj->is_InlineType()) {
+           obj_klass = makecon(TypeKlassPtr::make(_gvn.type(not_null_obj)->inline_klass()));
+         } else {
+           obj_klass = load_object_klass(not_null_obj);
+         }
          bool is_aastore = (java_bc() == Bytecodes::_aastore);
          Deoptimization::DeoptReason reason = is_aastore ?
            Deoptimization::Reason_array_check : Deoptimization::Reason_class_check;
          builtin_throw(reason);
        }

@@ -3388,11 +3626,126 @@
  
    // Return final merged results
    set_control( _gvn.transform(region) );
    record_for_igvn(region);
  
-   return record_profiled_receiver_for_speculation(res);
+   bool not_inline = !toop->can_be_inline_type();
+   bool not_flat_in_array = !UseFlatArray || not_inline || (toop->is_inlinetypeptr() && !toop->inline_klass()->flat_in_array());
+   if (EnableValhalla && not_flat_in_array) {
+     // Check if obj has been loaded from an array
+     obj = obj->isa_DecodeN() ? obj->in(1) : obj;
+     Node* array = nullptr;
+     if (obj->isa_Load()) {
+       Node* address = obj->in(MemNode::Address);
+       if (address->isa_AddP()) {
+         array = address->as_AddP()->in(AddPNode::Base);
+       }
+     } else if (obj->is_Phi()) {
+       Node* region = obj->in(0);
+       // TODO make this more robust (see JDK-8231346)
+       if (region->req() == 3 && region->in(2) != nullptr && region->in(2)->in(0) != nullptr) {
+         IfNode* iff = region->in(2)->in(0)->isa_If();
+         if (iff != nullptr) {
+           iff->is_flat_array_check(&_gvn, &array);
+         }
+       }
+     }
+     if (array != nullptr) {
+       const TypeAryPtr* ary_t = _gvn.type(array)->isa_aryptr();
+       if (ary_t != nullptr && !ary_t->is_flat()) {
+         if (!ary_t->is_not_null_free() && not_inline) {
+           // Casting array element to a non-inline-type, mark array as not null-free.
+           Node* cast = _gvn.transform(new CheckCastPPNode(control(), array, ary_t->cast_to_not_null_free()));
+           replace_in_map(array, cast);
+         } else if (!ary_t->is_not_flat()) {
+           // Casting array element to a non-flat type, mark array as not flat.
+           Node* cast = _gvn.transform(new CheckCastPPNode(control(), array, ary_t->cast_to_not_flat()));
+           replace_in_map(array, cast);
+         }
+       }
+     }
+   }
+ 
+   if (!stopped() && !res->is_InlineType()) {
+     res = record_profiled_receiver_for_speculation(res);
+     if (toop->is_inlinetypeptr()) {
+       Node* vt = InlineTypeNode::make_from_oop(this, res, toop->inline_klass(), !gvn().type(res)->maybe_null());
+       res = vt;
+       if (safe_for_replace) {
+         replace_in_map(obj, vt);
+         replace_in_map(not_null_obj, vt);
+         replace_in_map(res, vt);
+       }
+     }
+   }
+   return res;
+ }
+ 
+ Node* GraphKit::inline_type_test(Node* obj, bool is_inline) {
+   Node* mark_adr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
+   Node* mark = make_load(nullptr, mark_adr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered);
+   Node* mask = MakeConX(markWord::inline_type_pattern);
+   Node* masked = _gvn.transform(new AndXNode(mark, mask));
+   Node* cmp = _gvn.transform(new CmpXNode(masked, mask));
+   return _gvn.transform(new BoolNode(cmp, is_inline ? BoolTest::eq : BoolTest::ne));
+ }
+ 
+ Node* GraphKit::array_lh_test(Node* klass, jint mask, jint val, bool eq) {
+   Node* lh_adr = basic_plus_adr(klass, in_bytes(Klass::layout_helper_offset()));
+   // Make sure to use immutable memory here to enable hoisting the check out of loops
+   Node* lh_val = _gvn.transform(LoadNode::make(_gvn, nullptr, immutable_memory(), lh_adr, lh_adr->bottom_type()->is_ptr(), TypeInt::INT, T_INT, MemNode::unordered));
+   Node* masked = _gvn.transform(new AndINode(lh_val, intcon(mask)));
+   Node* cmp = _gvn.transform(new CmpINode(masked, intcon(val)));
+   return _gvn.transform(new BoolNode(cmp, eq ? BoolTest::eq : BoolTest::ne));
+ }
+ 
+ // TODO 8325106 With JEP 401, flatness is not a property of the Class anymore.
+ Node* GraphKit::flat_array_test(Node* array_or_klass, bool flat) {
+   // We can't use immutable memory here because the mark word is mutable.
+   // PhaseIdealLoop::move_flat_array_check_out_of_loop will make sure the
+   // check is moved out of loops (mainly to enable loop unswitching).
+   Node* mem = UseArrayMarkWordCheck ? memory(Compile::AliasIdxRaw) : immutable_memory();
+   Node* cmp = _gvn.transform(new FlatArrayCheckNode(C, mem, array_or_klass));
+   record_for_igvn(cmp); // Give it a chance to be optimized out by IGVN
+   return _gvn.transform(new BoolNode(cmp, flat ? BoolTest::eq : BoolTest::ne));
+ }
+ 
+ Node* GraphKit::null_free_array_test(Node* klass, bool null_free) {
+   return array_lh_test(klass, Klass::_lh_null_free_array_bit_inplace, 0, !null_free);
+ }
+ 
+ // Deoptimize if 'ary' is a null-free inline type array and 'val' is null
+ Node* GraphKit::inline_array_null_guard(Node* ary, Node* val, int nargs, bool safe_for_replace) {
+   RegionNode* region = new RegionNode(3);
+   Node* null_ctl = top();
+   null_check_oop(val, &null_ctl);
+   if (null_ctl != top()) {
+     PreserveJVMState pjvms(this);
+     set_control(null_ctl);
+     {
+       // Deoptimize if null-free array
+       BuildCutout unless(this, null_free_array_test(load_object_klass(ary), /* null_free = */ false), PROB_MAX);
+       inc_sp(nargs);
+       uncommon_trap(Deoptimization::Reason_null_check,
+                     Deoptimization::Action_none);
+     }
+     region->init_req(1, control());
+   }
+   region->init_req(2, control());
+   set_control(_gvn.transform(region));
+   record_for_igvn(region);
+   if (_gvn.type(val) == TypePtr::NULL_PTR) {
+     // Since we were just successfully storing null, the array can't be null free.
+     const TypeAryPtr* ary_t = _gvn.type(ary)->is_aryptr();
+     ary_t = ary_t->cast_to_not_null_free();
+     Node* cast = _gvn.transform(new CheckCastPPNode(control(), ary, ary_t));
+     if (safe_for_replace) {
+       replace_in_map(ary, cast);
+     }
+     ary = cast;
+   }
+   return ary;
  }
  
  //------------------------------next_monitor-----------------------------------
  // What number should be given to the next monitor?
  int GraphKit::next_monitor() {

@@ -3456,10 +3809,11 @@
    // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
    assert(SynchronizationEntryBCI == InvocationEntryBci, "");
  
    if( !GenerateSynchronizationCode )
      return nullptr;                // Not locking things?
+ 
    if (stopped())                // Dead monitor?
      return nullptr;
  
    assert(dead_locals_are_killed(), "should kill locals before sync. point");
  

@@ -3524,10 +3878,11 @@
      return;
    if (stopped()) {               // Dead monitor?
      map()->pop_monitor();        // Kill monitor from debug info
      return;
    }
+   assert(!obj->is_InlineType(), "should not unlock on inline type");
  
    // Memory barrier to avoid floating things down past the locked region
    insert_mem_bar(Op_MemBarReleaseLock);
  
    const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type();

@@ -3564,14 +3919,24 @@
  // almost always feature constant types.
  Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) {
    const TypeKlassPtr* klass_t = _gvn.type(klass_node)->isa_klassptr();
    if (!StressReflectiveCode && klass_t != nullptr) {
      bool xklass = klass_t->klass_is_exact();
-     if (xklass || (klass_t->isa_aryklassptr() && klass_t->is_aryklassptr()->elem() != Type::BOTTOM)) {
+     bool can_be_flat = false;
+     const TypeAryPtr* ary_type = klass_t->as_instance_type()->isa_aryptr();
+     if (UseFlatArray && !xklass && ary_type != nullptr && !ary_type->is_null_free()) {
+       // TODO 8325106 Fix comment
+       // The runtime type of [LMyValue might be [QMyValue due to [QMyValue <: [LMyValue. Don't constant fold.
+       const TypeOopPtr* elem = ary_type->elem()->make_oopptr();
+       can_be_flat = ary_type->can_be_inline_array() && (!elem->is_inlinetypeptr() || elem->inline_klass()->flat_in_array());
+     }
+     if (!can_be_flat && (xklass || (klass_t->isa_aryklassptr() && klass_t->is_aryklassptr()->elem() != Type::BOTTOM))) {
        jint lhelper;
-       if (klass_t->isa_aryklassptr()) {
-         BasicType elem = klass_t->as_instance_type()->isa_aryptr()->elem()->array_element_basic_type();
+       if (klass_t->is_flat()) {
+         lhelper = ary_type->flat_layout_helper();
+       } else if (klass_t->isa_aryklassptr()) {
+         BasicType elem = ary_type->elem()->array_element_basic_type();
          if (is_reference_type(elem, true)) {
            elem = T_OBJECT;
          }
          lhelper = Klass::array_layout_helper(elem);
        } else {

@@ -3596,11 +3961,13 @@
    DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory());
    assert(init_in_merge->memory_at(alias_idx) == init_in_raw, "");
  
    Node* prevmem = kit.memory(alias_idx);
    init_in_merge->set_memory_at(alias_idx, prevmem);
-   kit.set_memory(init_out_raw, alias_idx);
+   if (init_out_raw != nullptr) {
+     kit.set_memory(init_out_raw, alias_idx);
+   }
  }
  
  //---------------------------set_output_for_allocation-------------------------
  Node* GraphKit::set_output_for_allocation(AllocateNode* alloc,
                                            const TypeOopPtr* oop_type,

@@ -3635,21 +4002,48 @@
      // and link them properly (as a group) to the InitializeNode.
      assert(init->in(InitializeNode::Memory) == malloc, "");
      MergeMemNode* minit_in = MergeMemNode::make(malloc);
      init->set_req(InitializeNode::Memory, minit_in);
      record_for_igvn(minit_in); // fold it up later, if possible
+     _gvn.set_type(minit_in, Type::MEMORY);
      Node* minit_out = memory(rawidx);
      assert(minit_out->is_Proj() && minit_out->in(0) == init, "");
      // Add an edge in the MergeMem for the header fields so an access
      // to one of those has correct memory state
      set_memory(minit_out, C->get_alias_index(oop_type->add_offset(oopDesc::mark_offset_in_bytes())));
      set_memory(minit_out, C->get_alias_index(oop_type->add_offset(oopDesc::klass_offset_in_bytes())));
      if (oop_type->isa_aryptr()) {
-       const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot);
-       int            elemidx  = C->get_alias_index(telemref);
-       hook_memory_on_init(*this, elemidx, minit_in, minit_out);
+       const TypeAryPtr* arytype = oop_type->is_aryptr();
+       if (arytype->is_flat()) {
+         // Initially all flat array accesses share a single slice
+         // but that changes after parsing. Prepare the memory graph so
+         // it can optimize flat array accesses properly once they
+         // don't share a single slice.
+         assert(C->flat_accesses_share_alias(), "should be set at parse time");
+         C->set_flat_accesses_share_alias(false);
+         ciInlineKlass* vk = arytype->elem()->inline_klass();
+         for (int i = 0, len = vk->nof_nonstatic_fields(); i < len; i++) {
+           ciField* field = vk->nonstatic_field_at(i);
+           if (field->offset_in_bytes() >= TrackedInitializationLimit * HeapWordSize)
+             continue;  // do not bother to track really large numbers of fields
+           int off_in_vt = field->offset_in_bytes() - vk->first_field_offset();
+           const TypePtr* adr_type = arytype->with_field_offset(off_in_vt)->add_offset(Type::OffsetBot);
+           int fieldidx = C->get_alias_index(adr_type, true);
+           // Pass nullptr for init_out. Having per flat array element field memory edges as uses of the Initialize node
+           // can result in per flat array field Phis to be created which confuses the logic of
+           // Compile::adjust_flat_array_access_aliases().
+           hook_memory_on_init(*this, fieldidx, minit_in, nullptr);
+         }
+         C->set_flat_accesses_share_alias(true);
+         hook_memory_on_init(*this, C->get_alias_index(TypeAryPtr::INLINES), minit_in, minit_out);
+       } else {
+         const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot);
+         int            elemidx  = C->get_alias_index(telemref);
+         hook_memory_on_init(*this, elemidx, minit_in, minit_out);
+       }
      } else if (oop_type->isa_instptr()) {
+       set_memory(minit_out, C->get_alias_index(oop_type)); // mark word
        ciInstanceKlass* ik = oop_type->is_instptr()->instance_klass();
        for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) {
          ciField* field = ik->nonstatic_field_at(i);
          if (field->offset_in_bytes() >= TrackedInitializationLimit * HeapWordSize)
            continue;  // do not bother to track really large numbers of fields

@@ -3696,18 +4090,19 @@
  //  - If 'return_size_val', report the total object size to the caller.
  //  - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize)
  Node* GraphKit::new_instance(Node* klass_node,
                               Node* extra_slow_test,
                               Node* *return_size_val,
-                              bool deoptimize_on_exception) {
+                              bool deoptimize_on_exception,
+                              InlineTypeNode* inline_type_node) {
    // Compute size in doublewords
    // The size is always an integral number of doublewords, represented
    // as a positive bytewise size stored in the klass's layout_helper.
    // The layout_helper also encodes (in a low bit) the need for a slow path.
    jint  layout_con = Klass::_lh_neutral_value;
    Node* layout_val = get_layout_helper(klass_node, layout_con);
-   int   layout_is_con = (layout_val == nullptr);
+   bool  layout_is_con = (layout_val == nullptr);
  
    if (extra_slow_test == nullptr)  extra_slow_test = intcon(0);
    // Generate the initial go-slow test.  It's either ALWAYS (return a
    // Node for 1) or NEVER (return a null) or perhaps (in the reflective
    // case) a computed value derived from the layout_helper.

@@ -3754,24 +4149,24 @@
    const TypeOopPtr* oop_type = tklass->as_instance_type();
  
    // Now generate allocation code
  
    // The entire memory state is needed for slow path of the allocation
-   // since GC and deoptimization can happened.
+   // since GC and deoptimization can happen.
    Node *mem = reset_memory();
    set_all_memory(mem); // Create new memory state
  
    AllocateNode* alloc = new AllocateNode(C, AllocateNode::alloc_type(Type::TOP),
                                           control(), mem, i_o(),
                                           size, klass_node,
-                                          initial_slow_test);
+                                          initial_slow_test, inline_type_node);
  
    return set_output_for_allocation(alloc, oop_type, deoptimize_on_exception);
  }
  
  //-------------------------------new_array-------------------------------------
- // helper for both newarray and anewarray
+ // helper for newarray and anewarray
  // The 'length' parameter is (obviously) the length of the array.
  // The optional arguments are for specialized use by intrinsics:
  //  - If 'return_size_val', report the non-padded array size (sum of header size
  //    and array body) to the caller.
  //  - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize)

@@ -3780,11 +4175,11 @@
                            int   nargs,          // number of arguments to push back for uncommon trap
                            Node* *return_size_val,
                            bool deoptimize_on_exception) {
    jint  layout_con = Klass::_lh_neutral_value;
    Node* layout_val = get_layout_helper(klass_node, layout_con);
-   int   layout_is_con = (layout_val == nullptr);
+   bool  layout_is_con = (layout_val == nullptr);
  
    if (!layout_is_con && !StressReflectiveCode &&
        !too_many_traps(Deoptimization::Reason_class_check)) {
      // This is a reflective array creation site.
      // Optimistically assume that it is a subtype of Object[],

@@ -3810,11 +4205,11 @@
    int fast_size_limit = FastAllocateSizeLimit;
    if (layout_is_con) {
      assert(!StressReflectiveCode, "stress mode does not use these paths");
      // Increase the size limit if we have exact knowledge of array type.
      int log2_esize = Klass::layout_helper_log2_element_size(layout_con);
-     fast_size_limit <<= (LogBytesPerLong - log2_esize);
+     fast_size_limit <<= MAX2(LogBytesPerLong - log2_esize, 0);
    }
  
    Node* initial_slow_cmp  = _gvn.transform( new CmpUNode( length, intcon( fast_size_limit ) ) );
    Node* initial_slow_test = _gvn.transform( new BoolNode( initial_slow_cmp, BoolTest::gt ) );
  

@@ -3827,13 +4222,14 @@
    Node* header_size = nullptr;
    // (T_BYTE has the weakest alignment and size restrictions...)
    if (layout_is_con) {
      int       hsize  = Klass::layout_helper_header_size(layout_con);
      int       eshift = Klass::layout_helper_log2_element_size(layout_con);
+     bool is_flat_array = Klass::layout_helper_is_flatArray(layout_con);
      if ((round_mask & ~right_n_bits(eshift)) == 0)
        round_mask = 0;  // strength-reduce it if it goes away completely
-     assert((hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded");
+     assert(is_flat_array || (hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded");
      int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
      assert(header_size_min <= hsize, "generic minimum is smallest");
      header_size = intcon(hsize);
    } else {
      Node* hss   = intcon(Klass::_lh_header_size_shift);

@@ -3916,20 +4312,49 @@
    // else if round_mask == 0, the size computation is self-rounding
  
    // Now generate allocation code
  
    // The entire memory state is needed for slow path of the allocation
-   // since GC and deoptimization can happened.
+   // since GC and deoptimization can happen.
    Node *mem = reset_memory();
    set_all_memory(mem); // Create new memory state
  
    if (initial_slow_test->is_Bool()) {
      // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick.
      initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn);
    }
  
-   const TypeOopPtr* ary_type = _gvn.type(klass_node)->is_klassptr()->as_instance_type();
+   const TypeKlassPtr* ary_klass = _gvn.type(klass_node)->isa_klassptr();
+   const TypeOopPtr* ary_type = ary_klass->as_instance_type();
+   const TypeAryPtr* ary_ptr = ary_type->isa_aryptr();
+ 
+   // TODO 8325106 Fix comment
+   // Inline type array variants:
+   // - null-ok:              MyValue.ref[] (ciObjArrayKlass "[LMyValue")
+   // - null-free:            MyValue.val[] (ciObjArrayKlass "[QMyValue")
+   // - null-free, flat     : MyValue.val[] (ciFlatArrayKlass "[QMyValue")
+   // Check if array is a null-free, non-flat inline type array
+   // that needs to be initialized with the default inline type.
+   Node* default_value = nullptr;
+   Node* raw_default_value = nullptr;
+   if (ary_ptr != nullptr && ary_ptr->klass_is_exact()) {
+     // Array type is known
+     if (ary_ptr->is_null_free() && !ary_ptr->is_flat()) {
+       ciInlineKlass* vk = ary_ptr->elem()->inline_klass();
+       default_value = InlineTypeNode::default_oop(gvn(), vk);
+       if (UseCompressedOops) {
+         // With compressed oops, the 64-bit init value is built from two 32-bit compressed oops
+         default_value = _gvn.transform(new EncodePNode(default_value, default_value->bottom_type()->make_narrowoop()));
+         Node* lower = _gvn.transform(new CastP2XNode(control(), default_value));
+         Node* upper = _gvn.transform(new LShiftLNode(lower, intcon(32)));
+         raw_default_value = _gvn.transform(new OrLNode(lower, upper));
+       } else {
+         raw_default_value = _gvn.transform(new CastP2XNode(control(), default_value));
+       }
+     }
+   }
+ 
    Node* valid_length_test = _gvn.intcon(1);
    if (ary_type->isa_aryptr()) {
      BasicType bt = ary_type->isa_aryptr()->elem()->array_element_basic_type();
      jint max = TypeAryPtr::max_array_length(bt);
      Node* valid_length_cmp  = _gvn.transform(new CmpUNode(length, intcon(max)));

@@ -3940,12 +4365,12 @@
    AllocateArrayNode* alloc
      = new AllocateArrayNode(C, AllocateArrayNode::alloc_type(TypeInt::INT),
                              control(), mem, i_o(),
                              size, klass_node,
                              initial_slow_test,
-                             length, valid_length_test);
- 
+                             length, valid_length_test,
+                             default_value, raw_default_value);
    // Cast to correct type.  Note that the klass_node may be constant or not,
    // and in the latter case the actual array type will be inexact also.
    // (This happens via a non-constant argument to inline_native_newArray.)
    // In any case, the value of klass_node provides the desired array type.
    const TypeInt* length_type = _gvn.find_int_type(length);

@@ -4088,15 +4513,15 @@
  }
  
  Node* GraphKit::load_String_value(Node* str, bool set_ctrl) {
    int value_offset = java_lang_String::value_offset();
    const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
-                                                      false, nullptr, 0);
+                                                      false, nullptr, Type::Offset(0));
    const TypePtr* value_field_type = string_type->add_offset(value_offset);
    const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::NotNull,
-                                                   TypeAry::make(TypeInt::BYTE, TypeInt::POS),
-                                                   ciTypeArrayKlass::make(T_BYTE), true, 0);
+                                                   TypeAry::make(TypeInt::BYTE, TypeInt::POS, false, false, true, true),
+                                                   ciTypeArrayKlass::make(T_BYTE), true, Type::Offset(0));
    Node* p = basic_plus_adr(str, str, value_offset);
    Node* load = access_load_at(str, p, value_field_type, value_type, T_OBJECT,
                                IN_HEAP | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0) | MO_UNORDERED);
    return load;
  }

@@ -4105,11 +4530,11 @@
    if (!CompactStrings) {
      return intcon(java_lang_String::CODER_UTF16);
    }
    int coder_offset = java_lang_String::coder_offset();
    const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
-                                                      false, nullptr, 0);
+                                                      false, nullptr, Type::Offset(0));
    const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
  
    Node* p = basic_plus_adr(str, str, coder_offset);
    Node* load = access_load_at(str, p, coder_field_type, TypeInt::BYTE, T_BYTE,
                                IN_HEAP | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0) | MO_UNORDERED);

@@ -4117,21 +4542,21 @@
  }
  
  void GraphKit::store_String_value(Node* str, Node* value) {
    int value_offset = java_lang_String::value_offset();
    const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
-                                                      false, nullptr, 0);
+                                                      false, nullptr, Type::Offset(0));
    const TypePtr* value_field_type = string_type->add_offset(value_offset);
  
    access_store_at(str,  basic_plus_adr(str, value_offset), value_field_type,
                    value, TypeAryPtr::BYTES, T_OBJECT, IN_HEAP | MO_UNORDERED);
  }
  
  void GraphKit::store_String_coder(Node* str, Node* value) {
    int coder_offset = java_lang_String::coder_offset();
    const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
-                                                      false, nullptr, 0);
+                                                      false, nullptr, Type::Offset(0));
    const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
  
    access_store_at(str, basic_plus_adr(str, coder_offset), coder_field_type,
                    value, TypeInt::BYTE, T_BYTE, IN_HEAP | MO_UNORDERED);
  }

@@ -4240,9 +4665,24 @@
      }
    }
    const Type* con_type = Type::make_constant_from_field(field, holder, field->layout_type(),
                                                          /*is_unsigned_load=*/false);
    if (con_type != nullptr) {
-     return makecon(con_type);
+     Node* con = makecon(con_type);
+     if (field->type()->is_inlinetype()) {
+       con = InlineTypeNode::make_from_oop(this, con, field->type()->as_inline_klass(), field->is_null_free());
+     } else if (con_type->is_inlinetypeptr()) {
+       con = InlineTypeNode::make_from_oop(this, con, con_type->inline_klass(), field->is_null_free());
+     }
+     return con;
    }
    return nullptr;
  }
+ 
+ //---------------------------load_mirror_from_klass----------------------------
+ // Given a klass oop, load its java mirror (a java.lang.Class oop).
+ Node* GraphKit::load_mirror_from_klass(Node* klass) {
+   Node* p = basic_plus_adr(klass, in_bytes(Klass::java_mirror_offset()));
+   Node* load = make_load(nullptr, p, TypeRawPtr::NOTNULL, T_ADDRESS, MemNode::unordered);
+   // mirror = ((OopHandle)mirror)->resolve();
+   return access_load(load, TypeInstPtr::MIRROR, T_OBJECT, IN_NATIVE);
+ }
< prev index next >