1 
   2 
   3 /*
   4  * Copyright (c) 1997, 2020, Oracle and/or its affiliates. All rights reserved.
   5  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   6  *
   7  * This code is free software; you can redistribute it and/or modify it
   8  * under the terms of the GNU General Public License version 2 only, as
   9  * published by the Free Software Foundation.
  10  *
  11  * This code is distributed in the hope that it will be useful, but WITHOUT
  12  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  13  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  14  * version 2 for more details (a copy is included in the LICENSE file that
  15  * accompanied this code).
  16  *
  17  * You should have received a copy of the GNU General Public License version
  18  * 2 along with this work; if not, write to the Free Software Foundation,
  19  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
  20  *
  21  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
  22  * or visit www.oracle.com if you need additional information or have any
  23  * questions.
  24  *
  25  */
  26 
  27 #include "precompiled.hpp"
  28 #include "jvm.h"
  29 #include "classfile/javaClasses.inline.hpp"
  30 #include "classfile/symbolTable.hpp"
  31 #include "classfile/systemDictionary.hpp"
  32 #include "code/codeCache.hpp"
  33 #include "code/debugInfoRec.hpp"
  34 #include "code/nmethod.hpp"
  35 #include "code/pcDesc.hpp"
  36 #include "code/scopeDesc.hpp"
  37 #include "compiler/compilationPolicy.hpp"
  38 #include "interpreter/bytecode.hpp"
  39 #include "interpreter/interpreter.hpp"
  40 #include "interpreter/oopMapCache.hpp"
  41 #include "memory/allocation.inline.hpp"
  42 #include "memory/oopFactory.hpp"
  43 #include "memory/resourceArea.hpp"
  44 #include "memory/universe.hpp"
  45 #include "oops/constantPool.hpp"
  46 #include "oops/method.hpp"
  47 #include "oops/objArrayKlass.hpp"
  48 #include "oops/objArrayOop.inline.hpp"
  49 #include "oops/oop.inline.hpp"
  50 #include "oops/fieldStreams.inline.hpp"
  51 #include "oops/typeArrayOop.inline.hpp"
  52 #include "oops/verifyOopClosure.hpp"
  53 #include "prims/jvmtiThreadState.hpp"
  54 #include "runtime/atomic.hpp"
  55 #include "runtime/biasedLocking.hpp"
  56 #include "runtime/deoptimization.hpp"
  57 #include "runtime/fieldDescriptor.hpp"
  58 #include "runtime/fieldDescriptor.inline.hpp"
  59 #include "runtime/frame.inline.hpp"
  60 #include "runtime/handles.inline.hpp"
  61 #include "runtime/interfaceSupport.inline.hpp"
  62 #include "runtime/jniHandles.inline.hpp"
  63 #include "runtime/objectMonitor.inline.hpp"
  64 #include "runtime/safepointVerifiers.hpp"
  65 #include "runtime/sharedRuntime.hpp"
  66 #include "runtime/signature.hpp"
  67 #include "runtime/stubRoutines.hpp"
  68 #include "runtime/thread.hpp"
  69 #include "runtime/threadSMR.hpp"
  70 #include "runtime/vframe.hpp"
  71 #include "runtime/vframeArray.hpp"
  72 #include "runtime/vframe_hp.hpp"
  73 #include "utilities/events.hpp"
  74 #include "utilities/macros.hpp"
  75 #include "utilities/preserveException.hpp"
  76 #include "utilities/xmlstream.hpp"
  77 #if INCLUDE_JFR
  78 #include "jfr/jfrEvents.hpp"
  79 #include "jfr/metadata/jfrSerializer.hpp"
  80 #endif
  81 
  82 bool DeoptimizationMarker::_is_active = false;
  83 
  84 Deoptimization::UnrollBlock::UnrollBlock(int  size_of_deoptimized_frame,
  85                                          int  caller_adjustment,
  86                                          int  caller_actual_parameters,
  87                                          int  number_of_frames,
  88                                          intptr_t* frame_sizes,
  89                                          address* frame_pcs,
  90                                          BasicType return_type,
  91                                          int exec_mode) {
  92   _size_of_deoptimized_frame = size_of_deoptimized_frame;
  93   _caller_adjustment         = caller_adjustment;
  94   _caller_actual_parameters  = caller_actual_parameters;
  95   _number_of_frames          = number_of_frames;
  96   _frame_sizes               = frame_sizes;
  97   _frame_pcs                 = frame_pcs;
  98   _register_block            = NEW_C_HEAP_ARRAY(intptr_t, RegisterMap::reg_count * 2, mtCompiler);
  99   _return_type               = return_type;
 100   _initial_info              = 0;
 101   // PD (x86 only)
 102   _counter_temp              = 0;
 103   _unpack_kind               = exec_mode;
 104   _sender_sp_temp            = 0;
 105 
 106   _total_frame_sizes         = size_of_frames();
 107   assert(exec_mode >= 0 && exec_mode < Unpack_LIMIT, "Unexpected exec_mode");
 108 }
 109 
 110 
 111 Deoptimization::UnrollBlock::~UnrollBlock() {
 112   FREE_C_HEAP_ARRAY(intptr_t, _frame_sizes);
 113   FREE_C_HEAP_ARRAY(intptr_t, _frame_pcs);
 114   FREE_C_HEAP_ARRAY(intptr_t, _register_block);
 115 }
 116 
 117 
 118 intptr_t* Deoptimization::UnrollBlock::value_addr_at(int register_number) const {
 119   assert(register_number < RegisterMap::reg_count, "checking register number");
 120   return &_register_block[register_number * 2];
 121 }
 122 
 123 
 124 
 125 int Deoptimization::UnrollBlock::size_of_frames() const {
 126   // Acount first for the adjustment of the initial frame
 127   int result = _caller_adjustment;
 128   for (int index = 0; index < number_of_frames(); index++) {
 129     result += frame_sizes()[index];
 130   }
 131   return result;
 132 }
 133 
 134 
 135 void Deoptimization::UnrollBlock::print() {
 136   ttyLocker ttyl;
 137   tty->print_cr("UnrollBlock");
 138   tty->print_cr("  size_of_deoptimized_frame = %d", _size_of_deoptimized_frame);
 139   tty->print(   "  frame_sizes: ");
 140   for (int index = 0; index < number_of_frames(); index++) {
 141     tty->print(INTX_FORMAT " ", frame_sizes()[index]);
 142   }
 143   tty->cr();
 144 }
 145 
 146 
 147 // In order to make fetch_unroll_info work properly with escape
 148 // analysis, The method was changed from JRT_LEAF to JRT_BLOCK_ENTRY and
 149 // ResetNoHandleMark and HandleMark were removed from it. The actual reallocation
 150 // of previously eliminated objects occurs in realloc_objects, which is
 151 // called from the method fetch_unroll_info_helper below.
 152 JRT_BLOCK_ENTRY(Deoptimization::UnrollBlock*, Deoptimization::fetch_unroll_info(JavaThread* thread, int exec_mode))
 153   // It is actually ok to allocate handles in a leaf method. It causes no safepoints,
 154   // but makes the entry a little slower. There is however a little dance we have to
 155   // do in debug mode to get around the NoHandleMark code in the JRT_LEAF macro
 156 
 157   // fetch_unroll_info() is called at the beginning of the deoptimization
 158   // handler. Note this fact before we start generating temporary frames
 159   // that can confuse an asynchronous stack walker. This counter is
 160   // decremented at the end of unpack_frames().
 161   if (TraceDeoptimization) {
 162     tty->print_cr("Deoptimizing thread " INTPTR_FORMAT, p2i(thread));
 163   }
 164   thread->inc_in_deopt_handler();
 165 
 166   return fetch_unroll_info_helper(thread, exec_mode);
 167 JRT_END
 168 
 169 #if COMPILER2_OR_JVMCI
 170 static bool eliminate_allocations(JavaThread* thread, int exec_mode, CompiledMethod* compiled_method,
 171                                   frame& deoptee, RegisterMap& map, GrowableArray<compiledVFrame*>* chunk,
 172                                   bool& deoptimized_objects) {
 173   bool realloc_failures = false;
 174   assert (chunk->at(0)->scope() != NULL,"expect only compiled java frames");
 175 
 176   JavaThread* deoptee_thread = chunk->at(0)->thread();
 177   assert(exec_mode == Deoptimization::Unpack_none || (deoptee_thread == thread), "a frame can only be deoptimized by the owner thread");
 178 
 179   GrowableArray<ScopeValue*>* objects = chunk->at(0)->scope()->objects();
 180 
 181   // The flag return_oop() indicates call sites which return oop
 182   // in compiled code. Such sites include java method calls,
 183   // runtime calls (for example, used to allocate new objects/arrays
 184   // on slow code path) and any other calls generated in compiled code.
 185   // It is not guaranteed that we can get such information here only
 186   // by analyzing bytecode in deoptimized frames. This is why this flag
 187   // is set during method compilation (see Compile::Process_OopMap_Node()).
 188   // If the previous frame was popped or if we are dispatching an exception,
 189   // we don't have an oop result.
 190   bool save_oop_result = chunk->at(0)->scope()->return_oop() && !thread->popframe_forcing_deopt_reexecution() && (exec_mode == Deoptimization::Unpack_deopt);
 191   Handle return_value;
 192   if (save_oop_result) {
 193     // Reallocation may trigger GC. If deoptimization happened on return from
 194     // call which returns oop we need to save it since it is not in oopmap.
 195     oop result = deoptee.saved_oop_result(&map);
 196     assert(oopDesc::is_oop_or_null(result), "must be oop");
 197     return_value = Handle(thread, result);
 198     assert(Universe::heap()->is_in_or_null(result), "must be heap pointer");
 199     if (TraceDeoptimization) {
 200       ttyLocker ttyl;
 201       tty->print_cr("SAVED OOP RESULT " INTPTR_FORMAT " in thread " INTPTR_FORMAT, p2i(result), p2i(thread));
 202     }
 203   }
 204   if (objects != NULL) {
 205     if (exec_mode == Deoptimization::Unpack_none) {
 206       assert(thread->thread_state() == _thread_in_vm, "assumption");
 207       Thread* THREAD = thread;
 208       // Clear pending OOM if reallocation fails and return true indicating allocation failure
 209       realloc_failures = Deoptimization::realloc_objects(thread, &deoptee, &map, objects, CHECK_AND_CLEAR_(true));
 210       deoptimized_objects = true;
 211     } else {
 212       JRT_BLOCK
 213       realloc_failures = Deoptimization::realloc_objects(thread, &deoptee, &map, objects, THREAD);
 214       JRT_END
 215     }
 216     bool skip_internal = (compiled_method != NULL) && !compiled_method->is_compiled_by_jvmci();
 217     Deoptimization::reassign_fields(&deoptee, &map, objects, realloc_failures, skip_internal);
 218 #ifndef PRODUCT
 219     if (TraceDeoptimization) {
 220       ttyLocker ttyl;
 221       tty->print_cr("REALLOC OBJECTS in thread " INTPTR_FORMAT, p2i(deoptee_thread));
 222       Deoptimization::print_objects(objects, realloc_failures);
 223     }
 224 #endif
 225   }
 226   if (save_oop_result) {
 227     // Restore result.
 228     deoptee.set_saved_oop_result(&map, return_value());
 229   }
 230   return realloc_failures;
 231 }
 232 
 233 static void eliminate_locks(JavaThread* thread, GrowableArray<compiledVFrame*>* chunk, bool realloc_failures,
 234                             frame& deoptee, int exec_mode, bool& deoptimized_objects) {
 235   JavaThread* deoptee_thread = chunk->at(0)->thread();
 236   assert(!EscapeBarrier::objs_are_deoptimized(deoptee_thread, deoptee.id()), "must relock just once");
 237 
 238 #ifndef PRODUCT
 239   bool first = true;
 240 #endif
 241   for (int i = 0; i < chunk->length(); i++) {
 242     compiledVFrame* cvf = chunk->at(i);
 243     assert (cvf->scope() != NULL,"expect only compiled java frames");
 244     GrowableArray<MonitorInfo*>* monitors = cvf->monitors();
 245     if (monitors->is_nonempty()) {
 246       bool relocked = Deoptimization::relock_objects(thread, monitors, deoptee_thread, deoptee, exec_mode, realloc_failures);
 247       deoptimized_objects = deoptimized_objects || relocked;
 248 #ifndef PRODUCT
 249       if (PrintDeoptimizationDetails) {
 250         ttyLocker ttyl;
 251         for (int j = 0; j < monitors->length(); j++) {
 252           MonitorInfo* mi = monitors->at(j);
 253           if (mi->eliminated()) {
 254             if (first) {
 255               first = false;
 256               tty->print_cr("RELOCK OBJECTS in thread " INTPTR_FORMAT, p2i(thread));
 257             }
 258             if (exec_mode == Deoptimization::Unpack_none) {
 259               ObjectMonitor* monitor = deoptee_thread->current_waiting_monitor();
 260               if (monitor != NULL && (oop)monitor->object() == mi->owner()) {
 261                 tty->print_cr("     object <" INTPTR_FORMAT "> DEFERRED relocking after wait", p2i(mi->owner()));
 262                 continue;
 263               }
 264             }
 265             if (mi->owner_is_scalar_replaced()) {
 266               Klass* k = java_lang_Class::as_Klass(mi->owner_klass());
 267               tty->print_cr("     failed reallocation for klass %s", k->external_name());
 268             } else {
 269               tty->print_cr("     object <" INTPTR_FORMAT "> locked", p2i(mi->owner()));
 270             }
 271           }
 272         }
 273       }
 274 #endif // !PRODUCT
 275     }
 276   }
 277 }
 278 
 279 // Deoptimize objects, that is reallocate and relock them, just before they escape through JVMTI.
 280 // The given vframes cover one physical frame.
 281 bool Deoptimization::deoptimize_objects_internal(JavaThread* thread, GrowableArray<compiledVFrame*>* chunk, bool& realloc_failures) {
 282   frame deoptee = chunk->at(0)->fr();
 283   JavaThread* deoptee_thread = chunk->at(0)->thread();
 284   CompiledMethod* cm = deoptee.cb()->as_compiled_method_or_null();
 285   RegisterMap map(chunk->at(0)->register_map());
 286   bool deoptimized_objects = false;
 287 
 288   bool const jvmci_enabled = JVMCI_ONLY(UseJVMCICompiler) NOT_JVMCI(false);
 289 
 290   // Reallocate the non-escaping objects and restore their fields.
 291   if (jvmci_enabled COMPILER2_PRESENT(|| (DoEscapeAnalysis && EliminateAllocations))) {
 292     realloc_failures = eliminate_allocations(thread, Unpack_none, cm, deoptee, map, chunk, deoptimized_objects);
 293   }
 294 
 295   // Revoke biases of objects with eliminated locks in the given frame.
 296   Deoptimization::revoke_for_object_deoptimization(deoptee_thread, deoptee, &map, thread);
 297 
 298   // MonitorInfo structures used in eliminate_locks are not GC safe.
 299   NoSafepointVerifier no_safepoint;
 300 
 301   // Now relock objects if synchronization on them was eliminated.
 302   if (jvmci_enabled COMPILER2_PRESENT(|| ((DoEscapeAnalysis || EliminateNestedLocks) && EliminateLocks))) {
 303     eliminate_locks(thread, chunk, realloc_failures, deoptee, Unpack_none, deoptimized_objects);
 304   }
 305   return deoptimized_objects;
 306 }
 307 #endif // COMPILER2_OR_JVMCI
 308 
 309 // This is factored, since it is both called from a JRT_LEAF (deoptimization) and a JRT_ENTRY (uncommon_trap)
 310 Deoptimization::UnrollBlock* Deoptimization::fetch_unroll_info_helper(JavaThread* thread, int exec_mode) {
 311 
 312   // Note: there is a safepoint safety issue here. No matter whether we enter
 313   // via vanilla deopt or uncommon trap we MUST NOT stop at a safepoint once
 314   // the vframeArray is created.
 315   //
 316 
 317   // Allocate our special deoptimization ResourceMark
 318   DeoptResourceMark* dmark = new DeoptResourceMark(thread);
 319   assert(thread->deopt_mark() == NULL, "Pending deopt!");
 320   thread->set_deopt_mark(dmark);
 321 
 322   frame stub_frame = thread->last_frame(); // Makes stack walkable as side effect
 323   RegisterMap map(thread, true);
 324   RegisterMap dummy_map(thread, false);
 325   // Now get the deoptee with a valid map
 326   frame deoptee = stub_frame.sender(&map);
 327   // Set the deoptee nmethod
 328   assert(thread->deopt_compiled_method() == NULL, "Pending deopt!");
 329   CompiledMethod* cm = deoptee.cb()->as_compiled_method_or_null();
 330   thread->set_deopt_compiled_method(cm);
 331 
 332   if (VerifyStack) {
 333     thread->validate_frame_layout();
 334   }
 335 
 336   // Create a growable array of VFrames where each VFrame represents an inlined
 337   // Java frame.  This storage is allocated with the usual system arena.
 338   assert(deoptee.is_compiled_frame(), "Wrong frame type");
 339   GrowableArray<compiledVFrame*>* chunk = new GrowableArray<compiledVFrame*>(10);
 340   vframe* vf = vframe::new_vframe(&deoptee, &map, thread);
 341   while (!vf->is_top()) {
 342     assert(vf->is_compiled_frame(), "Wrong frame type");
 343     chunk->push(compiledVFrame::cast(vf));
 344     vf = vf->sender();
 345   }
 346   assert(vf->is_compiled_frame(), "Wrong frame type");
 347   chunk->push(compiledVFrame::cast(vf));
 348 
 349   bool realloc_failures = false;
 350 
 351 #if COMPILER2_OR_JVMCI
 352 #if INCLUDE_JVMCI
 353   bool jvmci_enabled = true;
 354 #else
 355   bool jvmci_enabled = false;
 356 #endif
 357 
 358   // Reallocate the non-escaping objects and restore their fields. Then
 359   // relock objects if synchronization on them was eliminated.
 360   if (jvmci_enabled COMPILER2_PRESENT( || (DoEscapeAnalysis && EliminateAllocations) )) {
 361     bool unused;
 362     realloc_failures = eliminate_allocations(thread, exec_mode, cm, deoptee, map, chunk, unused);
 363   }
 364 #endif // COMPILER2_OR_JVMCI
 365 
 366   // Revoke biases, done with in java state.
 367   // No safepoints allowed after this
 368   revoke_from_deopt_handler(thread, deoptee, &map);
 369 
 370   // Ensure that no safepoint is taken after pointers have been stored
 371   // in fields of rematerialized objects.  If a safepoint occurs from here on
 372   // out the java state residing in the vframeArray will be missed.
 373   // Locks may be rebaised in a safepoint.
 374   NoSafepointVerifier no_safepoint;
 375 
 376 #if COMPILER2_OR_JVMCI
 377   if ((jvmci_enabled COMPILER2_PRESENT( || ((DoEscapeAnalysis || EliminateNestedLocks) && EliminateLocks) ))
 378       && !EscapeBarrier::objs_are_deoptimized(thread, deoptee.id())) {
 379     bool unused;
 380     eliminate_locks(thread, chunk, realloc_failures, deoptee, exec_mode, unused);
 381   }
 382 #endif // COMPILER2_OR_JVMCI
 383 
 384   ScopeDesc* trap_scope = chunk->at(0)->scope();
 385   Handle exceptionObject;
 386   if (trap_scope->rethrow_exception()) {
 387     if (PrintDeoptimizationDetails) {
 388       tty->print_cr("Exception to be rethrown in the interpreter for method %s::%s at bci %d", trap_scope->method()->method_holder()->name()->as_C_string(), trap_scope->method()->name()->as_C_string(), trap_scope->bci());
 389     }
 390     GrowableArray<ScopeValue*>* expressions = trap_scope->expressions();
 391     guarantee(expressions != NULL && expressions->length() > 0, "must have exception to throw");
 392     ScopeValue* topOfStack = expressions->top();
 393     exceptionObject = StackValue::create_stack_value(&deoptee, &map, topOfStack)->get_obj();
 394     guarantee(exceptionObject() != NULL, "exception oop can not be null");
 395   }
 396 
 397   vframeArray* array = create_vframeArray(thread, deoptee, &map, chunk, realloc_failures);
 398 #if COMPILER2_OR_JVMCI
 399   if (realloc_failures) {
 400     pop_frames_failed_reallocs(thread, array);
 401   }
 402 #endif
 403 
 404   assert(thread->vframe_array_head() == NULL, "Pending deopt!");
 405   thread->set_vframe_array_head(array);
 406 
 407   // Now that the vframeArray has been created if we have any deferred local writes
 408   // added by jvmti then we can free up that structure as the data is now in the
 409   // vframeArray
 410 
 411   if (JvmtiDeferredUpdates::deferred_locals(thread) != NULL) {
 412     GrowableArray<jvmtiDeferredLocalVariableSet*>* list = JvmtiDeferredUpdates::deferred_locals(thread);
 413     int i = 0;
 414     do {
 415       // Because of inlining we could have multiple vframes for a single frame
 416       // and several of the vframes could have deferred writes. Find them all.
 417       if (list->at(i)->id() == array->original().id()) {
 418         jvmtiDeferredLocalVariableSet* dlv = list->at(i);
 419         list->remove_at(i);
 420         // individual jvmtiDeferredLocalVariableSet are CHeapObj's
 421         delete dlv;
 422       } else {
 423         i++;
 424       }
 425     } while ( i < list->length() );
 426     if (list->length() == 0) {
 427       JvmtiDeferredUpdates* updates = thread->deferred_updates();
 428       thread->set_deferred_updates(NULL);
 429       // free deferred updates.
 430       delete updates;
 431     }
 432 
 433   }
 434 
 435   // Compute the caller frame based on the sender sp of stub_frame and stored frame sizes info.
 436   CodeBlob* cb = stub_frame.cb();
 437   // Verify we have the right vframeArray
 438   assert(cb->frame_size() >= 0, "Unexpected frame size");
 439   intptr_t* unpack_sp = stub_frame.sp() + cb->frame_size();
 440 
 441   // If the deopt call site is a MethodHandle invoke call site we have
 442   // to adjust the unpack_sp.
 443   nmethod* deoptee_nm = deoptee.cb()->as_nmethod_or_null();
 444   if (deoptee_nm != NULL && deoptee_nm->is_method_handle_return(deoptee.pc()))
 445     unpack_sp = deoptee.unextended_sp();
 446 
 447 #ifdef ASSERT
 448   assert(cb->is_deoptimization_stub() ||
 449          cb->is_uncommon_trap_stub() ||
 450          strcmp("Stub<DeoptimizationStub.deoptimizationHandler>", cb->name()) == 0 ||
 451          strcmp("Stub<UncommonTrapStub.uncommonTrapHandler>", cb->name()) == 0,
 452          "unexpected code blob: %s", cb->name());
 453 #endif
 454 
 455   // This is a guarantee instead of an assert because if vframe doesn't match
 456   // we will unpack the wrong deoptimized frame and wind up in strange places
 457   // where it will be very difficult to figure out what went wrong. Better
 458   // to die an early death here than some very obscure death later when the
 459   // trail is cold.
 460   // Note: on ia64 this guarantee can be fooled by frames with no memory stack
 461   // in that it will fail to detect a problem when there is one. This needs
 462   // more work in tiger timeframe.
 463   guarantee(array->unextended_sp() == unpack_sp, "vframe_array_head must contain the vframeArray to unpack");
 464 
 465   int number_of_frames = array->frames();
 466 
 467   // Compute the vframes' sizes.  Note that frame_sizes[] entries are ordered from outermost to innermost
 468   // virtual activation, which is the reverse of the elements in the vframes array.
 469   intptr_t* frame_sizes = NEW_C_HEAP_ARRAY(intptr_t, number_of_frames, mtCompiler);
 470   // +1 because we always have an interpreter return address for the final slot.
 471   address* frame_pcs = NEW_C_HEAP_ARRAY(address, number_of_frames + 1, mtCompiler);
 472   int popframe_extra_args = 0;
 473   // Create an interpreter return address for the stub to use as its return
 474   // address so the skeletal frames are perfectly walkable
 475   frame_pcs[number_of_frames] = Interpreter::deopt_entry(vtos, 0);
 476 
 477   // PopFrame requires that the preserved incoming arguments from the recently-popped topmost
 478   // activation be put back on the expression stack of the caller for reexecution
 479   if (JvmtiExport::can_pop_frame() && thread->popframe_forcing_deopt_reexecution()) {
 480     popframe_extra_args = in_words(thread->popframe_preserved_args_size_in_words());
 481   }
 482 
 483   // Find the current pc for sender of the deoptee. Since the sender may have been deoptimized
 484   // itself since the deoptee vframeArray was created we must get a fresh value of the pc rather
 485   // than simply use array->sender.pc(). This requires us to walk the current set of frames
 486   //
 487   frame deopt_sender = stub_frame.sender(&dummy_map); // First is the deoptee frame
 488   deopt_sender = deopt_sender.sender(&dummy_map);     // Now deoptee caller
 489 
 490   // It's possible that the number of parameters at the call site is
 491   // different than number of arguments in the callee when method
 492   // handles are used.  If the caller is interpreted get the real
 493   // value so that the proper amount of space can be added to it's
 494   // frame.
 495   bool caller_was_method_handle = false;
 496   if (deopt_sender.is_interpreted_frame()) {
 497     methodHandle method(thread, deopt_sender.interpreter_frame_method());
 498     Bytecode_invoke cur = Bytecode_invoke_check(method, deopt_sender.interpreter_frame_bci());
 499     if (cur.is_invokedynamic() || cur.is_invokehandle()) {
 500       // Method handle invokes may involve fairly arbitrary chains of
 501       // calls so it's impossible to know how much actual space the
 502       // caller has for locals.
 503       caller_was_method_handle = true;
 504     }
 505   }
 506 
 507   //
 508   // frame_sizes/frame_pcs[0] oldest frame (int or c2i)
 509   // frame_sizes/frame_pcs[1] next oldest frame (int)
 510   // frame_sizes/frame_pcs[n] youngest frame (int)
 511   //
 512   // Now a pc in frame_pcs is actually the return address to the frame's caller (a frame
 513   // owns the space for the return address to it's caller).  Confusing ain't it.
 514   //
 515   // The vframe array can address vframes with indices running from
 516   // 0.._frames-1. Index  0 is the youngest frame and _frame - 1 is the oldest (root) frame.
 517   // When we create the skeletal frames we need the oldest frame to be in the zero slot
 518   // in the frame_sizes/frame_pcs so the assembly code can do a trivial walk.
 519   // so things look a little strange in this loop.
 520   //
 521   int callee_parameters = 0;
 522   int callee_locals = 0;
 523   for (int index = 0; index < array->frames(); index++ ) {
 524     // frame[number_of_frames - 1 ] = on_stack_size(youngest)
 525     // frame[number_of_frames - 2 ] = on_stack_size(sender(youngest))
 526     // frame[number_of_frames - 3 ] = on_stack_size(sender(sender(youngest)))
 527     frame_sizes[number_of_frames - 1 - index] = BytesPerWord * array->element(index)->on_stack_size(callee_parameters,
 528                                                                                                     callee_locals,
 529                                                                                                     index == 0,
 530                                                                                                     popframe_extra_args);
 531     // This pc doesn't have to be perfect just good enough to identify the frame
 532     // as interpreted so the skeleton frame will be walkable
 533     // The correct pc will be set when the skeleton frame is completely filled out
 534     // The final pc we store in the loop is wrong and will be overwritten below
 535     frame_pcs[number_of_frames - 1 - index ] = Interpreter::deopt_entry(vtos, 0) - frame::pc_return_offset;
 536 
 537     callee_parameters = array->element(index)->method()->size_of_parameters();
 538     callee_locals = array->element(index)->method()->max_locals();
 539     popframe_extra_args = 0;
 540   }
 541 
 542   // Compute whether the root vframe returns a float or double value.
 543   BasicType return_type;
 544   {
 545     methodHandle method(thread, array->element(0)->method());
 546     Bytecode_invoke invoke = Bytecode_invoke_check(method, array->element(0)->bci());
 547     return_type = invoke.is_valid() ? invoke.result_type() : T_ILLEGAL;
 548   }
 549 
 550   // Compute information for handling adapters and adjusting the frame size of the caller.
 551   int caller_adjustment = 0;
 552 
 553   // Compute the amount the oldest interpreter frame will have to adjust
 554   // its caller's stack by. If the caller is a compiled frame then
 555   // we pretend that the callee has no parameters so that the
 556   // extension counts for the full amount of locals and not just
 557   // locals-parms. This is because without a c2i adapter the parm
 558   // area as created by the compiled frame will not be usable by
 559   // the interpreter. (Depending on the calling convention there
 560   // may not even be enough space).
 561 
 562   // QQQ I'd rather see this pushed down into last_frame_adjust
 563   // and have it take the sender (aka caller).
 564 
 565   if (deopt_sender.is_compiled_frame() || caller_was_method_handle) {
 566     caller_adjustment = last_frame_adjust(0, callee_locals);
 567   } else if (callee_locals > callee_parameters) {
 568     // The caller frame may need extending to accommodate
 569     // non-parameter locals of the first unpacked interpreted frame.
 570     // Compute that adjustment.
 571     caller_adjustment = last_frame_adjust(callee_parameters, callee_locals);
 572   }
 573 
 574   // If the sender is deoptimized the we must retrieve the address of the handler
 575   // since the frame will "magically" show the original pc before the deopt
 576   // and we'd undo the deopt.
 577 
 578   frame_pcs[0] = deopt_sender.raw_pc();
 579 
 580   assert(CodeCache::find_blob_unsafe(frame_pcs[0]) != NULL, "bad pc");
 581 
 582 #if INCLUDE_JVMCI
 583   if (exceptionObject() != NULL) {
 584     thread->set_exception_oop(exceptionObject());
 585     exec_mode = Unpack_exception;
 586   }
 587 #endif
 588 
 589   if (thread->frames_to_pop_failed_realloc() > 0 && exec_mode != Unpack_uncommon_trap) {
 590     assert(thread->has_pending_exception(), "should have thrown OOME");
 591     thread->set_exception_oop(thread->pending_exception());
 592     thread->clear_pending_exception();
 593     exec_mode = Unpack_exception;
 594   }
 595 
 596 #if INCLUDE_JVMCI
 597   if (thread->frames_to_pop_failed_realloc() > 0) {
 598     thread->set_pending_monitorenter(false);
 599   }
 600 #endif
 601 
 602   UnrollBlock* info = new UnrollBlock(array->frame_size() * BytesPerWord,
 603                                       caller_adjustment * BytesPerWord,
 604                                       caller_was_method_handle ? 0 : callee_parameters,
 605                                       number_of_frames,
 606                                       frame_sizes,
 607                                       frame_pcs,
 608                                       return_type,
 609                                       exec_mode);
 610   // On some platforms, we need a way to pass some platform dependent
 611   // information to the unpacking code so the skeletal frames come out
 612   // correct (initial fp value, unextended sp, ...)
 613   info->set_initial_info((intptr_t) array->sender().initial_deoptimization_info());
 614 
 615   if (array->frames() > 1) {
 616     if (VerifyStack && TraceDeoptimization) {
 617       ttyLocker ttyl;
 618       tty->print_cr("Deoptimizing method containing inlining");
 619     }
 620   }
 621 
 622   array->set_unroll_block(info);
 623   return info;
 624 }
 625 
 626 // Called to cleanup deoptimization data structures in normal case
 627 // after unpacking to stack and when stack overflow error occurs
 628 void Deoptimization::cleanup_deopt_info(JavaThread *thread,
 629                                         vframeArray *array) {
 630 
 631   // Get array if coming from exception
 632   if (array == NULL) {
 633     array = thread->vframe_array_head();
 634   }
 635   thread->set_vframe_array_head(NULL);
 636 
 637   // Free the previous UnrollBlock
 638   vframeArray* old_array = thread->vframe_array_last();
 639   thread->set_vframe_array_last(array);
 640 
 641   if (old_array != NULL) {
 642     UnrollBlock* old_info = old_array->unroll_block();
 643     old_array->set_unroll_block(NULL);
 644     delete old_info;
 645     delete old_array;
 646   }
 647 
 648   // Deallocate any resource creating in this routine and any ResourceObjs allocated
 649   // inside the vframeArray (StackValueCollections)
 650 
 651   delete thread->deopt_mark();
 652   thread->set_deopt_mark(NULL);
 653   thread->set_deopt_compiled_method(NULL);
 654 
 655 
 656   if (JvmtiExport::can_pop_frame()) {
 657     // Regardless of whether we entered this routine with the pending
 658     // popframe condition bit set, we should always clear it now
 659     thread->clear_popframe_condition();
 660   }
 661 
 662   // unpack_frames() is called at the end of the deoptimization handler
 663   // and (in C2) at the end of the uncommon trap handler. Note this fact
 664   // so that an asynchronous stack walker can work again. This counter is
 665   // incremented at the beginning of fetch_unroll_info() and (in C2) at
 666   // the beginning of uncommon_trap().
 667   thread->dec_in_deopt_handler();
 668 }
 669 
 670 // Moved from cpu directories because none of the cpus has callee save values.
 671 // If a cpu implements callee save values, move this to deoptimization_<cpu>.cpp.
 672 void Deoptimization::unwind_callee_save_values(frame* f, vframeArray* vframe_array) {
 673 
 674   // This code is sort of the equivalent of C2IAdapter::setup_stack_frame back in
 675   // the days we had adapter frames. When we deoptimize a situation where a
 676   // compiled caller calls a compiled caller will have registers it expects
 677   // to survive the call to the callee. If we deoptimize the callee the only
 678   // way we can restore these registers is to have the oldest interpreter
 679   // frame that we create restore these values. That is what this routine
 680   // will accomplish.
 681 
 682   // At the moment we have modified c2 to not have any callee save registers
 683   // so this problem does not exist and this routine is just a place holder.
 684 
 685   assert(f->is_interpreted_frame(), "must be interpreted");
 686 }
 687 
 688 // Return BasicType of value being returned
 689 JRT_LEAF(BasicType, Deoptimization::unpack_frames(JavaThread* thread, int exec_mode))
 690 
 691   // We are already active in the special DeoptResourceMark any ResourceObj's we
 692   // allocate will be freed at the end of the routine.
 693 
 694   // It is actually ok to allocate handles in a leaf method. It causes no safepoints,
 695   // but makes the entry a little slower. There is however a little dance we have to
 696   // do in debug mode to get around the NoHandleMark code in the JRT_LEAF macro
 697   ResetNoHandleMark rnhm; // No-op in release/product versions
 698   HandleMark hm;
 699 
 700   frame stub_frame = thread->last_frame();
 701 
 702   // Since the frame to unpack is the top frame of this thread, the vframe_array_head
 703   // must point to the vframeArray for the unpack frame.
 704   vframeArray* array = thread->vframe_array_head();
 705 
 706 #ifndef PRODUCT
 707   if (TraceDeoptimization) {
 708     ttyLocker ttyl;
 709     tty->print_cr("DEOPT UNPACKING thread " INTPTR_FORMAT " vframeArray " INTPTR_FORMAT " mode %d",
 710                   p2i(thread), p2i(array), exec_mode);
 711   }
 712 #endif
 713   Events::log_deopt_message(thread, "DEOPT UNPACKING pc=" INTPTR_FORMAT " sp=" INTPTR_FORMAT " mode %d",
 714               p2i(stub_frame.pc()), p2i(stub_frame.sp()), exec_mode);
 715 
 716   UnrollBlock* info = array->unroll_block();
 717 
 718   // We set the last_Java frame. But the stack isn't really parsable here. So we
 719   // clear it to make sure JFR understands not to try and walk stacks from events
 720   // in here.
 721   intptr_t* sp = thread->frame_anchor()->last_Java_sp();
 722   thread->frame_anchor()->set_last_Java_sp(NULL);
 723 
 724   // Unpack the interpreter frames and any adapter frame (c2 only) we might create.
 725   array->unpack_to_stack(stub_frame, exec_mode, info->caller_actual_parameters());
 726 
 727   thread->frame_anchor()->set_last_Java_sp(sp);
 728 
 729   BasicType bt = info->return_type();
 730 
 731   // If we have an exception pending, claim that the return type is an oop
 732   // so the deopt_blob does not overwrite the exception_oop.
 733 
 734   if (exec_mode == Unpack_exception)
 735     bt = T_OBJECT;
 736 
 737   // Cleanup thread deopt data
 738   cleanup_deopt_info(thread, array);
 739 
 740 #ifndef PRODUCT
 741   if (VerifyStack) {
 742     ResourceMark res_mark;
 743     // Clear pending exception to not break verification code (restored afterwards)
 744     PRESERVE_EXCEPTION_MARK;
 745 
 746     thread->validate_frame_layout();
 747 
 748     // Verify that the just-unpacked frames match the interpreter's
 749     // notions of expression stack and locals
 750     vframeArray* cur_array = thread->vframe_array_last();
 751     RegisterMap rm(thread, false);
 752     rm.set_include_argument_oops(false);
 753     bool is_top_frame = true;
 754     int callee_size_of_parameters = 0;
 755     int callee_max_locals = 0;
 756     for (int i = 0; i < cur_array->frames(); i++) {
 757       vframeArrayElement* el = cur_array->element(i);
 758       frame* iframe = el->iframe();
 759       guarantee(iframe->is_interpreted_frame(), "Wrong frame type");
 760 
 761       // Get the oop map for this bci
 762       InterpreterOopMap mask;
 763       int cur_invoke_parameter_size = 0;
 764       bool try_next_mask = false;
 765       int next_mask_expression_stack_size = -1;
 766       int top_frame_expression_stack_adjustment = 0;
 767       methodHandle mh(thread, iframe->interpreter_frame_method());
 768       OopMapCache::compute_one_oop_map(mh, iframe->interpreter_frame_bci(), &mask);
 769       BytecodeStream str(mh, iframe->interpreter_frame_bci());
 770       int max_bci = mh->code_size();
 771       // Get to the next bytecode if possible
 772       assert(str.bci() < max_bci, "bci in interpreter frame out of bounds");
 773       // Check to see if we can grab the number of outgoing arguments
 774       // at an uncommon trap for an invoke (where the compiler
 775       // generates debug info before the invoke has executed)
 776       Bytecodes::Code cur_code = str.next();
 777       if (Bytecodes::is_invoke(cur_code)) {
 778         Bytecode_invoke invoke(mh, iframe->interpreter_frame_bci());
 779         cur_invoke_parameter_size = invoke.size_of_parameters();
 780         if (i != 0 && !invoke.is_invokedynamic() && MethodHandles::has_member_arg(invoke.klass(), invoke.name())) {
 781           callee_size_of_parameters++;
 782         }
 783       }
 784       if (str.bci() < max_bci) {
 785         Bytecodes::Code next_code = str.next();
 786         if (next_code >= 0) {
 787           // The interpreter oop map generator reports results before
 788           // the current bytecode has executed except in the case of
 789           // calls. It seems to be hard to tell whether the compiler
 790           // has emitted debug information matching the "state before"
 791           // a given bytecode or the state after, so we try both
 792           if (!Bytecodes::is_invoke(cur_code) && cur_code != Bytecodes::_athrow) {
 793             // Get expression stack size for the next bytecode
 794             InterpreterOopMap next_mask;
 795             OopMapCache::compute_one_oop_map(mh, str.bci(), &next_mask);
 796             next_mask_expression_stack_size = next_mask.expression_stack_size();
 797             if (Bytecodes::is_invoke(next_code)) {
 798               Bytecode_invoke invoke(mh, str.bci());
 799               next_mask_expression_stack_size += invoke.size_of_parameters();
 800             }
 801             // Need to subtract off the size of the result type of
 802             // the bytecode because this is not described in the
 803             // debug info but returned to the interpreter in the TOS
 804             // caching register
 805             BasicType bytecode_result_type = Bytecodes::result_type(cur_code);
 806             if (bytecode_result_type != T_ILLEGAL) {
 807               top_frame_expression_stack_adjustment = type2size[bytecode_result_type];
 808             }
 809             assert(top_frame_expression_stack_adjustment >= 0, "stack adjustment must be positive");
 810             try_next_mask = true;
 811           }
 812         }
 813       }
 814 
 815       // Verify stack depth and oops in frame
 816       // This assertion may be dependent on the platform we're running on and may need modification (tested on x86 and sparc)
 817       if (!(
 818             /* SPARC */
 819             (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_size_of_parameters) ||
 820             /* x86 */
 821             (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + callee_max_locals) ||
 822             (try_next_mask &&
 823              (iframe->interpreter_frame_expression_stack_size() == (next_mask_expression_stack_size -
 824                                                                     top_frame_expression_stack_adjustment))) ||
 825             (is_top_frame && (exec_mode == Unpack_exception) && iframe->interpreter_frame_expression_stack_size() == 0) ||
 826             (is_top_frame && (exec_mode == Unpack_uncommon_trap || exec_mode == Unpack_reexecute || el->should_reexecute()) &&
 827              (iframe->interpreter_frame_expression_stack_size() == mask.expression_stack_size() + cur_invoke_parameter_size))
 828             )) {
 829         {
 830           ttyLocker ttyl;
 831 
 832           // Print out some information that will help us debug the problem
 833           tty->print_cr("Wrong number of expression stack elements during deoptimization");
 834           tty->print_cr("  Error occurred while verifying frame %d (0..%d, 0 is topmost)", i, cur_array->frames() - 1);
 835           tty->print_cr("  Fabricated interpreter frame had %d expression stack elements",
 836                         iframe->interpreter_frame_expression_stack_size());
 837           tty->print_cr("  Interpreter oop map had %d expression stack elements", mask.expression_stack_size());
 838           tty->print_cr("  try_next_mask = %d", try_next_mask);
 839           tty->print_cr("  next_mask_expression_stack_size = %d", next_mask_expression_stack_size);
 840           tty->print_cr("  callee_size_of_parameters = %d", callee_size_of_parameters);
 841           tty->print_cr("  callee_max_locals = %d", callee_max_locals);
 842           tty->print_cr("  top_frame_expression_stack_adjustment = %d", top_frame_expression_stack_adjustment);
 843           tty->print_cr("  exec_mode = %d", exec_mode);
 844           tty->print_cr("  cur_invoke_parameter_size = %d", cur_invoke_parameter_size);
 845           tty->print_cr("  Thread = " INTPTR_FORMAT ", thread ID = %d", p2i(thread), thread->osthread()->thread_id());
 846           tty->print_cr("  Interpreted frames:");
 847           for (int k = 0; k < cur_array->frames(); k++) {
 848             vframeArrayElement* el = cur_array->element(k);
 849             tty->print_cr("    %s (bci %d)", el->method()->name_and_sig_as_C_string(), el->bci());
 850           }
 851           cur_array->print_on_2(tty);
 852         } // release tty lock before calling guarantee
 853         guarantee(false, "wrong number of expression stack elements during deopt");
 854       }
 855       VerifyOopClosure verify;
 856       iframe->oops_interpreted_do(&verify, &rm, false);
 857       callee_size_of_parameters = mh->size_of_parameters();
 858       callee_max_locals = mh->max_locals();
 859       is_top_frame = false;
 860     }
 861   }
 862 #endif /* !PRODUCT */
 863 
 864   return bt;
 865 JRT_END
 866 
 867 class DeoptimizeMarkedClosure : public HandshakeClosure {
 868  public:
 869   DeoptimizeMarkedClosure() : HandshakeClosure("Deoptimize") {}
 870   void do_thread(Thread* thread) {
 871     JavaThread* jt = (JavaThread*)thread;
 872     jt->deoptimize_marked_methods();
 873   }
 874 };
 875 
 876 void Deoptimization::deoptimize_all_marked(nmethod* nmethod_only) {
 877   ResourceMark rm;
 878   DeoptimizationMarker dm;
 879 
 880   // Make the dependent methods not entrant
 881   if (nmethod_only != NULL) {
 882     nmethod_only->mark_for_deoptimization();
 883     nmethod_only->make_not_entrant();
 884   } else {
 885     MutexLocker mu(SafepointSynchronize::is_at_safepoint() ? NULL : CodeCache_lock, Mutex::_no_safepoint_check_flag);
 886     CodeCache::make_marked_nmethods_not_entrant();
 887   }
 888 
 889   DeoptimizeMarkedClosure deopt;
 890   if (SafepointSynchronize::is_at_safepoint()) {
 891     Threads::java_threads_do(&deopt);
 892   } else {
 893     Handshake::execute(&deopt);
 894   }
 895 }
 896 
 897 Deoptimization::DeoptAction Deoptimization::_unloaded_action
 898   = Deoptimization::Action_reinterpret;
 899 
 900 
 901 
 902 #if INCLUDE_JVMCI || INCLUDE_AOT
 903 template<typename CacheType>
 904 class BoxCacheBase : public CHeapObj<mtCompiler> {
 905 protected:
 906   static InstanceKlass* find_cache_klass(Symbol* klass_name, TRAPS) {
 907     ResourceMark rm;
 908     char* klass_name_str = klass_name->as_C_string();
 909     Klass* k = SystemDictionary::find(klass_name, Handle(), Handle(), THREAD);
 910     guarantee(k != NULL, "%s must be loaded", klass_name_str);
 911     InstanceKlass* ik = InstanceKlass::cast(k);
 912     guarantee(ik->is_initialized(), "%s must be initialized", klass_name_str);
 913     CacheType::compute_offsets(ik);
 914     return ik;
 915   }
 916 };
 917 
 918 template<typename PrimitiveType, typename CacheType, typename BoxType> class BoxCache  : public BoxCacheBase<CacheType> {
 919   PrimitiveType _low;
 920   PrimitiveType _high;
 921   jobject _cache;
 922 protected:
 923   static BoxCache<PrimitiveType, CacheType, BoxType> *_singleton;
 924   BoxCache(Thread* thread) {
 925     InstanceKlass* ik = BoxCacheBase<CacheType>::find_cache_klass(CacheType::symbol(), thread);
 926     objArrayOop cache = CacheType::cache(ik);
 927     assert(cache->length() > 0, "Empty cache");
 928     _low = BoxType::value(cache->obj_at(0));
 929     _high = _low + cache->length() - 1;
 930     _cache = JNIHandles::make_global(Handle(thread, cache));
 931   }
 932   ~BoxCache() {
 933     JNIHandles::destroy_global(_cache);
 934   }
 935 public:
 936   static BoxCache<PrimitiveType, CacheType, BoxType>* singleton(Thread* thread) {
 937     if (_singleton == NULL) {
 938       BoxCache<PrimitiveType, CacheType, BoxType>* s = new BoxCache<PrimitiveType, CacheType, BoxType>(thread);
 939       if (!Atomic::replace_if_null(&_singleton, s)) {
 940         delete s;
 941       }
 942     }
 943     return _singleton;
 944   }
 945   oop lookup(PrimitiveType value) {
 946     if (_low <= value && value <= _high) {
 947       int offset = value - _low;
 948       return objArrayOop(JNIHandles::resolve_non_null(_cache))->obj_at(offset);
 949     }
 950     return NULL;
 951   }
 952   oop lookup_raw(intptr_t raw_value) {
 953     // Have to cast to avoid little/big-endian problems.
 954     if (sizeof(PrimitiveType) > sizeof(jint)) {
 955       jlong value = (jlong)raw_value;
 956       return lookup(value);
 957     }
 958     PrimitiveType value = (PrimitiveType)*((jint*)&raw_value);
 959     return lookup(value);
 960   }
 961 };
 962 
 963 typedef BoxCache<jint, java_lang_Integer_IntegerCache, java_lang_Integer> IntegerBoxCache;
 964 typedef BoxCache<jlong, java_lang_Long_LongCache, java_lang_Long> LongBoxCache;
 965 typedef BoxCache<jchar, java_lang_Character_CharacterCache, java_lang_Character> CharacterBoxCache;
 966 typedef BoxCache<jshort, java_lang_Short_ShortCache, java_lang_Short> ShortBoxCache;
 967 typedef BoxCache<jbyte, java_lang_Byte_ByteCache, java_lang_Byte> ByteBoxCache;
 968 
 969 template<> BoxCache<jint, java_lang_Integer_IntegerCache, java_lang_Integer>* BoxCache<jint, java_lang_Integer_IntegerCache, java_lang_Integer>::_singleton = NULL;
 970 template<> BoxCache<jlong, java_lang_Long_LongCache, java_lang_Long>* BoxCache<jlong, java_lang_Long_LongCache, java_lang_Long>::_singleton = NULL;
 971 template<> BoxCache<jchar, java_lang_Character_CharacterCache, java_lang_Character>* BoxCache<jchar, java_lang_Character_CharacterCache, java_lang_Character>::_singleton = NULL;
 972 template<> BoxCache<jshort, java_lang_Short_ShortCache, java_lang_Short>* BoxCache<jshort, java_lang_Short_ShortCache, java_lang_Short>::_singleton = NULL;
 973 template<> BoxCache<jbyte, java_lang_Byte_ByteCache, java_lang_Byte>* BoxCache<jbyte, java_lang_Byte_ByteCache, java_lang_Byte>::_singleton = NULL;
 974 
 975 class BooleanBoxCache : public BoxCacheBase<java_lang_Boolean> {
 976   jobject _true_cache;
 977   jobject _false_cache;
 978 protected:
 979   static BooleanBoxCache *_singleton;
 980   BooleanBoxCache(Thread *thread) {
 981     InstanceKlass* ik = find_cache_klass(java_lang_Boolean::symbol(), thread);
 982     _true_cache = JNIHandles::make_global(Handle(thread, java_lang_Boolean::get_TRUE(ik)));
 983     _false_cache = JNIHandles::make_global(Handle(thread, java_lang_Boolean::get_FALSE(ik)));
 984   }
 985   ~BooleanBoxCache() {
 986     JNIHandles::destroy_global(_true_cache);
 987     JNIHandles::destroy_global(_false_cache);
 988   }
 989 public:
 990   static BooleanBoxCache* singleton(Thread* thread) {
 991     if (_singleton == NULL) {
 992       BooleanBoxCache* s = new BooleanBoxCache(thread);
 993       if (!Atomic::replace_if_null(&_singleton, s)) {
 994         delete s;
 995       }
 996     }
 997     return _singleton;
 998   }
 999   oop lookup_raw(intptr_t raw_value) {
1000     // Have to cast to avoid little/big-endian problems.
1001     jboolean value = (jboolean)*((jint*)&raw_value);
1002     return lookup(value);
1003   }
1004   oop lookup(jboolean value) {
1005     if (value != 0) {
1006       return JNIHandles::resolve_non_null(_true_cache);
1007     }
1008     return JNIHandles::resolve_non_null(_false_cache);
1009   }
1010 };
1011 
1012 BooleanBoxCache* BooleanBoxCache::_singleton = NULL;
1013 
1014 oop Deoptimization::get_cached_box(AutoBoxObjectValue* bv, frame* fr, RegisterMap* reg_map, TRAPS) {
1015    Klass* k = java_lang_Class::as_Klass(bv->klass()->as_ConstantOopReadValue()->value()());
1016    BasicType box_type = SystemDictionary::box_klass_type(k);
1017    if (box_type != T_OBJECT) {
1018      StackValue* value = StackValue::create_stack_value(fr, reg_map, bv->field_at(box_type == T_LONG ? 1 : 0));
1019      switch(box_type) {
1020        case T_INT:     return IntegerBoxCache::singleton(THREAD)->lookup_raw(value->get_int());
1021        case T_CHAR:    return CharacterBoxCache::singleton(THREAD)->lookup_raw(value->get_int());
1022        case T_SHORT:   return ShortBoxCache::singleton(THREAD)->lookup_raw(value->get_int());
1023        case T_BYTE:    return ByteBoxCache::singleton(THREAD)->lookup_raw(value->get_int());
1024        case T_BOOLEAN: return BooleanBoxCache::singleton(THREAD)->lookup_raw(value->get_int());
1025        case T_LONG:    return LongBoxCache::singleton(THREAD)->lookup_raw(value->get_int());
1026        default:;
1027      }
1028    }
1029    return NULL;
1030 }
1031 #endif // INCLUDE_JVMCI || INCLUDE_AOT
1032 
1033 #if COMPILER2_OR_JVMCI
1034 bool Deoptimization::realloc_objects(JavaThread* thread, frame* fr, RegisterMap* reg_map, GrowableArray<ScopeValue*>* objects, TRAPS) {
1035   Handle pending_exception(THREAD, thread->pending_exception());
1036   const char* exception_file = thread->exception_file();
1037   int exception_line = thread->exception_line();
1038   thread->clear_pending_exception();
1039 
1040   bool failures = false;
1041 
1042   for (int i = 0; i < objects->length(); i++) {
1043     assert(objects->at(i)->is_object(), "invalid debug information");
1044     ObjectValue* sv = (ObjectValue*) objects->at(i);
1045 
1046     Klass* k = java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()());
1047     oop obj = NULL;
1048 
1049     if (k->is_instance_klass()) {
1050 #if INCLUDE_JVMCI || INCLUDE_AOT
1051       CompiledMethod* cm = fr->cb()->as_compiled_method_or_null();
1052       if (cm->is_compiled_by_jvmci() && sv->is_auto_box()) {
1053         AutoBoxObjectValue* abv = (AutoBoxObjectValue*) sv;
1054         obj = get_cached_box(abv, fr, reg_map, THREAD);
1055         if (obj != NULL) {
1056           // Set the flag to indicate the box came from a cache, so that we can skip the field reassignment for it.
1057           abv->set_cached(true);
1058         }
1059       }
1060 #endif // INCLUDE_JVMCI || INCLUDE_AOT
1061       InstanceKlass* ik = InstanceKlass::cast(k);
1062       if (obj == NULL) {
1063         obj = ik->allocate_instance(THREAD);
1064       }
1065     } else if (k->is_typeArray_klass()) {
1066       TypeArrayKlass* ak = TypeArrayKlass::cast(k);
1067       assert(sv->field_size() % type2size[ak->element_type()] == 0, "non-integral array length");
1068       int len = sv->field_size() / type2size[ak->element_type()];
1069       obj = ak->allocate(len, THREAD);
1070     } else if (k->is_objArray_klass()) {
1071       ObjArrayKlass* ak = ObjArrayKlass::cast(k);
1072       obj = ak->allocate(sv->field_size(), THREAD);
1073     }
1074 
1075     if (obj == NULL) {
1076       failures = true;
1077     }
1078 
1079     assert(sv->value().is_null(), "redundant reallocation");
1080     assert(obj != NULL || HAS_PENDING_EXCEPTION, "allocation should succeed or we should get an exception");
1081     CLEAR_PENDING_EXCEPTION;
1082     sv->set_value(obj);
1083   }
1084 
1085   if (failures) {
1086     THROW_OOP_(Universe::out_of_memory_error_realloc_objects(), failures);
1087   } else if (pending_exception.not_null()) {
1088     thread->set_pending_exception(pending_exception(), exception_file, exception_line);
1089   }
1090 
1091   return failures;
1092 }
1093 
1094 #if INCLUDE_JVMCI
1095 /**
1096  * For primitive types whose kind gets "erased" at runtime (shorts become stack ints),
1097  * we need to somehow be able to recover the actual kind to be able to write the correct
1098  * amount of bytes.
1099  * For that purpose, this method assumes that, for an entry spanning n bytes at index i,
1100  * the entries at index n + 1 to n + i are 'markers'.
1101  * For example, if we were writing a short at index 4 of a byte array of size 8, the
1102  * expected form of the array would be:
1103  *
1104  * {b0, b1, b2, b3, INT, marker, b6, b7}
1105  *
1106  * Thus, in order to get back the size of the entry, we simply need to count the number
1107  * of marked entries
1108  *
1109  * @param virtualArray the virtualized byte array
1110  * @param i index of the virtual entry we are recovering
1111  * @return The number of bytes the entry spans
1112  */
1113 static int count_number_of_bytes_for_entry(ObjectValue *virtualArray, int i) {
1114   int index = i;
1115   while (++index < virtualArray->field_size() &&
1116            virtualArray->field_at(index)->is_marker()) {}
1117   return index - i;
1118 }
1119 
1120 /**
1121  * If there was a guarantee for byte array to always start aligned to a long, we could
1122  * do a simple check on the parity of the index. Unfortunately, that is not always the
1123  * case. Thus, we check alignment of the actual address we are writing to.
1124  * In the unlikely case index 0 is 5-aligned for example, it would then be possible to
1125  * write a long to index 3.
1126  */
1127 static jbyte* check_alignment_get_addr(typeArrayOop obj, int index, int expected_alignment) {
1128     jbyte* res = obj->byte_at_addr(index);
1129     assert((((intptr_t) res) % expected_alignment) == 0, "Non-aligned write");
1130     return res;
1131 }
1132 
1133 static void byte_array_put(typeArrayOop obj, intptr_t val, int index, int byte_count) {
1134   switch (byte_count) {
1135     case 1:
1136       obj->byte_at_put(index, (jbyte) *((jint *) &val));
1137       break;
1138     case 2:
1139       *((jshort *) check_alignment_get_addr(obj, index, 2)) = (jshort) *((jint *) &val);
1140       break;
1141     case 4:
1142       *((jint *) check_alignment_get_addr(obj, index, 4)) = (jint) *((jint *) &val);
1143       break;
1144     case 8:
1145       *((jlong *) check_alignment_get_addr(obj, index, 8)) = (jlong) *((jlong *) &val);
1146       break;
1147     default:
1148       ShouldNotReachHere();
1149   }
1150 }
1151 #endif // INCLUDE_JVMCI
1152 
1153 
1154 // restore elements of an eliminated type array
1155 void Deoptimization::reassign_type_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, typeArrayOop obj, BasicType type) {
1156   int index = 0;
1157   intptr_t val;
1158 
1159   for (int i = 0; i < sv->field_size(); i++) {
1160     StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i));
1161     switch(type) {
1162     case T_LONG: case T_DOUBLE: {
1163       assert(value->type() == T_INT, "Agreement.");
1164       StackValue* low =
1165         StackValue::create_stack_value(fr, reg_map, sv->field_at(++i));
1166 #ifdef _LP64
1167       jlong res = (jlong)low->get_int();
1168 #else
1169       jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int());
1170 #endif
1171       obj->long_at_put(index, res);
1172       break;
1173     }
1174 
1175     // Have to cast to INT (32 bits) pointer to avoid little/big-endian problem.
1176     case T_INT: case T_FLOAT: { // 4 bytes.
1177       assert(value->type() == T_INT, "Agreement.");
1178       bool big_value = false;
1179       if (i + 1 < sv->field_size() && type == T_INT) {
1180         if (sv->field_at(i)->is_location()) {
1181           Location::Type type = ((LocationValue*) sv->field_at(i))->location().type();
1182           if (type == Location::dbl || type == Location::lng) {
1183             big_value = true;
1184           }
1185         } else if (sv->field_at(i)->is_constant_int()) {
1186           ScopeValue* next_scope_field = sv->field_at(i + 1);
1187           if (next_scope_field->is_constant_long() || next_scope_field->is_constant_double()) {
1188             big_value = true;
1189           }
1190         }
1191       }
1192 
1193       if (big_value) {
1194         StackValue* low = StackValue::create_stack_value(fr, reg_map, sv->field_at(++i));
1195   #ifdef _LP64
1196         jlong res = (jlong)low->get_int();
1197   #else
1198         jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int());
1199   #endif
1200         obj->int_at_put(index, (jint)*((jint*)&res));
1201         obj->int_at_put(++index, (jint)*(((jint*)&res) + 1));
1202       } else {
1203         val = value->get_int();
1204         obj->int_at_put(index, (jint)*((jint*)&val));
1205       }
1206       break;
1207     }
1208 
1209     case T_SHORT:
1210       assert(value->type() == T_INT, "Agreement.");
1211       val = value->get_int();
1212       obj->short_at_put(index, (jshort)*((jint*)&val));
1213       break;
1214 
1215     case T_CHAR:
1216       assert(value->type() == T_INT, "Agreement.");
1217       val = value->get_int();
1218       obj->char_at_put(index, (jchar)*((jint*)&val));
1219       break;
1220 
1221     case T_BYTE: {
1222       assert(value->type() == T_INT, "Agreement.");
1223       // The value we get is erased as a regular int. We will need to find its actual byte count 'by hand'.
1224       val = value->get_int();
1225 #if INCLUDE_JVMCI
1226       int byte_count = count_number_of_bytes_for_entry(sv, i);
1227       byte_array_put(obj, val, index, byte_count);
1228       // According to byte_count contract, the values from i + 1 to i + byte_count are illegal values. Skip.
1229       i += byte_count - 1; // Balance the loop counter.
1230       index += byte_count;
1231       // index has been updated so continue at top of loop
1232       continue;
1233 #else
1234       obj->byte_at_put(index, (jbyte)*((jint*)&val));
1235       break;
1236 #endif // INCLUDE_JVMCI
1237     }
1238 
1239     case T_BOOLEAN: {
1240       assert(value->type() == T_INT, "Agreement.");
1241       val = value->get_int();
1242       obj->bool_at_put(index, (jboolean)*((jint*)&val));
1243       break;
1244     }
1245 
1246       default:
1247         ShouldNotReachHere();
1248     }
1249     index++;
1250   }
1251 }
1252 
1253 // restore fields of an eliminated object array
1254 void Deoptimization::reassign_object_array_elements(frame* fr, RegisterMap* reg_map, ObjectValue* sv, objArrayOop obj) {
1255   for (int i = 0; i < sv->field_size(); i++) {
1256     StackValue* value = StackValue::create_stack_value(fr, reg_map, sv->field_at(i));
1257     assert(value->type() == T_OBJECT, "object element expected");
1258     obj->obj_at_put(i, value->get_obj()());
1259   }
1260 }
1261 
1262 class ReassignedField {
1263 public:
1264   int _offset;
1265   BasicType _type;
1266 public:
1267   ReassignedField() {
1268     _offset = 0;
1269     _type = T_ILLEGAL;
1270   }
1271 };
1272 
1273 int compare(ReassignedField* left, ReassignedField* right) {
1274   return left->_offset - right->_offset;
1275 }
1276 
1277 // Restore fields of an eliminated instance object using the same field order
1278 // returned by HotSpotResolvedObjectTypeImpl.getInstanceFields(true)
1279 static int reassign_fields_by_klass(InstanceKlass* klass, frame* fr, RegisterMap* reg_map, ObjectValue* sv, int svIndex, oop obj, bool skip_internal) {
1280   GrowableArray<ReassignedField>* fields = new GrowableArray<ReassignedField>();
1281   InstanceKlass* ik = klass;
1282   while (ik != NULL) {
1283     for (AllFieldStream fs(ik); !fs.done(); fs.next()) {
1284       if (!fs.access_flags().is_static() && (!skip_internal || !fs.access_flags().is_internal())) {
1285         ReassignedField field;
1286         field._offset = fs.offset();
1287         field._type = Signature::basic_type(fs.signature());
1288         fields->append(field);
1289       }
1290     }
1291     ik = ik->superklass();
1292   }
1293   fields->sort(compare);
1294   for (int i = 0; i < fields->length(); i++) {
1295     intptr_t val;
1296     ScopeValue* scope_field = sv->field_at(svIndex);
1297     StackValue* value = StackValue::create_stack_value(fr, reg_map, scope_field);
1298     int offset = fields->at(i)._offset;
1299     BasicType type = fields->at(i)._type;
1300     switch (type) {
1301       case T_OBJECT: case T_ARRAY:
1302         assert(value->type() == T_OBJECT, "Agreement.");
1303         obj->obj_field_put(offset, value->get_obj()());
1304         break;
1305 
1306       // Have to cast to INT (32 bits) pointer to avoid little/big-endian problem.
1307       case T_INT: case T_FLOAT: { // 4 bytes.
1308         assert(value->type() == T_INT, "Agreement.");
1309         bool big_value = false;
1310         if (i+1 < fields->length() && fields->at(i+1)._type == T_INT) {
1311           if (scope_field->is_location()) {
1312             Location::Type type = ((LocationValue*) scope_field)->location().type();
1313             if (type == Location::dbl || type == Location::lng) {
1314               big_value = true;
1315             }
1316           }
1317           if (scope_field->is_constant_int()) {
1318             ScopeValue* next_scope_field = sv->field_at(svIndex + 1);
1319             if (next_scope_field->is_constant_long() || next_scope_field->is_constant_double()) {
1320               big_value = true;
1321             }
1322           }
1323         }
1324 
1325         if (big_value) {
1326           i++;
1327           assert(i < fields->length(), "second T_INT field needed");
1328           assert(fields->at(i)._type == T_INT, "T_INT field needed");
1329         } else {
1330           val = value->get_int();
1331           obj->int_field_put(offset, (jint)*((jint*)&val));
1332           break;
1333         }
1334       }
1335         /* no break */
1336 
1337       case T_LONG: case T_DOUBLE: {
1338         assert(value->type() == T_INT, "Agreement.");
1339         StackValue* low = StackValue::create_stack_value(fr, reg_map, sv->field_at(++svIndex));
1340 #ifdef _LP64
1341         jlong res = (jlong)low->get_int();
1342 #else
1343         jlong res = jlong_from((jint)value->get_int(), (jint)low->get_int());
1344 #endif
1345         obj->long_field_put(offset, res);
1346         break;
1347       }
1348 
1349       case T_SHORT:
1350         assert(value->type() == T_INT, "Agreement.");
1351         val = value->get_int();
1352         obj->short_field_put(offset, (jshort)*((jint*)&val));
1353         break;
1354 
1355       case T_CHAR:
1356         assert(value->type() == T_INT, "Agreement.");
1357         val = value->get_int();
1358         obj->char_field_put(offset, (jchar)*((jint*)&val));
1359         break;
1360 
1361       case T_BYTE:
1362         assert(value->type() == T_INT, "Agreement.");
1363         val = value->get_int();
1364         obj->byte_field_put(offset, (jbyte)*((jint*)&val));
1365         break;
1366 
1367       case T_BOOLEAN:
1368         assert(value->type() == T_INT, "Agreement.");
1369         val = value->get_int();
1370         obj->bool_field_put(offset, (jboolean)*((jint*)&val));
1371         break;
1372 
1373       default:
1374         ShouldNotReachHere();
1375     }
1376     svIndex++;
1377   }
1378   return svIndex;
1379 }
1380 
1381 // restore fields of all eliminated objects and arrays
1382 void Deoptimization::reassign_fields(frame* fr, RegisterMap* reg_map, GrowableArray<ScopeValue*>* objects, bool realloc_failures, bool skip_internal) {
1383   for (int i = 0; i < objects->length(); i++) {
1384     ObjectValue* sv = (ObjectValue*) objects->at(i);
1385     Klass* k = java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()());
1386     Handle obj = sv->value();
1387     assert(obj.not_null() || realloc_failures, "reallocation was missed");
1388     if (PrintDeoptimizationDetails) {
1389       tty->print_cr("reassign fields for object of type %s!", k->name()->as_C_string());
1390     }
1391     if (obj.is_null()) {
1392       continue;
1393     }
1394 #if INCLUDE_JVMCI || INCLUDE_AOT
1395     // Don't reassign fields of boxes that came from a cache. Caches may be in CDS.
1396     if (sv->is_auto_box() && ((AutoBoxObjectValue*) sv)->is_cached()) {
1397       continue;
1398     }
1399 #endif // INCLUDE_JVMCI || INCLUDE_AOT
1400     if (k->is_instance_klass()) {
1401       InstanceKlass* ik = InstanceKlass::cast(k);
1402       reassign_fields_by_klass(ik, fr, reg_map, sv, 0, obj(), skip_internal);
1403     } else if (k->is_typeArray_klass()) {
1404       TypeArrayKlass* ak = TypeArrayKlass::cast(k);
1405       reassign_type_array_elements(fr, reg_map, sv, (typeArrayOop) obj(), ak->element_type());
1406     } else if (k->is_objArray_klass()) {
1407       reassign_object_array_elements(fr, reg_map, sv, (objArrayOop) obj());
1408     }
1409   }
1410 }
1411 
1412 
1413 // relock objects for which synchronization was eliminated
1414 bool Deoptimization::relock_objects(JavaThread* thread, GrowableArray<MonitorInfo*>* monitors,
1415                                     JavaThread* deoptee_thread, frame& fr, int exec_mode, bool realloc_failures) {
1416   bool relocked_objects = false;
1417   for (int i = 0; i < monitors->length(); i++) {
1418     MonitorInfo* mon_info = monitors->at(i);
1419     if (mon_info->eliminated()) {
1420       assert(!mon_info->owner_is_scalar_replaced() || realloc_failures, "reallocation was missed");
1421       relocked_objects = true;
1422       if (!mon_info->owner_is_scalar_replaced()) {
1423         Handle obj(thread, mon_info->owner());
1424         markWord mark = obj->mark();
1425         if (UseBiasedLocking && mark.has_bias_pattern()) {
1426           // New allocated objects may have the mark set to anonymously biased.
1427           // Also the deoptimized method may called methods with synchronization
1428           // where the thread-local object is bias locked to the current thread.
1429           assert(mark.is_biased_anonymously() ||
1430                  mark.biased_locker() == deoptee_thread, "should be locked to current thread");
1431           // Reset mark word to unbiased prototype.
1432           markWord unbiased_prototype = markWord::prototype().set_age(mark.age());
1433           obj->set_mark(unbiased_prototype);
1434         } else if (exec_mode == Unpack_none) {
1435           if (mark.has_locker() && fr.sp() > (intptr_t*)mark.locker()) {
1436             // With exec_mode == Unpack_none obj may be thread local and locked in
1437             // a callee frame. In this case the bias was revoked before in revoke_for_object_deoptimization().
1438             // Make the lock in the callee a recursive lock and restore the displaced header.
1439             markWord dmw = mark.displaced_mark_helper();
1440             mark.locker()->set_displaced_header(markWord::encode((BasicLock*) NULL));
1441             obj->set_mark(dmw);
1442           }
1443           if (mark.has_monitor()) {
1444             // defer relocking if the deoptee thread is currently waiting for obj
1445             ObjectMonitor* waiting_monitor = deoptee_thread->current_waiting_monitor();
1446             if (waiting_monitor != NULL && (oop)waiting_monitor->object() == obj()) {
1447               assert(fr.is_deoptimized_frame(), "frame must be scheduled for deoptimization");
1448               mon_info->lock()->set_displaced_header(markWord::unused_mark());
1449               JvmtiDeferredUpdates::inc_relock_count_after_wait(deoptee_thread);
1450               continue;
1451             }
1452           }
1453         }
1454         BasicLock* lock = mon_info->lock();
1455         ObjectSynchronizer::enter(obj, lock, deoptee_thread);
1456         assert(mon_info->owner()->is_locked(), "object must be locked now");
1457       }
1458     }
1459   }
1460   return relocked_objects;
1461 }
1462 
1463 
1464 #ifndef PRODUCT
1465 // print information about reallocated objects
1466 void Deoptimization::print_objects(GrowableArray<ScopeValue*>* objects, bool realloc_failures) {
1467   fieldDescriptor fd;
1468 
1469   for (int i = 0; i < objects->length(); i++) {
1470     ObjectValue* sv = (ObjectValue*) objects->at(i);
1471     Klass* k = java_lang_Class::as_Klass(sv->klass()->as_ConstantOopReadValue()->value()());
1472     Handle obj = sv->value();
1473 
1474     tty->print("     object <" INTPTR_FORMAT "> of type ", p2i(sv->value()()));
1475     k->print_value();
1476     assert(obj.not_null() || realloc_failures, "reallocation was missed");
1477     if (obj.is_null()) {
1478       tty->print(" allocation failed");
1479     } else {
1480       tty->print(" allocated (%d bytes)", obj->size() * HeapWordSize);
1481     }
1482     tty->cr();
1483 
1484     if (Verbose && !obj.is_null()) {
1485       k->oop_print_on(obj(), tty);
1486     }
1487   }
1488 }
1489 #endif
1490 #endif // COMPILER2_OR_JVMCI
1491 
1492 vframeArray* Deoptimization::create_vframeArray(JavaThread* thread, frame fr, RegisterMap *reg_map, GrowableArray<compiledVFrame*>* chunk, bool realloc_failures) {
1493   Events::log_deopt_message(thread, "DEOPT PACKING pc=" INTPTR_FORMAT " sp=" INTPTR_FORMAT, p2i(fr.pc()), p2i(fr.sp()));
1494 
1495 #ifndef PRODUCT
1496   if (PrintDeoptimizationDetails) {
1497     ttyLocker ttyl;
1498     tty->print("DEOPT PACKING thread " INTPTR_FORMAT " ", p2i(thread));
1499     fr.print_on(tty);
1500     tty->print_cr("     Virtual frames (innermost first):");
1501     for (int index = 0; index < chunk->length(); index++) {
1502       compiledVFrame* vf = chunk->at(index);
1503       tty->print("       %2d - ", index);
1504       vf->print_value();
1505       int bci = chunk->at(index)->raw_bci();
1506       const char* code_name;
1507       if (bci == SynchronizationEntryBCI) {
1508         code_name = "sync entry";
1509       } else {
1510         Bytecodes::Code code = vf->method()->code_at(bci);
1511         code_name = Bytecodes::name(code);
1512       }
1513       tty->print(" - %s", code_name);
1514       tty->print_cr(" @ bci %d ", bci);
1515       if (Verbose) {
1516         vf->print();
1517         tty->cr();
1518       }
1519     }
1520   }
1521 #endif
1522 
1523   // Register map for next frame (used for stack crawl).  We capture
1524   // the state of the deopt'ing frame's caller.  Thus if we need to
1525   // stuff a C2I adapter we can properly fill in the callee-save
1526   // register locations.
1527   frame caller = fr.sender(reg_map);
1528   int frame_size = caller.sp() - fr.sp();
1529 
1530   frame sender = caller;
1531 
1532   // Since the Java thread being deoptimized will eventually adjust it's own stack,
1533   // the vframeArray containing the unpacking information is allocated in the C heap.
1534   // For Compiler1, the caller of the deoptimized frame is saved for use by unpack_frames().
1535   vframeArray* array = vframeArray::allocate(thread, frame_size, chunk, reg_map, sender, caller, fr, realloc_failures);
1536 
1537   // Compare the vframeArray to the collected vframes
1538   assert(array->structural_compare(thread, chunk), "just checking");
1539 
1540 #ifndef PRODUCT
1541   if (PrintDeoptimizationDetails) {
1542     ttyLocker ttyl;
1543     tty->print_cr("     Created vframeArray " INTPTR_FORMAT, p2i(array));
1544   }
1545 #endif // PRODUCT
1546 
1547   return array;
1548 }
1549 
1550 #if COMPILER2_OR_JVMCI
1551 void Deoptimization::pop_frames_failed_reallocs(JavaThread* thread, vframeArray* array) {
1552   // Reallocation of some scalar replaced objects failed. Record
1553   // that we need to pop all the interpreter frames for the
1554   // deoptimized compiled frame.
1555   assert(thread->frames_to_pop_failed_realloc() == 0, "missed frames to pop?");
1556   thread->set_frames_to_pop_failed_realloc(array->frames());
1557   // Unlock all monitors here otherwise the interpreter will see a
1558   // mix of locked and unlocked monitors (because of failed
1559   // reallocations of synchronized objects) and be confused.
1560   for (int i = 0; i < array->frames(); i++) {
1561     MonitorChunk* monitors = array->element(i)->monitors();
1562     if (monitors != NULL) {
1563       for (int j = 0; j < monitors->number_of_monitors(); j++) {
1564         BasicObjectLock* src = monitors->at(j);
1565         if (src->obj() != NULL) {
1566           ObjectSynchronizer::exit(src->obj(), src->lock(), thread);
1567         }
1568       }
1569       array->element(i)->free_monitors(thread);
1570 #ifdef ASSERT
1571       array->element(i)->set_removed_monitors();
1572 #endif
1573     }
1574   }
1575 }
1576 #endif
1577 
1578 static void collect_monitors(compiledVFrame* cvf, GrowableArray<Handle>* objects_to_revoke, bool only_eliminated) {
1579   GrowableArray<MonitorInfo*>* monitors = cvf->monitors();
1580   Thread* thread = Thread::current();
1581   for (int i = 0; i < monitors->length(); i++) {
1582     MonitorInfo* mon_info = monitors->at(i);
1583     if ((mon_info->eliminated() == only_eliminated) && !mon_info->owner_is_scalar_replaced() && mon_info->owner() != NULL) {
1584       objects_to_revoke->append(Handle(thread, mon_info->owner()));
1585     }
1586   }
1587 }
1588 
1589 static void get_monitors_from_stack(GrowableArray<Handle>* objects_to_revoke, JavaThread* thread, frame fr, RegisterMap* map,
1590                                     bool only_eliminated) {
1591   // Unfortunately we don't have a RegisterMap available in most of
1592   // the places we want to call this routine so we need to walk the
1593   // stack again to update the register map.
1594   if (map == NULL || !map->update_map()) {
1595     StackFrameStream sfs(thread, true);
1596     bool found = false;
1597     while (!found && !sfs.is_done()) {
1598       frame* cur = sfs.current();
1599       sfs.next();
1600       found = cur->id() == fr.id();
1601     }
1602     assert(found, "frame to be deoptimized not found on target thread's stack");
1603     map = sfs.register_map();
1604   }
1605 
1606   vframe* vf = vframe::new_vframe(&fr, map, thread);
1607   compiledVFrame* cvf = compiledVFrame::cast(vf);
1608   // Revoke monitors' biases in all scopes
1609   while (!cvf->is_top()) {
1610     collect_monitors(cvf, objects_to_revoke, only_eliminated);
1611     cvf = compiledVFrame::cast(cvf->sender());
1612   }
1613   collect_monitors(cvf, objects_to_revoke, only_eliminated);
1614 }
1615 
1616 void Deoptimization::revoke_from_deopt_handler(JavaThread* thread, frame fr, RegisterMap* map) {
1617   if (!UseBiasedLocking) {
1618     return;
1619   }
1620   GrowableArray<Handle>* objects_to_revoke = new GrowableArray<Handle>();
1621   get_monitors_from_stack(objects_to_revoke, thread, fr, map, false);
1622 
1623   int len = objects_to_revoke->length();
1624   for (int i = 0; i < len; i++) {
1625     oop obj = (objects_to_revoke->at(i))();
1626     BiasedLocking::revoke_own_lock(objects_to_revoke->at(i), thread);
1627     assert(!obj->mark().has_bias_pattern(), "biases should be revoked by now");
1628   }
1629 }
1630 
1631 // Revoke the bias of objects with eliminated locking to prepare subsequent relocking.
1632 void Deoptimization::revoke_for_object_deoptimization(JavaThread* deoptee_thread, frame fr, RegisterMap* map, JavaThread* thread) {
1633   if (!UseBiasedLocking) {
1634     return;
1635   }
1636   GrowableArray<Handle>* objects_to_revoke = new GrowableArray<Handle>();
1637   // Collect monitors, but only those with eliminated locking.
1638   get_monitors_from_stack(objects_to_revoke, deoptee_thread, fr, map, true);
1639 
1640   int len = objects_to_revoke->length();
1641   for (int i = 0; i < len; i++) {
1642     oop obj = (objects_to_revoke->at(i))();
1643     markWord mark = obj->mark();
1644     if (!mark.has_bias_pattern() ||
1645         mark.is_biased_anonymously() || // eliminated locking does not bias an object if it wasn't before
1646         !obj->klass()->prototype_header().has_bias_pattern() || // bulk revoke ignores eliminated monitors
1647         (obj->klass()->prototype_header().bias_epoch() != mark.bias_epoch())) { // bulk rebias ignores eliminated monitors
1648       // We reach here regularly if there's just eliminated locking on obj.
1649       // We must not call BiasedLocking::revoke_own_lock() in this case, as we would hit assertions, because it is a
1650       // prerequisite that there has to be non-eliminated locking on obj by deoptee_thread.
1651       // Luckily we don't have to revoke here, because obj has to be a  non-escaping obj and can be relocked without
1652       // revoking the bias. See Deoptimization::relock_objects().
1653       continue;
1654     }
1655     BiasedLocking::revoke(objects_to_revoke->at(i), thread);
1656     assert(!objects_to_revoke->at(i)->mark().has_bias_pattern(), "biases should be revoked by now");
1657   }
1658 }
1659 
1660 void Deoptimization::deoptimize_single_frame(JavaThread* thread, frame fr, Deoptimization::DeoptReason reason) {
1661   assert(fr.can_be_deoptimized(), "checking frame type");
1662 
1663   gather_statistics(reason, Action_none, Bytecodes::_illegal);
1664 
1665   if (LogCompilation && xtty != NULL) {
1666     CompiledMethod* cm = fr.cb()->as_compiled_method_or_null();
1667     assert(cm != NULL, "only compiled methods can deopt");
1668 
1669     ttyLocker ttyl;
1670     xtty->begin_head("deoptimized thread='" UINTX_FORMAT "' reason='%s' pc='" INTPTR_FORMAT "'",(uintx)thread->osthread()->thread_id(), trap_reason_name(reason), p2i(fr.pc()));
1671     cm->log_identity(xtty);
1672     xtty->end_head();
1673     for (ScopeDesc* sd = cm->scope_desc_at(fr.pc()); ; sd = sd->sender()) {
1674       xtty->begin_elem("jvms bci='%d'", sd->bci());
1675       xtty->method(sd->method());
1676       xtty->end_elem();
1677       if (sd->is_top())  break;
1678     }
1679     xtty->tail("deoptimized");
1680   }
1681 
1682   // Patch the compiled method so that when execution returns to it we will
1683   // deopt the execution state and return to the interpreter.
1684   fr.deoptimize(thread);
1685 }
1686 
1687 void Deoptimization::deoptimize(JavaThread* thread, frame fr, DeoptReason reason) {
1688   // Deoptimize only if the frame comes from compile code.
1689   // Do not deoptimize the frame which is already patched
1690   // during the execution of the loops below.
1691   if (!fr.is_compiled_frame() || fr.is_deoptimized_frame()) {
1692     return;
1693   }
1694   ResourceMark rm;
1695   DeoptimizationMarker dm;
1696   deoptimize_single_frame(thread, fr, reason);
1697 }
1698 
1699 #if INCLUDE_JVMCI
1700 address Deoptimization::deoptimize_for_missing_exception_handler(CompiledMethod* cm) {
1701   // there is no exception handler for this pc => deoptimize
1702   cm->make_not_entrant();
1703 
1704   // Use Deoptimization::deoptimize for all of its side-effects:
1705   // gathering traps statistics, logging...
1706   // it also patches the return pc but we do not care about that
1707   // since we return a continuation to the deopt_blob below.
1708   JavaThread* thread = JavaThread::current();
1709   RegisterMap reg_map(thread, false);
1710   frame runtime_frame = thread->last_frame();
1711   frame caller_frame = runtime_frame.sender(&reg_map);
1712   assert(caller_frame.cb()->as_compiled_method_or_null() == cm, "expect top frame compiled method");
1713   Deoptimization::deoptimize(thread, caller_frame, Deoptimization::Reason_not_compiled_exception_handler);
1714 
1715   MethodData* trap_mdo = get_method_data(thread, methodHandle(thread, cm->method()), true);
1716   if (trap_mdo != NULL) {
1717     trap_mdo->inc_trap_count(Deoptimization::Reason_not_compiled_exception_handler);
1718   }
1719 
1720   return SharedRuntime::deopt_blob()->unpack_with_exception_in_tls();
1721 }
1722 #endif
1723 
1724 void Deoptimization::deoptimize_frame_internal(JavaThread* thread, intptr_t* id, DeoptReason reason) {
1725   assert(thread == Thread::current() || SafepointSynchronize::is_at_safepoint(),
1726          "can only deoptimize other thread at a safepoint");
1727   // Compute frame and register map based on thread and sp.
1728   RegisterMap reg_map(thread, false);
1729   frame fr = thread->last_frame();
1730   while (fr.id() != id) {
1731     fr = fr.sender(&reg_map);
1732   }
1733   deoptimize(thread, fr, reason);
1734 }
1735 
1736 
1737 void Deoptimization::deoptimize_frame(JavaThread* thread, intptr_t* id, DeoptReason reason) {
1738   if (thread == Thread::current()) {
1739     Deoptimization::deoptimize_frame_internal(thread, id, reason);
1740   } else {
1741     VM_DeoptimizeFrame deopt(thread, id, reason);
1742     VMThread::execute(&deopt);
1743   }
1744 }
1745 
1746 void Deoptimization::deoptimize_frame(JavaThread* thread, intptr_t* id) {
1747   deoptimize_frame(thread, id, Reason_constraint);
1748 }
1749 
1750 // JVMTI PopFrame support
1751 JRT_LEAF(void, Deoptimization::popframe_preserve_args(JavaThread* thread, int bytes_to_save, void* start_address))
1752 {
1753   thread->popframe_preserve_args(in_ByteSize(bytes_to_save), start_address);
1754 }
1755 JRT_END
1756 
1757 MethodData*
1758 Deoptimization::get_method_data(JavaThread* thread, const methodHandle& m,
1759                                 bool create_if_missing) {
1760   Thread* THREAD = thread;
1761   MethodData* mdo = m()->method_data();
1762   if (mdo == NULL && create_if_missing && !HAS_PENDING_EXCEPTION) {
1763     // Build an MDO.  Ignore errors like OutOfMemory;
1764     // that simply means we won't have an MDO to update.
1765     Method::build_interpreter_method_data(m, THREAD);
1766     if (HAS_PENDING_EXCEPTION) {
1767       assert((PENDING_EXCEPTION->is_a(SystemDictionary::OutOfMemoryError_klass())), "we expect only an OOM error here");
1768       CLEAR_PENDING_EXCEPTION;
1769     }
1770     mdo = m()->method_data();
1771   }
1772   return mdo;
1773 }
1774 
1775 #if COMPILER2_OR_JVMCI
1776 void Deoptimization::load_class_by_index(const constantPoolHandle& constant_pool, int index, TRAPS) {
1777   // In case of an unresolved klass entry, load the class.
1778   // This path is exercised from case _ldc in Parse::do_one_bytecode,
1779   // and probably nowhere else.
1780   // Even that case would benefit from simply re-interpreting the
1781   // bytecode, without paying special attention to the class index.
1782   // So this whole "class index" feature should probably be removed.
1783 
1784   if (constant_pool->tag_at(index).is_unresolved_klass()) {
1785     Klass* tk = constant_pool->klass_at_ignore_error(index, CHECK);
1786     return;
1787   }
1788 
1789   assert(!constant_pool->tag_at(index).is_symbol(),
1790          "no symbolic names here, please");
1791 }
1792 
1793 
1794 void Deoptimization::load_class_by_index(const constantPoolHandle& constant_pool, int index) {
1795   EXCEPTION_MARK;
1796   load_class_by_index(constant_pool, index, THREAD);
1797   if (HAS_PENDING_EXCEPTION) {
1798     // Exception happened during classloading. We ignore the exception here, since it
1799     // is going to be rethrown since the current activation is going to be deoptimized and
1800     // the interpreter will re-execute the bytecode.
1801     CLEAR_PENDING_EXCEPTION;
1802     // Class loading called java code which may have caused a stack
1803     // overflow. If the exception was thrown right before the return
1804     // to the runtime the stack is no longer guarded. Reguard the
1805     // stack otherwise if we return to the uncommon trap blob and the
1806     // stack bang causes a stack overflow we crash.
1807     assert(THREAD->is_Java_thread(), "only a java thread can be here");
1808     JavaThread* thread = (JavaThread*)THREAD;
1809     bool guard_pages_enabled = thread->stack_guards_enabled();
1810     if (!guard_pages_enabled) guard_pages_enabled = thread->reguard_stack();
1811     assert(guard_pages_enabled, "stack banging in uncommon trap blob may cause crash");
1812   }
1813 }
1814 
1815 #if INCLUDE_JFR
1816 
1817 class DeoptReasonSerializer : public JfrSerializer {
1818  public:
1819   void serialize(JfrCheckpointWriter& writer) {
1820     writer.write_count((u4)(Deoptimization::Reason_LIMIT + 1)); // + Reason::many (-1)
1821     for (int i = -1; i < Deoptimization::Reason_LIMIT; ++i) {
1822       writer.write_key((u8)i);
1823       writer.write(Deoptimization::trap_reason_name(i));
1824     }
1825   }
1826 };
1827 
1828 class DeoptActionSerializer : public JfrSerializer {
1829  public:
1830   void serialize(JfrCheckpointWriter& writer) {
1831     static const u4 nof_actions = Deoptimization::Action_LIMIT;
1832     writer.write_count(nof_actions);
1833     for (u4 i = 0; i < Deoptimization::Action_LIMIT; ++i) {
1834       writer.write_key(i);
1835       writer.write(Deoptimization::trap_action_name((int)i));
1836     }
1837   }
1838 };
1839 
1840 static void register_serializers() {
1841   static int critical_section = 0;
1842   if (1 == critical_section || Atomic::cmpxchg(&critical_section, 0, 1) == 1) {
1843     return;
1844   }
1845   JfrSerializer::register_serializer(TYPE_DEOPTIMIZATIONREASON, true, new DeoptReasonSerializer());
1846   JfrSerializer::register_serializer(TYPE_DEOPTIMIZATIONACTION, true, new DeoptActionSerializer());
1847 }
1848 
1849 static void post_deoptimization_event(CompiledMethod* nm,
1850                                       const Method* method,
1851                                       int trap_bci,
1852                                       int instruction,
1853                                       Deoptimization::DeoptReason reason,
1854                                       Deoptimization::DeoptAction action) {
1855   assert(nm != NULL, "invariant");
1856   assert(method != NULL, "invariant");
1857   if (EventDeoptimization::is_enabled()) {
1858     static bool serializers_registered = false;
1859     if (!serializers_registered) {
1860       register_serializers();
1861       serializers_registered = true;
1862     }
1863     EventDeoptimization event;
1864     event.set_compileId(nm->compile_id());
1865     event.set_compiler(nm->compiler_type());
1866     event.set_method(method);
1867     event.set_lineNumber(method->line_number_from_bci(trap_bci));
1868     event.set_bci(trap_bci);
1869     event.set_instruction(instruction);
1870     event.set_reason(reason);
1871     event.set_action(action);
1872     event.commit();
1873   }
1874 }
1875 
1876 #endif // INCLUDE_JFR
1877 
1878 JRT_ENTRY(void, Deoptimization::uncommon_trap_inner(JavaThread* thread, jint trap_request)) {
1879   HandleMark hm;
1880 
1881   // uncommon_trap() is called at the beginning of the uncommon trap
1882   // handler. Note this fact before we start generating temporary frames
1883   // that can confuse an asynchronous stack walker. This counter is
1884   // decremented at the end of unpack_frames().
1885   thread->inc_in_deopt_handler();
1886 
1887   // We need to update the map if we have biased locking.
1888 #if INCLUDE_JVMCI
1889   // JVMCI might need to get an exception from the stack, which in turn requires the register map to be valid
1890   RegisterMap reg_map(thread, true);
1891 #else
1892   RegisterMap reg_map(thread, UseBiasedLocking);
1893 #endif
1894   frame stub_frame = thread->last_frame();
1895   frame fr = stub_frame.sender(&reg_map);
1896   // Make sure the calling nmethod is not getting deoptimized and removed
1897   // before we are done with it.
1898   nmethodLocker nl(fr.pc());
1899 
1900   // Log a message
1901   Events::log_deopt_message(thread, "Uncommon trap: trap_request=" PTR32_FORMAT " fr.pc=" INTPTR_FORMAT " relative=" INTPTR_FORMAT,
1902               trap_request, p2i(fr.pc()), fr.pc() - fr.cb()->code_begin());
1903 
1904   {
1905     ResourceMark rm;
1906 
1907     DeoptReason reason = trap_request_reason(trap_request);
1908     DeoptAction action = trap_request_action(trap_request);
1909 #if INCLUDE_JVMCI
1910     int debug_id = trap_request_debug_id(trap_request);
1911 #endif
1912     jint unloaded_class_index = trap_request_index(trap_request); // CP idx or -1
1913 
1914     vframe*  vf  = vframe::new_vframe(&fr, &reg_map, thread);
1915     compiledVFrame* cvf = compiledVFrame::cast(vf);
1916 
1917     CompiledMethod* nm = cvf->code();
1918 
1919     ScopeDesc*      trap_scope  = cvf->scope();
1920 
1921     if (TraceDeoptimization) {
1922       ttyLocker ttyl;
1923       tty->print_cr("  bci=%d pc=" INTPTR_FORMAT ", relative_pc=" INTPTR_FORMAT ", method=%s" JVMCI_ONLY(", debug_id=%d"), trap_scope->bci(), p2i(fr.pc()), fr.pc() - nm->code_begin(), trap_scope->method()->name_and_sig_as_C_string()
1924 #if INCLUDE_JVMCI
1925           , debug_id
1926 #endif
1927           );
1928     }
1929 
1930     methodHandle    trap_method(THREAD, trap_scope->method());
1931     int             trap_bci    = trap_scope->bci();
1932 #if INCLUDE_JVMCI
1933     jlong           speculation = thread->pending_failed_speculation();
1934     if (nm->is_compiled_by_jvmci() && nm->is_nmethod()) { // Exclude AOTed methods
1935       nm->as_nmethod()->update_speculation(thread);
1936     } else {
1937       assert(speculation == 0, "There should not be a speculation for methods compiled by non-JVMCI compilers");
1938     }
1939 
1940     if (trap_bci == SynchronizationEntryBCI) {
1941       trap_bci = 0;
1942       thread->set_pending_monitorenter(true);
1943     }
1944 
1945     if (reason == Deoptimization::Reason_transfer_to_interpreter) {
1946       thread->set_pending_transfer_to_interpreter(true);
1947     }
1948 #endif
1949 
1950     Bytecodes::Code trap_bc     = trap_method->java_code_at(trap_bci);
1951     // Record this event in the histogram.
1952     gather_statistics(reason, action, trap_bc);
1953 
1954     // Ensure that we can record deopt. history:
1955     // Need MDO to record RTM code generation state.
1956     bool create_if_missing = ProfileTraps || UseCodeAging RTM_OPT_ONLY( || UseRTMLocking );
1957 
1958     methodHandle profiled_method;
1959 #if INCLUDE_JVMCI
1960     if (nm->is_compiled_by_jvmci()) {
1961       profiled_method = methodHandle(THREAD, nm->method());
1962     } else {
1963       profiled_method = trap_method;
1964     }
1965 #else
1966     profiled_method = trap_method;
1967 #endif
1968 
1969     MethodData* trap_mdo =
1970       get_method_data(thread, profiled_method, create_if_missing);
1971 
1972     JFR_ONLY(post_deoptimization_event(nm, trap_method(), trap_bci, trap_bc, reason, action);)
1973 
1974     // Log a message
1975     Events::log_deopt_message(thread, "Uncommon trap: reason=%s action=%s pc=" INTPTR_FORMAT " method=%s @ %d %s",
1976                               trap_reason_name(reason), trap_action_name(action), p2i(fr.pc()),
1977                               trap_method->name_and_sig_as_C_string(), trap_bci, nm->compiler_name());
1978 
1979     // Print a bunch of diagnostics, if requested.
1980     if (TraceDeoptimization || LogCompilation) {
1981       ResourceMark rm;
1982       ttyLocker ttyl;
1983       char buf[100];
1984       if (xtty != NULL) {
1985         xtty->begin_head("uncommon_trap thread='" UINTX_FORMAT "' %s",
1986                          os::current_thread_id(),
1987                          format_trap_request(buf, sizeof(buf), trap_request));
1988 #if INCLUDE_JVMCI
1989         if (speculation != 0) {
1990           xtty->print(" speculation='" JLONG_FORMAT "'", speculation);
1991         }
1992 #endif
1993         nm->log_identity(xtty);
1994       }
1995       Symbol* class_name = NULL;
1996       bool unresolved = false;
1997       if (unloaded_class_index >= 0) {
1998         constantPoolHandle constants (THREAD, trap_method->constants());
1999         if (constants->tag_at(unloaded_class_index).is_unresolved_klass()) {
2000           class_name = constants->klass_name_at(unloaded_class_index);
2001           unresolved = true;
2002           if (xtty != NULL)
2003             xtty->print(" unresolved='1'");
2004         } else if (constants->tag_at(unloaded_class_index).is_symbol()) {
2005           class_name = constants->symbol_at(unloaded_class_index);
2006         }
2007         if (xtty != NULL)
2008           xtty->name(class_name);
2009       }
2010       if (xtty != NULL && trap_mdo != NULL && (int)reason < (int)MethodData::_trap_hist_limit) {
2011         // Dump the relevant MDO state.
2012         // This is the deopt count for the current reason, any previous
2013         // reasons or recompiles seen at this point.
2014         int dcnt = trap_mdo->trap_count(reason);
2015         if (dcnt != 0)
2016           xtty->print(" count='%d'", dcnt);
2017         ProfileData* pdata = trap_mdo->bci_to_data(trap_bci);
2018         int dos = (pdata == NULL)? 0: pdata->trap_state();
2019         if (dos != 0) {
2020           xtty->print(" state='%s'", format_trap_state(buf, sizeof(buf), dos));
2021           if (trap_state_is_recompiled(dos)) {
2022             int recnt2 = trap_mdo->overflow_recompile_count();
2023             if (recnt2 != 0)
2024               xtty->print(" recompiles2='%d'", recnt2);
2025           }
2026         }
2027       }
2028       if (xtty != NULL) {
2029         xtty->stamp();
2030         xtty->end_head();
2031       }
2032       if (TraceDeoptimization) {  // make noise on the tty
2033         tty->print("Uncommon trap occurred in");
2034         nm->method()->print_short_name(tty);
2035         tty->print(" compiler=%s compile_id=%d", nm->compiler_name(), nm->compile_id());
2036 #if INCLUDE_JVMCI
2037         if (nm->is_nmethod()) {
2038           const char* installed_code_name = nm->as_nmethod()->jvmci_name();
2039           if (installed_code_name != NULL) {
2040             tty->print(" (JVMCI: installed code name=%s) ", installed_code_name);
2041           }
2042         }
2043 #endif
2044         tty->print(" (@" INTPTR_FORMAT ") thread=" UINTX_FORMAT " reason=%s action=%s unloaded_class_index=%d" JVMCI_ONLY(" debug_id=%d"),
2045                    p2i(fr.pc()),
2046                    os::current_thread_id(),
2047                    trap_reason_name(reason),
2048                    trap_action_name(action),
2049                    unloaded_class_index
2050 #if INCLUDE_JVMCI
2051                    , debug_id
2052 #endif
2053                    );
2054         if (class_name != NULL) {
2055           tty->print(unresolved ? " unresolved class: " : " symbol: ");
2056           class_name->print_symbol_on(tty);
2057         }
2058         tty->cr();
2059       }
2060       if (xtty != NULL) {
2061         // Log the precise location of the trap.
2062         for (ScopeDesc* sd = trap_scope; ; sd = sd->sender()) {
2063           xtty->begin_elem("jvms bci='%d'", sd->bci());
2064           xtty->method(sd->method());
2065           xtty->end_elem();
2066           if (sd->is_top())  break;
2067         }
2068         xtty->tail("uncommon_trap");
2069       }
2070     }
2071     // (End diagnostic printout.)
2072 
2073     // Load class if necessary
2074     if (unloaded_class_index >= 0) {
2075       constantPoolHandle constants(THREAD, trap_method->constants());
2076       load_class_by_index(constants, unloaded_class_index);
2077     }
2078 
2079     // Flush the nmethod if necessary and desirable.
2080     //
2081     // We need to avoid situations where we are re-flushing the nmethod
2082     // because of a hot deoptimization site.  Repeated flushes at the same
2083     // point need to be detected by the compiler and avoided.  If the compiler
2084     // cannot avoid them (or has a bug and "refuses" to avoid them), this
2085     // module must take measures to avoid an infinite cycle of recompilation
2086     // and deoptimization.  There are several such measures:
2087     //
2088     //   1. If a recompilation is ordered a second time at some site X
2089     //   and for the same reason R, the action is adjusted to 'reinterpret',
2090     //   to give the interpreter time to exercise the method more thoroughly.
2091     //   If this happens, the method's overflow_recompile_count is incremented.
2092     //
2093     //   2. If the compiler fails to reduce the deoptimization rate, then
2094     //   the method's overflow_recompile_count will begin to exceed the set
2095     //   limit PerBytecodeRecompilationCutoff.  If this happens, the action
2096     //   is adjusted to 'make_not_compilable', and the method is abandoned
2097     //   to the interpreter.  This is a performance hit for hot methods,
2098     //   but is better than a disastrous infinite cycle of recompilations.
2099     //   (Actually, only the method containing the site X is abandoned.)
2100     //
2101     //   3. In parallel with the previous measures, if the total number of
2102     //   recompilations of a method exceeds the much larger set limit
2103     //   PerMethodRecompilationCutoff, the method is abandoned.
2104     //   This should only happen if the method is very large and has
2105     //   many "lukewarm" deoptimizations.  The code which enforces this
2106     //   limit is elsewhere (class nmethod, class Method).
2107     //
2108     // Note that the per-BCI 'is_recompiled' bit gives the compiler one chance
2109     // to recompile at each bytecode independently of the per-BCI cutoff.
2110     //
2111     // The decision to update code is up to the compiler, and is encoded
2112     // in the Action_xxx code.  If the compiler requests Action_none
2113     // no trap state is changed, no compiled code is changed, and the
2114     // computation suffers along in the interpreter.
2115     //
2116     // The other action codes specify various tactics for decompilation
2117     // and recompilation.  Action_maybe_recompile is the loosest, and
2118     // allows the compiled code to stay around until enough traps are seen,
2119     // and until the compiler gets around to recompiling the trapping method.
2120     //
2121     // The other actions cause immediate removal of the present code.
2122 
2123     // Traps caused by injected profile shouldn't pollute trap counts.
2124     bool injected_profile_trap = trap_method->has_injected_profile() &&
2125                                  (reason == Reason_intrinsic || reason == Reason_unreached);
2126 
2127     bool update_trap_state = (reason != Reason_tenured) && !injected_profile_trap;
2128     bool make_not_entrant = false;
2129     bool make_not_compilable = false;
2130     bool reprofile = false;
2131     switch (action) {
2132     case Action_none:
2133       // Keep the old code.
2134       update_trap_state = false;
2135       break;
2136     case Action_maybe_recompile:
2137       // Do not need to invalidate the present code, but we can
2138       // initiate another
2139       // Start compiler without (necessarily) invalidating the nmethod.
2140       // The system will tolerate the old code, but new code should be
2141       // generated when possible.
2142       break;
2143     case Action_reinterpret:
2144       // Go back into the interpreter for a while, and then consider
2145       // recompiling form scratch.
2146       make_not_entrant = true;
2147       // Reset invocation counter for outer most method.
2148       // This will allow the interpreter to exercise the bytecodes
2149       // for a while before recompiling.
2150       // By contrast, Action_make_not_entrant is immediate.
2151       //
2152       // Note that the compiler will track null_check, null_assert,
2153       // range_check, and class_check events and log them as if they
2154       // had been traps taken from compiled code.  This will update
2155       // the MDO trap history so that the next compilation will
2156       // properly detect hot trap sites.
2157       reprofile = true;
2158       break;
2159     case Action_make_not_entrant:
2160       // Request immediate recompilation, and get rid of the old code.
2161       // Make them not entrant, so next time they are called they get
2162       // recompiled.  Unloaded classes are loaded now so recompile before next
2163       // time they are called.  Same for uninitialized.  The interpreter will
2164       // link the missing class, if any.
2165       make_not_entrant = true;
2166       break;
2167     case Action_make_not_compilable:
2168       // Give up on compiling this method at all.
2169       make_not_entrant = true;
2170       make_not_compilable = true;
2171       break;
2172     default:
2173       ShouldNotReachHere();
2174     }
2175 
2176     // Setting +ProfileTraps fixes the following, on all platforms:
2177     // 4852688: ProfileInterpreter is off by default for ia64.  The result is
2178     // infinite heroic-opt-uncommon-trap/deopt/recompile cycles, since the
2179     // recompile relies on a MethodData* to record heroic opt failures.
2180 
2181     // Whether the interpreter is producing MDO data or not, we also need
2182     // to use the MDO to detect hot deoptimization points and control
2183     // aggressive optimization.
2184     bool inc_recompile_count = false;
2185     ProfileData* pdata = NULL;
2186     if (ProfileTraps && !is_client_compilation_mode_vm() && update_trap_state && trap_mdo != NULL) {
2187       assert(trap_mdo == get_method_data(thread, profiled_method, false), "sanity");
2188       uint this_trap_count = 0;
2189       bool maybe_prior_trap = false;
2190       bool maybe_prior_recompile = false;
2191       pdata = query_update_method_data(trap_mdo, trap_bci, reason, true,
2192 #if INCLUDE_JVMCI
2193                                    nm->is_compiled_by_jvmci() && nm->is_osr_method(),
2194 #endif
2195                                    nm->method(),
2196                                    //outputs:
2197                                    this_trap_count,
2198                                    maybe_prior_trap,
2199                                    maybe_prior_recompile);
2200       // Because the interpreter also counts null, div0, range, and class
2201       // checks, these traps from compiled code are double-counted.
2202       // This is harmless; it just means that the PerXTrapLimit values
2203       // are in effect a little smaller than they look.
2204 
2205       DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason);
2206       if (per_bc_reason != Reason_none) {
2207         // Now take action based on the partially known per-BCI history.
2208         if (maybe_prior_trap
2209             && this_trap_count >= (uint)PerBytecodeTrapLimit) {
2210           // If there are too many traps at this BCI, force a recompile.
2211           // This will allow the compiler to see the limit overflow, and
2212           // take corrective action, if possible.  The compiler generally
2213           // does not use the exact PerBytecodeTrapLimit value, but instead
2214           // changes its tactics if it sees any traps at all.  This provides
2215           // a little hysteresis, delaying a recompile until a trap happens
2216           // several times.
2217           //
2218           // Actually, since there is only one bit of counter per BCI,
2219           // the possible per-BCI counts are {0,1,(per-method count)}.
2220           // This produces accurate results if in fact there is only
2221           // one hot trap site, but begins to get fuzzy if there are
2222           // many sites.  For example, if there are ten sites each
2223           // trapping two or more times, they each get the blame for
2224           // all of their traps.
2225           make_not_entrant = true;
2226         }
2227 
2228         // Detect repeated recompilation at the same BCI, and enforce a limit.
2229         if (make_not_entrant && maybe_prior_recompile) {
2230           // More than one recompile at this point.
2231           inc_recompile_count = maybe_prior_trap;
2232         }
2233       } else {
2234         // For reasons which are not recorded per-bytecode, we simply
2235         // force recompiles unconditionally.
2236         // (Note that PerMethodRecompilationCutoff is enforced elsewhere.)
2237         make_not_entrant = true;
2238       }
2239 
2240       // Go back to the compiler if there are too many traps in this method.
2241       if (this_trap_count >= per_method_trap_limit(reason)) {
2242         // If there are too many traps in this method, force a recompile.
2243         // This will allow the compiler to see the limit overflow, and
2244         // take corrective action, if possible.
2245         // (This condition is an unlikely backstop only, because the
2246         // PerBytecodeTrapLimit is more likely to take effect first,
2247         // if it is applicable.)
2248         make_not_entrant = true;
2249       }
2250 
2251       // Here's more hysteresis:  If there has been a recompile at
2252       // this trap point already, run the method in the interpreter
2253       // for a while to exercise it more thoroughly.
2254       if (make_not_entrant && maybe_prior_recompile && maybe_prior_trap) {
2255         reprofile = true;
2256       }
2257     }
2258 
2259     // Take requested actions on the method:
2260 
2261     // Recompile
2262     if (make_not_entrant) {
2263       if (!nm->make_not_entrant()) {
2264         return; // the call did not change nmethod's state
2265       }
2266 
2267       if (pdata != NULL) {
2268         // Record the recompilation event, if any.
2269         int tstate0 = pdata->trap_state();
2270         int tstate1 = trap_state_set_recompiled(tstate0, true);
2271         if (tstate1 != tstate0)
2272           pdata->set_trap_state(tstate1);
2273       }
2274 
2275 #if INCLUDE_RTM_OPT
2276       // Restart collecting RTM locking abort statistic if the method
2277       // is recompiled for a reason other than RTM state change.
2278       // Assume that in new recompiled code the statistic could be different,
2279       // for example, due to different inlining.
2280       if ((reason != Reason_rtm_state_change) && (trap_mdo != NULL) &&
2281           UseRTMDeopt && (nm->as_nmethod()->rtm_state() != ProfileRTM)) {
2282         trap_mdo->atomic_set_rtm_state(ProfileRTM);
2283       }
2284 #endif
2285       // For code aging we count traps separately here, using make_not_entrant()
2286       // as a guard against simultaneous deopts in multiple threads.
2287       if (reason == Reason_tenured && trap_mdo != NULL) {
2288         trap_mdo->inc_tenure_traps();
2289       }
2290     }
2291 
2292     if (inc_recompile_count) {
2293       trap_mdo->inc_overflow_recompile_count();
2294       if ((uint)trap_mdo->overflow_recompile_count() >
2295           (uint)PerBytecodeRecompilationCutoff) {
2296         // Give up on the method containing the bad BCI.
2297         if (trap_method() == nm->method()) {
2298           make_not_compilable = true;
2299         } else {
2300           trap_method->set_not_compilable("overflow_recompile_count > PerBytecodeRecompilationCutoff", CompLevel_full_optimization);
2301           // But give grace to the enclosing nm->method().
2302         }
2303       }
2304     }
2305 
2306     // Reprofile
2307     if (reprofile) {
2308       CompilationPolicy::policy()->reprofile(trap_scope, nm->is_osr_method());
2309     }
2310 
2311     // Give up compiling
2312     if (make_not_compilable && !nm->method()->is_not_compilable(CompLevel_full_optimization)) {
2313       assert(make_not_entrant, "consistent");
2314       nm->method()->set_not_compilable("give up compiling", CompLevel_full_optimization);
2315     }
2316 
2317   } // Free marked resources
2318 
2319 }
2320 JRT_END
2321 
2322 ProfileData*
2323 Deoptimization::query_update_method_data(MethodData* trap_mdo,
2324                                          int trap_bci,
2325                                          Deoptimization::DeoptReason reason,
2326                                          bool update_total_trap_count,
2327 #if INCLUDE_JVMCI
2328                                          bool is_osr,
2329 #endif
2330                                          Method* compiled_method,
2331                                          //outputs:
2332                                          uint& ret_this_trap_count,
2333                                          bool& ret_maybe_prior_trap,
2334                                          bool& ret_maybe_prior_recompile) {
2335   bool maybe_prior_trap = false;
2336   bool maybe_prior_recompile = false;
2337   uint this_trap_count = 0;
2338   if (update_total_trap_count) {
2339     uint idx = reason;
2340 #if INCLUDE_JVMCI
2341     if (is_osr) {
2342       idx += Reason_LIMIT;
2343     }
2344 #endif
2345     uint prior_trap_count = trap_mdo->trap_count(idx);
2346     this_trap_count  = trap_mdo->inc_trap_count(idx);
2347 
2348     // If the runtime cannot find a place to store trap history,
2349     // it is estimated based on the general condition of the method.
2350     // If the method has ever been recompiled, or has ever incurred
2351     // a trap with the present reason , then this BCI is assumed
2352     // (pessimistically) to be the culprit.
2353     maybe_prior_trap      = (prior_trap_count != 0);
2354     maybe_prior_recompile = (trap_mdo->decompile_count() != 0);
2355   }
2356   ProfileData* pdata = NULL;
2357 
2358 
2359   // For reasons which are recorded per bytecode, we check per-BCI data.
2360   DeoptReason per_bc_reason = reason_recorded_per_bytecode_if_any(reason);
2361   assert(per_bc_reason != Reason_none || update_total_trap_count, "must be");
2362   if (per_bc_reason != Reason_none) {
2363     // Find the profile data for this BCI.  If there isn't one,
2364     // try to allocate one from the MDO's set of spares.
2365     // This will let us detect a repeated trap at this point.
2366     pdata = trap_mdo->allocate_bci_to_data(trap_bci, reason_is_speculate(reason) ? compiled_method : NULL);
2367 
2368     if (pdata != NULL) {
2369       if (reason_is_speculate(reason) && !pdata->is_SpeculativeTrapData()) {
2370         if (LogCompilation && xtty != NULL) {
2371           ttyLocker ttyl;
2372           // no more room for speculative traps in this MDO
2373           xtty->elem("speculative_traps_oom");
2374         }
2375       }
2376       // Query the trap state of this profile datum.
2377       int tstate0 = pdata->trap_state();
2378       if (!trap_state_has_reason(tstate0, per_bc_reason))
2379         maybe_prior_trap = false;
2380       if (!trap_state_is_recompiled(tstate0))
2381         maybe_prior_recompile = false;
2382 
2383       // Update the trap state of this profile datum.
2384       int tstate1 = tstate0;
2385       // Record the reason.
2386       tstate1 = trap_state_add_reason(tstate1, per_bc_reason);
2387       // Store the updated state on the MDO, for next time.
2388       if (tstate1 != tstate0)
2389         pdata->set_trap_state(tstate1);
2390     } else {
2391       if (LogCompilation && xtty != NULL) {
2392         ttyLocker ttyl;
2393         // Missing MDP?  Leave a small complaint in the log.
2394         xtty->elem("missing_mdp bci='%d'", trap_bci);
2395       }
2396     }
2397   }
2398 
2399   // Return results:
2400   ret_this_trap_count = this_trap_count;
2401   ret_maybe_prior_trap = maybe_prior_trap;
2402   ret_maybe_prior_recompile = maybe_prior_recompile;
2403   return pdata;
2404 }
2405 
2406 void
2407 Deoptimization::update_method_data_from_interpreter(MethodData* trap_mdo, int trap_bci, int reason) {
2408   ResourceMark rm;
2409   // Ignored outputs:
2410   uint ignore_this_trap_count;
2411   bool ignore_maybe_prior_trap;
2412   bool ignore_maybe_prior_recompile;
2413   assert(!reason_is_speculate(reason), "reason speculate only used by compiler");
2414   // JVMCI uses the total counts to determine if deoptimizations are happening too frequently -> do not adjust total counts
2415   bool update_total_counts = true JVMCI_ONLY( && !UseJVMCICompiler);
2416   query_update_method_data(trap_mdo, trap_bci,
2417                            (DeoptReason)reason,
2418                            update_total_counts,
2419 #if INCLUDE_JVMCI
2420                            false,
2421 #endif
2422                            NULL,
2423                            ignore_this_trap_count,
2424                            ignore_maybe_prior_trap,
2425                            ignore_maybe_prior_recompile);
2426 }
2427 
2428 Deoptimization::UnrollBlock* Deoptimization::uncommon_trap(JavaThread* thread, jint trap_request, jint exec_mode) {
2429   if (TraceDeoptimization) {
2430     tty->print("Uncommon trap ");
2431   }
2432   // Still in Java no safepoints
2433   {
2434     // This enters VM and may safepoint
2435     uncommon_trap_inner(thread, trap_request);
2436   }
2437   return fetch_unroll_info_helper(thread, exec_mode);
2438 }
2439 
2440 // Local derived constants.
2441 // Further breakdown of DataLayout::trap_state, as promised by DataLayout.
2442 const int DS_REASON_MASK   = ((uint)DataLayout::trap_mask) >> 1;
2443 const int DS_RECOMPILE_BIT = DataLayout::trap_mask - DS_REASON_MASK;
2444 
2445 //---------------------------trap_state_reason---------------------------------
2446 Deoptimization::DeoptReason
2447 Deoptimization::trap_state_reason(int trap_state) {
2448   // This assert provides the link between the width of DataLayout::trap_bits
2449   // and the encoding of "recorded" reasons.  It ensures there are enough
2450   // bits to store all needed reasons in the per-BCI MDO profile.
2451   assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits");
2452   int recompile_bit = (trap_state & DS_RECOMPILE_BIT);
2453   trap_state -= recompile_bit;
2454   if (trap_state == DS_REASON_MASK) {
2455     return Reason_many;
2456   } else {
2457     assert((int)Reason_none == 0, "state=0 => Reason_none");
2458     return (DeoptReason)trap_state;
2459   }
2460 }
2461 //-------------------------trap_state_has_reason-------------------------------
2462 int Deoptimization::trap_state_has_reason(int trap_state, int reason) {
2463   assert(reason_is_recorded_per_bytecode((DeoptReason)reason), "valid reason");
2464   assert(DS_REASON_MASK >= Reason_RECORDED_LIMIT, "enough bits");
2465   int recompile_bit = (trap_state & DS_RECOMPILE_BIT);
2466   trap_state -= recompile_bit;
2467   if (trap_state == DS_REASON_MASK) {
2468     return -1;  // true, unspecifically (bottom of state lattice)
2469   } else if (trap_state == reason) {
2470     return 1;   // true, definitely
2471   } else if (trap_state == 0) {
2472     return 0;   // false, definitely (top of state lattice)
2473   } else {
2474     return 0;   // false, definitely
2475   }
2476 }
2477 //-------------------------trap_state_add_reason-------------------------------
2478 int Deoptimization::trap_state_add_reason(int trap_state, int reason) {
2479   assert(reason_is_recorded_per_bytecode((DeoptReason)reason) || reason == Reason_many, "valid reason");
2480   int recompile_bit = (trap_state & DS_RECOMPILE_BIT);
2481   trap_state -= recompile_bit;
2482   if (trap_state == DS_REASON_MASK) {
2483     return trap_state + recompile_bit;     // already at state lattice bottom
2484   } else if (trap_state == reason) {
2485     return trap_state + recompile_bit;     // the condition is already true
2486   } else if (trap_state == 0) {
2487     return reason + recompile_bit;          // no condition has yet been true
2488   } else {
2489     return DS_REASON_MASK + recompile_bit;  // fall to state lattice bottom
2490   }
2491 }
2492 //-----------------------trap_state_is_recompiled------------------------------
2493 bool Deoptimization::trap_state_is_recompiled(int trap_state) {
2494   return (trap_state & DS_RECOMPILE_BIT) != 0;
2495 }
2496 //-----------------------trap_state_set_recompiled-----------------------------
2497 int Deoptimization::trap_state_set_recompiled(int trap_state, bool z) {
2498   if (z)  return trap_state |  DS_RECOMPILE_BIT;
2499   else    return trap_state & ~DS_RECOMPILE_BIT;
2500 }
2501 //---------------------------format_trap_state---------------------------------
2502 // This is used for debugging and diagnostics, including LogFile output.
2503 const char* Deoptimization::format_trap_state(char* buf, size_t buflen,
2504                                               int trap_state) {
2505   assert(buflen > 0, "sanity");
2506   DeoptReason reason      = trap_state_reason(trap_state);
2507   bool        recomp_flag = trap_state_is_recompiled(trap_state);
2508   // Re-encode the state from its decoded components.
2509   int decoded_state = 0;
2510   if (reason_is_recorded_per_bytecode(reason) || reason == Reason_many)
2511     decoded_state = trap_state_add_reason(decoded_state, reason);
2512   if (recomp_flag)
2513     decoded_state = trap_state_set_recompiled(decoded_state, recomp_flag);
2514   // If the state re-encodes properly, format it symbolically.
2515   // Because this routine is used for debugging and diagnostics,
2516   // be robust even if the state is a strange value.
2517   size_t len;
2518   if (decoded_state != trap_state) {
2519     // Random buggy state that doesn't decode??
2520     len = jio_snprintf(buf, buflen, "#%d", trap_state);
2521   } else {
2522     len = jio_snprintf(buf, buflen, "%s%s",
2523                        trap_reason_name(reason),
2524                        recomp_flag ? " recompiled" : "");
2525   }
2526   return buf;
2527 }
2528 
2529 
2530 //--------------------------------statics--------------------------------------
2531 const char* Deoptimization::_trap_reason_name[] = {
2532   // Note:  Keep this in sync. with enum DeoptReason.
2533   "none",
2534   "null_check",
2535   "null_assert" JVMCI_ONLY("_or_unreached0"),
2536   "range_check",
2537   "class_check",
2538   "array_check",
2539   "intrinsic" JVMCI_ONLY("_or_type_checked_inlining"),
2540   "bimorphic" JVMCI_ONLY("_or_optimized_type_check"),
2541   "profile_predicate",
2542   "unloaded",
2543   "uninitialized",
2544   "initialized",
2545   "unreached",
2546   "unhandled",
2547   "constraint",
2548   "div0_check",
2549   "age",
2550   "predicate",
2551   "loop_limit_check",
2552   "speculate_class_check",
2553   "speculate_null_check",
2554   "speculate_null_assert",
2555   "rtm_state_change",
2556   "unstable_if",
2557   "unstable_fused_if",
2558 #if INCLUDE_JVMCI
2559   "aliasing",
2560   "transfer_to_interpreter",
2561   "not_compiled_exception_handler",
2562   "unresolved",
2563   "jsr_mismatch",
2564 #endif
2565   "tenured"
2566 };
2567 const char* Deoptimization::_trap_action_name[] = {
2568   // Note:  Keep this in sync. with enum DeoptAction.
2569   "none",
2570   "maybe_recompile",
2571   "reinterpret",
2572   "make_not_entrant",
2573   "make_not_compilable"
2574 };
2575 
2576 const char* Deoptimization::trap_reason_name(int reason) {
2577   // Check that every reason has a name
2578   STATIC_ASSERT(sizeof(_trap_reason_name)/sizeof(const char*) == Reason_LIMIT);
2579 
2580   if (reason == Reason_many)  return "many";
2581   if ((uint)reason < Reason_LIMIT)
2582     return _trap_reason_name[reason];
2583   static char buf[20];
2584   sprintf(buf, "reason%d", reason);
2585   return buf;
2586 }
2587 const char* Deoptimization::trap_action_name(int action) {
2588   // Check that every action has a name
2589   STATIC_ASSERT(sizeof(_trap_action_name)/sizeof(const char*) == Action_LIMIT);
2590 
2591   if ((uint)action < Action_LIMIT)
2592     return _trap_action_name[action];
2593   static char buf[20];
2594   sprintf(buf, "action%d", action);
2595   return buf;
2596 }
2597 
2598 // This is used for debugging and diagnostics, including LogFile output.
2599 const char* Deoptimization::format_trap_request(char* buf, size_t buflen,
2600                                                 int trap_request) {
2601   jint unloaded_class_index = trap_request_index(trap_request);
2602   const char* reason = trap_reason_name(trap_request_reason(trap_request));
2603   const char* action = trap_action_name(trap_request_action(trap_request));
2604 #if INCLUDE_JVMCI
2605   int debug_id = trap_request_debug_id(trap_request);
2606 #endif
2607   size_t len;
2608   if (unloaded_class_index < 0) {
2609     len = jio_snprintf(buf, buflen, "reason='%s' action='%s'" JVMCI_ONLY(" debug_id='%d'"),
2610                        reason, action
2611 #if INCLUDE_JVMCI
2612                        ,debug_id
2613 #endif
2614                        );
2615   } else {
2616     len = jio_snprintf(buf, buflen, "reason='%s' action='%s' index='%d'" JVMCI_ONLY(" debug_id='%d'"),
2617                        reason, action, unloaded_class_index
2618 #if INCLUDE_JVMCI
2619                        ,debug_id
2620 #endif
2621                        );
2622   }
2623   return buf;
2624 }
2625 
2626 juint Deoptimization::_deoptimization_hist
2627         [Deoptimization::Reason_LIMIT]
2628     [1 + Deoptimization::Action_LIMIT]
2629         [Deoptimization::BC_CASE_LIMIT]
2630   = {0};
2631 
2632 enum {
2633   LSB_BITS = 8,
2634   LSB_MASK = right_n_bits(LSB_BITS)
2635 };
2636 
2637 void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action,
2638                                        Bytecodes::Code bc) {
2639   assert(reason >= 0 && reason < Reason_LIMIT, "oob");
2640   assert(action >= 0 && action < Action_LIMIT, "oob");
2641   _deoptimization_hist[Reason_none][0][0] += 1;  // total
2642   _deoptimization_hist[reason][0][0]      += 1;  // per-reason total
2643   juint* cases = _deoptimization_hist[reason][1+action];
2644   juint* bc_counter_addr = NULL;
2645   juint  bc_counter      = 0;
2646   // Look for an unused counter, or an exact match to this BC.
2647   if (bc != Bytecodes::_illegal) {
2648     for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) {
2649       juint* counter_addr = &cases[bc_case];
2650       juint  counter = *counter_addr;
2651       if ((counter == 0 && bc_counter_addr == NULL)
2652           || (Bytecodes::Code)(counter & LSB_MASK) == bc) {
2653         // this counter is either free or is already devoted to this BC
2654         bc_counter_addr = counter_addr;
2655         bc_counter = counter | bc;
2656       }
2657     }
2658   }
2659   if (bc_counter_addr == NULL) {
2660     // Overflow, or no given bytecode.
2661     bc_counter_addr = &cases[BC_CASE_LIMIT-1];
2662     bc_counter = (*bc_counter_addr & ~LSB_MASK);  // clear LSB
2663   }
2664   *bc_counter_addr = bc_counter + (1 << LSB_BITS);
2665 }
2666 
2667 jint Deoptimization::total_deoptimization_count() {
2668   return _deoptimization_hist[Reason_none][0][0];
2669 }
2670 
2671 void Deoptimization::print_statistics() {
2672   juint total = total_deoptimization_count();
2673   juint account = total;
2674   if (total != 0) {
2675     ttyLocker ttyl;
2676     if (xtty != NULL)  xtty->head("statistics type='deoptimization'");
2677     tty->print_cr("Deoptimization traps recorded:");
2678     #define PRINT_STAT_LINE(name, r) \
2679       tty->print_cr("  %4d (%4.1f%%) %s", (int)(r), ((r) * 100.0) / total, name);
2680     PRINT_STAT_LINE("total", total);
2681     // For each non-zero entry in the histogram, print the reason,
2682     // the action, and (if specifically known) the type of bytecode.
2683     for (int reason = 0; reason < Reason_LIMIT; reason++) {
2684       for (int action = 0; action < Action_LIMIT; action++) {
2685         juint* cases = _deoptimization_hist[reason][1+action];
2686         for (int bc_case = 0; bc_case < BC_CASE_LIMIT; bc_case++) {
2687           juint counter = cases[bc_case];
2688           if (counter != 0) {
2689             char name[1*K];
2690             Bytecodes::Code bc = (Bytecodes::Code)(counter & LSB_MASK);
2691             if (bc_case == BC_CASE_LIMIT && (int)bc == 0)
2692               bc = Bytecodes::_illegal;
2693             sprintf(name, "%s/%s/%s",
2694                     trap_reason_name(reason),
2695                     trap_action_name(action),
2696                     Bytecodes::is_defined(bc)? Bytecodes::name(bc): "other");
2697             juint r = counter >> LSB_BITS;
2698             tty->print_cr("  %40s: " UINT32_FORMAT " (%.1f%%)", name, r, (r * 100.0) / total);
2699             account -= r;
2700           }
2701         }
2702       }
2703     }
2704     if (account != 0) {
2705       PRINT_STAT_LINE("unaccounted", account);
2706     }
2707     #undef PRINT_STAT_LINE
2708     if (xtty != NULL)  xtty->tail("statistics");
2709   }
2710 }
2711 
2712 // Returns true iff objects were reallocated and relocked because of access through JVMTI
2713 bool EscapeBarrier::objs_are_deoptimized(JavaThread* thread, intptr_t* fr_id) {
2714   // first/oldest update holds the flag
2715   GrowableArray<jvmtiDeferredLocalVariableSet*>* list = JvmtiDeferredUpdates::deferred_locals(thread);
2716   bool result = false;
2717   if (list != NULL) {
2718     for (int i = 0; i < list->length(); i++) {
2719       if (list->at(i)->matches(fr_id)) {
2720         result = list->at(i)->objects_are_deoptimized();
2721         break;
2722       }
2723     }
2724   }
2725   return result;
2726 }
2727 
2728 // Object references of frames up to the given depth are about to be accessed. Frames with
2729 // optimizations based on escape state that is potentially changed by the accesses need to be
2730 // deoptimized and the referenced objects need to be reallocated and relocked.
2731 // Up to depth this is done for frames with not escaping objects in scope. For deeper frames it is
2732 // done only, if they pass not escaping objects as arguments, because they potentially escape from
2733 // callee frames within the given depth.
2734 // The search for deeper frames is ended if an entry frame is found, because arguments to
2735 // native methods are considered to escape globally.
2736 bool EscapeBarrier::deoptimize_objects(int depth) {
2737   if (barrier_active() && deoptee_thread()->has_last_Java_frame()) {
2738     ResourceMark rm(calling_thread());
2739     HandleMark   hm;
2740     RegisterMap  reg_map(deoptee_thread());
2741     vframe* vf = deoptee_thread()->last_java_vframe(&reg_map);
2742     int cur_depth = 0;
2743     while (vf != NULL && ((cur_depth <= depth) || !vf->is_entry_frame())) {
2744       if (vf->is_compiled_frame()) {
2745         compiledVFrame* cvf = compiledVFrame::cast(vf);
2746         // Deoptimize frame and local objects if any exist.
2747         // If cvf is deeper than depth, then we deoptimize iff local objects are passed as args.
2748         bool should_deopt = cur_depth <= depth ? cvf->not_global_escape_in_scope() : cvf->arg_escape();
2749         if (should_deopt && !deoptimize_objects(cvf->fr().id())) {
2750           // reallocation of scalar replaced objects failed, because heap is exhausted
2751           return false;
2752         }
2753 
2754         // move to top frame
2755         while(!vf->is_top()) {
2756           cur_depth++;
2757           vf = vf->sender();
2758         }
2759       }
2760 
2761       // move to next physical frame
2762       cur_depth++;
2763       vf = vf->sender();
2764     }
2765   }
2766   return true;
2767 }
2768 
2769 bool EscapeBarrier::deoptimize_objects_all_threads() {
2770   if (!barrier_active()) return true;
2771   ResourceMark rm(calling_thread());
2772   for (JavaThreadIteratorWithHandle jtiwh; JavaThread *jt = jtiwh.next(); ) {
2773     if (jt->has_last_Java_frame()) {
2774       RegisterMap reg_map(jt);
2775       vframe* vf = jt->last_java_vframe(&reg_map);
2776       assert(jt->frame_anchor()->walkable(),
2777              "The stack of JavaThread " PTR_FORMAT " is not walkable. Thread state is %d",
2778              p2i(jt), jt->thread_state());
2779       while (vf != NULL) {
2780         if (vf->is_compiled_frame()) {
2781           compiledVFrame* cvf = compiledVFrame::cast(vf);
2782           if ((cvf->not_global_escape_in_scope() || cvf->arg_escape()) &&
2783               !deoptimize_objects_internal(jt, cvf->fr().id())) {
2784             return false; // reallocation failure
2785           }
2786           // move to top frame
2787           while(!vf->is_top()) {
2788             vf = vf->sender();
2789           }
2790         }
2791         // move to next physical frame
2792         vf = vf->sender();
2793       }
2794     }
2795   }
2796   return true; // success
2797 }
2798 
2799 bool EscapeBarrier::_deoptimizing_objects_for_all_threads = false;
2800 bool EscapeBarrier::_self_deoptimization_in_progress      = false;
2801 
2802 class EscapeBarrierSuspendHandshake : public HandshakeClosure {
2803   JavaThread* _excluded_thread;
2804  public:
2805   EscapeBarrierSuspendHandshake(JavaThread* excluded_thread, const char* name) : HandshakeClosure(name), _excluded_thread(excluded_thread) { }
2806   void do_thread(Thread* th) {
2807     if (th->is_Java_thread() && !th->is_hidden_from_external_view() && (th != _excluded_thread)) {
2808       th->set_obj_deopt_flag();
2809     }
2810   }
2811 };
2812 
2813 void EscapeBarrier::sync_and_suspend_one() {
2814   assert(_calling_thread != NULL, "calling thread must not be NULL");
2815   assert(_deoptee_thread != NULL, "deoptee thread must not be NULL");
2816   assert(barrier_active(), "should not call");
2817 
2818   // Sync with other threads that might be doing deoptimizations
2819   {
2820     // Need to switch to _thread_blocked for the wait() call
2821     ThreadBlockInVM tbivm(_calling_thread);
2822     MonitorLocker ml(_calling_thread, EscapeBarrier_lock, Mutex::_no_safepoint_check_flag);
2823     while (_self_deoptimization_in_progress || _deoptee_thread->is_obj_deopt_suspend()) {
2824       ml.wait();
2825     }
2826 
2827     if (self_deopt()) {
2828       _self_deoptimization_in_progress = true;
2829       return;
2830     }
2831 
2832     // set suspend flag for target thread
2833     _deoptee_thread->set_obj_deopt_flag();
2834   }
2835 
2836   // suspend target thread
2837   EscapeBarrierSuspendHandshake sh(NULL, "EscapeBarrierSuspendOne");
2838   Handshake::execute_direct(&sh, _deoptee_thread);
2839   assert(!_deoptee_thread->has_last_Java_frame() || _deoptee_thread->frame_anchor()->walkable(),
2840          "stack should be walkable now");
2841 }
2842 
2843 void EscapeBarrier::sync_and_suspend_all() {
2844   assert(barrier_active(), "should not call");
2845   assert(_calling_thread != NULL, "calling thread must not be NULL");
2846   assert(all_threads(), "sanity");
2847 
2848   // Sync with other threads that might be doing deoptimizations
2849   {
2850     // Need to switch to _thread_blocked for the wait() call
2851     ThreadBlockInVM tbivm(_calling_thread);
2852     MonitorLocker ml(_calling_thread, EscapeBarrier_lock, Mutex::_no_safepoint_check_flag);
2853 
2854     bool deopt_in_progress;
2855     do {
2856       deopt_in_progress = _self_deoptimization_in_progress;
2857       for (JavaThreadIteratorWithHandle jtiwh; JavaThread *jt = jtiwh.next(); ) {
2858         deopt_in_progress = (deopt_in_progress || jt->is_obj_deopt_suspend());
2859         if (deopt_in_progress) {
2860           break;
2861         }
2862       }
2863       if (deopt_in_progress) {
2864         ml.wait(); // then check again
2865       }
2866     } while(deopt_in_progress);
2867 
2868     _self_deoptimization_in_progress = true;
2869     _deoptimizing_objects_for_all_threads = true;
2870   }
2871 
2872   EscapeBarrierSuspendHandshake sh(_calling_thread, "EscapeBarrierSuspendAll");
2873   Handshake::execute(&sh);
2874 #ifdef ASSERT
2875   for (JavaThreadIteratorWithHandle jtiwh; JavaThread *jt = jtiwh.next(); ) {
2876     if (jt->is_hidden_from_external_view()) continue;
2877     assert(!jt->has_last_Java_frame() || jt->frame_anchor()->walkable(),
2878            "The stack of JavaThread " PTR_FORMAT " is not walkable. Thread state is %d",
2879            p2i(jt), jt->thread_state());
2880   }
2881 #endif // ASSERT
2882 }
2883 
2884 void EscapeBarrier::resume_one() {
2885   assert(barrier_active(), "should not call");
2886   assert(!all_threads(), "use resume_all()");
2887   MonitorLocker ml(_calling_thread, EscapeBarrier_lock, Mutex::_no_safepoint_check_flag);
2888   if (self_deopt()) {
2889     assert(_self_deoptimization_in_progress, "incorrect synchronization");
2890     _self_deoptimization_in_progress = false;
2891   } else {
2892     _deoptee_thread->clear_obj_deopt_flag();
2893   }
2894   ml.notify_all();
2895 }
2896 
2897 void EscapeBarrier::resume_all() {
2898   assert(barrier_active(), "should not call");
2899   assert(all_threads(), "use resume_one()");
2900   MonitorLocker ml(_calling_thread, EscapeBarrier_lock, Mutex::_no_safepoint_check_flag);
2901   assert(_self_deoptimization_in_progress, "incorrect synchronization");
2902   _deoptimizing_objects_for_all_threads = false;
2903   _self_deoptimization_in_progress = false;
2904   for (JavaThreadIteratorWithHandle jtiwh; JavaThread *jt = jtiwh.next(); ) {
2905     jt->clear_obj_deopt_flag();
2906   }
2907   ml.notify_all();
2908 }
2909 
2910 void EscapeBarrier::thread_added(JavaThread* jt) {
2911   if (!jt->is_hidden_from_external_view()) {
2912     MutexLocker ml(EscapeBarrier_lock, Mutex::_no_safepoint_check_flag);
2913     if (_deoptimizing_objects_for_all_threads) {
2914       jt->set_obj_deopt_flag();
2915     }
2916   }
2917 }
2918 
2919 void EscapeBarrier::thread_removed(JavaThread* jt) {
2920   MonitorLocker ml(EscapeBarrier_lock, Mutex::_no_safepoint_check_flag);
2921   if (jt->is_obj_deopt_suspend()) {
2922     // jt terminated before it self suspended.
2923     // Other threads might be waiting to perform deoptimizations for it.
2924     jt->clear_obj_deopt_flag();
2925     ml.notify_all();
2926   }
2927 }
2928 
2929 // Remember that objects were reallocated and relocked for the compiled frame with the given id
2930 static void set_objs_are_deoptimized(JavaThread* thread, intptr_t* fr_id) {
2931   // set in first/oldest update
2932   GrowableArray<jvmtiDeferredLocalVariableSet*>* list = JvmtiDeferredUpdates::deferred_locals(thread);
2933   DEBUG_ONLY(bool found = false);
2934   if (list != NULL) {
2935     for (int i = 0; i < list->length(); i++) {
2936       if (list->at(i)->matches(fr_id)) {
2937         DEBUG_ONLY(found = true);
2938         list->at(i)->set_objs_are_deoptimized();
2939         break;
2940       }
2941     }
2942   }
2943   assert(found, "variable set should exist at least for one vframe");
2944 }
2945 
2946 // Deoptimize the given frame and deoptimize objects with optimizations based on escape analysis,
2947 // i.e. reallocate scalar replaced objects on the heap and relock objects if locking has been
2948 // eliminated.
2949 // Deoptimized objects are kept as JVMTI deferred updates until the compiled frame is replaced with interpreter frames.
2950 // Returns false iff at least one reallocation failed.
2951 bool EscapeBarrier::deoptimize_objects_internal(JavaThread* deoptee, intptr_t* fr_id) {
2952   if (!barrier_active()) return true;
2953 
2954   JavaThread* ct = calling_thread();
2955   bool realloc_failures = false;
2956 
2957   if (!objs_are_deoptimized(deoptee, fr_id)) {
2958     // Make sure the frame identified by fr_id is deoptimized and fetch its last vframe
2959     compiledVFrame* last_cvf;
2960     bool fr_is_deoptimized;
2961     do {
2962       StackFrameStream fst(deoptee);
2963       while (fst.current()->id() != fr_id && !fst.is_done()) {
2964         fst.next();
2965       }
2966       assert(fst.current()->id() == fr_id, "frame not found");
2967       assert(fst.current()->is_compiled_frame(), "only compiled frames can contain stack allocated objects");
2968       fr_is_deoptimized = fst.current()->is_deoptimized_frame();
2969       if (!fr_is_deoptimized) {
2970         // Execution must not continue in the compiled method, so we deoptimize the frame.
2971         Deoptimization::deoptimize_frame(deoptee, fr_id);
2972       } else {
2973         last_cvf = compiledVFrame::cast(vframe::new_vframe(fst.current(), fst.register_map(), deoptee));
2974       }
2975     } while(!fr_is_deoptimized);
2976 
2977     // collect inlined frames
2978     compiledVFrame* cvf = last_cvf;
2979     GrowableArray<compiledVFrame*>* vfs = new GrowableArray<compiledVFrame*>(10);
2980     while (!cvf->is_top()) {
2981       vfs->push(cvf);
2982       cvf = compiledVFrame::cast(cvf->sender());
2983     }
2984     vfs->push(cvf);
2985 
2986     // reallocate and relock optimized objects
2987     bool deoptimized_objects = Deoptimization::deoptimize_objects_internal(ct, vfs, realloc_failures);
2988     if (!realloc_failures && deoptimized_objects) {
2989       // now do the updates
2990       for (int frame_index = 0; frame_index < vfs->length(); frame_index++) {
2991         cvf = vfs->at(frame_index);
2992 
2993         // locals
2994         GrowableArray<ScopeValue*>* scopeLocals = cvf->scope()->locals();
2995         StackValueCollection* locals = cvf->locals();
2996         if (locals != NULL) {
2997           for (int i2 = 0; i2 < locals->size(); i2++) {
2998             StackValue* var = locals->at(i2);
2999             if (var->type() == T_OBJECT && scopeLocals->at(i2)->is_object()) {
3000               jvalue val;
3001               val.l = cast_from_oop<jobject>(locals->at(i2)->get_obj()());
3002               cvf->update_local(T_OBJECT, i2, val);
3003             }
3004           }
3005         }
3006 
3007         // expressions
3008         GrowableArray<ScopeValue*>* scopeExpressions = cvf->scope()->expressions();
3009         StackValueCollection* expressions = cvf->expressions();
3010         if (expressions != NULL) {
3011           for (int i2 = 0; i2 < expressions->size(); i2++) {
3012             StackValue* var = expressions->at(i2);
3013             if (var->type() == T_OBJECT && scopeExpressions->at(i2)->is_object()) {
3014               jvalue val;
3015               val.l = cast_from_oop<jobject>(expressions->at(i2)->get_obj()());
3016               cvf->update_stack(T_OBJECT, i2, val);
3017             }
3018           }
3019         }
3020 
3021         // monitors
3022         GrowableArray<MonitorInfo*>* monitors = cvf->monitors();
3023         if (monitors != NULL) {
3024           for (int i2 = 0; i2 < monitors->length(); i2++) {
3025             if (monitors->at(i2)->eliminated()) {
3026               assert(!monitors->at(i2)->owner_is_scalar_replaced(), "reallocation failure, should not update");
3027               cvf->update_monitor(i2, monitors->at(i2));
3028             }
3029           }
3030         }
3031       }
3032       set_objs_are_deoptimized(deoptee, fr_id);
3033     }
3034   }
3035   return !realloc_failures;
3036 }
3037 
3038 #else // COMPILER2_OR_JVMCI
3039 
3040 
3041 // Stubs for C1 only system.
3042 bool Deoptimization::trap_state_is_recompiled(int trap_state) {
3043   return false;
3044 }
3045 
3046 const char* Deoptimization::trap_reason_name(int reason) {
3047   return "unknown";
3048 }
3049 
3050 void Deoptimization::print_statistics() {
3051   // no output
3052 }
3053 
3054 void
3055 Deoptimization::update_method_data_from_interpreter(MethodData* trap_mdo, int trap_bci, int reason) {
3056   // no udpate
3057 }
3058 
3059 int Deoptimization::trap_state_has_reason(int trap_state, int reason) {
3060   return 0;
3061 }
3062 
3063 void Deoptimization::gather_statistics(DeoptReason reason, DeoptAction action,
3064                                        Bytecodes::Code bc) {
3065   // no update
3066 }
3067 
3068 const char* Deoptimization::format_trap_state(char* buf, size_t buflen,
3069                                               int trap_state) {
3070   jio_snprintf(buf, buflen, "#%d", trap_state);
3071   return buf;
3072 }
3073 
3074 #endif // COMPILER2_OR_JVMCI